Address Management in GPU Superclusters

Abstract

[0003] Described herein is a network fabric including multiple graphical processing unit (GPU) clusters communicatively coupled to each other via multiple switches arranged in a hierarchical structure including a first tier of switches, a second tier of switches, and a third tier of switches. One or more switches are selected from the third tier of switches to form a set of target switches, and each target switch receives address information for each GPU included in the multiple GPU clusters. Each target switch generates multiple sets of address information by filtering the received address information based on a condition and transmits the multiple sets of address information to each switch included in the first tier of switches, and the switch stores a subset of the multiple sets of address information according to the condition.

Description

[Technical Field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This application is a non-provisional application of, and claims the benefit of, each of the following provisional applications, the entire contents of each of which are incorporated herein by reference for all purposes: (1) U.S. Provisional Patent Application No. 63 / 422,650, filed November 4, 2022; (2) U.S. Provisional Patent Application No. 63 / 424,282, filed November 10, 2022; (3) U.S. Provisional Patent Application No. 63 / 425,646, filed November 15, 2022; (4) U.S. Provisional Patent Application No. 63 / 460,766, filed April 20, 2023; (5) U.S. Provisional Patent Application No. 63 / 583,512, filed September 18, 2023.

[0002] Field The present disclosure relates generally to cloud architectures, and more particularly to a supercluster architecture for graphical processing units (GPUs). More particularly, the present disclosure relates to a network architecture that provides hybrid clusters of GPUs (e.g., GPUs of different generations or operating at different speeds) for coexistence within the same network fabric. The supercluster architecture provides seamless scaling of GPUs with ever-increasing customer demands. [Background technology]

[0003] background Organizations continue to migrate their business applications and databases to the cloud to reduce the costs of purchasing, updating, and maintaining on-premises hardware and software. High-performance computer applications consistently consume all available computing power to achieve a specific outcome or result. Such applications require dedicated network performance, fast storage, high computing power, and large amounts of memory—resources that are in short supply in the virtualized infrastructure that makes up today's commodity clouds.

[0004] Cloud infrastructure service providers offer newer and faster graphical processing units (GPUs) to address the requirements of such applications. GPU workloads typically run on one or more host machines. Typically, such workloads cannot achieve the expected level of throughput. One factor contributing to this problem is the lack of flow entropy, e.g., equal cost multi-path (ECMP) flow entropy. Furthermore, the problem is exacerbated by the fact that host machines (i.e., hosts) exchange traffic without considering other hosts in their local network vicinity.

[0005] Furthermore, conventional GPU clusters typically scale within the range of 1,000 to 4,000 GPUs. The limitations on scaling the number of GPUs are due to limitations imposed by the network topology built to support the GPU cluster. The network topology built to support the GPU cluster leads to significant oversubscription, thus posing challenges to scaling the cluster. Furthermore, conventional GPU clusters impose severe limitations on the routing policies employed within the cluster. For example, conventional GPU clusters do not support standard custom routing protocols. Furthermore, conventional GPU clusters are built in a manner that supports a single transmission rate for all GPUs in the cluster. Therefore, there is a need to build a GPU cluster that can scale at a much higher level than conventional GPU clusters and can also support communication between different GPU clusters operating at different transmission rates. The embodiments described herein address these and other issues. Summary of the Invention

[0006] overview The present disclosure relates to cloud architectures, and more particularly to supercluster architectures for graphical processing units (GPUs). More particularly, the present disclosure relates to network architectures that provide hybrid clusters of GPUs (e.g., GPUs of different generations or operating at different speeds) to coexist within the same network fabric. The supercluster architecture provides seamless scaling of GPUs with ever-increasing customer demands. Various embodiments are described herein, including methods, systems, non-transitory computer-readable storage media that store programs, code, or instructions executable by one or more processors, and the like. Some embodiments may be implemented using a computer program product that includes computer programs / instructions that, when executed by a processor, cause the processor to perform any of the methods described in this disclosure.

[0007] One aspect of the present disclosure provides a method including: providing a plurality of graphical processing unit (GPU) clusters, the plurality of GPU clusters being communicatively coupled to each other via a plurality of switches arranged in a hierarchical structure, the hierarchical structure including a first tier of switches, a second tier of switches, and a third tier of switches; selecting one or more switches from the third tier of switches to form a set of target switches; receiving, by each target switch in the set of target switches, address information of each GPU included in the plurality of GPU clusters; generating, by each target switch in the set of target switches, a plurality of sets of address information by filtering the received address information based on a condition; and transmitting, by each target switch, the plurality of sets of address information to each switch included in the first tier of switches, the switch storing and transmitting a subset of the plurality of sets of address information according to the condition.

[0008] Aspects of the present disclosure provide a computing device comprising one or more data processors and a non-transitory computer-readable storage medium containing instructions that, when executed on the one or more data processors, cause the computing device to perform some or all of one or more methods disclosed herein.

[0009] Another aspect of the present disclosure provides a computer program product, tangibly embodied in a non-transitory machine-readable storage medium, comprising instructions configured to cause one or more data processors to perform some or all of one or more of the methods disclosed herein.

[0010] The foregoing, together with other features and embodiments, will become more apparent upon reference to the following specification, claims, and accompanying drawings.

[0011] The features, embodiments, and advantages of the present disclosure will be better understood when the following detailed description is read in conjunction with the accompanying drawings. [Brief explanation of the drawings]

[0012] [Figure 1] 1 is a high-level diagram of a distributed environment illustrating a virtual or overlay cloud network hosted by a cloud service provider infrastructure, according to an embodiment. [Figure 2] FIG. 2 illustrates a simplified architectural diagram of physical components in a physical network within a CSPI, according to an embodiment. [Figure 3] FIG. 1 illustrates an exemplary arrangement within CSPI in which a host machine is connected to multiple network virtualization devices (NVDs), according to one embodiment. [Figure 4]A diagram illustrating connections between a host machine and an NVD to achieve I / O virtualization to support multi-tenancy functionality in accordance with one embodiment. [Figure 5] FIG. 2 illustrates a simplified block diagram of a physical network provided by CSPI, according to one embodiment. [Figure 6A] FIG. 1 illustrates an architecture of a hybrid GPU cluster, according to an embodiment. [Figure 6B] FIG. 2 illustrates another architecture of a hybrid GPU cluster, according to an embodiment. [Figure 7] FIG. 1 illustrates an exemplary flowchart illustrating steps performed in provisioning a request using a hybrid GPU cluster, according to an embodiment. [Figure 8] FIG. 1 illustrates a simplified block diagram of a cloud infrastructure incorporating a CLOS network deployment, according to an embodiment. [Figure 9] FIG. 1 illustrates a logical topology constructed without using locality information, according to an embodiment. [Figure 10] FIG. 1 illustrates a logical topology constructed using locality information, according to an embodiment. [Figure 11] FIG. 1 illustrates a simplified block diagram of a hybrid GPU cluster illustrating locality of network components, according to some embodiments. [Figure 12] FIG. 1 illustrates an exemplary flowchart illustrating steps performed in provisioning a request using locality information, according to an embodiment. [Figure 13] FIG. 1 illustrates a policy-based routing mechanism implemented in a hybrid GPU cluster, according to an embodiment. [Figure 14] FIG. 2 shows a flowchart illustrating steps performed by a network device in routing a packet, according to an embodiment. [Figure 15A]FIG. 1 illustrates a hybrid GPU cluster architecture illustrating the placement of route reflectors, according to some embodiments. [Figure 15B] FIG. 2 illustrates a flowchart depicting steps performed by a route reflector in managing the size of address tables stored by a switch, according to an embodiment. [Figure 16] FIG. 1 is a block diagram illustrating one pattern for implementing a cloud infrastructure as a service system, according to at least one embodiment. [Figure 17] FIG. 1 is a block diagram illustrating another pattern for implementing a cloud infrastructure as a service system, according to at least one embodiment. [Figure 18] FIG. 1 is a block diagram illustrating another pattern for implementing a cloud infrastructure as a service system, according to at least one embodiment. [Figure 19] FIG. 1 is a block diagram illustrating another pattern for implementing a cloud infrastructure as a service system, according to at least one embodiment. [Figure 20] FIG. 1 is a block diagram illustrating an exemplary computer system according to at least one embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0013] Detailed Description In the following description, for purposes of explanation, specific details are set forth in order to provide a thorough understanding of certain embodiments. However, it will be apparent that various embodiments may be practiced without these specific details. The figures and descriptions are not intended to be limiting. The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any embodiment or design described herein as "exemplary" is not necessarily to be construed as preferred or advantageous over other embodiments or designs.

[0014] FIELD OF THE DISCLOSURE

[0002] Embodiments of the present disclosure relate to cloud architectures, and more particularly to a supercluster architecture for graphical processing units (GPUs). More particularly, the present disclosure relates to a network architecture that provides hybrid clusters of GPUs (e.g., GPUs of different generations or operating at different speeds) for coexistence within the same network fabric. The supercluster architecture provides seamless scaling of GPUs with ever-increasing customer demands.

[0015] Cloud Network Example The term cloud services is typically used to refer to services made available on-demand (e.g., through a subscription model) by a cloud service provider (CSP) to users or customers using systems and infrastructure (cloud infrastructure) provided by the CSP. Typically, the servers and systems that make up the CSP's infrastructure are separate from the customer's own on-premises servers and systems. Thus, customers can use cloud services provided by the CSP without having to purchase separate hardware and software resources for the service. Cloud services are designed to provide subscribing customers with easy and scalable access to applications and computing resources without requiring the customer to invest in procuring the infrastructure used to deliver the service.

[0016] There are multiple cloud service providers offering different types of cloud services. There are various types or models of cloud services, including Software-as-a-Service (SaaS), Platform-as-a-Service (PaaS), Infrastructure-as-a-Service (IaaS), etc.

[0017] A customer can subscribe to one or more cloud services offered by a CSP. A customer can be any entity, such as an individual, an organization, or a business. When a customer subscribes or registers for a service offered by a CSP, a tenancy or account is created for that customer. The account then enables the customer to access one or more subscribed cloud resources associated with the account.

[0018] As mentioned above, infrastructure as a service (IaaS) is a specific type of cloud computing service. In the IaaS model, a CSP provides infrastructure (called cloud services provider infrastructure or CSPI) that can be used by customers to build their own customizable networks and deploy their resources. Therefore, the customer's resources and network are hosted in a distributed environment by the infrastructure provided by the CSP. This differs from traditional computing, where the customer's resources and network are hosted by the infrastructure provided by the customer.

[0019] CSPI can comprise interconnected, high-performance computing resources, including various host machines, memory resources, and network resources that form a physical network, also known as a substrate network or underlay network. Resources in CSPI can be distributed across one or more data centers, which can be geographically distributed across one or more geographic regions. Virtualization software can be run by these physical resources to provide a virtualized, distributed environment. This virtualization creates an overlay network (also known as a software-based network, software-defined network, or virtual network) on top of the physical network. The CSPI physical network provides the underlying foundation upon which one or more overlay or virtual networks can be created on top of the physical network. The physical network (or substrate or underlay network) includes physical network devices such as physical switches, routers, computers, and host machines. An overlay network is a logical (or virtual) network that operates on top of the physical substrate network. A particular physical network can support one or more overlay networks. Overlay networks typically use encapsulation techniques to distinguish traffic belonging to different overlay networks. A virtual network or overlay network is also called a virtual cloud network (VCN). A virtual network is implemented using software virtualization technologies (e.g., hypervisors, network virtualization devices (NVDs) (e.g., smart NICs), top-of-rack (TOR) switches, virtualization functions performed by smart TORs that perform one or more functions performed by NVDs, and other mechanisms) to create a layer of network abstraction that can run on top of a physical network. Virtual networks can take many forms, including peer-to-peer networks, IP networks, etc.Virtual networks are typically either Layer 3 IP networks or Layer 2 VLANs. This method of virtual or overlay networking is often called a virtual Layer 3 network or an overlay Layer 3 network. Examples of protocols developed for virtual networks include IP-in-IP (or Generic Routing Encapsulation (GRE)), Virtual Extensible LAN (VXLAN - IETF RFC 7348), Virtual Private Networks (VPN) (e.g., MPLS Layer 3 Virtual Private Network (RFC 4364)), VMware's NSX, and Generic Network Virtualization Encapsulation (GENEVE).

[0020] In the case of IaaS, the infrastructure provided by the CSP (CSPI) may be configured to provide virtualized computing resources over a public network (e.g., the Internet). In the IaaS model, a cloud computing service provider may host infrastructure components (e.g., servers, storage devices, network nodes (e.g., hardware), deployment software, platform virtualization (e.g., hypervisor layer), etc.). In some cases, the IaaS provider may offer various services (e.g., billing, monitoring, logging, security, load balancing, clustering, etc.) incidental to those infrastructure components. Accordingly, these services may be policy-driven, allowing IaaS users to implement policies to drive load balancing and maintain application availability and performance. CSPI provides infrastructure and a set of complementary cloud services that enable customers to build and run a wide range of applications and services within a highly available, hosted, distributed environment. CSPI provides high-performance computing resources and computing power as well as storage capacity within a flexible virtual network that can be securely accessed from various networked locations, such as from the customer's on-premises network. When a customer subscribes or registers for an IaaS service offered by a CSP, the tenancy created for that customer is a secure, isolated partition within the CSP where the customer can create, organize, and manage their cloud resources.

[0021] Customers can build their own virtual networks using the compute, memory, and network resources provided by CSPI. One or more customer resources or workloads, such as compute instances, can be deployed into these virtual networks. For example, customers can build one or more customizable private virtual networks called virtual cloud networks (VCNs) using resources provided by CSPI. Customers can deploy one or more customer resources, such as compute instances, into their VCNs. Compute instances can take the form of virtual machines, bare metal instances, etc. Thus, CSPI provides infrastructure and a set of complementary cloud services that enable customers to build and run a wide range of applications and services within a highly available, hosted virtual environment. While customers do not manage or control the underlying physical resources provided by CSPI, they have control over the operating system, storage, deployed applications, and in some cases, limited control over selected network components (e.g., firewalls).

[0022] The CSP may provide a console that allows customers and network administrators to configure, access, and manage resources deployed in the cloud using CSPI resources. In one embodiment, the console provides a web-based user interface that can be used to access and manage the CSPI. In some implementations, the console is a web-based application provided by the CSP.

[0023] CSPI may support single-tenancy or multi-tenancy architectures. In a single-tenancy architecture, a software component (e.g., an application, a database) or a hardware component (e.g., a host machine or server) serves a single customer or tenant. In a multi-tenancy architecture, a software component or hardware component serves multiple customers or tenants. Thus, in a multi-tenancy architecture, CSPI resources are shared among multiple customers or tenants. In a multi-tenancy situation, precautions are taken and safeguards are implemented within CSPI to ensure that each tenant's data remains isolated and invisible to other tenants.

[0024] In a physical network, a network endpoint (“endpoint”) refers to a computing device or system that is connected to the physical network and communicates with the connected network. Network endpoints in a physical network may be connected to a local area network (LAN), a wide area network (WAN), or other types of physical networks. Examples of traditional endpoints in a physical network include modems, hubs, bridges, switches, routers, and other network devices, physical computers (or host machines), and the like. Each physical device in a physical network has a fixed network address that can be used to communicate with the device. This fixed network address can be a Layer 2 address (e.g., a MAC address), a fixed Layer 3 address (e.g., an IP address), and the like. In a virtual environment or virtual network, endpoints can include various virtual endpoints, such as virtual machines hosted by components of the physical network (e.g., hosted by a physical host machine). These endpoints in the virtual network are addressed by overlay addresses, such as overlay Layer 2 addresses (e.g., an overlay MAC address) and overlay Layer 3 addresses (e.g., an overlay IP address). Network overlays enable flexibility by allowing network administrators to move between overlay addresses associated with network endpoints using software management (e.g., by software implementing the virtual network's control plane). Thus, unlike physical networks, in virtual networks, overlay addresses (e.g., overlay IP addresses) can be moved from one endpoint to another using network management software. Because virtual networks are built on top of physical networks, communication between components in a virtual network involves both the virtual network and the underlying physical network.To facilitate such communications, CSPI components are configured to learn and store mappings that map overlay addresses in the virtual network to actual physical addresses in the substrate network, and vice versa. These mappings are then used to facilitate communications. Customer traffic is encapsulated to facilitate routing within the virtual network.

[0025] Thus, a physical address (e.g., a physical IP address) is associated with a component in a physical network, and an overlay address (e.g., an overlay IP address) is associated with an entity in a virtual or overlay network. A physical IP address is an IP address associated with a physical device (e.g., a network device) in a substrate or physical network. For example, each NVD has an associated physical IP address. An overlay IP address is an overlay address associated with an entity in an overlay network, such as associated with a compute instance in a customer's virtual cloud network (VCN). Two different customers or tenants, each with their own private VCN, could potentially use the same overlay IP address in their VCN without any knowledge of each other. Both physical IP addresses and overlay IP addresses are types of real IP addresses. These IP addresses exist separately from virtual IP addresses. A virtual IP address is typically a single IP address that represents or maps to multiple real IP addresses. A virtual IP address provides a one-to-many mapping between a virtual IP address and multiple real IP addresses. For example, a load balancer may use a VIP to map to or represent multiple servers, each with its own real IP address.

[0026] A cloud infrastructure or CSPI is physically hosted in one or more data centers in one or more regions around the world. The CSPI may include components in a physical or substrate network and virtual components (e.g., virtual networks, compute instances, virtual machines, etc.) in a virtual network built on top of the physical network components. In one embodiment, the CSPI is organized and hosted in realms, regions, and availability domains. A region is a local geographic area that typically contains one or more data centers. Regions are generally independent of each other and may be separated by vast distances, for example, spanning multiple countries or continents. For example, a first region may be in Australia, another region may be in Japan, yet another region may be in India, etc. CSPI resources are divided among regions so that each region contains its own independent subset of CSPI resources. Each region may provide a set of core infrastructure services and resources, such as compute resources (e.g., bare metal servers, virtual machines, containers, and related infrastructure), storage resources (e.g., block volume storage, file storage, object storage, archive storage), network resources (e.g., virtual cloud networks (VCNs), load balancing resources, connectivity to on-premises networks), database resources, edge network resources (e.g., DNS), and access management and monitoring resources. Each region typically has multiple paths connecting it to other regions within the realm.

[0027] Because using nearby resources is faster than using resources that are farther away, applications are typically deployed in the region where they are most frequently used (i.e., deployed to the infrastructure associated with that region). Applications may also be deployed in different regions for a variety of reasons, such as redundancy to mitigate the risk of region-wide events such as large weather systems or earthquakes, to meet changing requirements for legal jurisdictions, tax areas, and other business or societal criteria.

[0028] Data centers within a region may be further organized and subdivided into availability domains (ADs). An availability domain may correspond to one or more data centers located within a region. A region may be composed of one or more availability domains. In such a distributed environment, CSPI resources are either specific to a region, such as a virtual cloud network (VCN), or specific to an availability domain, such as a compute instance.

[0029] ADs within a region are isolated from each other, fault-tolerant, and configured to be highly unlikely to fail simultaneously. This is achieved by ADs that do not share critical infrastructure resources, such as networks, physical cables, cable routes, or cable entry points, so that a failure in one AD within a region is unlikely to affect the availability of other ADs in the same region. ADs within the same region may be connected to each other by low-latency, high-bandwidth networks that provide highly available connectivity to other networks (e.g., the Internet, customer on-premises networks, etc.) and enable the creation of replicated systems in multiple ADs for both high availability and disaster recovery. Cloud services use multiple ADs to ensure high availability and protect against resource failures. As the infrastructure provided by an IaaS provider grows, more regions and ADs with additional capacity may be added. Traffic between availability domains is typically encrypted.

[0030] In one embodiment, regions are grouped into realms. A realm is a logical collection of regions. Realms are isolated from each other and do not share any data. Regions within the same realm may communicate with each other, but regions in different realms cannot. A customer's tenancy or account, along with a CSP, exists within a single realm and can be distributed across one or more regions belonging to that realm. Typically, when a customer subscribes to an IaaS service, a tenancy or account is created for the customer in a region designated by the customer within the realm (called the "home" region). The customer can extend their tenancy across one or more other regions within the realm. A customer cannot access regions that are not within the realm in which the customer's tenancy resides.

[0031] An IaaS provider may offer multiple realms, each catering to the requirements of a particular set of customers or users. For example, a commercial realm may be offered to commercial customers. As another example, a realm may be offered to a particular country for customers in that country. As yet another example, a government realm may be offered to a government, etc. For example, the government realm may cater to specific government requirements and may have higher security than the commercial realm. For example, Oracle Cloud Infrastructure (OCI) currently offers two realms: one for a commercial region and one for a government cloud region (e.g., FedRAMP-certified and IL5-certified).

[0032] In one embodiment, an AD can be subdivided into one or more failure domains. A failure domain is a group of infrastructure resources within an AD to provide anti-affinity. Fault domains enable the distribution of compute instances so that multiple compute instances do not reside on the same physical hardware within a single AD. This distribution is known as anti-affinity. A failure domain refers to a set of hardware components (computers, switches, etc.) that share a single point of failure. A compute pool is logically divided into failure domains. Therefore, a hardware failure or compute hardware maintenance event that affects one failure domain does not affect instances in other failure domains. Depending on the embodiment, the number of failure domains per AD may vary. For example, in one embodiment, each AD includes three failure domains. Fault domains act as logical data centers within an AD.

[0033] When a customer subscribes to an IaaS service, resources from CSPI are provisioned for the customer and associated with the customer's tenancy. The customer can use these provisioned resources to build private networks and deploy resources into these networks. A customer's network hosted in the cloud by CSPI is called a virtual cloud network (VCN). A customer can configure one or more virtual cloud networks (VCNs) using the CSPI resources allocated to the customer. A VCN is a virtual private network or software-defined private network. The customer's resources deployed within a customer's VCN can include compute instances (e.g., virtual machines, bare metal instances) and other resources. These compute instances may represent various customer workloads, such as applications, load balancers, databases, etc. Compute instances deployed in a VCN can communicate with endpoints publicly accessible over public networks such as the Internet ("public endpoints"), with other instances in the same VCN or other VCNs (e.g., other VCNs of the customer or VCNs not belonging to the customer), with the customer's on-premises data center or network, and with service endpoints and other types of endpoints.

