Data access by virtual processor in distributed system
By introducing virtual I/O interfaces and distributed virtual hypervisors in distributed computing systems, the delay and resource consumption problems caused by data access requests across computer nodes are solved, local processing and data consistency are achieved, and system performance is optimized and efficiency.
Patent Information
- Application Number
- CN202410310925.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-06-02
- Filing Date
- 2024-03-19
- Publication Date
- 2024-12-03
AI Technical Summary
In distributed computing systems, data access requests across computer nodes lead to processing delays and resource consumption, especially when data and metadata need to be processed and transmitted between multiple computer nodes, affecting system performance.
Introduce virtual I/O interface in a software-defined server, data requests and responses are made through the queue data structure between the virtual processor and the virtual I/O processor, ensuring that data access is processed on the local computer nodes and reducing cross-node traffic. And maintain the consistency of data structures through a distributed hypervisor, and use machine learning modules to optimize VCPU and page migration to reduce delays.
It effectively reduces data transmission and processing delays between computer nodes, improves system performance, reduces resource consumption, and improves system efficiency by optimizing data access paths.
Smart Images

Figure CN119065780A_ABST
Abstract
Description
Background Art
[0001] A distributed computing arrangement may include a cluster of computer nodes to provide increased processing throughput compared to a single node system. The computer nodes may execute corresponding programs for performing corresponding operations. In some cases, a distributed computing arrangement may include a virtualized distributed system including a virtualized environment in which programs may be executed. BRIEF DESCRIPTION OF THE DRAWINGS
[0002] Some implementations of the disclosure are described with reference to the following figures.
[0003] Figure 1 is a block diagram of an arrangement including a cluster of computer nodes coupled to a shared storage system, according to some examples.
[0004] Figure 2 is a block diagram of an arrangement including a cluster of computer nodes, virtual machines (VMs) across the computer nodes, and hyperkernels across the computer nodes, according to some examples.
[0005] Figure 3 and Figure 4 is a flow diagram of a process according to some examples.
[0006] Figure 5 is a block diagram of an arrangement including a cluster of computer nodes, a shared storage system, and a management console according to some examples.
[0007] Figure 6 is a block diagram of a distributed system according to some examples.
[0008] Figure 7 is a block diagram of a storage medium storing machine-readable instructions according to some examples.
[0009] Figure 8 is a flow diagram of a process according to some examples.
[0010] Throughout the drawings, the same reference numerals denote similar, but not necessarily identical, elements. The drawings are not necessarily drawn to scale, and the dimensions of certain components may be exaggerated to more clearly illustrate the examples shown. In addition, the drawings provide examples and / or implementations consistent with the description; however, the description is not limited to the examples and / or implementations provided in the drawings. DETAILED DESCRIPTION
[0011] An example of a virtualized environment of a virtualized distributed system is a virtual machine (VM) in which an operating system (OS) (referred to as a "guest OS") and one or more application programs are executed. In some examples, a computer node can execute one or more VMs. In such an example, multiple VMs are executed across a cluster of computer nodes.
[0012] In other examples, a VM can be executed across multiple computer nodes. In such an example, instead of executing multiple VMs in a computer node, a VM can be executed on multiple computer nodes. A virtualized distributed system in which a VM can be executed across multiple computer nodes can be referred to as a software-defined server. A software-defined server includes a distributed virtual hypervisor that aggregates the physical resources of the computer nodes of the virtualized distributed system and presents the aggregation of physical resources to the guest OS of the VM. A virtual hypervisor is also referred to as a hypervisor. A distributed virtual hypervisor is referred to as a hyperkernel (discussed further below).
[0013] Programs (e.g., guest OS, applications, etc.) running in a VM of a software-defined server can request access to data in a storage system. In some examples, the computer nodes of a software-defined server can each be independently connected to a corresponding different storage system. A "storage system" includes one or more storage devices. In such an example, if a program in a first computer node requests access to data in a storage system connected to a different second computer node (read access or write access), the access request must be forwarded from the first computer node to the second computer node, which performs the requested data access and returns the result of the data access to the first computer node. The transmission of access requests and access requested data (read data or write data) between computer nodes increases the latency of completing access requests and consumes resources (processing resources, communication resources, memory resources, etc.) of a virtualized distributed system.
[0014] In other examples, a network attached storage system may be connected to a computer node of a software-defined server via a network. In such an example, any computer node may access the network attached storage system via the network. The network attached storage system may be referred to as a "shared storage system" because the storage system is accessible from any computer node of the software-defined server. However, even in the example where the shared storage system is connected to a software-defined server, the distributed virtual hypervisor in the software-defined server may not be able to efficiently process data access requests originating from different computer nodes. For example, in order to process a data access request originating from or received at a first computer node, the first computer node may distribute work to multiple other computer nodes to access data associated with the data access request. Such a distribution of work across multiple computer nodes may cause a large amount of traffic (both data and metadata) between multiple computer nodes.
[0015] According to some implementations of the present disclosure, a virtual input / output (I / O) interface is deployed in a software-defined server and includes a data structure having a corresponding queue used by a virtual processor in a VM of the software-defined server to submit a request for access to data of a network-attached storage system. The virtual I / O interface is deployed between the VM and the shared storage system (or more specifically, between the VM and a distributed virtual hypervisor that manages access to the shared storage system). A given virtual processor in a given computer node uses a corresponding data structure in the virtual I / O interface to submit a request for access to data in the shared storage system and receive the result of the request. The data access in response to the request uses a local I / O connection from a given computer node to the shared storage system so that access to the data that satisfies the request can be performed locally at the given computer node without involving any other computer node. Local processing of requests at