[0034] CSPs may offer various services using CSPI. In some cases, customers of a CSPI may act as service providers themselves and offer services using CSPI resources. Service providers may expose service endpoints characterized by identifying information (e.g., IP addresses, DNS names, and DNS ports). Customer resources (e.g., compute instances) can consume a particular service by accessing the service endpoint exposed by the service for that service. These service endpoints are generally publicly accessible by users over a public communications network, such as the Internet, using the public IP address associated with the endpoint. Publicly accessible network endpoints are sometimes referred to as public endpoints.

[0035] In one embodiment, a service provider may expose a service via an endpoint for the service (sometimes referred to as a service endpoint). Customers of the service can then access the service using this service endpoint. In some implementations, a service endpoint provided for a service can be accessed by multiple customers wishing to consume the service. In other implementations, a dedicated service endpoint may be provided to a customer, allowing only that customer to access the service using that dedicated service endpoint.

[0036] In one embodiment, when a VCN is created, it is associated with a private overlay Classless Inter-Domain Routing (CIDR) address space, which is a range of private overlay IP addresses (e.g., 10.0 / 16) that is assigned to the VCN. A VCN includes associated subnets, route tables, and gateways. A VCN exists within a single region but can span one, more, or all of the region's availability domains. A gateway is a virtual interface configured for a VCN that enables traffic to and from the VCN to one or more endpoints outside the VCN. One or more different types of gateways may be configured for a VCN to enable communication with different types of endpoints.

[0037] A VCN can be subdivided into one or more subnetworks, such as one or more subnets. A subnet is thus a unit of configuration or subdivision that can be created within a VCN. A VCN can contain one or more subnets. Each subnet in a VCN is associated with a contiguous range of overlay IP addresses (e.g., 10.0.0.0 / 24 and 10.0.1.0 / 24) that represents a subset of address space within the VCN's address space and does not overlap with other subnets in that VCN.

[0038] Each compute instance is associated with a virtual network interface card (VNIC) that allows the compute instance to participate in a subnet of a VCN. A VNIC is a logical representation of a physical network interface card (NIC). In general, a VNIC is an interface between an entity (e.g., a compute instance, a service) and a virtual network. A VNIC resides in a subnet and has one or more associated IP addresses and associated security rules or policies. A VNIC is equivalent to a Layer 2 port on a switch. A VNIC is connected to a compute instance and to a subnet within a VCN. A VNIC associated with a compute instance allows the compute instance to become part of a subnet of a VCN and enables the compute instance to communicate (e.g., send and receive packets) with endpoints on the same subnet as the compute instance, endpoints in a different subnet within the VCN, or endpoints outside the VCN. Thus, the VNIC associated with a compute instance determines how the compute instance connects with endpoints inside and outside the VCN. A VNIC for a compute instance is created and associated with the compute instance when the compute instance is created and added to a subnet in a VCN. For a subnet containing a set of compute instances, the subnet contains VNICs corresponding to the set of compute instances, and each VNIC connects to one compute instance in the set of compute instances.

[0039] Each compute instance is assigned a private overlay IP address through the VNIC associated with the compute instance. This private overlay IP address is assigned to the VNIC associated with the compute instance when the compute instance is created and is used to route traffic to and from the compute instance. All VNICs within a particular subnet use the same route table, security lists, and DHCP options. As previously mentioned, each subnet within a VCN is associated with a contiguous range of overlay IP addresses (e.g., 10.0.0.0 / 24 and 10.0.1.0 / 24) that represents a subset of address space within the VCN's address space that does not overlap with other subnets within that VCN. For a VNIC on a particular subnet of a VCN, the private overlay IP address assigned to the VNIC is an address from the contiguous range of overlay IP addresses assigned to that subnet.

[0040] In one embodiment, a compute instance may optionally be assigned an additional overlay IP address in addition to the private overlay IP address, such as one or more public IP addresses if it is in a public subnet. These multiple addresses may be assigned to the same VNIC or across multiple VNICs associated with the compute instance. However, each instance has a primary VNIC that is created during instance launch and associated with the private overlay IP address assigned to the instance, and this primary VNIC cannot be removed. Additional VNICs, called secondary VNICs, may be added to an existing instance in the same availability domain as the primary VNIC. All VNICs are in the same availability domain as the instance. The secondary VNICs can be in a subnet in the same VCN as the primary VNIC or in a different subnet, either in the same VCN or in a different VCN.

[0041] If a compute instance is in a public subnet, the compute instance may optionally be assigned a public IP address. When a subnet is created, it can be specified as either a public subnet or a private subnet. A private subnet means that resources (e.g., compute instances) and associated VNICs in the subnet cannot have public overlay IP addresses. A public subnet means that resources and associated VNICs in the subnet can have public IP addresses. Customers can specify a subnet to exist in a single availability domain or across multiple availability domains within a region or realm.

[0042] As mentioned above, a VCN may be subdivided into one or more subnets. In one embodiment, a virtual router (VR) configured for a VCN (referred to as a VCN VR or simply VR) enables communication between subnets of the VCN. For a subnet within a VCN, the VR represents the logical gateway for that subnet, allowing the subnet (i.e., the compute instances on that subnet) to communicate with endpoints on other subnets within the VCN and with other endpoints outside the VCN. A VCN VR is a logical entity configured to route traffic between VNICs within a VCN and a virtual gateway ("gateway") associated with the VCN. Gateways are further described below with respect to FIG. 1. A VCN VR is a Layer 3 / IP layer concept. In one embodiment, there is one VCN VR per VCN, and the VCN VR has a potentially unlimited number of ports addressed by IP addresses, one port for each subnet of the VCN. In this way, the VCN VR has a different IP address for each subnet within the VCN to which the VCN VR is connected. The VRs are also connected to various gateways configured for the VCN. In one embodiment, a specific overlay IP address from a subnet's overlay IP address range is reserved for ports in that subnet's VCN VR. For example, consider a VCN that includes two subnets with associated address ranges 10.0 / 16 and 10.1 / 16, respectively. For the first subnet in the VCN with address range 10.0 / 16, an address from this range is reserved for ports in that subnet's VCN VR. In some cases, the first IP address from this range may be reserved for a VCN VR. For example, for a subnet with overlay IP address range 10.0 / 16, IP address 10.0.0.1 may be reserved for ports in that subnet's VCN VR.For a second subnet in the same VCN with address range 10.1 / 16, the VCN VR may have a port in that second subnet with IP address 10.1.0.1. The VCN VR has a different IP address for each of the subnets in the VCN.

[0043] In some other embodiments, each subnet in a VCN may include a VR associated with it that is addressable by the subnet using a reserved or default IP address associated with the VR. The reserved or default IP address may, for example, be the first IP address from a range of IP addresses associated with the subnet. VNICs in a subnet can use this default or reserved IP address to communicate with (e.g., send and receive packets from) the VR associated with the subnet. In such embodiments, a VR is an ingress / egress point for that subnet. VRs associated with a subnet in a VCN can communicate with other VRs associated with other subnets in the VCN. VRs can also communicate with gateways associated with the VCN. The VR functions for a subnet are running on or performed by one or more NVDs that are performing the VNIC functions for VNICs in the subnet.

[0044] Route tables, security rules, and DHCP options may be configured for a VCN. A route table is a virtual route table for a VCN and contains rules for routing traffic from subnets within the VCN to destinations outside the VCN via gateways or specially configured instances. A VCN's route table can be customized to control how packets are forwarded / routed to and from the VCN. DHCP options refer to configuration information that is automatically provided to instances when they launch.

[0045] Security rules configured for a VCN represent the overlay firewall rules for the VCN. Security rules include ingress and egress rules and can specify the type of traffic (e.g., based on protocol and port) allowed in and out of instances within the VCN. Customers can choose whether a particular rule is stateful or stateless. For example, a customer can allow incoming SSH traffic from any location to a set of instances by configuring a stateful ingress rule with source CIDR 0.0.0.0 / 0 and destination TCP port 22. Security rules can be implemented using network security groups or security lists. A network security group consists of a set of security rules that apply only to resources within that group. A security list, on the other hand, contains rules that apply to all resources in any subnet that uses the security list. A VCN may have a default security list that contains default security rules. DHCP options configured for a VCN provide configuration information that is automatically provided to instances within the VCN when the instances launch.

[0046] In one embodiment, configuration information for a VCN is determined and stored by a VCN control plane. The configuration information for a VCN may include, for example, information about address ranges associated with the VCN, subnets and associated information within the VCN, one or more VRs associated with the VCN, compute instances and associated VNICs within the VCN, NVDs (e.g., VNICs, VRs, gateways) performing various virtualized network functions associated with the VCN, VCN state information, and other VCN-related information. In one embodiment, a VCN distribution service publishes the configuration information stored by the VCN control plane or portions thereof to the NVD. The distributed information may be used to update information (e.g., forwarding tables, routing tables, etc.) stored and used by the NVD to forward packets to and from compute instances within the VCN.

[0047] In one embodiment, VCN and subnet creation is handled by a VCN Control Plane (CP), and compute instance launch is handled by the Compute Control Plane. The Compute Control Plane is responsible for allocating physical resources to compute instances and then calls the VCN Control Plane to create and attach VNICs to the compute instances. The VCN CP also sends VCN data mappings to a VCN Data Plane configured to perform packet forwarding and routing functions. In one embodiment, the VCN CP provides a distribution service that is responsible for providing updates to the VCN Data Plane. Examples of VCN Control Planes are also shown in Figures 12, 13, 14, and 15 (see reference numbers 1216, 1316, 1416, and 1516) and are described below.

[0048] A customer may create one or more VCNs using resources hosted by CSPI. Compute instances deployed in a customer's VCN may communicate with various endpoints. These endpoints may include endpoints hosted by CSPI and endpoints external to CSPI.

[0049] Various different architectures for implementing cloud-based services using CSPI are shown in Figures 1, 2, 3, 4, 5, and 12-16 and are described below. Figure 1 is a high-level diagram of a distributed environment 100 illustrating an overlay VCN or customer VCN hosted by CSPI according to one embodiment. The distributed environment shown in Figure 1 includes multiple components within an overlay network. The distributed environment 100 shown in Figure 1 is merely an example and is not intended to unduly limit the scope of the claimed embodiments. Many variations, alternatives, and modifications are possible. For example, in some implementations, the distributed environment shown in Figure 1 may include more or fewer systems or components than those shown in Figure 1, may combine two or more subsystems, or may include a different configuration or arrangement of systems.

[0050] As shown in the example depicted in FIG. 1 , distributed environment 100 includes CSPI 101, which provides services and resources that customers can subscribe to and use to build their own virtual cloud networks (VCNs). In one embodiment, CSPI 101 provides IaaS services to subscribing customers. Data centers within CSPI 101 may be organized into one or more regions. One exemplary region, “Region US” 102, is shown in FIG. 1 . A customer has configured a customer VCN with Oracle International Corporation for region 102. A customer may deploy various compute instances into VCN 104, which may include virtual machines or bare metal instances. Example instances include applications, databases, load balancers, etc.

[0051] In the embodiment shown in FIG. 1 , customer VCN 104 includes two subnets, "Subnet 1" and "Subnet 2," each with its own CIDR IP address range. In FIG. 1 , Subnet 1's overlay IP address range is 10.0 / 16, and Subnet 2's address range is 10.1 / 16. VCN virtual router 105 represents the VCN's logical gateway, enabling communication between subnets in VCN 104 and with other endpoints outside the VCN. VCN VR 105 is configured to route traffic between VNICs in VCN 104 and the gateway associated with VCN 104. VCN VR 105 provides a port for each subnet in VCN 104. For example, VR 105 may provide a port with IP address 10.0.0.1 for Subnet 1 and a port with IP address 10.1.0.1 for Subnet 2.

[0052] Multiple compute instances may be deployed in each subnet, and the compute instances can be virtual machine instances and / or bare metal instances. The compute instances in a subnet may be hosted by one or more host machines in CSPI101. A compute instance joins a subnet through a VNIC associated with the compute instance. For example, as shown in FIG. 1, compute instance C1 becomes part of subnet 1 through a VNIC associated with the compute instance. Similarly, compute instance C2 becomes part of subnet 1 through a VNIC associated with C2. In a similar manner, multiple compute instances, which may be virtual machine instances or bare metal instances, may become part of subnet 1. Each compute instance is assigned a private overlay IP address and a MAC address through its associated VNIC. For example, in FIG. 1, compute instance C1 has an overlay IP address of 10.0.0.2 and a MAC address of M1, while compute instance C2 has a private overlay IP address of 10.0.0.3 and a MAC address of M2. Each compute instance in Subnet 1, including compute instances C1 and C2, has a default route to VCN VR105 using IP address 10.0.0.1, which is the IP address of a port in VCN VR105 in Subnet 1.

[0053] Subnet2 may have multiple compute instances deployed, including virtual machine instances and / or bare metal instances. For example, as shown in FIG. 1, compute instances D1 and D2 become part of Subnet2 through VNICs associated with the respective compute instances. In the embodiment shown in FIG. 1, compute instance D1 has an overlay IP address of 10.1.0.2 and a MAC address of MM1, while compute instance D2 has a private overlay IP address of 10.1.0.3 and a MAC address of MM2. Each compute instance in Subnet2, including compute instances D1 and D2, has a default route to VCN VR105 using IP address 10.1.0.1, which is the IP address of a port in VCN VR105 in Subnet2.

[0054] VCN A 104 may include one or more load balancers. For example, a load balancer may be provided for a subnet and configured to load balance traffic across multiple compute instances on the subnet. A load balancer may be provided to load balance traffic across multiple subnets within a VCN.

[0055] A particular compute instance deployed in VCN 104 can communicate with various endpoints. These endpoints may include endpoints hosted by CSPI 200 and endpoints outside of CSPI 200. Endpoints hosted by CSPI 101 may include endpoints on the same subnet as the particular compute instance (e.g., communication between two compute instances in Subnet 1), endpoints on a different subnet but within the same VCN (e.g., communication between a compute instance in Subnet 1 and a compute instance in Subnet 2), endpoints in a different VCN within the same region (e.g., communication between a compute instance in Subnet 1 and an endpoint in a VCN within the same region 106 or 110, communication between a compute instance in Subnet 1 and an endpoint within the service network 110 within the same region), or endpoints in a VCN in a different region (e.g., communication between a compute instance in Subnet 1 and an endpoint in a VCN in a different region 108). Compute instances in a subnet hosted by CSPI 101 may communicate with endpoints not hosted by CSPI 101 (i.e., outside of CSPI 101). These external endpoints include endpoints within the customer's on-premise network 116, endpoints within other remote cloud-hosted networks 118, public endpoints 114 accessible via public networks such as the Internet, and other endpoints.

[0056] Communication between compute instances on the same subnet is facilitated using VNICs associated with the source and destination compute instances. For example, compute instance C1 in Subnet 1 may want to send a packet to compute instance C2 in Subnet 1. For a packet originating from the source compute instance and destined for another compute instance in the same subnet, the packet is first processed by the VNIC associated with the source compute instance. The processing performed by the VNIC associated with the source compute instance may include determining the packet's destination information from the packet header, identifying any policies (e.g., security lists) configured for the VNIC associated with the source compute instance, determining the packet's next hop, performing any packet encapsulation / decapsulation functions as needed, and then forwarding / routing the packet to the next hop to facilitate communication of the packet with its intended destination. If the destination compute instance is in the same subnet as the source compute instance, the VNIC associated with the source compute instance is configured to identify the VNIC associated with the destination compute instance and forward the packet to that VNIC for processing. The VNIC associated with the destination compute instance then executes and forwards the packet to the destination compute instance.

[0057] For packets traveling from a compute instance in a subnet to an endpoint in a different subnet within the same VCN, this communication is facilitated by the VNICs and VCN VRs associated with the source and destination compute instances. For example, if compute instance C1 in Subnet 1 in Figure 1 wants to send a packet to compute instance D1 in Subnet 2, the packet is first processed by the VNIC associated with compute instance C1. The VNIC associated with compute instance C1 is configured to route the packet to VCN VR105 using the VCN VR's default route or port 10.0.0.1. VCN VR105 is configured to route the packet to Subnet 2 using port 10.1.0.1. The packet is then received and processed by the VNIC associated with D1, which forwards the packet to compute instance D1.

[0058] For packets traveling from a compute instance within VCN 104 to an endpoint outside VCN 104, the communication is facilitated by a VNIC associated with the source compute instance, VCN VR 105, and a gateway associated with VCN 104. One or more types of gateways may be associated with VCN 104. A gateway is an interface between a VCN and another endpoint, where the other endpoint is outside the VCN. A gateway is a Layer 3 / IP layer concept that allows a VCN to communicate with endpoints outside the VCN. Thus, a gateway facilitates traffic flow between a VCN and other VCNs or networks. Different types of gateways may be configured for a VCN to facilitate different types of communication with different types of endpoints. Depending on the gateway, the communication may go over a public network (e.g., the Internet) or a private network. Various communication protocols may be used for these communications.

[0059] For example, compute instance C1 may wish to communicate with an endpoint outside VCN 104. The packet may first be processed by a VNIC associated with the source compute instance C1. This VNIC processing determines that the packet's destination is outside of Subnet 1 of C1. The VNIC associated with C1 may forward the packet to VCN VR105 of VCN 104. VCN VR105 then processes the packet and, as part of this processing, determines a particular gateway associated with VCN 104 as the packet's next hop based on the packet's destination. VCN VR105 may then forward the packet to the particular identified gateway. For example, if the destination is an endpoint within a customer's on-premises network, VCN VR105 may forward the packet to a dynamic routing gateway (DRG) gateway 122 configured for VCN 104. The packet may then be forwarded from the gateway to the next hop to facilitate propagation of the packet to its final intended destination.

[0060] Various types of gateways may be configured for a VCN. Examples of gateways that may be configured for a VCN are shown in FIG. 1 and described below. Examples of gateways associated with a VCN are also shown in FIGS. 12, 13, 14, and 15 (e.g., gateways referenced by reference numbers 1234, 1236, 1238, 1334, 1336, 1338, 1434, 1436, 1438, 1534, 1536, and 1538) and described below. As shown in the embodiment shown in FIG. 1, a dynamic routing gateway (DRG) 122 may be added to or associated with a customer's VCN 104 to provide a path for private network traffic communication between the customer's VCN 104 and another endpoint, which could be the customer's on-premises network 116, a VCN 108 in a different region of CSPI 101, or another remote cloud network 118 not hosted by CSPI 101. The customer on-premises network 116 may be a customer network or a customer data center built using the customer's resources. Access to the customer on-premises network 116 is typically highly restricted. For customers with both a customer on-premises network 116 and one or more VCNs 104 deployed or hosted in the cloud by CSPI 101, the customer may want the customer on-premises network 116 and the customer's cloud-based VCNs 104 to be able to communicate with each other. This allows the customer to build an extended hybrid environment that encompasses the customer VCNs 104 hosted by CSPI 101 and the customer on-premises network 116. The DRG 122 enables this communication. To enable such communication, a communication channel 124 is set up, with one endpoint of the channel in the customer on-premises network 116 and the other endpoint in CSPI 101 connected to the customer VCN 104. The communication channel 124 can traverse a public or private communication network, such as the Internet.Various communication protocols may be used, such as IPsec VPN technology over a public communication network such as the Internet, or Oracle's FastConnect technology, which uses a private network instead of a public network. A device or equipment in the customer's on-premises network 116 that forms one endpoint of the communication channel 124 is called customer premise equipment (CPE), such as CPE 126 shown in Figure 1. On the CSPI 101 side, the endpoint may be a host machine running DRG 122.

[0061] In one embodiment, a Remote Peering Connection (RPC) can be added to a DRG, allowing a customer to peer one VCN with another VCN in a different region. Using such an RPC, a customer's VCN 104 can connect with a VCN 108 in another region using the DRG 122. The DRG 122 may also be used to communicate with other remote cloud networks 118 not hosted by CSPI 101, such as the Microsoft Azure cloud, the Amazon AWS cloud, etc.

[0062] 1, an Internet Gateway (IGW) 120 may be configured for a customer's VCN 104, enabling compute instances on VCN 104 to communicate with public endpoints 114 accessible over a public network, such as the Internet. IGW 120 is a gateway that connects a VCN to a public network, such as the Internet. IGW 120 enables direct access for public subnets within a VCN, such as VCN 104 (resources within the public subnet have public overlay IP addresses) to public endpoints 112 on the public network 114, such as the Internet. Using IGW 120, connections can be initiated from subnets within VCN 104 or from the Internet.

[0063] A Network Address Translation (NAT) gateway 128 is configured for customer VCN 104 to enable access to the Internet for cloud resources in the customer's VCN that do not have dedicated public overlay IP addresses, without exposing those resources to direct incoming Internet connections (e.g., L4-L7 connections). This allows private subnets in the VCN, such as Private Subnet 1 in VCN 104, to private endpoints on the Internet. The NAT gateway only allows connections to be initiated from the private subnet to the public Internet; connections cannot be initiated from the Internet to the private subnet.

[0064] In one embodiment, a service gateway (SGW) 126 may be configured for a customer's VCN 104 and provides a pathway for private network traffic between VCN 104 and supported service endpoints in service network 110. In one embodiment, service network 110 may be provided by a CSP and may offer a variety of services. An example of such a service network is Oracle's Services Network, which offers a variety of services that may be used by customers. For example, a compute instance (e.g., a database system) in a private subnet of a customer's VCN 104 can back up data to a service endpoint (e.g., object storage) without requiring a public IP address or access to the Internet. In one embodiment, a VCN may include only one SGW, and connections can be initiated only from subnets within the VCN, not from service network 110. When a VCN is peered with another VCN, resources in the other VCN typically do not have access to the SGW. Resources in an on-premises network connected to a VCN using a FastConnect or VPN connection can also use a service gateway configured for that VCN.

[0065] In one implementation, SGW 126 uses the concept of a service classless inter-domain routing (CIDR) label, which is a string that represents the public IP address range of all regions for a service or group of services of interest. A customer uses the service CIDR label when configuring the SGW and associated route rules to control traffic to the service. A customer can optionally utilize the service CIDR label when configuring security rules, avoiding the need to adjust those security rules if the service's public IP address changes in the future.

[0066] A Local Peering Gateway (LPG) 132 is a gateway that can be added to a customer's VCN 104, allowing the VCN 104 to peer with another VCN in the same region. Peering means that the VCNs communicate using private IP addresses without the traffic traversing a public network such as the Internet or routing the traffic through the customer's on-premises network 116. In a preferred embodiment, a VCN includes a separate LPG for each peering it establishes. Local peering or VCN peering is a common method used to establish network connectivity between different applications or infrastructure management functions.

[0067] A service provider, such as a provider of a service in service network 110, may provide access to the service using various access models. According to a public access model, the service may be exposed as a public endpoint, which may be publicly accessible by compute instances in the customer's VCN over a public network such as the Internet, and / or privately accessible through SGW 126. According to a specific private access model, the service is made accessible as a private IP endpoint in a private subnet in the customer's VCN. This access is called Private Endpoint (PE) access and allows service providers to expose services as instances in the customer's private network. A private endpoint resource represents a service in a customer's VCN. Each PE appears as a VNIC (referred to as a PE-VNIC with one or more private IPs) in a customer-selected subnet in the customer's VCN. Thus, the PE provides a way to present services in a subnet of the private customer's VCN using VNICs. Because the endpoints are exposed as VNICs, all features associated with the VNIC, such as routing rules, security lists, etc., are available to the PE VNIC.

[0068] A service provider can register a service to make it accessible through the PE. The provider can associate a policy with the service that limits the visibility of the service to the customer's tenancy. The provider can register multiple services under a single virtual IP address (VIP), especially for multi-tenant services. There can be multiple such private endpoints (in multiple VCNs) representing the same service.

[0069] Compute instances in the private subnet can then access the service using the private IP address of the PE VNIC or the DNS name of the service. Compute instances in a customer's VCN can access the service by sending traffic to the private IP address of the PE in the customer's VCN. A Private Access Gateway (PAGW) 130 is a gateway resource that can be connected to a service provider's VCN (e.g., a VCN in service network 110) and serves as the ingress / egress point for all traffic to and from private endpoints in customer subnets. The PAGW 130 allows providers to scale the number of PE connections without utilizing internal IP address resources. A provider needs to configure only one PAGW for any number of services registered within a single VCN. A provider can represent services as private endpoints in multiple VCNs for one or more customers. From the customer's perspective, the PE VNIC appears to be connected to the service with which the customer wants to interact, instead of to a customer instance. Traffic going to the private endpoint is routed to the service through the PAGW 130. These are called customer-to-service private connections (C2S (customer-to-service) connections).

[0070] The PE concept can also be used to extend private access of services to customer on-premises networks and data centers by allowing traffic to flow through FastConnect / IPsec links and private endpoints in the customer's VCN. Private access of services can also be extended to customer peered VCNs by allowing traffic to flow between LPG 132 and PEs in the customer's VCN.

[0071] A customer can control routing within a VCN at the subnet level, allowing the customer to specify which subnets within a customer's VCN, such as VCN 104, use each gateway. A VCN's route tables are used to determine whether traffic is allowed to exit a VCN through a particular gateway. For example, in a particular example, a route table for a public subnet within a customer's VCN 104 may send non-local traffic through IGW 120. A route table for a private subnet within the same customer's VCN 104 may send traffic going to a CSP service through SGW 126. All remaining traffic may be sent through NAT gateway 128. Route tables only control traffic that exits a VCN.

[0072] Security lists associated with a VCN are used to control traffic entering the VCN through the gateway via inbound connections. All resources within a subnet use the same route table and security lists. Security lists may be used to control the specific types of traffic allowed into and out of instances within a VCN's subnet. Security list rules may include ingress (inbound) rules and egress (outbound) rules. For example, ingress rules may specify allowed source address ranges, while egress rules may specify allowed destination address ranges. Security rules may specify a specific protocol (e.g., TCP, ICMP), a specific port (e.g., 22 for SSH, 3389 for Windows RDP), etc. In some implementations, the instance's operating system may enforce its own firewall rules that match the security list rules. Rules may be stateful (e.g., connections are tracked and responses are automatically allowed without explicit security list rules for the response traffic) or stateless.

[0073] Access from a customer's VCN (i.e., by resources or compute instances deployed in VCN 104) can be categorized as public access, private access, or dedicated access. Public access refers to an access model in which public IP addresses or NATs are used to access public endpoints. Private access allows customer workloads in VCN 104 with private IP addresses (e.g., resources in a private subnet) to access services without traversing a public network such as the Internet. In one embodiment, CSPI 101 enables workloads in a customer's VCN with private IP addresses to access (public service endpoints of) services using a service gateway. Thus, the service gateway provides a private access model by establishing a virtual link between the customer's VCN and the service's public endpoints, which reside outside the customer's private network.

[0074] Additionally, CSPI may provide dedicated public access using technologies such as FastConnect public peering, where a customer's on-premises instances can use a FastConnect connection to access one or more services in the customer's VCN without traversing a public network such as the internet. CSPI may also provide dedicated private access using FastConnect private peering, where a customer's on-premises instances with private IP addresses can use a FastConnect connection to access workloads in the customer's VCN. FastConnect is a network connectivity alternative to using the public internet for connecting a customer's on-premises network to CSPI and its services. FastConnect provides an easy, resilient, and economical way to create dedicated private connections with higher bandwidth options and a more reliable and consistent network experience when compared to internet-based connections.

[0075] FIG. 1 and the accompanying discussion above describe various virtual components within an exemplary virtual network. As previously mentioned, a virtual network is built on top of an underlying physical or substrate network. FIG. 2 illustrates a simplified architectural diagram of physical components in a physical network within CSPI 200 that serves as the foundation for a virtual network, according to one embodiment. As illustrated, CSPI 200 provides a distributed environment that includes components and resources (e.g., compute, memory, and network resources) provided by a cloud service provider (CSP). These components and resources are used to provide cloud services (e.g., IaaS services) to subscribing customers, i.e., customers who subscribe to one or more services offered by the CSP. Based on the services to which the customer subscribes, a subset of CSPI 200's resources (e.g., compute, memory, and network resources) is provisioned for the customer. The customer can then build their own cloud-based (i.e., CSPI-hosted), customizable private virtual network using the physical compute, memory, and network resources provided by CSPI 200. As previously indicated, these customer networks are called virtual cloud networks (VCNs). Customers can deploy one or more customer resources, such as compute instances, into these customer VCNs. The compute instances can be in the form of virtual machines, bare metal instances, etc. CSPI 200 provides infrastructure and a set of complementary cloud services that enable customers to build and run a wide range of applications and services within a highly available hosted environment.

[0076] In the example embodiment shown in FIG. 2, the physical components of CSPI 200 include one or more physical host machines or servers (e.g., 202, 206, 208), network virtualization devices (NVDs) (e.g., 210, 212), top-of-rack (TOR) switches (e.g., 214, 216), and a physical network (e.g., 218), as well as switches within physical network 218. The physical host machines or servers may host and execute various compute instances that participate in one or more subnets of the VCN. The compute instances may include virtual machine instances and bare metal instances. For example, the various compute instances shown in FIG. 1 may be hosted by the physical host machines shown in FIG. 2. The virtual machine compute instances in the VCN may be executed by one host machine or by multiple different host machines. The physical host machines may host virtual host machines, container-based hosts or functions, etc. The VNICs and VCN VRs shown in FIG. 1 may be executed by the NVDs shown in FIG. 2. The gateway shown in FIG. 1 may be executed by a host machine and / or by the NVD shown in FIG.

[0077] A host machine or server may run a hypervisor (also called a virtual machine monitor or VMM) that creates and enables a virtual environment on the host machine. The virtualized environment facilitates cloud-based computing. One or more compute instances may be created, run, and managed on the host machine by the hypervisor on the host machine. The hypervisor on the host machine enables the host machine's physical computing resources (e.g., compute, memory, and network resources) to be shared among the various compute instances executed by the host machine.

[0078] For example, as shown in FIG. 2, host machines 202 and 208 execute hypervisors 260 and 266, respectively. These hypervisors may be implemented using software, firmware, or hardware, or a combination thereof. Typically, a hypervisor is a process or software layer that resides on a host machine's operating system (OS), which executes on the host machine's hardware processor. A hypervisor provides a virtual environment by allowing the host machine's physical computing resources (e.g., processing resources such as processors / cores, memory resources, and network resources) to be shared among various virtual machine computing instances executed by the host machine. For example, in FIG. 2, hypervisor 260 may reside on top of host machine 202's OS and allow host machine 202's computing resources (e.g., processing resources, memory resources, and network resources) to be shared among computing instances (e.g., virtual machines) executed by host machine 202. A virtual machine can have its own operating system (referred to as a guest operating system), which may be the same as or different from the host machine's OS. The operating system of a virtual machine executed by a host machine may be the same as or different from the operating system of another virtual machine executed by the same host machine. Thus, a hypervisor allows multiple operating systems to run side by side with each other while sharing the same computing resources of the host machine. The host machines shown in Figure 2 may have the same or different types of hypervisors.

[0079] A compute instance can be a virtual machine instance or a bare metal instance. In Figure 2, compute instance 268 on host machine 202 and compute instance 274 on host machine 208 are examples of virtual machine instances. Host machine 206 is an example of a bare metal instance being offered to a customer.

[0080] In some examples, an entire host machine may be provisioned to a single customer, and one or more compute instances (either virtual machines or bare metal instances) hosted by that host machine all belong to that same customer. In other examples, a host machine may be shared among multiple customers (i.e., multiple tenants). In such a multi-tenancy situation, a host machine may host virtual machine compute instances belonging to different customers. These compute instances may be members of different VCNs for different customers. In some embodiments, bare metal compute instances are hosted by bare metal servers that do not have a hypervisor. When a bare metal compute instance is provisioned, a single customer or tenant maintains control of the physical CPU, memory, and network interfaces of the host machine hosting the bare metal instance, and the host machine is not shared with other customers or tenants.

[0081] As previously mentioned, each compute instance that is part of a VCN is associated with a VNIC that enables the compute instance to be a member of a subnet of the VCN. The VNIC associated with a compute instance facilitates communication of packets or frames to and from the compute instance. The VNIC is associated with the compute instance when the compute instance is created. In one embodiment, for a compute instance executed by a host machine, the VNIC associated with the compute instance is executed by an NVD connected to the host machine. For example, in FIG. 2, host machine 202 executes virtual machine compute instance 268 that is associated with VNIC 276, which is executed by NVD 210 connected to host machine 202. As another example, bare metal instance 272 hosted by host machine 206 is associated with VNIC 280, which is executed by NVD 212 connected to host machine 206. As yet another example, VNIC 284 is associated with compute instance 274 executed by host machine 208, which is executed by NVD 212 connected to host machine 208.

[0082] For compute instances hosted by a host machine, the NVD connected to that host machine also executes VCN VRs corresponding to the VCNs of which those compute instances are members. For example, in the embodiment shown in Figure 2, NVD 210 executes VCN VR 277 corresponding to the VCN of which compute instance 268 is a member. NVD 212 may execute one or more VCN VRs 283 corresponding to the VCNs corresponding to the compute instances hosted by host machines 206 and 208.

[0083] A host machine may include one or more network interface cards (NICs) that allow the host machine to be connected to other devices. The NICs on the host machine may provide one or more ports (or interfaces) that allow the host machine to be communicatively connected to another device. For example, a host machine may be connected to an NVD using one or more ports (or interfaces) provided on the host machine and the NVD. A host machine may also be connected to other devices, such as another host machine.

[0084] 2, host machine 202 is connected to NVD 210 using link 220 extending between port 234 provided by NIC 232 of host machine 202 and port 236 of NVD 210. Host machine 206 is connected to NVD 212 using link 224 extending between port 246 provided by NIC 244 of host machine 206 and port 248 of NVD 212. Host machine 208 is connected to NVD 212 using link 226 extending between port 252 provided by NIC 250 of host machine 208 and port 254 of NVD 212.

[0085] The NVDs are then connected via communication links to top-of-rack (TOR) switches (also called switch fabrics) that are connected to a physical network 218. In one embodiment, the links between the host machines and the NVDs and between the NVDs and the TOR switches are Ethernet links. For example, in Figure 2, links 228 and 230 are used to connect NVDs 210 and 212 to TOR switches 214 and 216, respectively. In one embodiment, links 220, 224, 226, 228, and 230 are Ethernet links. The collection of host machines and NVDs connected to a TOR may be referred to as a rack.

[0086] The physical network 218 provides a communications fabric that allows the TOR switches to communicate with each other. The physical network 218 can be a multi-tier network. In one implementation, the physical network 218 is a multi-tier Clos network of switches, with the TOR switches 214 and 216 representing leaf-level nodes of the multi-tier, multi-node physical switching network 218. Various Clos network configurations are possible, including, but not limited to, 2-tier networks, 3-tier networks, 4-tier networks, 5-tier networks, and generally, "n"-tier networks. An example Clos network is shown in FIG. 5 and described below.

[0087] Various connection configurations are possible between host machines and NVDs, such as one-to-one, many-to-one, and one-to-many configurations. In a one-to-one implementation, each host machine is connected to its own separate NVD. For example, in FIG. 2, host machine 202 is connected to NVD 210 via NIC 232 on host machine 202. In a many-to-one configuration, multiple host machines are connected to a single NVD. For example, in FIG. 2, host machines 206 and 208 are connected to the same NVD 212 via NICs 244 and 250, respectively.

[0088] In a one-to-many configuration, one host machine is connected to multiple NVDs. FIG. 3 illustrates an example of a CSPI 300 in which a host machine is connected to multiple NVDs. As illustrated in FIG. 3, a host machine 302 includes a network interface card (NIC) 304 including multiple ports 306 and 308. The host machine 300 is connected to a first NVD 310 via port 306 and link 320, and to a second NVD 312 via port 308 and link 322. Ports 306 and 308 may be Ethernet ports, and links 320 and 322 between the host machine 302 and the NVDs 310 and 312 may be Ethernet links. The NVD 310 is then connected to a first TOR switch 314, and the NVD 312 is connected to a second TOR switch 316. The links between the NVDs 310 and 312 and the TOR switches 314 and 316 may be Ethernet links. TOR switches 314 and 316 represent layer 0 switching devices within a multi-tier physical network 318 .

[0089] 3 provides two separate physical network paths between the physical switch network 318 and the host machine 302: a first path traversing the TOR switch 314, through the NVD 310, and to the host machine 302, and a second path traversing the TOR switch 316, through the NVD 312, and to the host machine 302. The separate paths result in improved availability (referred to as high availability) of the host machine 302. If there is a problem with one of the paths (e.g., a link in one of the paths fails) or devices (e.g., a particular NVD is not functioning), the other path may be used for communication to and from the host machine 302.

[0090] In the configuration shown in Figure 3, the host machine is connected to two different NVDs using two different ports provided by the host machine's NIC. In other embodiments, the host machine may include multiple NICs that allow the host machine to be connected to multiple NVDs.

[0091] Referring again to Figure 2, an NVD is a physical device or component that performs one or more network and / or storage virtualization functions. An NVD may be any device that has one or more processing units (e.g., a CPU, Network Processing Units (NPUs), FPGAs, packet processing pipelines, etc.), memory including cache, and ports. Various virtualization functions may be performed by software / firmware executed by the one or more processing units of the NVD.

[0092] The NVD may be implemented in various forms. For example, in one embodiment, the NVD is implemented as an interface card called a smart NIC or intelligent NIC that contains an embedded processor. A smart NIC is a separate device from the NIC on the host machine. In Figure 2, NVDs 210 and 212 may be implemented as smart NICs connected to host machine 202 and host machines 206 and 208, respectively.

[0093] However, a smart NIC is just one example of an implementation of an NVD. Various other implementations are possible. For example, in some other implementations, the NVD or one or more functions performed by the NVD may be incorporated into or performed by one or more host machines, one or more TOR switches, and other components of CSPI200. For example, the NVD may be embodied in a host machine, and the functions performed by the NVD may be performed by the host machine. As another example, the NVD may be part of a TOR switch, or a TOR switch may be configured to perform the functions performed by the NVD, allowing the TOR switch to perform various complex packet transformations used in public clouds. A TOR that performs the functions of an NVD may be referred to as a smart TOR. In yet other implementations where customers are provided with virtual machine (VM) instances rather than bare metal (BM) instances, the functions performed by the NVD may be implemented inside the hypervisor of the host machine. In some other implementations, some of the NVD's functions may be offloaded to a centralized service running on a set of host machines.

[0094] In one embodiment, such as when implemented as a smart NIC as shown in FIG. 2, the NVD may include multiple physical ports that allow the NVD to be connected to one or more host machines and one or more TOR switches. Ports on the NVD may be classified as host-facing ports (also referred to as "south ports") or network-facing or TOR-facing ports (also referred to as "north ports"). Host-facing ports of an NVD are ports used to connect the NVD to host machines. Examples of host-facing ports in FIG. 2 include port 236 on NVD 210 and ports 248 and 254 on NVD 212. Network-facing ports of an NVD are ports used to connect the NVD to TOR switches. Examples of network-facing ports in FIG. 2 include port 256 on NVD 210 and port 258 on NVD 212. As shown in FIG. 2, the NVD 210 is connected to the TOR switch 214 using link 228 extending from port 256 of the NVD 210 to the TOR switch 214. Similarly, the NVD 212 is connected to the TOR switch 216 using a link 230 that extends from a port 258 of the NVD 212 to the TOR switch 216 .

[0095] The NVD may receive packets and frames from the host machine (e.g., packets and frames generated by compute instances hosted by the host machine) via a host-facing port, and after performing any necessary packet processing, may forward those packets and frames to the TOR switch via the NVD's network-facing port. The NVD may receive packets and frames from the TOR switch via the NVD's network-facing port, and after performing any necessary packet processing, may forward those packets and frames to the host machine via the NVD's host-facing port.

[0096] In one embodiment, there may be multiple ports and associated links between the NVD and the TOR switch. These ports and links may be aggregated to form a link aggregator group (called a LAG) of multiple ports or links. Link aggregation allows multiple physical links between two endpoints (e.g., between the NVD and the TOR switch) to be treated as a single logical link. All physical links within a particular LAG may operate in full-duplex mode at the same speed. LAGs help increase bandwidth and improve the reliability of the connection between two endpoints. If one of the physical links in the LAG fails, traffic is dynamically and transparently reassigned to one of the other physical links in the LAG. The aggregated physical link provides higher bandwidth than an individual link. Multiple ports associated with a LAG are treated as a single logical port. Traffic can be load-balanced across the multiple physical links in a LAG. One or more LAGs may be configured between two endpoints. Two endpoints may exist between the NVD and the TOR switch, between a host machine and the NVD, etc.

[0097] The NVD implements or performs network virtualization functions. These functions are performed by software / firmware executed by the NVD. Examples of network virtualization functions include, but are not limited to, packet encapsulation and decapsulation functions, functions for creating VCN networks, functions for enforcing network policies such as VCN security list (firewall) functions, and functions for facilitating routing and forwarding of packets to and from compute instances in the VCN. In one embodiment, upon receiving a packet, the NVD is configured to execute a packet processing pipeline to process the packet and determine how the packet should be forwarded or routed. As part of this packet processing pipeline, the NVD may perform one or more virtual functions associated with the overlay network, such as running VNICs associated with compute instances in the VCN, running virtual routers (VRs) associated with the VCN, encapsulating and decapsulating packets to facilitate forwarding or routing within the virtual network, running certain gateways (e.g., local peering gateways), enforcing security lists, network security groups, network address translation (NAT) functions (e.g., per-host public IP to private IP translation), bandwidth throttling functions, and other functions.

[0098] In one embodiment, the packet processing data path within the NVD may comprise multiple packet pipelines, each consisting of a series of packet transformation stages. In one implementation, upon receipt of a packet, the packet is parsed and sorted into a single pipeline. The packet is then processed in a linear fashion, one stage at a time, until the packet is either dropped or transmitted through an interface of the NVD. These stages provide basic functional packet processing building blocks (e.g., header validation, performing bandwidth throttling, inserting a new Layer 2 header, performing L4 firewalling, VCN encapsulation / decapsulation, etc.), such that new pipelines can be constructed by assembling existing stages, and new functionality can be added by creating and inserting new stages into existing pipelines.

[0099] The NVD may perform both control plane and data plane functions corresponding to the control and data planes of a VCN. Examples of a VCN control plane are also shown in Figures 12, 13, 14, and 15 (see reference numbers 1216, 1316, 1416, and 1516) and are described below. Examples of a VCN data plane are shown in Figures 12, 13, 14, and 15 (see reference numbers 1218, 1318, 1418, and 1518) and are described below. Control plane functions include functions used to configure the network (e.g., set up routes and route tables, configure VNICs, etc.) that control how data is forwarded. In one embodiment, a VCN control plane is provided that centrally computes mappings between all overlays and substrates and publishes these mappings to the NVD and to virtual network edge devices such as various gateways, such as DRGs, SGWs, and IGWs. Firewall rules may also be published using the same mechanism. In one embodiment, the NVD retrieves only mappings that are relevant to the NVD. The data plane functions include functionality for the actual routing / forwarding of packets based on configuration settings using the control plane. The VCN data plane is implemented by encapsulating customer network packets before they traverse the substrate network. The encapsulation / decapsulation functions are implemented in the NVD. In one embodiment, the NVD is configured to intercept all network packets entering and leaving the host machine and perform network virtualization functions.

[0100] As indicated above, the NVD performs various virtualization functions, including VNICs and VCN VRs. The NVD may execute VNICs associated with compute instances hosted by one or more host machines connected to the VNICs. For example, as shown in FIG. 2, NVD 210 executes the functions of VNIC 276 associated with compute instance 268 hosted by host machine 202 connected to NVD 210. As another example, NVD 212 executes VNIC 280 associated with bare metal compute instance 272 hosted by host machine 206 and VNIC 284 associated with compute instance 274 hosted by host machine 208. The host machines may host compute instances belonging to different VCNs that belong to different customers, and the NVDs connected to the host machines may execute VNICs (i.e., perform functions related to the VNICs) corresponding to the compute instances.

[0101] NVDs also execute VCN virtual routers corresponding to the VCNs of the compute instances. For example, in the embodiment shown in FIG. 2, NVD 210 executes VCN VR 277 corresponding to the VCN to which compute instance 268 belongs. NVD 212 executes one or more VCN VRs 283 corresponding to one or more VCNs to which compute instances hosted by host machines 206 and 208 belong. In one embodiment, a VCN VR corresponding to that VCN is executed by all NVDs connected to a host machine that hosts at least one compute instance belonging to that VCN. If a host machine hosts compute instances that belong to different VCNs, the NVDs connected to that host machine may execute VCN VRs corresponding to those different VCNs.

[0102] In addition to VNICs and VCN VRs, the NVD may run various software (e.g., daemons) and may include one or more hardware components that facilitate the various network virtualization functions performed by the NVD. For simplicity, these various components are grouped together as a “packet processing component” shown in FIG. 2 . For example, NVD 210 includes packet processing component 286, and NVD 212 includes packet processing component 288. For example, the packet processing component of the NVD may include a packet processor configured to interact with the NVD's ports and hardware interfaces, monitor all packets received by and transmitted using the NVD, and store network information. The network information may include, for example, network flow information and per-flow information (e.g., per-flow statistics) that identify the various network flows processed by the NVD. In one embodiment, the network flow information may be stored per VNIC. In addition to performing per-packet operations, the packet processor may implement a stateful NAT and an L4 firewall (FW). As another example, the packet processing component may include a replication agent configured to replicate information stored by the NVD to one or more different replication target stores. As yet another example, the packet processing component may include a logging agent configured to perform the logging functions of the NVD. The packet processing component may also include software for monitoring the performance and health of the NVD, and possibly software for monitoring the status and health of other components connected to the NVD.

[0103] FIG. 1 illustrates components of an exemplary virtual network or overlay network, including a VCN, subnets within the VCN, compute instances deployed to the subnets, VNICs associated with the compute instances, VRs for the VCN, and a set of gateways configured for the VCN. The overlay components illustrated in FIG. 1 may be executed or hosted by one or more of the physical components illustrated in FIG. 2. For example, compute instances within a VCN may be executed or hosted by one or more host machines illustrated in FIG. 2. For compute instances hosted by a host machine, the VNICs associated with the compute instances are typically executed by an NVD connected to the host machine (i.e., the VNIC functionality is provided by the NVD connected to the host machine). The VCN VR functionality for the VCN is performed by all NVDs connected to host machines hosting or running compute instances that are part of the VCN. The gateways associated with a VCN may be executed by one or more different types of NVDs. For example, some gateways may be executed by smart NICs, while other gateways may be executed by one or more host machines or other implementations of NVDs.

[0104] As previously mentioned, compute instances within a customer's VCN may communicate with various endpoints, which can be in the same subnet as the source compute instance, or in a different subnet within the same VCN as the source compute instance, or the endpoints are outside the VCN of the source compute instance. These communications are facilitated using VNICs associated with the compute instances, VCN VRs, and gateways associated with the VCN.

[0105] For communication between two compute instances on the same subnet within a VCN, the communication is facilitated using VNICs associated with the source and destination compute instances. The source and destination compute instances may be hosted by the same host machine or by different host machines. A packet originating from the source compute instance may be forwarded from the host machine hosting the source compute instance to an NVD connected to that host machine. In the NVD, the packet is processed using a packet processing pipeline, which may include execution of the VNIC associated with the source compute instance. Because the packet's destination endpoint is within the same subnet, execution of the VNIC associated with the source compute instance causes the packet to be forwarded to an NVD running the VNIC associated with the destination compute instance, which then processes the packet and forwards it to the destination compute instance. The VNICs associated with the source and destination compute instances may run on the same NVD (e.g., when the source and destination compute instances are both hosted by the same host machine) or on different NVDs (e.g., when the source and destination compute instances are hosted by different host machines connected to different NVDs). The VNICs may use routing / forwarding tables stored by the NVDs to determine the next hop of a packet.

[0106] For packets traveling from a compute instance in a subnet to an endpoint in a different subnet within the same VCN, the packet originating from the source compute instance travels from the host machine hosting the source compute instance to the NVD connected to that host machine. In the NVD, the packet is processed using a packet processing pipeline, which may include running one or more VNICs and VRs associated with the VCN. For example, as part of the packet processing pipeline, the NVD executes or invokes a function corresponding to the VNIC associated with the source compute instance (also referred to as executing the VNIC). The function executed by the VNIC may include examining the VLAN tag on the packet. Because the packet's destination is outside the subnet, a VCN VR function is then invoked and executed by the NVD. The VCN VR then routes the packet to the NVD running the VNIC associated with the destination compute instance. The VNIC associated with the destination compute instance then processes the packet and forwards the packet to the destination compute instance. The VNICs associated with the source compute instance and the destination compute instance may run on the same NVD (e.g., when the source compute instance and the destination compute instance are both hosted by the same host machine) or on different NVDs (e.g., when the source compute instance and the destination compute instance are hosted by different host machines connected to different NVDs).

[0107] If the packet's destination is outside the VCN of the source compute instance, the packet originating from the source compute instance is propagated from the host machine hosting the source compute instance to the NVD connected to that host machine. The NVD runs the VNIC associated with the source compute instance. Because the packet's destination endpoint is outside the VCN, the packet is then processed by the VCN VR for that VCN. The NVD may invoke a VCN VR function to cause the packet to be forwarded to an NVD running the appropriate gateway associated with the VCN. For example, if the destination is an endpoint in a customer's on-premises network, the packet may be forwarded by the VCN VR to an NVD running a DRG gateway configured for the VCN. The VCN VR may run on the same NVD as the NVD running the VNIC associated with the source compute instance or by a different NVD. The gateway may be run by the NVD, which can be a smart NIC, a host machine, or another NVD implementation. The packet is then processed by the gateway and forwarded to the next hop, which facilitates the packet's propagation to the intended destination endpoint. 2, a packet originating from compute instance 268 may be communicated from host machine 202 (using NIC 232) over link 220 to NVD 210. At NVD 210, VNIC 276 is invoked because it is the VNIC associated with source compute instance 268. VNIC 276 is configured to examine information encapsulated within the packet, determine a next hop for forwarding the packet in order to facilitate communication of the packet to its intended destination endpoint, and then forward the packet to the determined next hop.

[0108] Compute instances deployed in a VCN can communicate with various endpoints. These endpoints may include endpoints hosted by CSPI200 and endpoints external to CSPI200. Endpoints hosted by CSPI200 may include instances in the same VCN or other VCNs, which may be the customer's VCN or VCNs not belonging to the customer. Communication between endpoints hosted by CSPI200 may occur via physical network 218. Compute instances may communicate with endpoints not hosted by CSPI200 or external to CSPI200. Examples of these endpoints include endpoints within a customer's on-premises network or data center, or public endpoints accessible via a public network such as the Internet. Communication with endpoints external to CSPI200 may occur via a public network (e.g., the Internet) (not shown in FIG. 2) or a private network (not shown in FIG. 2) using various communication protocols.

[0109] The architecture of CSPI 200 shown in FIG. 2 is merely exemplary and is not intended to be limiting. Variations, substitutions, and modifications are possible in alternative embodiments. For example, in some implementations, CSPI 200 may include more or fewer systems or components than those shown in FIG. 2, may combine two or more systems, or may include a different configuration or arrangement of systems. The systems, subsystems, and other components shown in FIG. 2 may be implemented in software (e.g., code, instructions, programs) executed by one or more processing units (e.g., processors, cores) of the respective systems, using hardware, or a combination thereof. The software may be stored in a non-transitory storage medium (e.g., a memory device).

[0110] FIG. 4 illustrates connections between host machines and an NVD to achieve I / O virtualization to support multitenancy functionality, according to one embodiment. As shown in FIG. 4, host machine 402 runs hypervisor 404, which provides a virtual environment. Host machine 402 runs two virtual machine instances: VM1 406 belonging to customer / tenant #1 and VM2 408 belonging to customer / tenant #2. Host machine 402 includes a physical NIC 410 connected to an NVD 412 via link 414. Each of the compute instances is connected to a VNIC run by NVD 412. In the embodiment of FIG. 4, VM1 406 is connected to VNIC-VM1 420, and VM2 408 is connected to VNIC-VM2 422.

[0111] 4, NIC 410 includes two logical NICs: logical NIC A 416 and logical NIC B 418. Each virtual machine is connected to and configured to function with its own logical NIC. For example, VM1 406 is connected to logical NIC A 416, and VM2 408 is connected to logical NIC B 418. Because of the logical NICs, each tenant's virtual machine believes it has its own host machine and NIC, even though host machine 402 includes only one physical NIC 410 that is shared by multiple tenants.

[0112] In one embodiment, each logical NIC is assigned its own VLAN ID. Thus, a particular VLAN ID is assigned to logical NIC A 416 for Tenant 1, and another VLAN ID is assigned to logical NIC B 418 for Tenant 2. When a packet is communicated from VM1 406, a tag assigned to Tenant 1 is attached to the packet by the hypervisor, and the packet is then communicated from host machine 402 to NVD 412 via link 414. In a similar manner, when a packet is communicated from VM2 408, a tag assigned to Tenant 2 is attached to the packet by the hypervisor, and the packet is then communicated from host machine 402 to NVD 412 via link 414. Thus, a packet 424 communicated from host machine 402 to NVD 412 has an associated tag 426 that identifies the particular tenant and associated VM. At the NVD, for a packet 424 received from host machine 402, a tag 426 associated with the packet is used to determine whether the packet should be processed by VNIC-VM1 420 or VNIC-VM2 422. The packet is then processed by the corresponding VNIC. The configuration shown in Figure 4 allows each tenant's compute instance to be confident that it owns its own host machine and NIC. The setup shown in Figure 4 enables I / O virtualization to support multi-tenancy capabilities.

[0113] FIG. 5 illustrates a simplified block diagram of a physical network 500, according to one embodiment. The embodiment illustrated in FIG. 5 is structured as a Clos network. A Clos network is a specific type of network topology designed to provide connection redundancy while maintaining high bisection bandwidth and maximum resource utilization. A Clos network is a type of nonblocking multi-stage or multi-layer switching network, and the number of stages or layers can be two, three, four, five, etc. The embodiment illustrated in FIG. 5 is a three-layer network, including layers 1, 2, and 3. TOR switch 504 represents a layer 0 switch in the Clos network. One or more NVDs are connected to the TOR switch. Layer 0 switches are also referred to as edge devices of the physical network. Layer 0 switches are connected to layer 1 switches, also referred to as leaf switches. In the embodiment illustrated in FIG. 5, a set of “n” layer 0 TOR switches are connected to a set of “n” layer 1 switches, together forming a pod. Each layer 0 switch in a pod is interconnected to all layer 1 switches within the pod, but there are no switch connections between pods. In one implementation, the two pods are referred to as blocks. Each block is served by or connected to a set of "n" layer 2 switches (sometimes called spine switches). There can be multiple blocks in a physical network topology. The layer 2 switches are then connected to "n" layer 3 switches (sometimes called super spine switches). Communication of packets through the physical network 500 is typically performed using one or more layer 3 communication protocols. Typically, all layers of the physical network except the TOR layer have n-way redundancy, thus enabling high availability. Policies may be specified for pods and blocks to control the visibility of switches to each other within the physical network to enable scaling of the physical network.

[0114] A characteristic of Clos networks is that the maximum number of hops required to reach from one tier-0 switch to another tier-0 switch (or from an NVD connected to a tier-0 switch to another NVD connected to a tier-0 switch) is fixed. For example, in a three-tier Clos network, a maximum of seven hops are required for a packet to reach from one NVD to another, with the source and target NVDs connected to the leaf layers of the Clos network. Similarly, in a four-tier Clos network, a maximum of nine hops are required for a packet to reach from one NVD to another, with the source and target NVDs connected to the leaf layers of the Clos network. Therefore, the Clos network architecture maintains consistent latency throughout the network, which is important for intra- and inter-datacenter communications. Clos topologies scale horizontally and are cost-effective. The network's bandwidth / throughput capacity can be easily increased by adding more switches (e.g., more leaf switches and spine switches) to various tiers and by increasing the number of links between switches at adjacent tiers.

[0115] In one embodiment, each resource in CSPI is assigned a unique identifier called a Cloud Identifier (CID). This identifier is included as part of the resource's information and can be used to manage the resource, for example, via a console or via an API. An exemplary syntax for a CID is as follows: ocid1.<RESOURCE TYPE> . <realm>.[REGION][.FUTURE USE].<UNIQUE ID> where: ocid1: A literal column that indicates the version of the CID. RESOURCE TYPE: The type of resource (e.g., instance, volume, VCN, subnet, user, group, etc.). REALM: The realm in which the resource resides. Example values ​​are "c1" for the commercial realm, "c2" for the government cloud realm, or "c3" for the federal cloud realm. Each realm may have its own domain name. REGION: The region the resource is in. This part may be blank if a region is not applicable to the resource. FUTURE USE: Reserved for future use. UNIQUE ID: The unique part of the ID. This format may vary depending on the type of resource or service.

[0116] Hybrid GPU Architecture (GPU Supercluster) FIG. 6A illustrates an architecture of a hybrid general processing unit (GPU) cluster, according to an embodiment. The hybrid GPU cluster includes multiple GPU clusters communicatively coupled to each other by a hierarchy of switches (e.g., a CLOS architecture of hierarchically arranged switches). Note that the CLOS architecture enables GPU scaling to a much greater extent than traditional GPU clusters. The hybrid GPU architecture allows multiple GPU clusters to coexist within the same network fabric, referred to herein as a GPU fabric. In some implementations, at least some of the GPU clusters operate at a different speed than some other GPU clusters included in the hybrid GPU architecture. For example, a hybrid GPU cluster may include one or more GPU clusters operating at 100G and one or more GPU clusters operating at 400G. The hybrid GPU cluster architecture of FIG. 6A is also referred to herein as a GPU supercluster architecture.

[0117] The GPU supercluster architecture 600 includes multiple blocks that host multiple GPU clusters. For example, as shown in FIG. 6A , the GPU supercluster architecture 600 includes “K” blocks (e.g., Block 1 605 through Block K 625), each configured to host a specific GPU cluster. The first block (e.g., Block 1 605) hosts the first GPU cluster 631, while the Kth block (i.e., Block K 625) hosts the second GPU cluster 633. Note that each GPU cluster included within a block is hosted on one or more racks. For example, the first GPU cluster 631 includes a first set of GPUs hosted on one or more racks (e.g., Rack 1 (619A) through Rack K (619B)). Similarly, a second GPU cluster 633 includes a second set of GPUs hosted on one or more racks (eg, rack 1 (629A) through rack M (629B)).

[0118] According to some embodiments, the first GPU cluster 631 is a GPU cluster operating at a first speed (e.g., 100G). More specifically, the first GPU cluster 631 includes one or more GPUs (i.e., a first set of GPUs), and a network link from the GPUs operates at the first speed. In other words, in the host machine, the GPU card is paired with a network interface card (NIC), and the NIC operates at the first speed. Similarly, the second GPU cluster 633 is a GPU cluster operating at a second speed (e.g., 400G). More specifically, the second GPU cluster 633 includes one or more GPUs (i.e., a second set of GPUs), and the second speed is different from the first speed.

[0119] The CLOS architecture includes a layered or hierarchical structure of switches including a first tier of switches (T1), a second tier of switches (T2), and a third tier of switches (T3). In one implementation, the first tier of switches and the second tier of switches are each included in a block of the supercluster architecture. For example, as shown in FIG. 6A, block 1 605 includes multiple first tier switches 617A-617B, while block K 625 includes multiple first tier switches 628A-628B. In a similar manner, block 1 605 includes multiple second tier switches 615A-615B, while block K 625 includes multiple second tier switches 627A-627B. The network fabric further includes a third tier of switches. In some implementations, the switches in the third tier may be divided to form third tier groups of switches, e.g., the switches in the group labeled 601 (i.e., switches 613A-613B) and the switches in the group labeled 621 (i.e., switches 623A-623B), respectively. Note that the switches in the third tier are also referred to herein as upper tier switches. Each group of switches in the third tier of switches communicatively couples blocks, as described below.

[0120] In some implementations, the tier / hierarchical levels of switches are arranged as follows: (i) tier 3 (or upper-tier switches) with ports operating at a second speed, e.g., 400G. These switches (e.g., switches 613A, 613B, 623A, and 623B) can support 100G, 200G, or 400G switches to be connected; (ii) tier 2 (or mid-level switches, e.g., switches 615A, 615B, 627A, and 627B) configured to support 400G connections as either 4 x 100G connections (i.e., multiple 100G connections) or 1 x 400G connection (i.e., a single 400G connection); and (iii) tier 1 (or lowest-level switches, e.g., switches 617A, 617B, 628A, and 628B). The type of switches deployed in tier 1 is selected based on the type of GPU cluster to which these switches are expected to be connected. For example, if one wishes to connect a rack of GPUs (e.g., GPUs in cluster 1 631) where each GPU is running at a speed of 100G, then the tier 1 switch selected is a 100G switch, whereas if one wishes to connect a rack of GPUs (e.g., GPUs in cluster 633) where each GPU is running at a speed of 400G, then the tier 1 switch selected is a 400G switch.

[0121] As described above, the fabric of the GPU supercluster is arranged as multiple blocks, with each block supporting a GPU cluster operating at a certain speed. The GPU supercluster includes at least a first GPU cluster operating at a first speed and a second GPU cluster operating at a second speed different from the first speed. In each block, a tier 1 switch is communicatively coupled (at one end) to a cluster of GPUs hosted on one or more racks and to a tier 2 switch (at the other end). The tier 2 switch then communicatively couples the tier 1 switch to a tier 3 switch. The tier 3 switch is configured to couple different blocks together, i.e., communicatively couple different clusters of GPUs.

[0122] Referring to FIG. 6A , block 1 605 includes K tier 1 switches (e.g., K=64 switches), e.g., switches labeled 617A-617B, each tier 1 switch having M upstream ports (i.e., ports facing toward the tier 2 switches) and M downstream ports (i.e., ports facing toward the GPU racks), e.g., M=32 ports. Each upstream and downstream port of these tier 1 switches operates at a first speed, e.g., 100G (because these switches are coupled to GPU cluster 631, which operates at a first speed of 100G). The downstream ports of the tier 1 switches are coupled to the GPU racks. For example, a particular tier 1 switch (e.g., switch 617A) in block 1 605 is coupled to a rack of GPUs (619A), while switch 617B is coupled to rack 619B. The M downstream ports of switch 617A are communicatively coupled to a rack (619A) that includes P (e.g., P=M / 2) GPUs, each connected to the switch via two links (i.e., utilizing two ports of switch 617A), thereby providing a pair of 2×100G connections.

[0123] The M upstream ports (e.g., M=32 ports) of switch 617A are communicatively coupled to M tier 2 switches (e.g., switches 615A, 615B). Note that the tier 2 switches have a dimension of P (e.g., P=16) downstream ports and P upstream ports, each operating at a second speed, e.g., 400G (i.e., a speed different from the operating speed of the upstream ports of the tier 1 switches). For example, as shown in FIG. 6A , switch 615A has P=16 downstream ports labeled 630. In this case, each downstream port of the tier 2 switch (e.g., switch 615A) operating at 400G is split (e.g., optically) into multiple (e.g., four) 100G ports (referred to herein as subports). Each subport operating at 100G can then be communicatively coupled to a corresponding upstream port of the tier 1 switch (operating at 100G). It will be appreciated that the division of downstream ports of the tier 2 switches can be achieved via optical module connectors possessing breakout functionality, capable of evenly distributing a large bandwidth on one end to multiple lower-speed connections on the other end. P upstream ports of switch 615A (tier 2), each operating at 400G, are communicatively coupled to P=16 tier 3 switches. For example, as shown in FIG. 6A, the P=16 upstream ports of switch 615A are communicatively coupled to the first tier 3 switch / upper tier switch in each group, e.g., switch 613A (in upper tier 1) and switch 623A (in upper tier P).

[0124] Referring now to block K 625, M tier 2 switches (e.g., switches 627A, 627B) are included, each of which has dimensions of P (e.g., P=M / 2) upstream ports and P downstream ports. Each of these switches has ports operating at a second speed. The upstream ports of the tier 2 switches in block K 625 are communicatively coupled to the tier 3 switches in a manner similar to the coupling of the tier 2 switches and tier 3 switches in block 1 605. However, the P downstream ports of the tier 2 switches are communicatively coupled to the P tier 1 switches. Specifically, in contrast to the coupling described above with respect to downstream port 630 of tier 2 switch 615A (in block 1), the downstream ports of the tier 2 switches in block K do not require any form of partitioning.

[0125] According to some embodiments, the coupling of the downstream ports of the tier 1 switch in block K 625 to the GPU clusters 633 hosted on racks 629A-629B can take one of two forms. In one implementation, as shown in FIG. 6A , for a particular tier 1 switch (e.g., switch 628A), the M downstream ports of this switch are connected to the M GPUs hosted in the rack. Thus, there is a one-to-one correspondence between the downstream ports of the switch and the GPUs hosted in the rack. Note that such coupling provides a single connection 632 between the GPUs hosted in the rack and the downstream ports of switch 628A. In another implementation 650 (as shown in FIG. 6B ), link aggregation techniques (e.g., link bonding) are utilized to provide further host flexibility; for example, a single 400G interface of a port of switch 628A is presented to applications running on GPUs hosted on rack 629A as two different links (labeled 640 in FIG. 6B ). Thus, the above implementation provides the ability to present one of 1 x 400G links (i.e., a single connection) or 2 x 200G links (i.e., a bonded or multiple connections) to an application running on the GPU.

[0126] Thus, as previously mentioned, the supercluster architecture of FIGS. 6A and 6B provides ultra-high performance at scale. The multi-tier CLOS topology provides a non-blocking network fabric that can scale to tens of thousands of GPUs. Note that while a traditional GPU cluster might fit into a few rows within a single data center room, for example, a large supercluster might span multiple rooms within a building (i.e., data halls) or even multiple adjacent buildings within a data center complex. The cable distance between two GPUs may be long, resulting in some packets passing through these data halls and resulting in slightly higher latency. Below, techniques for counteracting potentially high latency that may occur are described with reference to FIGS. 9-12. Furthermore, it is understood that the architectures described above with reference to FIGS. 6A and 6B are merely examples and are not intended to limit the scope of the present disclosure. Variations, alternatives, and modifications are possible in alternative embodiments. For example, in some implementations, an architecture may include more or fewer systems or components than those shown in Figures 6A and 6B, may combine two or more systems, or may have a different configuration (e.g., switch dimensions, number of tiered layers in a CLOS topology, number of blocks, type (i.e., speed) of GPU clusters, etc.) or arrangement of systems. The systems, subsystems, and other components shown in Figures 6A and 6B may be implemented in software (e.g., code, instructions, programs) executed by one or more processing units (e.g., processors, cores) of the respective systems, using hardware, or a combination thereof. The software may be stored in a non-transitory storage medium (e.g., in a memory device).

[0127] FIG. 7 shows an exemplary flowchart 700 illustrating steps performed in provisioning a request using a hybrid GPU cluster. The process shown in FIG. 7 may be implemented in software (e.g., code, instructions, programs) executed by one or more processing units (e.g., processors, cores) of the respective systems, hardware, or a combination thereof. The software may be stored in a non-transitory storage medium (e.g., a memory device). The method presented in FIG. 7 and described below is intended to be exemplary and non-limiting. The steps shown in FIG. 7 may be performed by a control plane of a network fabric. While FIG. 7 shows various processing steps occurring in a particular sequence or order, this is not intended to be limiting. In an alternative embodiment, these steps may be performed in some different order, or some steps may be performed in parallel.

[0128] The process begins in step 705, where a network fabric is provided, the network fabric including a plurality of GPU clusters. Each GPU cluster includes one or more GPUs hosted on one or more racks. The plurality of GPU clusters includes at least a first GPU cluster operating at a first speed and a second GPU cluster operating at a second speed (different from the first speed). In some implementations, providing the network fabric in step 705 may include substeps 710-720, as described below.

[0129] In step 710, multiple blocks are instantiated, each block including one or more racks for hosting GPUs belonging to a particular GPU cluster. Note that a single GPU cluster is housed within a block. Thus, a first block may host a first GPU cluster operating at a first speed, and a second block may host a second GPU cluster operating at a second speed. In step 715, multiple switches are arranged in a hierarchical manner (i.e., a CLOS architecture). In some implementations, the hierarchical structure of switches may correspond to a three-tier switch architecture. In this case, switches belonging to tier 1 and tier 2 may be provided within each block of the network fabric, as shown in FIG. 6A. In step 720, multiple groups of upper-tier switches may be provided. Note that this layer of switches may correspond to the tier 3 layer of switches in the CLOS architecture. The tier 3 layer of switches communicatively couples different blocks included in the network fabric. In step 725, the request may be received by a control plane of the network fabric. The request may correspond to a customer request for execution of a workload. In response to receiving the request, at step 730, the control plane may allocate one or more GPUs (based on constraints associated with the request) from multiple GPU clusters to execute the workload.

[0130] FIG. 8 illustrates a block diagram of a cloud infrastructure 800 incorporating a CLOS network arrangement of switches, according to one embodiment. The cloud infrastructure 800 includes multiple blocks, e.g., Block 1 605 through Block K 625. Each block may include multiple racks, and each rack hosts multiple host machines (also referred to herein as hosts). For simplicity, the blocks illustrated in FIG. 8 are illustrated as including a single rack. However, it should be noted that a block may include multiple racks. Block 1 includes a rack 619A that is shown to include two host machines, namely, Host 1-A 812 and Host 1-B 814. Similarly, Block K 625 includes a rack 629A that is shown to include two other host machines, namely, Host 2-A 822 and Host 2-B 824. It is understood that the example of FIG. 8 (i.e., each rack including two host machines) is intended to be illustrative and non-limiting. For example, the cloud infrastructure may include three or more racks, and each rack may include three or more host machines. Further, note that each rack is not constrained to have the same number of hosts. Rather, a rack may contain a greater or lesser number of host machines compared to the number of host machines contained in another rack.

[0131] Each host machine includes multiple graphical processing units (GPUs). For example, host machine 1-A 812 includes N GPUs, e.g., GPU1 813. It is further understood that the example of FIG. 8 , including each host machine including the same number of GPUs, i.e., N GPUs, is intended to be illustrative and non-limiting; that is, each host machine may include a different number of GPUs. Each rack is associated with a tier 1 switch (also referred to herein as a top-of-rack (TOR) switch) communicatively coupled to the GPUs hosted on the host machines in the rack. For example, Rack 1 619A is associated with a TOR switch 617A communicatively coupled to host machines Host 1-A 812 and Host 1-B 814, while Rack 2 629A is associated with a TOR switch 627A communicatively coupled to host machines Host 2-A 622 and Host 2-K 624. The TOR switches shown in Figure 6 (i.e., switches 617A and 627A) are each understood to include N ports used to communicatively couple the TOR switch to N GPUs hosted on each host machine included in the corresponding rack. The coupling of the TOR switch to the GPUs as shown in Figure 8 is intended to be exemplary and non-limiting. For example, in some embodiments, the TOR switch may include multiple ports, each corresponding to a GPU on each host machine, i.e., a GPU on a host machine may be connected to a unique port of the TOR via a communication link.

[0132] The TOR switch associated with each rack is communicatively coupled to one or more tier 2 switches in each block. For example, switch 617A is communicatively coupled to tier 2 switches 615A-615B, while switch 627A is communicatively coupled to switches 623A-623B. Similar to the configuration of FIG. 6A, the tier 2 switches are communicatively coupled to upper tier switches 613A-613B and 623A-623B, as shown in FIG. 8. Information sent from a particular tier 1 switch to a tier 2 switch (or from a tier 2 switch to a tier 3 switch) is referred to herein as communication implemented via an uplink, while information sent from a tier 1 switch to a host machine (or from a tier 3 switch to a tier 2 switch) is referred to herein as communication implemented via a downlink. According to some embodiments, the tier switches of FIG. 8 form a CLOS network arrangement (e.g., a multi-stage switching network), and each tier 1 switch may be considered to form a “leaf” node in the CLOS network.

[0133] According to some embodiments, GPUs included in a host machine execute tasks related to machine learning. In such a setting, a single task may be executed / distributed across multiple GPUs, which may be distributed across multiple host machines, multiple racks, and / or different blocks. Because all these GPUs are working on the same task (i.e., workload), they all need to communicate with each other in a time-synchronized manner. Furthermore, at any particular time, the GPUs are in one of a computation mode or a communication mode, i.e., the GPUs communicate with each other at approximately the same time. The speed of the workload is determined by the speed of the slowest GPU.

[0134] Equal-cost multipath (ECMP) routing is typically used to route packets from a source GPU to a destination GPU. ECMP routing uses a selection technique to select a specific path when there are multiple equal-cost paths available for routing traffic from a sender to a receiver. Thus, a selection algorithm is used at a network device receiving traffic (e.g., a TOR switch) to select an outgoing link to be used to forward traffic from the network device to subsequent devices. This outgoing link selection occurs at each network device in the path from the sender to the receiver. Hash-based selection is a widely used ECMP selection technique, and the hash can be based, for example, on the packet's 4-tuple (e.g., source port, destination port, source IP, destination IP).

[0135] ECMP routing is a flow-aware routing technique in which each flow (i.e., a stream of data packets) hashes to the same path for the duration of the flow. Packets within a flow are therefore forwarded from a network device using a specific outgoing port / link. This is typically done to ensure that packets within a flow arrive in order, i.e., no packet reordering is required. However, ECMP routing is not aware of bandwidth (or throughput). In other words, the switch performs statistical flow-aware (not throughput-aware) ECMP load balancing of flows across parallel links.

[0136] In standard ECMP routing (i.e., flow-only aware routing), a problem is that flows received by a network device over two separate incoming links may hash to the same outgoing link, thereby causing flow collisions. For example, consider a situation where two flows arrive over two separate 100G incoming links and each of the flows hashes to the same 100G outgoing link. Such a situation results in congestion (i.e., flow collisions), resulting in packets being dropped because the incoming bandwidth is 200G but the outgoing bandwidth is 100G. As shown in FIG. 8, there are two flows: flow 1 841, which is directed from a first GPU on host machine Host 1-A 812 to switch 617A, and flow 2 843, which is directed from another GPU on host machine Host 1-B 814 to switch 617A. Note that these two flows are directed to the switch on separate links. For purposes of explanation, the links shown in FIG. 8 are assumed to have the same capacity (i.e., bandwidth) of 100G. When switch 617A runs the ECMP routing algorithm, two flows may hash to use the same outgoing link of the switch, for example, link 850 connecting switch 617A to an upper layer switch (e.g., switch 615A), which would result in a collision between the two flows (represented by an "X" mark), resulting in the packet being dropped.

[0137] Such collision situations are typically problematic for all types of traffic, regardless of protocol. For example, TCP is intelligent in that if a packet is dropped and the sender does not receive an acknowledgment for the dropped packet, the packet will be retransmitted. However, the situation is exacerbated for remote direct memory access (RDMA) type traffic. RDMA networks do not use TCP for various reasons (e.g., TCP does not have high performance). RDMA networks use protocols such as RDMA over Infiniband or RDMA over converged Ethernet (RoCE). RoCE has congestion control algorithms that allow the sender to slow down the rate at which packets are sent when it identifies congestion or a dropped packet. In the case of a dropped packet, not only the dropped packet but also multiple packets around the dropped packet are retransmitted, further consuming available bandwidth and degrading performance.

[0138] The flow collision problem is a significant issue for the supercluster GPU architecture of FIG. 6A due to its strict time-synchronization requirements. For example, as mentioned above, GPUs may execute machine learning tasks (i.e., workloads) in which all GPUs communicate with each other in a time-synchronized manner. For machine learning tasks and other types of tasks, a logical topology (e.g., a ring topology, a tree topology, etc.) is constructed by the host machine to enable communication between GPUs. GPUs connect to each other using a logical topology that can be multilevel or multidimensional. In some implementations, to execute a workload, an application constructs a virtual (or logical) topology for interconnecting GPUs. Typically, such applications are unaware of the underlying physical topology of the host machines and therefore attempt to construct the logical topology in a random (i.e., arbitrary) manner. Such a randomly constructed logical topology causes GPU host machines to exchange traffic regardless of other GPU hosts present in the local network vicinity. This increases the likelihood of traffic congestion, resulting in reduced GPU throughput. For example, consider a case in which a pair of host machines is required to execute a machine learning task. In this case, if a random selection of a pair of host machines is performed, and one of the host machines is in a first local neighborhood (e.g., a first rack) and the other host machine is in another local neighborhood (different from the first local neighborhood) (e.g., a second rack), the execution of the machine learning task may not only result in a certain amount of latency (e.g., a delay occurring in communication between the first host machine and the second host machine), but also increase the possibility of traffic congestion. In contrast, if a pair of host machines selected to execute a machine learning task are in the same local neighborhood (e.g., in the same rack or the same block), it is understood that communication between the host machines not only results in minimal latency, but also improves the possibility of avoiding traffic congestion.

[0139] Techniques for overcoming the aforementioned problems are described below. Specifically, the techniques described herein utilize GPU hierarchical locality information in the process of building a logical topology, thereby avoiding unnecessary traffic congestion. Furthermore, embodiments of the present disclosure enable customers to reduce application-to-service latency by "placing" workloads on nearby host machines. Furthermore, customers can use the locality information to place workloads in a manner that achieves higher anti-affinity, thus achieving greater resilience by reducing resource common destiny.

[0140] According to some embodiments, host machines are unaware of the physical topology of the network, i.e., a particular host machine is unaware of the physical location / position of other host machines in the network. For example, referring to FIG. 8 , host machine 1-A 812 is unaware that host machine 1-B 814 are actually contained in the same rack (i.e., rack 619A) and located behind the same TOR switch, i.e., switch 617A. However, the network control plane is aware of the overall physical topology of the host machines. In one implementation, the network control plane exposes such locality information (e.g., hierarchical locality information identifying at least the rack containing the host machine, the block hosting the rack, etc.) to the host machines to achieve traffic locality and avoid unnecessary traffic congestion. Doing so has a significant impact on the performance of GPU workloads, as described below with reference to FIGS. 9 and 10.

[0141] In some embodiments, the network control plane utilizes an instance metadata service (IMDS) to expose (and store) metadata information (e.g., hierarchical locality information) to host machines. Such metadata information may be exposed to the host machines via a network virtualization device (NVD) associated with the host machines. It is understood that the hierarchical locality information may include metadata indicating a rack identifier for the rack containing the host machine as well as a block identifier for the block hosting the rack. It is understood that each of the host machines may query the IMDS to obtain the metadata information associated with the host machine. As described below, the exposed locality information is utilized to construct an optimal logical topology to achieve higher GPU workload throughput.

[0142] Referring to FIG. 9, a logical topology constructed without considering host machine locality information is shown according to one embodiment. The logical topology shown in FIG. 9 corresponds to a situation where there are eight host machines: host machine 1-A 901, host machine 1-B 903, host machine 2-A 905, host machine 2-B 907, host machine 3-A 909, host machine 3-B 911, host machine 4-A 913, and host machine 4-B 915. Note that, as shown in FIG. 9, host machines 1-A 901 and 1-B 903 are located in the same rack, i.e., behind the same TOR switch. Similarly, host machine pairs (host machines 2-A 905 and 2-B 907), (host machines 3-A 909 and 3-B 911), and (host machines 4-A 913 and 4-B 915) are each contained in a different rack. This is represented in FIG. 9 as "TOR local traffic." Further, note that host machines 1-A, 1-B, 2-A, 2-B, 3-A, and 3-B are contained in the same block (i.e., different racks within the same block), while host machines 4-A and 4-B are located in another block. This is represented as "block local traffic" in Figure 9.

[0143] The logical topology shown in Figure 9 is a ring topology constructed without using hierarchical locality information, i.e., this ring topology is constructed in a random (arbitrary) manner. As shown in Figure 9, this ring is constructed in such a way that Host 1-A is directly connected to Host 4-B via the link labeled 941. Furthermore, Host 4-B is connected to Host 3-B via the link labeled 942, Host 3-B is connected to Host 4-A via the link labeled 943, Host 4-A is connected to Host 3-A via the link labeled 944, Host 3-A is connected to Host 2-B via the link labeled 945, Host 2-B is connected to Host 2-A via the link labeled 946, Host 2-A is connected to Host 1-B via the link labeled 947, and Host 1-B is connected to Host 1-A via the link labeled 948.

[0144] A logical topology constructed in the manner shown in FIG. 9 is prone to network flow collisions due to ECMP traffic distribution. Note that although host 4-A 913 and host 4-B 915 are located behind the same TOR switch, i.e., contained in the same rack, traffic originating from host 4-B and destined for host 4-A traverses the following route: The traffic is first routed from host 4-B to host 3-B (i.e., via virtual link 942), and then from host 3-B to host 4-A (i.e., via virtual link 943). Thus, traffic intended for a destination host (e.g., host 4-A) located in the same rack as the source host (i.e., host 4-B) is unnecessarily routed to a host machine (i.e., host 3-B) that is not only outside the rack but also located in an entirely different block (compared to the source and destination host machines). In other words, traffic that could be routed locally within block 930 is routed to a host machine outside the block, only to be rerouted back to block 930. This occurs due to the arbitrarily constructed logical topology in Figure 9, leading to an increased probability of flow collisions, resulting in reduced throughput for GPU workloads (e.g., higher latency, jitter loss, etc.).

[0145] Referring now to FIG. 10 , another logical topology constructed taking into account hierarchical locality information of host machines is shown, according to one embodiment. The logical topology shown in FIG. 10 corresponds to the same eight host machines as shown in FIG. 9 . The logical topology shown in FIG. 10 is a ring topology constructed using hierarchical locality information, i.e., the ring topology is configured based on hierarchical locality information obtained, for example, from the IMDS. According to some embodiments, the logical topology may be constructed by a constituent host machine, which may be one of the host machines shown in FIG. 10 . As shown in FIG. 10 , the ring is constructed in such a way that host 1-A is directly connected to host 4-B via the link labeled 1041. Additionally, Host 4-B is connected to Host 4-A via a link labeled 1042, Host 4-A is connected to Host 3-B via a link labeled 1043, Host 3-B is connected to Host 3-A via a link labeled 1044, Host 3-A is connected to Host 2-B via a link labeled 1045, Host 2-B is connected to Host 2-A via a link labeled 1046, Host 2-A is connected to Host 1-B via a link labeled 1047, and Host 1-B is connected to Host 1-A via a link labeled 1048.

[0146] The logical topology shown in FIG. 10 prevents traffic between two host machines within the same rack (i.e., located behind the same TOR switch) from being unnecessarily routed outside the rack and / or block. For example, considering the same example of FIG. 9, if Host 4-B attempts to send traffic to Host 4-A, the traffic may be routed via link 1042 (via a single hop in the virtual layer). This contrasts with the situation shown in FIG. 9, where traffic is routed outside the block (i.e., to Host 3-B) and then rerouted back into the block (i.e., to Host 4-A). FIG. 11 shows the eight-node physical topology 1100 considered in the above examples of FIGS. 9 and 10. In particular, Host 1-A and Host 1-B are shown to be included in the same rack 1105A (connected by tier 1 switch 1102A), Host 2-A and Host 2-B are shown to be included in the same rack 1105B (connected by tier 1 switch 1102B), Host 3-A and Host 3-B are shown to be included in the same rack 1105C (connected by tier 1 switch 1102C), and Host 4-A and Host 4-B are shown to be included in the same rack 1107A (connected by tier 1 switch 1104A). Note that tier 1 switches 1102A, 1102B, and 1102C are located in the same block 920, while tier 1 switch 1104A is located in another block 930. In this manner, building a logical topology based on hierarchical locality information leads to a reduced likelihood of flow collisions and improves GPU workload throughput.

[0147] FIG. 12 shows an exemplary flowchart 1200 illustrating steps performed in provisioning requests using hierarchical locality information, according to an embodiment. The process shown in FIG. 12 may be implemented in software (e.g., code, instructions, programs) executed by one or more processing units (e.g., processors, cores) of the respective systems, hardware, or a combination thereof. The software may be stored in a non-transitory storage medium (e.g., a memory device). The method presented in FIG. 12 and described below is intended to be exemplary and non-limiting. While FIG. 12 shows various processing steps occurring in a particular sequence or order, this is not intended to be limiting. In an alternative embodiment, these steps may be performed in some different order, or some steps may be performed in parallel.

[0148] The process begins at step 1205, where, for each host machine among a plurality of host machines (e.g., host machines included in a GPU cluster), hierarchical locality information of the host machine is stored. The hierarchical locality information of the host machine includes, for example, information identifying a rack containing the host machine, a block in which the rack is located, etc. In some embodiments, an instance metadata service may be utilized to store the hierarchical locality information on the host machine via a network virtualization device (NVD) associated with the host machine. Note that the hierarchical locality information may correspond to information indicating, for example, an identifier of the rack containing the host machine (i.e., rack ID), an identifier of a TOR switch associated with the rack, a block identifier of a block in which the rack is located (i.e., block ID), etc.

[0149] At step 1210, the control plane receives a request (e.g., from a customer) to execute a workload. Note that the workload corresponds to one or more processes to be executed using a GPU associated with a host machine. The process then moves to step 1215, where one or more host machines from a plurality of host machines are identified as available for executing the workload. The identification of available host machines may be performed in multiple ways. For example, according to one embodiment, the control plane may maintain a current load (i.e., processing workload) being processed by the host machines. Based on the capabilities of each host machine and the amount of current load being processed by the host machine, the control plane may select one or more host machines available for executing the customer's workload. Furthermore, according to another embodiment, the control plane may pre-allocate a certain number of host machines per customer. Furthermore, the available host machines may be determined from this certain number of host machines.

[0150] According to some embodiments, the request received in step 1210 may include one or more constraints. For example, the request may include a first constraint associated with a latency threshold, i.e., the customer may desire that the workload be executed by lowering the latency below a predetermined threshold. The second constraint may correspond to an anti-affinity constraint. Such a constraint corresponds to the customer desiring that there be some degree of availability of host machines, i.e., that at least some of the host machines selected to execute the workload must be located in different racks and / or different blocks. Such constraints are typically incorporated by the customer to address rack failure issues. Furthermore, the customer request may include a constraint targeting the type of GPUs required to execute the workload. For example, the request may indicate that the customer desires a first number of GPUs operating at a first speed (e.g., 100G) and a second number of GPUs operating at a second speed (e.g., 400G). Such constraints may be taken into account by the control plane in allocating / identifying a certain number of host machines for the customer.

[0151] Next, the process moves to step 1220, where hierarchical locality information for each of the one or more host machines identified in step 1215 is obtained. Note that the locality information (for each host machine) may be stored, for example, by an instance metadata service in the corresponding host machine. In step 1225, the process identifies linkage information for the one or more host machines. It is understood that the linkage information for the one or more host machines corresponds to a logical topology formed by the one or more host machines (e.g., a logical topology such as that shown in FIG. 10 ). The process then moves to step 1230, where the hierarchical locality information and linkage information for the one or more host machines are provided in response to a request from the customer. According to some embodiments, once the customer has obtained the hierarchical locality information and linkage information, the customer may select a subset of one or more host machines for executing the workload (step 1240). Note that the selection of host machines may be performed based on one or more constraints associated with the workload.

[0152] As mentioned above, ECMP routing is a flow-aware routing technique, in which each flow (i.e., a stream of data packets) is hashed to a specific path for the duration of the flow. Therefore, packets within a flow are forwarded from a network device using a specific outgoing port / link. This is typically done to ensure that packets within a flow arrive in order, i.e., no packet reordering is required. However, ECMP routing is not aware of bandwidth (or throughput). In other words, a switch performs statistical flow-aware (not throughput-aware) ECMP load balancing of flows across parallel links. In standard ECMP routing (i.e., flow-only aware routing), the problem is that flows received by a network device via two separate incoming links may be hashed to the same outgoing link, thereby causing flow collisions.

[0153] Described below is a routing technique (referred to herein as a GPU-based routing policy mechanism or a GPU-based traffic routing mechanism) for overcoming the aforementioned flow collision problem. It is understood that the flow collision problem affects traffic from the CPU and the GPU. However, the flow collision problem is a much bigger problem for the GPU due to its strict time synchronization requirements. Furthermore, it should be understood that the standard ECMP routing mechanism, due to the inherent nature of routing information in a statistical bandwidth-unaware manner, leads to a flow collision situation regardless of whether the network is oversubscribed or undersubscribed.

[0154] Referring now to FIG. 13, a GPU-based policy routing mechanism implemented in a hybrid GPU cluster according to an embodiment is illustrated. For convenience and explanation purposes, the architecture illustrated in FIG. 13 is the same as the architecture illustrated in FIG. 8. According to some embodiments, data packets from a sender to a receiver are routed within the network fabric in a hop-by-hop manner. A routing policy is configured at each network device that binds an incoming port link (of the network device) to an outgoing port link (of the network device). The network device may be any one of the switches included in the switch hierarchy, i.e., layer 1, 2, or 3. Referring to FIG. 13, two flows are shown: flow 1 from GPU 1 on host machine 812, with an intended destination being GPU 1 on host machine 822, and flow 2 from GPU N on host machine 814, with an intended destination being GPU N on host machine 824. In one implementation, all network devices (i.e., switches included in the network fabric) are configured to bind (or match) incoming port links to outgoing port links. The correspondence of incoming port links to outgoing port links is maintained in each network device (eg, in a policy table).

[0155] 13, for flow 1 (i.e., the flow shown by the solid line), it can be observed that when a first layer switch (617A) receives a packet on link 841, the first layer switch is configured to forward the received packet on outgoing link 850. Similarly, when a second layer switch (615A) receives a packet via link 850, the second layer switch is configured to forward the packet on outgoing link 851 to a layer 3 switch (613A). The layer 3 switch (613A) is then configured to forward the packet received on link 851 to outgoing link 852 for transmission to a layer 2 switch 623A (contained in another block). The switch 623A then forwards the packet on its outgoing link 853 to the switch 627A, which then forwards the packet on outgoing link 854 for delivery to its intended destination. Note that the route of flow 2 (i.e., the flow shown with a dashed line) is forwarded by intermediate network devices in a manner similar to that described above (with reference to flow 1) so that it is delivered to its intended destination. Note that in contrast to the situation shown in Figure 8 (which utilizes ECMP routing), the situation shown in Figure 13 avoids route collisions.

[0156] Thus, in one implementation of the GPU policy-based routing mechanism, each network device in the network fabric of FIG. 14 is configured to bind incoming ports / links to outgoing ports / links to avoid collisions. It is understood that at each network device in the cloud infrastructure, there is a one-to-one correspondence between incoming port links and outgoing port links, i.e., the mapping between incoming port links and outgoing port links is performed independently of the flow and / or the protocol executed by the flow. Furthermore, if an outgoing link of a particular network device fails, according to some embodiments, the network device is configured to switch its routing policy from GPU policy-based routing to standard ECMP routing, obtain a new available output link (from the various available output links), and transmit the flow to the new output link. Note that this may result in flow collisions and, as a result, congestion.

[0157] FIG. 14 shows a flowchart 1400 illustrating steps performed by a network device in routing packets, according to an embodiment. The process shown in FIG. 14 may be implemented in software (e.g., code, instructions, programs) executed by one or more processing units (e.g., processors, cores) of the respective system, hardware, or a combination thereof. The software may be stored in a non-transitory storage medium (e.g., a memory device). The method presented in FIG. 14 and described below is intended to be exemplary and non-limiting. While FIG. 14 shows various processing steps occurring in a particular sequence or order, this is not intended to be limiting. In an alternative embodiment, these steps may be performed in some different order, or some steps may be performed in parallel.

[0158] The process starts at step 1405, where a plurality of GPU clusters are provided in a network fabric, communicatively coupled to each other via a plurality of network devices (e.g., switches) arranged in a hierarchical manner. It is understood that a first GPU cluster of the plurality of GPU clusters operates at a first speed (e.g., 100G), while a second GPU cluster of the plurality of GPU clusters operates at a second speed (e.g., 400G) different from the first speed. At step 1410, the process performs a pre-configuration step in which a routing policy is configured for each network device included in the network fabric. It is noted that the routing policy corresponds to mapping incoming port links of the network device to outgoing port links.

[0159] In step 1415, a network device (e.g., a first network device) receives a data packet transmitted by a graphical processing unit (GPU) of a host machine. In step 1420, the network device determines the incoming port / link on which the packet was received. In step 1425, the network device identifies the outgoing port / link corresponding to the incoming port / link (on which the packet was received) based on policy routing information, i.e., the routing information preconfigured in step 1410. According to some embodiments, the policy routing information corresponds to a preconfigured GPU routing table of the network device, which links each incoming port link of the network device to a unique outgoing link port of the network device.

[0160] The process then moves to step 1430, where a query is performed to determine whether the outgoing port link is in a functional state, e.g., whether the outgoing link is active. If the response to this query is affirmative (i.e., the link is active), the process moves to step 1435; otherwise, if the response to the query is negative (i.e., the link is in a failed / inactive state), the process moves to step 1450.

[0161] At step 1435, the network device forwards the received data packet to another network device utilizing the outgoing port link (identified at step 1425). Further, the process performs another query at step 1440 to determine whether the packet reached its intended destination. If the query is answered affirmatively, the process simply ends (step 1460). Otherwise, if the query at step 1440 is answered negatively, the process moves to step 1445, where the next network device (i.e., the next-hop network device on the path from the source host machine to the destination host machine) processes the packet. Specifically, the next network device repeats steps 1420, 1425, 1430, 1435, and 1440.

[0162] If the query of step 1430 is answered negatively, then in step 1450, the network device obtains flow information for the data packet; for example, the flow information may correspond to a 4-tuple (i.e., source port, destination port, source IP address, destination IP address) associated with the packet. Based on the obtained flow information, the network device utilizes ECMP routing to identify a new outgoing port link, i.e., an available outgoing port link. The process then moves to step 1455, where the network device forwards the data packet using the newly obtained outgoing port link (in step 1450). The process then returns to step 1440 and repeats the process until the packet is delivered to its intended destination.

[0163] According to some embodiments of the present disclosure, in another implementation of the GPU-based policy routing framework, only a subset of network devices in the network fabric is configured to bind incoming ports / links to outgoing ports / links. In this implementation, the subset of network devices that implement the GPU-based policy routing mechanism corresponds to switches included in the tier 1 and tier 2 levels of the switch hierarchy. For example, with reference to FIG. 13 , switches 617A and 627A (included in tier 1) and switches 615A, 615B, 623A, and 623B (included in tier 2) implement policy-based routing. In this implementation, switches at the tier 1 and tier 2 levels bind incoming uplinks / ports of the switches to outgoing uplinks / ports. Note that in this implementation, switches included in the tier 3 level do not implement GPU policy-based routing. Rather, switches at this level may implement standard routing protocols (e.g., ECMP routing) to route traffic. It is further understood that in both implementations of GPU policy-based routing, when a particular switch receives a packet to be forwarded, the switch determines whether the packet's next hop is on one of the switch's downlinks. If the packet's next hop is on one of the switch's downlinks, in some implementations, the packet is sent without using policy-based routing, i.e., forwarded using a standard routing protocol. If the packet is to be forwarded on an uplink, the switch may utilize policy-based routing, i.e., routing depending on the tier in which the switch is included.

[0164] According to some embodiments, deployment of the GPU supercluster architecture of FIG. 6A or FIG. 6B presents challenges with managing global address space (e.g., MAC addresses for each GPU included in multiple clusters of GPUs). Specifically, in large-scale GPU cluster deployments, each tier 1 switch is required to manage and maintain (i.e., store) a MAC address table (e.g., forwarding table) containing the addresses of each GPU included in the cluster. Doing so can cause the switch's forwarding table to overflow. There are two problems associated with this situation: A switch is associated with a control plane (i.e., where BGP and other routing protocols reside) and a data plane (i.e., where the forwarding table resides). The problem of maintaining an efficient address space exists in both the control plane and the data plane.

[0165] It is understood that the forwarding table is required to store MAC addresses (i.e., for the overlay customer network) and IP addresses (i.e., for the underlay physical network). The problem is exacerbated in the context of deploying the GPU supercluster architecture of FIG. 6A or FIG. 6B because of the limited storage space in the forwarding table. In other words, the problem posed by the limited storage space of the forwarding table is how to limit the size of the forwarding table so that the network can be scaled without needing to scale the size of the forwarding table. Below, techniques are described that can be utilized in both the control plane and the data plane, respectively, to provide a mechanism for efficiently managing the storage space of the forwarding table.

[0166] FIG. 15A illustrates a hybrid GPU cluster architecture illustrating the placement of route reflectors, according to some embodiments. For illustrative purposes, FIG. 15A illustrates the architecture as including two blocks: a first block 1501 and a second block 1521. Each of the first and second blocks includes a hierarchy of deployed switches. For example, as shown in FIG. 15A, the first block 1501 includes a first tier of switches labeled 1503A (i.e., tier 1 switches) and a second tier of switches labeled 1503B (i.e., tier 2 switches). The second block 1521 includes a first tier of switches labeled 1523A and a second tier of switches labeled 1523B.

[0167] Similar to the architecture of FIG. 6A, the GPU supercluster architecture of FIG. 15A includes a third tier of switches (i.e., tier 3 switches), labeled 1503C and 1523C, respectively. It is understood that the multiple switches included in the third tier of switches may be divided into multiple groups in the third tier of switches (e.g., groups 1503C, 1523C, respectively). Further, similar to the architecture of FIG. 6A, it is noted that the first tier of switches is communicatively coupled to multiple GPU clusters on one end and to a second tier of switches on another end. The second tier of switches then communicatively couples the first tier of switches to a third tier of switches, which communicatively couples the different blocks included in the GPU supercluster architecture.

[0168] According to some embodiments, one or more switches from a third tier of switches are selected to form a set of target switches. Such target switches are referred to herein as route reflectors. Note that the selection of the set of target switches may be performed randomly (in some implementations). In other implementations, a route reflector may be selected from each group of multiple groups of third tier switches. For example, the first switch included in each group of third tier switches may be designated to perform the function of a route reflector, as described below. According to some embodiments, the total number of target switches included in the set of target switches is in the range of 4 to 16.

[0169] In some embodiments, each switch (1503A, 1523A) included in the first layer of switches forms a peering connection (e.g., BGP peering) with each of one or more target switches, i.e., a route reflector (1503C, 1523C) included in the third layer of switches. The route reflector is configured to reduce address information (e.g., MAC addresses) maintained in a forwarding table by the layer 1 switch, as described below. It is understood that when a particular layer 1 switch receives a packet from a GPU (i.e., a GPU coupled to the particular layer 1 switch), the layer 1 switch may transmit the GPU's address information to each target switch via the peering connection. In this way, each target switch included in the layer 3 layer of switches receives address information for each GPU included in a multiple-GPU cluster.

[0170] Upon receiving address information for each GPU included in the multiple GPU clusters, each target switch processes the received address information to generate multiple sets of address information. For example, according to one embodiment, a particular target switch may filter the received address information to generate one or more sets of address information based on certain conditions. In one example, the target switch may filter the received address information for a GPU based on the customer's VLAN to which the GPU belongs. In other words, the set of address information generated by the target switch corresponds to grouping together GPUs associated with a customer. Furthermore, each target switch may advertise / transmit the generated set or sets of address information to each switch included in the tier 1 layer of switches. Then, the particular tier 1 switch stores (in its forwarding table) only a subset of the one or more sets of received address information according to the conditions. For example, if a particular tier 1 switch is associated with a particular customer's VLAN, the particular tier 1 switch may store only the MAC addresses of GPUs that belong to the same customer's VLAN. In this manner, a particular tier switch may ignore (e.g., discard) other sets of received address information corresponding to GPUs associated with other customer VLANs. It is understood that the target switch may utilize other criteria to generate one or more sets of address information. For example, the target switch may filter GPU address information based on GPU type (i.e., blocks of GPUs operating at different speeds). In this manner, the route reflector reduces the MAC address space in the forwarding tables in the control plane.

[0171] According to some embodiments, techniques may be utilized at the data plane layer to achieve further reductions in the number of MAC addresses maintained in the forwarding table. For example, according to one technique, switches in the layer 1 level may be further configured to remove entries from the address table based on a timer associated with the entry. Additionally, a least recently used mechanism may be employed, in which entries may be removed from the forwarding table based on the usage (e.g., least used) of a particular entry. Yet another approach may include aggregating entries in the IP table (corresponding to addresses of underlay physical network components), thereby providing more storage space for MAC entries in the forwarding table. Thus, according to the aforementioned techniques, embodiments of the present disclosure limit the size of the forwarding table in order to scale the GPU cluster network without having to make compromises with respect to scaling the size of the forwarding table.

[0172] FIG. 15B shows a flowchart 1550 illustrating steps performed by a route reflector in managing the size of address tables stored by a switch, according to an embodiment. The process shown in FIG. 15B may be implemented in software (e.g., code, instructions, programs) executed by one or more processing units (e.g., processors, cores) of the respective system, hardware, or a combination thereof. The software may be stored in a non-transitory storage medium (e.g., a memory device). The method presented in FIG. 15B and described below is intended to be exemplary and non-limiting. While FIG. 15B shows various processing steps occurring in a particular sequence or order, this is not intended to be limiting. In an alternative embodiment, these steps may be performed in some different order, or some steps may be performed in parallel.

[0173] The process begins at step 1555, where a plurality of GPU clusters are provided in a network fabric, communicatively coupled to each other via a plurality of network devices (e.g., switches) arranged in a hierarchical manner. It is understood that a first GPU cluster of the plurality of GPU clusters operates at a first speed (e.g., 100G), while a second GPU cluster of the plurality of GPU clusters operates at a second speed (e.g., 400G) different from the first speed. The hierarchical structure of the switches includes at least a layer 1 level of switches, a layer 2 level of switches, and a layer 3 level of switches.

[0174] At step 1560, the process selects one or more switches from the tier 3 layer of switches to form a set of target switches (i.e., route reflectors). At step 1565, each target switch included in the set of target switches receives address information (e.g., MAC address) of each GPU included in the multiple-GPU cluster. As described above, each tier 1 switch is configured to establish a peering connection (e.g., BGP connection) with each of the target switches included in the tier 3 layer of switches. In some embodiments, the GPU's address information may be sent to the target switch via the peered connection (e.g., by the tier 1 switch upon receiving a packet from the GPU).

[0175] The process then moves to step 1570, where each target switch generates multiple sets of address information. In some embodiments, such sets of address information may be generated by filtering / grouping the received address information of GPUs based on certain conditions. For example, as described above, the target switch may filter the received address information of GPUs based on the customer's VLAN to which the GPU belongs. In other words, the sets of address information generated by the target switch correspond to grouping together GPUs associated with a customer. In step 1575, the target switch advertises / transmits the multiple sets of address information to each switch included in the tier 1 level of switches. Upon receiving the sets of address information, the switches included in the tier 1 level of switches store a subset of the multiple sets of address information according to the conditions to conserve storage space in their forwarding tables.

[0176] Cloud Infrastructure Examples As mentioned above, infrastructure as a service (IaaS) is a specific type of cloud computing. IaaS can be configured to provide virtualized computing resources over a public network (e.g., the Internet). In the IaaS model, a cloud computing provider can host infrastructure components (e.g., servers, storage devices, network nodes (e.g., hardware), deployment software, platform virtualization (e.g., hypervisor layer), etc.). In some cases, an IaaS provider may offer various services (e.g., billing, monitoring, logging, security, load balancing, clustering, etc.) incidental to those infrastructure components. Accordingly, these services can be policy-driven, allowing IaaS users to implement policies to drive load balancing and maintain application availability and performance.

[0177] In some cases, IaaS customers may access resources and services over a wide area network (WAN), such as the Internet, and can use the cloud provider's services to install the remaining elements of their application stack. For example, a user may log into an IaaS platform to create virtual machines (VMs), install an operating system (OS) on each VM, deploy middleware such as databases, create storage buckets for workloads and backups, and even install enterprise software on the VMs. The customer can then use the provider's services to perform a variety of functions, including balancing network traffic, troubleshooting application issues, monitoring performance, managing disaster recovery, etc.

[0178] In most cases, the cloud computing model requires the participation of a cloud provider. A cloud provider may be, but need not be, a third-party service that specializes in providing (e.g., offering, renting, selling) IaaS. An entity may choose to deploy a private cloud and become its own provider of infrastructure services.

[0179] In some examples, IaaS deployment is the process of connecting a new application or a new version of an application to a prepared application server or the like. This process may include the process of preparing the server (e.g., installing libraries, daemons, etc.). This process is often managed by the cloud provider below the hypervisor layer (e.g., server, storage, network hardware, and virtualization). Thus, the customer may be responsible for handling the deployment of the OS, middleware, and / or application (e.g., on self-service virtual machines (e.g., that can be spun up on demand)).

[0180] In some examples, IaaS provisioning may also refer to obtaining computers or virtual hosts for use and installing needed libraries or services on those computers or virtual hosts. In most cases, deployment does not include provisioning, which may need to be performed first.

[0181] In some cases, IaaS provisioning presents two distinct challenges. First, there is the initial challenge of provisioning an initial set of infrastructure before anything can be run. Second, there is the challenge of evolving the existing infrastructure (e.g., adding new services, modifying services, removing services, etc.) after everything has been provisioned. In some cases, these two challenges may be addressed by allowing the configuration of the infrastructure to be defined declaratively. In other words, the infrastructure (e.g., which components are needed and how those components interact) may be defined by one or more configuration files. In this way, the entire topology of the infrastructure (e.g., which resources depend on which resources and how each of those resources works together) may be described declaratively. In some cases, after the topology is defined, workflows may be generated to create and / or manage the various components described in the configuration files.

[0182] In some examples, the infrastructure may include many interconnected elements. For example, there may be one or more virtual private clouds (VPCs) (e.g., configurable and / or shared, possibly on-demand pools of computing resources), also known as a core network. In some examples, there may be one or more security group rules and one or more virtual machines (VMs) provisioned to define how the network is secured. Other infrastructure elements, such as load balancers, databases, etc., may also be provisioned. The infrastructure can evolve over time as more infrastructure elements are desired and / or added.

[0183] In some cases, continuous deployment techniques may be employed to enable deployment of infrastructure code across various virtual computing environments. Additionally, the described techniques may enable infrastructure management within these environments. In some examples, a service team may write code that is desired to be deployed to one or more, but often many, different production environments (e.g., across various geographic locations, sometimes across the world). However, in some examples, the infrastructure onto which the code will be deployed must first be set up. In some cases, provisioning can be done manually, and provisioning tools may be utilized to provision the resources and / or deployment tools may be utilized to deploy the code after the infrastructure has been provisioned.

[0184] 16 is a block diagram 1600 illustrating an example pattern of an IaaS architecture according to at least one embodiment. A service operator 1602 may be communicatively coupled to a secure host tenancy 1604, which may include a virtual cloud network (VCN) 1606 and a secure host subnet 1608. In some examples, the service operator 1602 may employ one or more client computing devices, which may be portable handheld devices (e.g., iPhones, mobile phones, iPads, computing tablets, personal digital assistants (PDAs)) or wearable devices (e.g., Google Glass head-mounted displays) that are Internet, email, short message service (SMS), Blackberry, or other communication protocol enabled, running software such as Microsoft Windows Mobile, and / or various mobile operating systems such as iOS, Windows Phone, Android, BlackBerry 8, Palm OS, etc. Alternatively, the client computing devices may be general-purpose personal computers, including, by way of example, personal and / or laptop computers running various versions of the Microsoft Windows, Apple Macintosh, and / or Linux operating systems. The client computing devices may be workstation computers running any of a variety of commercially available UNIX or UNIX-like operating systems, including, but not limited to, various GNU / Linux operating systems, such as Google Chrome OS.Alternatively or additionally, the client computing device may be any other electronic device, such as a thin client computer, an Internet-enabled gaming system (e.g., a Microsoft Xbox gaming console with or without a Kinect® gesture input device), and / or a personal messaging device, that can communicate over a network accessible to VCN 1606 and / or the Internet.

[0185] VCN 1606 may include a local peering gateway (LPG) 1610, which may be communicatively coupled to a secure shell (SSH) VCN 1612 via an LPG 1610 included in SSH VCN 1612. SSH VCN 1612 may include an SSH subnet 1614, which may be communicatively coupled to a control plane VCN 1616 via an LPG 1610 included in control plane VCN 1616. SSH VCN 1612 may also be communicatively coupled to a data plane VCN 1618 via LPG 1610. The control plane VCN 1616 and the data plane VCN 1618 may be included in a service tenancy 1619, which may be owned and / or operated by the IaaS provider.

[0186] The control plane VCN 1616 may include a control plane demilitarized zone (DMZ) tier 1620 that serves as a perimeter network (e.g., a portion of an enterprise network between the enterprise intranet and an external network). Servers based in the DMZ may have limited responsibilities and help keep security breaches contained. Additionally, the DMZ tier 1620 may include one or more load balancer (LB) subnets 1622, a control plane app tier 1624 that may include an app subnet 1626, and a control plane data tier 1628 that may include a database (DB) subnet 1630 (e.g., a front-end DB subnet and / or a back-end DB subnet). LB subnet 1622 included in control plane DMZ layer 1620 can be communicatively coupled to app subnet 1626 and Internet gateway 1634 included in control plane app layer 1624, which may be included in control plane VCN 1616, and app subnet 1626 can be communicatively coupled to DB subnet 1630, as well as service gateway 1636 and network address translation (NAT) gateway 1638 included in control plane data layer 1628. Control plane VCN 1616 can include service gateway 1636 and NAT gateway 1638.

[0187] The control plane VCN 1616 can include a data plane mirror app layer 1640, which can include an app subnet 1626. The app subnet 1626 included in the data plane mirror app layer 1640 can include a virtual network interface controller (VNIC) 1642, which can run a compute instance 1644. The compute instance 1644 can communicatively couple the app subnet 1626 of the data plane mirror app layer 1640 to the app subnet 1626, which can be included in the data plane app layer 1646.

[0188] The data plane VCN 1618 may include a data plane app layer 1646, a data plane DMZ layer 1648, and a data plane data layer 1650. The data plane DMZ layer 1648 may include a LB subnet 1622, which may be communicatively coupled to an app subnet 1626 of the data plane app layer 1646 and an internet gateway 1634 of the data plane VCN 1618. The app subnet 1626 may be communicatively coupled to a service gateway 1636 of the data plane VCN 1618 and a NAT gateway 1638 of the data plane VCN 1618. The data plane data layer 1650 may also include a DB subnet 1630, which may be communicatively coupled to the app subnet 1626 of the data plane app layer 1646.

[0189] The internet gateways 1634 of the control plane VCNs 1616 and of the data plane VCNs 1618 may be communicatively coupled to a metadata management service 1652, which may be communicatively coupled to the public internet 1654. The public internet 1654 may be communicatively coupled to NAT gateways 1638 of the control plane VCNs 1616 and of the data plane VCNs 1618. The service gateways 1636 of the control plane VCNs 1616 and of the data plane VCNs 1618 may be communicatively coupled to cloud services 1656.

[0190] In some examples, a service gateway 1636 in the control plane VCN 1616 or in the data plane VCN 1618 can make application programming interface (API) calls to cloud services 1656 without traversing the public internet 1654. API calls from the service gateway 1636 to the cloud services 1656 can be one-way: the service gateway 1636 can make the API call to the cloud services 1656, and the cloud services 1656 can send the requested data to the service gateway 1636. However, the cloud services 1656 may not initiate the API call to the service gateway 1636.

[0191] In some examples, secure host tenancy 1604 can be directly connected to service tenancy 1619 or may be otherwise separate. Secure host subnet 1608 can communicate with SSH subnet 1614 through LPG 1610, which may enable bidirectional communication on otherwise separate systems. Connecting secure host subnet 1608 to SSH subnet 1614 may give secure host subnet 1608 access to other entities within service tenancy 1619.

[0192] The control plane VCN 1616 may enable users of the service tenancy 1619 to configure or otherwise provision desired resources. The desired resources provisioned in the control plane VCN 1616 may be deployed or otherwise used in the data plane VCN 1618. In some examples, the control plane VCN 1616 may be separate from the data plane VCN 1618, and the data plane mirror app layer 1640 of the control plane VCN 1616 may communicate with the data plane app layer 1646 of the data plane VCN 1618 via a VNIC 1642, which may be included in the data plane mirror app layer 1640 and the data plane app layer 1646.

[0193] In some examples, a user or customer of the system may make a request, for example, a create, read, update, or delete (CRUD) operation, via the public internet 1654, which may communicate the request to a metadata management service 1652. The metadata management service 1652 may communicate the request to the control plane VCN 1616 via an internet gateway 1634. The request may be received by a LB subnet 1622 included in the control plane DMZ tier 1620. The LB subnet 1622 may determine that the request is valid, and in response to this determination, the LB subnet 1622 may send the request to an app subnet 1626 included in the control plane app tier 1624. If the validity of the request is confirmed and the request requires a call to the public internet 1654, the call to the public internet 1654 may be sent to a NAT gateway 1638, which may make the call to the public internet 1654. Memory that may be desirable to store with the request may be stored within the DB subnet 1630.

[0194] In some examples, the data plane mirror app layer 1640 can facilitate direct communication between the control plane VCN 1616 and the data plane VCN 1618. For example, it may be desirable for changes, updates, or other appropriate modifications to the configuration to be applied to resources included in the data plane VCN 1618. Via the VNIC 1642, the control plane VCN 1616 can communicate directly with the resources included in the data plane VCN 1618, thereby performing changes, updates, or other appropriate modifications to the configuration of the resources.

[0195] In some embodiments, the control plane VCN 1616 and the data plane VCN 1618 may be included in the service tenancy 1619. In this case, a user or customer of the system may not own or operate either the control plane VCN 1616 or the data plane VCN 1618. Instead, an IaaS provider may own or operate the control plane VCN 1616 and the data plane VCN 1618, which may both be included in the service tenancy 1619. This embodiment may enable network isolation that can prevent users or customers from interacting with other users' or other customers' resources. This embodiment may also enable users or customers of the system to store databases privately without having to rely on the public internet 1654 for storage, which may not have a desirable level of security.

[0196] In another embodiment, the LB subnet 1622 included in the control plane VCN 1616 may be configured to receive signals from the service gateway 1636. In this embodiment, the control plane VCN 1616 and the data plane VCN 1618 may be configured to be called by the IaaS provider's customers without calling the public Internet 1654. The IaaS provider's customers may desire this embodiment because databases used by the customers may be stored in the service tenancy 1619, which may be controlled by the IaaS provider and isolated from the public Internet 1654.

[0197] 17 is a block diagram 1700 illustrating another example pattern of an IaaS architecture, according to at least one embodiment. A service operator 1702 (e.g., service operator 1602 of FIG. 16 ) may be communicatively coupled to a secure host tenancy 1704 (e.g., secure host tenancy 1604 of FIG. 16 ), which may include a virtual cloud network (VCN) 1706 (e.g., VCN 1606 of FIG. 16 ) and a secure host subnet 1708 (e.g., secure host subnet 1608 of FIG. 16 ). VCN 1706 may include a local peering gateway (LPG) 1710 (e.g., LPG 1610 of FIG. 16 ), which may be communicatively coupled to a secure shell (SSH) VCN 1712 (e.g., SSH VCN 1612 of FIG. 16 ) via an LPG 1710 included in an SSH VCN 1712. SSH VCN 1712 can include SSH subnet 1714 (e.g., SSH subnet 1614 in FIG. 16 ), and SSH VCN 1712 can be communicatively coupled to control plane VCN 1716 (e.g., control plane VCN 1616 in FIG. 16 ) via LPG 1710 included in control plane VCN 1716. Control plane VCN 1716 can be included in service tenancy 1719 (e.g., service tenancy 1619 in FIG. 16 ), and data plane VCN 1718 (e.g., data plane VCN 1618 in FIG. 16 ) can be included in customer tenancy 1721, which can be owned or operated by a user or customer of the system.

[0198] The control plane VCN 1716 may include a control plane DMZ tier 1720 (e.g., the control plane DMZ tier 1620 of FIG. 16 ) that may include a LB subnet 1722 (e.g., the LB subnet 1622 of FIG. 16 ), a control plane app tier 1724 (e.g., the control plane app tier 1624 of FIG. 16 ) that may include an app subnet 1726 (e.g., the app subnet 1626 of FIG. 16 ), and a control plane data tier 1728 (e.g., the control plane data tier 1628 of FIG. 16 ) that may include a database (DB) subnet 1730 (e.g., similar to the database (DB) subnet 1630 of FIG. 16 ). The LB subnet 1722 included in the control plane DMZ layer 1720 can be communicatively coupled to an app subnet 1726 and an Internet gateway 1734 (e.g., Internet gateway 1634 in FIG. 16 ) included in the control plane app layer 1724, which may be included in the control plane VCN 1716, and the app subnet 1726 can be communicatively coupled to a DB subnet 1730 and a service gateway 1736 (e.g., service gateway in FIG. 16 ) and a network address translation (NAT) gateway 1738 (e.g., NAT gateway 1638 in FIG. 16 ) included in the control plane data layer 1728. The control plane VCN 1716 can include the service gateway 1736 and the NAT gateway 1738.

[0199] The control plane VCN 1716 can include a data plane mirror app layer 1740 (e.g., data plane mirror app layer 1640 of FIG. 16 ), which can include an app subnet 1726. The app subnet 1726 included in the data plane mirror app layer 1740 can include a virtual network interface controller (VNIC) 1742 (e.g., VNIC 1642) that can run a compute instance 1744 (e.g., similar to compute instance 1644 of FIG. 16 ). The compute instance 1744 can facilitate communication between the app subnet 1726 of the data plane mirror app layer 1740 and the app subnet 1726, which can be included in the data plane app layer 1746 (e.g., data plane app layer 1646 of FIG. 16 ), via the VNIC 1742 included in the data plane mirror app layer 1740 and the VNIC 1742 included in the data plane app layer 1746.

[0200] The internet gateway 1734 included in the control plane VCN 1716 may be communicatively coupled to a metadata management service 1752 (e.g., metadata management service 1652 of FIG. 16 ), which may be communicatively coupled to the public internet 1754 (e.g., public internet 1654 of FIG. 16 ). The public internet 1754 may be communicatively coupled to a NAT gateway 1738 included in the control plane VCN 1716. The service gateway 1736 included in the control plane VCN 1716 may be communicatively coupled to cloud services 1756 (e.g., cloud services 1656 of FIG. 16 ).

[0201] In some examples, data plane VCN 1718 may be included in customer tenancy 1721. In this case, the IaaS provider may provide a control plane VCN 1716 for each customer, and the IaaS provider may configure a unique compute instance 1744 for each customer that is included in service tenancy 1719. Each compute instance 1744 may enable communication between the control plane VCN 1716 included in service tenancy 1719 and the data plane VCN 1718 included in customer tenancy 1721. The compute instance 1744 may enable resources provisioned in the control plane VCN 1716 included in service tenancy 1719 to be deployed or otherwise used in the data plane VCN 1718 included in customer tenancy 1721.

[0202] In another example, an IaaS provider customer may have a database that resides in customer tenancy 1721. In this example, control plane VCN 1716 may include data plane mirror app tier 1740, which may include app subnet 1726. Data plane mirror app tier 1740 may reside in data plane VCN 1718, but data plane mirror app tier 1740 may not reside in data plane VCN 1718. That is, data plane mirror app tier 1740 may have access to customer tenancy 1721, but data plane mirror app tier 1740 may not reside in data plane VCN 1718 and may not be owned or operated by the IaaS provider customer. Data plane mirror app tier 1740 may be configured to make calls to data plane VCN 1718, but may not be configured to make calls to any entities included in control plane VCN 1716. A customer may desire to deploy or otherwise use resources in the data plane VCN 1718 that have been provisioned in the control plane VCN 1716, and the data plane mirror app layer 1740 can facilitate the desired deployment or other use of the customer's resources.

[0203] In some embodiments, the IaaS provider's customer can apply filters to the data plane VCN 1718. In this embodiment, the customer can determine which data plane VCNs 1718 are accessible, and the customer may restrict access from the data plane VCN 1718 to the public internet 1754. The IaaS provider may not be able to apply filters or otherwise control the data plane VCN 1718's access to any external networks or databases. Applying filters and controls by the customer to the data plane VCN 1718 contained in the customer's tenancy 1721 can help to isolate the data plane VCN 1718 from other customers and from the public internet 1754.

[0204] In some embodiments, cloud services 1756 may be called by service gateway 1736 to access services that may not reside on public internet 1754, control plane VCN 1716, or data plane VCN 1718. The connection between cloud services 1756 and control plane VCN 1716 or data plane VCN 1718 may not be up and running or continuous. Cloud services 1756 may reside on different networks owned or operated by the IaaS provider. Cloud services 1756 may be configured to receive calls from service gateway 1736 and may not be configured to receive calls from public internet 1754. Some cloud services 1756 may be isolated from other cloud services 1756, and control plane VCN 1716 may be isolated from cloud services 1756 that may not be in the same region as control plane VCN 1716. For example, control plane VCN 1716 may be located in "Region 1," and cloud service "deployment 16" may be located in Region 1 and Region 2. If a call is made to deployment 16 by service gateway 1736 included in control plane VCN 1716 located in Region 1, the call may be sent to deployment 16 in Region 1. In this example, control plane VCN 1716, or deployment 16 in Region 1, may not be communicatively coupled to or otherwise in communication with deployment 16 in Region 2.

[0205] 18 is a block diagram 1800 illustrating another example pattern of an IaaS architecture, according to at least one embodiment. A service operator 1802 (e.g., service operator 1602 of FIG. 16 ) may be communicatively coupled to a secure host tenancy 1804 (e.g., secure host tenancy 1604 of FIG. 16 ), which may include a virtual cloud network (VCN) 1806 (e.g., VCN 1606 of FIG. 16 ) and a secure host subnet 1808 (e.g., secure host subnet 1608 of FIG. 16 ). VCN 1806 may include an LPG 1810 (e.g., LPG 1610 of FIG. 16 ), which may be communicatively coupled to an SSH VCN 1812 (e.g., SSH VCN 1612 of FIG. 16 ) via an LPG 1810 included in SSH VCN 1812. SSH VCN 1812 can include SSH subnet 1814 (e.g., SSH subnet 1614 in FIG. 16 ), and SSH VCN 1812 can be communicatively coupled to control plane VCN 1816 (e.g., control plane VCN 1616 in FIG. 16 ) via LPG 1810 included in control plane VCN 1816, and to data plane VCN 1818 (e.g., data plane 1618 in FIG. 16 ) via LPG 1810 included in data plane VCN 1818. Control plane VCN 1816 and data plane VCN 1818 can be included in service tenancy 1819 (e.g., service tenancy 1619 in FIG. 16 ).

[0206] The control plane VCN 1816 may include a control plane DMZ tier 1820 (e.g., the control plane DMZ tier 1620 of FIG. 16 ) that may include a load balancer (LB) subnet 1822 (e.g., the LB subnet 1622 of FIG. 16 ), a control plane app tier 1824 (e.g., the control plane app tier 1624 of FIG. 16 ) that may include an app subnet 1826 (e.g., similar to the app subnet 1626 of FIG. 16 ), and a control plane data tier 1828 (e.g., the control plane data tier 1628 of FIG. 16 ) that may include a DB subnet 1830. The LB subnet 1822 included in the control plane DMZ layer 1820 can be communicatively coupled to an app subnet 1826 included in a control plane app layer 1824 that may be included in the control plane VCN 1816, and to an Internet gateway 1834 (e.g., Internet gateway 1634 in FIG. 16 ), and the app subnet 1826 can be communicatively coupled to a DB subnet 1830 included in a control plane data layer 1828, as well as to a service gateway 1836 (e.g., service gateway in FIG. 16 ) and a network address translation (NAT) gateway 1838 (e.g., NAT gateway 1638 in FIG. 16 ). The control plane VCN 1816 can include the service gateway 1836 and the NAT gateway 1838.

[0207] Data plane VCN 1818 may include a data plane app layer 1846 (e.g., data plane app layer 1646 in FIG. 16 ), a data plane DMZ layer 1848 (e.g., data plane DMZ layer 1648 in FIG. 16 ), and a data plane data layer 1850 (e.g., data plane data layer 1650 in FIG. 16 ). Data plane DMZ layer 1848 may include a trusted app subnet 1860 and an untrusted app subnet 1862 of data plane app layer 1846 and an LB subnet 1822 that may be communicatively coupled to an Internet gateway 1834 included in data plane VCN 1818. Trusted app subnet 1860 may be communicatively coupled to a service gateway 1836 included in data plane VCN 1818, a NAT gateway 1838 included in data plane VCN 1818, and a DB subnet 1830 included in data plane data layer 1850. The untrusted app subnet 1862 may be communicatively coupled to a service gateway 1836 included in the data plane VCN 1818 and to a DB subnet 1830 included in the data plane data layer 1850. The data plane data layer 1850 may include a DB subnet 1830 that may be communicatively coupled to a service gateway 1836 included in the data plane VCN 1818.

[0208] The untrusted app subnet 1862 may include one or more primary VNICs 1864(1)-(N), which may be communicatively coupled to tenant virtual machines (VMs) 1866(1)-(N). Each tenant VM 1866(1)-(N) may be communicatively coupled to a respective app subnet 1867(1)-(N), which may be included in a respective container egress VCN 1868(1)-(N), which may be included in a respective customer's tenancy 1870(1)-(N). Each secondary VNIC 1872(1)-(N) may facilitate communication between the untrusted app subnet 1862 included in the data plane VCN 1818 and the app subnet included in the container egress VCN 1868(1)-(N). Each container egress VCN 1868(1)-(N) may include a NAT gateway 1838, which may be communicatively coupled to the public internet 1854 (e.g., public internet 1654 in FIG. 16 ).

[0209] An internet gateway 1834 included in the control plane VCN 1816 and included in the data plane VCN 1818 may be communicatively coupled to a metadata management service 1852 (e.g., metadata management system 1652 of FIG. 16 ), which may be communicatively coupled to the public internet 1854. The public internet 1854 may be communicatively coupled to a NAT gateway 1838 included in the control plane VCN 1816 and included in the data plane VCN 1818. A service gateway 1836 included in the control plane VCN 1816 and included in the data plane VCN 1818 may be communicatively coupled to cloud services 1856.

[0210] In some embodiments, data plane VCN 1818 may be integrated with a customer's tenancy 1870. This integration may be useful or desirable for an IaaS provider's customer in some cases, such as when they may want support when executing code. A customer may provide code for execution that may be destructive, may communicate with other customers' resources, or may otherwise cause undesirable effects. In response, the IaaS provider may determine whether to execute the code provided to the IaaS provider by the customer.

[0211] In some examples, a customer of an IaaS provider may grant temporary network access to the IaaS provider and request the ability to connect to the data plane app layer 1846. The code to perform this function may run in VMs 1866(1)-(N), which may not be configured to run elsewhere on the data plane VCN 1818. Each VM 1866(1)-(N) may be connected to one customer's tenancy 1870. Each container 1871(1)-(N) contained in a VM 1866(1)-(N) may be configured to run code. In this case, double isolation may exist (e.g., containers 1871(1)-(N) running code may be contained in at least VMs 1866(1)-(N) that are included in untrusted app subnet 1862), which may help prevent incorrect or otherwise unwanted code from causing damage to the IaaS provider's network or to a different customer's network. Containers 1871(1)-(N) may be communicatively coupled to customer tenancy 1870 and may be configured to send or receive data to or from customer tenancy 1870. Containers 1871(1)-(N) may not be configured to send or receive data to or from any other entities in data plane VCN 1818. Upon completion of code execution, the IaaS provider may kill or otherwise destroy containers 1871(1)-(N).

[0212] In some embodiments, trusted app subnet 1860 may execute code that may be owned or operated by the IaaS provider. In this embodiment, trusted app subnet 1860 may be communicatively coupled to DB subnet 1830 and may be configured to perform CRUD operations within DB subnet 1830. Untrusted app subnet 1862 may be communicatively coupled to DB subnet 1830, but in this embodiment, the untrusted app subnet may be configured to perform read operations within DB subnet 1830. Containers 1871(1)-(N) capable of executing code from the customer, which may be included in each customer's VMs 1866(1)-(N), may not be communicatively coupled to DB subnet 1830.

[0213] In other embodiments, the control plane VCN 1816 and the data plane VCN 1818 may not be directly communicatively coupled. In this embodiment, there may not be direct communication between the control plane VCN 1816 and the data plane VCN 1818. However, communication can occur indirectly through at least one method. The LPG 1810 may be established by an IaaS provider and can facilitate communication between the control plane VCN 1816 and the data plane VCN 1818. In another example, the control plane VCN 1816 or the data plane VCN 1818 can make a call to a cloud service 1856 via a service gateway 1836. For example, a call from the control plane VCN 1816 to the cloud service 1856 may include a request for a service that can communicate with the data plane VCN 1818.

[0214] 19 is a block diagram 1900 illustrating another example pattern of an IaaS architecture, according to at least one embodiment. A service operator 1902 (e.g., service operator 1602 of FIG. 16 ) may be communicatively coupled to a secure host tenancy 1904 (e.g., secure host tenancy 1604 of FIG. 16 ), which may include a virtual cloud network (VCN) 1906 (e.g., VCN 1606 of FIG. 16 ) and a secure host subnet 1908 (e.g., secure host subnet 1608 of FIG. 16 ). VCN 1906 may include an LPG 1910 (e.g., LPG 1610 of FIG. 16 ), which may be communicatively coupled to an SSH VCN 1912 (e.g., SSH VCN 1612 of FIG. 16 ) via an LPG 1910 included in SSH VCN 1912. SSH VCN 1912 can include SSH subnet 1914 (e.g., SSH subnet 1614 in FIG. 16 ), and SSH VCN 1912 can be communicatively coupled to control plane VCN 1916 (e.g., control plane VCN 1616 in FIG. 16 ) via LPG 1910 included in control plane VCN 1916, and to data plane VCN 1918 (e.g., data plane 1618 in FIG. 16 ) via LPG 1910 included in data plane VCN 1918. Control plane VCN 1916 and data plane VCN 1918 can be included in service tenancy 1919 (e.g., service tenancy 1619 in FIG. 16 ).

[0215] The control plane VCN 1916 may include a control plane DMZ layer 1920 (e.g., the control plane DMZ layer 1620 of FIG. 16 ) that may include a LB subnet 1922 (e.g., the LB subnet 1622 of FIG. 16 ), a control plane app layer 1924 (e.g., the control plane app layer 1624 of FIG. 16 ) that may include an app subnet 1926 (e.g., the app subnet 1626 of FIG. 16 ), and a control plane data layer 1928 (e.g., the control plane data layer 1628 of FIG. 16 ) that may include a DB subnet 1930 (e.g., the DB subnet 1830 of FIG. 18 ). LB subnet 1922 included in control plane DMZ tier 1920 can be communicatively coupled to app subnet 1926 included in control plane app tier 1924, which may be included in control plane VCN 1916, and to an Internet gateway 1934 (e.g., Internet gateway 1634 in FIG. 16 ), and app subnet 1926 can be communicatively coupled to DB subnet 1930 included in control plane data tier 1928, as well as to service gateway 1936 (e.g., service gateway in FIG. 16 ) and network address translation (NAT) gateway 1938 (e.g., NAT gateway 1638 in FIG. 16 ). Control plane VCN 1916 can include service gateway 1936 and NAT gateway 1938.

[0216] Data plane VCN 1918 may include a data plane app layer 1946 (e.g., data plane app layer 1646 in FIG. 16 ), a data plane DMZ layer 1948 (e.g., data plane DMZ layer 1648 in FIG. 16 ), and a data plane data layer 1950 (e.g., data plane data layer 1650 in FIG. 16 ). Data plane DMZ layer 1948 may include a trusted app subnet 1960 (e.g., trusted app subnet 1860 in FIG. 18 ) and an untrusted app subnet 1962 (e.g., untrusted app subnet 1862 in FIG. 18 ) of data plane app layer 1946, as well as an LB subnet 1922 that may be communicatively coupled to an Internet gateway 1934 included in data plane VCN 1918. The trusted app subnet 1960 may be communicatively coupled to a service gateway 1936 included in the data plane VCN 1918, a NAT gateway 1938 included in the data plane VCN 1918, and a DB subnet 1930 included in the data plane data layer 1950. The untrusted app subnet 1962 may be communicatively coupled to the service gateway 1936 included in the data plane VCN 1918 and the DB subnet 1930 included in the data plane data layer 1950. The data plane data layer 1950 may include a DB subnet 1930 that may be communicatively coupled to the service gateway 1936 included in the data plane VCN 1918.

[0217] The untrusted app subnet 1962 may include primary VNICs 1964(1)-(N), which may be communicatively coupled to tenant virtual machines (VMs) 1966(1)-(N) residing within the untrusted app subnet 1962. Each tenant VM 1966(1)-(N) may execute code within a respective container 1967(1)-(N), which may be communicatively coupled to an app subnet 1926, which may be included in a data plane app layer 1946, which may be included in a container egress VCN 1968. Each secondary VNIC 1972(1)-(N) may facilitate communication between the untrusted app subnet 1962, which is included in the data plane VCN 1918, and the app subnet included in the container egress VCN 1968. The container egress VCN may include a NAT gateway 1938, which may be communicatively coupled to the public internet 1954 (e.g., public internet 1654 in FIG. 16 ).

[0218] An internet gateway 1934 included in the control plane VCN 1916 and included in the data plane VCN 1918 may be communicatively coupled to a metadata management service 1952 (e.g., metadata management system 1652 of FIG. 16 ), which may be communicatively coupled to the public internet 1954. The public internet 1954 may be communicatively coupled to a NAT gateway 1938 included in the control plane VCN 1916 and included in the data plane VCN 1918. A service gateway 1936 included in the control plane VCN 1916 and included in the data plane VCN 1918 may be communicatively coupled to cloud services 1956.

[0219] In some examples, the pattern illustrated by the architecture of block diagram 1900 in FIG. 19 may be considered an exception to the pattern illustrated by the architecture of block diagram 1800 in FIG. 18 and may be desirable for an IaaS provider's customers when the IaaS provider cannot communicate directly with the customer (e.g., in a disconnected region). Each container 1967(1)-(N) contained in a VM 1966(1)-(N) for each customer may be accessed by the customer in real time. The containers 1967(1)-(N) may be configured to make calls to each secondary VNIC 1972(1)-(N) contained in the app subnet 1926 of the data plane app tier 1946, which may be included in a container egress VCN 1968. The secondary VNICs 1972(1)-(N) may send the calls to a NAT gateway 1938, which may send the calls to the public Internet 1954. In this example, containers 1967(1)-(N) that may be accessed in real time by customers may be isolated from control plane VCN 1916 and may be isolated from other entities included in data plane VCN 1918. Containers 1967(1)-(N) may be isolated from other customer resources.

[0220] In another example, a customer can invoke cloud service 1956 using container 1967(1)-(N). In this example, the customer may execute code in container 1967(1)-(N) that requests a service from cloud service 1956. Container 1967(1)-(N) can send the request to secondary VNICs 1972(1)-(N), which can send the request to a NAT gateway, which can send the request to public internet 1954. Public internet 1954 can send the request via internet gateway 1934 to LB subnet 1922, which is included in control plane VCN 1916. In response to determining that the request is valid, LB subnet 1926 can send the request to app subnet 1926, which can send the request to cloud service 1956 via service gateway 1936.

[0221] It should be understood that the IaaS architectures 1600, 1700, 1800, 1900 shown in the figures may include components other than those shown. Furthermore, the illustrated embodiments are merely some examples of cloud infrastructure systems that may incorporate embodiments of the present disclosure. In some other embodiments, the IaaS systems may include more or fewer components than those shown in the figures, may combine two or more components, or may have a different configuration or arrangement of components.

[0222] In one embodiment, the IaaS system described herein may include a suite of application, middleware, and database service offerings delivered to customers in a self-service, subscription-based, elastically scalable, reliable, highly available, and secure manner. An example of such an IaaS system is Oracle Cloud Infrastructure (OCI), offered by the present assignee.

[0223] 20 illustrates an exemplary computer system 2000 in which various embodiments may be implemented. System 2000 may be used to implement any of the computer systems described above. As shown, computer system 2000 includes a processing unit 2004 that communicates with multiple peripheral subsystems via a bus subsystem 2002. These peripheral subsystems may include a processing acceleration unit 2006, an I / O subsystem 2008, a storage subsystem 2018, and a communication subsystem 2024. Storage subsystem 2018 includes a tangible computer-readable storage medium 2022 and a system memory 2010.

[0224] Bus subsystem 2002 provides a mechanism for allowing the various components and subsystems of computer system 2000 to communicate with each other as intended. While bus subsystem 2002 is shown schematically as a single bus, alternative embodiments of the bus subsystem may utilize multiple buses. Bus subsystem 2002 may be any of several types of bus structures, including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures. For example, such architectures may include an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MCA) bus, an Enhanced ISA (EISA) bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus, which may be implemented as a mezzanine bus manufactured to the IEEE P1386.1 standard.

[0225] Processing unit 2004, which may be implemented as one or more integrated circuits (e.g., conventional microprocessors or microcontrollers), controls the operation of computer system 2000. One or more processors may be included in processing unit 2004. These processors may include single-core or multi-core processors. In one embodiment, processing unit 2004 may be implemented as one or more independent processing units 2032 and / or 2034, with a single-core or multi-core processor included in each processing unit. In other embodiments, processing unit 2004 may be implemented as a quad-core processing unit formed by integrating two dual-core processors onto a single chip.

[0226] In various embodiments, processing unit 2004 may execute various programs in response to program code and may maintain multiple simultaneously executing programs or processes. At any particular time, some or all of the program code being executed may reside in processor 2004 and / or in storage subsystem 2018. With appropriate programming, processor 2004 may provide the various functions described above. Computer system 2000 may further include a processing acceleration unit 2006, which may include a digital signal processor (DSP), a special purpose processor, and / or the like.

[0227] The I / O subsystem 2008 may include user interface input devices and user interface output devices. User interface input devices may include a keyboard, a pointing device such as a mouse or trackball, a touchpad or touchscreen integrated into a display, a scroll wheel, a click wheel, a dial, buttons, switches, a keypad, a voice input device with a voice command recognition system, a microphone, and other types of input devices. User interface input devices may include, for example, a motion detection device and / or gesture recognition device, such as a Microsoft Kinect® motion sensor, which allows a user to control an input device, such as a Microsoft Xbox® 360 game controller, to interact with information via a natural user interface using gestures and spoken commands. User interface input devices may also include an eye gesture recognition device, such as a Google Glass® blink detector, which detects a user's eye activity (e.g., "blinking" when taking a photo and / or selecting a menu) and translates the eye gesture as input to an input device (e.g., Google Glass®). Additionally, the user interface input devices may include a voice recognition detection device that allows a user to interact with a voice recognition system (e.g., Siri® Navigator) via voice commands.

[0228] User interface input devices may include, but are not limited to, three-dimensional (3D) mice, joysticks or pointing sticks, gamepads, and graphic tablets, as well as audio / visual devices such as speakers, digital cameras, digital video cameras, portable media players, webcams, image scanners, fingerprint scanners, barcode reader 3D scanners, 3D printers, laser range finders, and eye-tracking devices. Additionally, user interface input devices may include medical imaging input devices such as, for example, computed tomography, magnetic resonance imaging, position emission tomography, and medical ultrasound devices. User interface input devices may also include audio input devices such as, for example, MIDI keyboards, digital musical instruments, and the like.

[0229] User interface output devices may include non-visual displays such as a display subsystem, indicator lights, or audio output devices. The display subsystem may be a flat-panel device, such as a flat-panel device using a cathode ray tube (CRT), a liquid crystal display (LCD), or a plasma display, a projection device, a touch screen, or the like. In general, use of the term "output device" is intended to include all possible types of devices and mechanisms for outputting information from computer system 2000 to a user or to another computer. For example, user interface output devices may include, but are not limited to, various display devices that visually convey textual, graphical, and audio / video information, such as monitors, printers, speakers, headphones, navigation systems, plotters, audio output devices, and modems.

[0230] Computer system 2000 may include a storage subsystem 2018, which includes software elements shown as presently residing in system memory 2010. System memory 2010 may store program instructions readable and executable by processing unit 2004, as well as data generated during the execution of these programs.

[0231] Depending on the configuration and type of computer system 2000, the system memory 2010 may be volatile (such as random-access memory (RAM)) and / or non-volatile (such as read-only memory (ROM), flash memory, etc.). RAM typically contains data and / or program modules that are immediately accessible to and / or presently being operated on and executed by the processing unit 2004. In some implementations, the system memory 2010 may include several different types of memory, such as static random-access memory (SRAM) or dynamic random-access memory (DRAM). In some implementations, the basic input / output system (BIOS), containing the basic routines that help to transfer information between elements within the computer system 2000, such as during start-up, may typically be stored in ROM. By way of example and not limitation, system memory 2010 also illustrates application programs 2012, program data 2014, and an operating system 2016, which may include client applications, web browsers, mid-tier applications, relational database management systems (RDBMS), and the like.Examples of operating systems 2016 may include various versions of Microsoft Windows®, Apple Macintosh®, and / or Linux operating systems, various commercially available UNIX® or UNIX-like operating systems (including, but not limited to, various GNU / Linux operating systems, Google Chrome® OS, etc.), and / or mobile operating systems such as iOS, Windows® Phone, Android® OS, BlackBerry® 20 OS, and Palm® OS operating systems.

[0232] The storage subsystem 2018 may provide a tangible, computer-readable storage medium for storing the basic programming and data configurations that provide the functionality of some embodiments. Software (programs, code modules, instructions) that, when executed by a processor, provide the aforementioned functionality may be stored in the storage subsystem 2018. These software modules or instructions may be executed by the processing unit 2004. The storage subsystem 2018 may provide a repository for storing data used in accordance with the present disclosure.

[0233] Storage subsystem 2000 may include computer-readable storage medium reader 2020, which may further be connected to computer-readable storage medium 2022. In combination with system memory 2010, and together, optionally, computer-readable storage medium 2022 may comprehensively represent remote, local, fixed, and / or removable storage media, as well as storage media for temporarily and / or more permanently containing, storing, transmitting, and retrieving computer-readable information.

[0234] The computer-readable storage medium 2022 containing the code or portions of code may include any suitable medium known or used in the art, including storage and communication media, such as, but not limited to, volatile and nonvolatile, removable and non-removable media, implemented in any manner or technology for storing and / or transmitting information. The computer-readable storage medium 2022 may include tangible computer-readable storage media, such as RAM, ROM, electronically erasable programmable ROM (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device, or other tangible computer-readable medium. The computer-readable storage medium 2022 may also include non-tangible computer-readable media, such as a data signal, data transmission, or any other medium that can be used to transmit desired information and that can be accessed by the computing system 2000.

[0235] By way of example, computer-readable storage medium 2022 may include hard disk drives that read from or write to non-removable, nonvolatile magnetic media, magnetic disk drives that read from or write to removable, nonvolatile magnetic disks, and optical disk drives that read from or write to removable, nonvolatile optical disks such as CD-ROMs, DVDs, and Blu-ray® disks or other optical media. Computer-readable storage medium 2022 may include, but is not limited to, Zip® drives, flash memory cards, universal serial bus (USB) flash drives, secure digital (SD) cards, DVD disks, digital video tapes, and the like. The computer-readable storage media 2022 may include solid-state drives (SSDs) based on non-volatile memory such as flash memory-based SSDs, enterprise flash drives, semiconductor ROM, volatile memory-based SSDs such as semiconductor RAM, dynamic RAM, static RAM, DRAM-based SSDs, magneto-resistive RAM (MRAM) SSDs, and hybrid SSDs that use a combination of DRAM and flash memory-based SSDs. Disk drives and associated computer-readable media may provide non-volatile storage of computer-readable instructions, data structures, program modules, and other data for the computer system 2000.

[0236] The communications subsystem 2024 provides an interface to other computer systems and networks. The communications subsystem 2024 serves as an interface for receiving data from and transmitting data to other systems in the computer system 2000. For example, the communications subsystem 2024 may enable the computer system 2000 to connect to one or more devices via the Internet. In some embodiments, the communications subsystem 2024 may include radio frequency (RF) transceiver components for accessing wireless voice and / or data networks (e.g., using cellular technology, advanced data network technologies such as 3G, 4G, or EDGE (enhanced data rates for global evolution), Wi-Fi (using the IEEE 802.11 family of standards, or other mobile communications technologies, or any combination thereof), global positioning system (GPS) receiver components, and / or other components. In some embodiments, the communications subsystem 2024 may provide a wired network connection (e.g., Ethernet) in addition to or instead of a wireless interface.

[0237] In some embodiments, the communications subsystem 2024 may receive incoming communications in the form of structured and / or unstructured data feeds 2026, event streams 2028, event updates 2030, etc., on behalf of one or more users who may use the computer system 2000.

[0238] By way of example, the communications subsystem 2024 may be configured to receive data feeds 2026 in real time from users of social networks and / or other communications services, such as Twitter® feeds, Facebook® updates, web feeds, such as Rich Site Summary (RSS) feeds, and / or real-time updates from one or more third-party sources.

[0239] Additionally, the communications subsystem 2024 may be configured to receive data in the form of a continuous data stream, which may include an event stream 2028 of real-time events and / or event updates 2030 that may have no explicit end and may be continuous in nature or unbounded. Examples of applications that generate continuous data may include, for example, sensor data applications, financial tickers, network performance measurement tools (e.g., network monitoring and traffic management applications), clickstream analysis tools, automobile traffic monitoring, etc.

[0240] The communications subsystem 2024 may be configured to output structured and / or unstructured data feeds 2026, event streams 2028, event updates 2030, etc. to one or more databases that can communicate with one or more streaming data source computers coupled to the computer system 2000.

[0241] The computer system 2000 can be one of a variety of types, including a handheld portable device (e.g., an iPhone® mobile phone, an iPad® computing tablet, a PDA), a wearable device (e.g., a Google Glass® head-mounted display), a PC, a workstation, a mainframe, a ticket machine, a server rack, or any other data processing system.

[0242] Due to the ever-changing nature of computers and networks, the description of computer system 2000 shown in the figure is intended to be a specific example only. Many other configurations are possible, including more or fewer components than the system shown in the figure. For example, customized hardware may be used, and / or particular elements may be implemented in hardware, firmware, software (including applets), or a combination thereof. Furthermore, connections to other computing devices, such as network input / output devices, may be employed. Based on the disclosure and teachings provided herein, those skilled in the art will appreciate other ways and / or manners for implementing the various embodiments.

[0243] While specific embodiments have been described, various modifications, variations, alternative constructions, and equivalents are encompassed within the scope of the present disclosure. The embodiments are not limited to operation in one particular data processing environment, but can freely operate in multiple data processing environments. Furthermore, while the embodiments have been described using a particular sequence of transactions and steps, it should be apparent to those skilled in the art that the scope of the present disclosure is not limited to the sequence of transactions and steps described. Various features and aspects of the above-described embodiments may be used individually or together.

[0244] Furthermore, while embodiments have been described using particular combinations of hardware and software, it should be recognized that other combinations of hardware and software are within the scope of the present disclosure. Embodiments may be implemented exclusively in hardware, exclusively in software, or using a combination thereof. Various processes described herein may be performed on the same processor or on different processors in any combination. Thus, when a component or module is described as being configured to perform an operation, such configuration may be realized, for example, by designing electronic circuitry to perform the operation, by programming a programmable electronic circuit (such as a microprocessor) to perform the operation, or by any combination thereof. Processes may communicate using various techniques, including, but not limited to, conventional techniques for inter-process communication; different pairs of processes may use different techniques, or the same pair of processes may use different techniques at different times.

[0245] Accordingly, the specification and drawings should be regarded in an illustrative, rather than a limiting sense. However, it will be apparent that additions, subtractions, deletions, and other modifications and changes may be made to the specification and drawings without departing from the broader spirit and scope as set forth in the claims. Accordingly, while particular disclosed embodiments have been described, they are not intended to be limiting. Various modifications and equivalents are within the scope of the appended claims.

[0246] The use of the terms "a," "an," and "the" and similar referents in the context of describing the disclosed embodiments (particularly in the context of the appended claims) should be construed to cover both the singular and the plural, unless otherwise indicated herein or clearly contradicted by context. The terms "comprising," "having," "including," and "containing" should be construed as open-ended (i.e., meaning "including, but not limited to"), unless otherwise noted. The term "connected" should be construed as partially or fully contained within, connected to, or joined together, even if there is something intervening. The recitation of ranges of values ​​herein is merely intended to serve as a shorthand method of individually referring to each separate value included in the range, unless otherwise indicated herein, and each separate value is incorporated herein as if it were individually recited herein. All methods described herein can be performed in any suitable order unless otherwise indicated herein or otherwise clearly contradicted by context. Any and all examples provided herein, or the use of exemplary language (e.g., "etc.") are intended merely to better clarify the embodiments and do not impose limitations on the scope of the disclosure unless specifically claimed. No language in the specification should be construed as indicating any non-claimed element as essential to the practice of the disclosure.

[0247] Disjunctive language, such as the phrase "at least one of X, Y, or Z," is generally intended to be understood within the context as being used to state that an item, condition, etc. may be either X, Y, or Z, or any combination thereof (e.g., X, Y, and / or Z), unless expressly stated otherwise. Thus, such disjunctive language is generally not intended to, and should not, imply that an embodiment requires that at least one of X, at least one of Y, or at least one of Z, respectively, be present.

[0248] Preferred embodiments of the present disclosure are described herein, including the best mode known for carrying out the disclosure. Variations of such preferred embodiments may become apparent to those skilled in the art upon reading the foregoing description. Those skilled in the art will be able to adopt such variations as appropriate, and the present disclosure may be practiced other than as specifically described herein. Accordingly, this disclosure includes all modifications and equivalents of the subject matter recited in the claims appended hereto as permitted by applicable law. Moreover, any combination of the above-described elements in all possible variations of the embodiments is encompassed by the present disclosure, unless otherwise indicated herein.

[0249] All references cited herein, including publications, patent applications, and patents, are incorporated by reference herein to the same extent as if each reference were individually and specifically indicated as incorporated by reference and were set forth herein in its entirety. While the foregoing specification has described aspects of the present disclosure with reference to specific embodiments herein, those skilled in the art will recognize that the present disclosure is not limited thereto. Various features and aspects of the foregoing disclosure may be used individually or together. Furthermore, embodiments may be utilized in any number of environments and applications beyond those described herein without departing from the broader spirit and scope of the present specification. Accordingly, the specification and drawings should be considered illustrative and not restrictive.< / realm>

Claims

1. It is a method, The invention includes providing a plurality of graphical processing unit (GPU) clusters, wherein the plurality of GPU clusters are communicably coupled to each other via a plurality of switches arranged in a hierarchical structure, the hierarchical structure including a first layer of switches, a second layer of switches, and a third layer of switches. The method involves selecting one or more switches from the third layer of the switches to form a set of target switches. Each target switch in the set of target switches receives the address information of each GPU included in the plurality of GPU clusters, Each target switch in the set of target switches generates multiple sets of address information by filtering the received address information based on the conditions, A method further comprising each target switch transmitting a plurality of sets of the address information to each switch included in the first layer of the switch, wherein the switch stores a subset of the plurality of sets of the address information in accordance with the conditions.

2. The method according to claim 1, wherein the plurality of GPU clusters include at least a first GPU cluster operating at a first speed and a second GPU cluster operating at a second speed different from the first speed.

3. The method according to claim 1, further comprising configuring a connection between each target switch in the set of target switches and each switch included in a first layer of the switches, wherein each target switch receives address information of each GPU included in the plurality of GPU clusters from the first layer of the switches via the connection.

4. The method according to claim 3, wherein the connection is a BGP peering connection.

5. The method according to claim 1, wherein the switch included in the first layer of the switch discards other subsets of the multiple sets of address information in accordance with the conditions.

6. The method according to claim 1, wherein the address information of each GPU included in the plurality of GPU clusters corresponds to the MAC address of the GPU.

7. The method according to claim 1, wherein the first layer of the switch is communicatively coupled to the plurality of GPU clusters at one end and communicatively coupled to the second layer of the switch at the other end, and the second layer of the switch is communicatively coupled to the first layer of the switch and to the third layer of the switch.

8. The method according to claim 1, wherein the third layer of the switch is divided into a plurality of groups of the third layer of the switch, and the selection further comprises selecting at least one target switch from each of the plurality of groups of the third layer of the switch.

9. The method according to claim 1, wherein the total number of target switches included in the set of target switches is in the range of 4 to 16.

10. The method according to claim 1, wherein the condition corresponds to the target switch grouping the address information of the GPUs included in the plurality of GPU clusters based on the VLAN of the customer to which the GPU belongs.

11. The method according to claim 1, wherein the switch stores a subset of a plurality of sets of address information in an address table, and the switch is further configured to remove entries from the address table based on a timer associated with the entries.

12. A program comprising computer executable instructions, wherein, when the computer executable instructions are executed by one or more processors, the program causes the one or more processors to perform the method according to any one of claims 1 to 11.

13. One or more processors, A computing device comprising a memory containing instructions, wherein, when the instructions are executed using one or more processors, the computing device causes the computing device to perform the method according to any one of claims 1 to 11.

Images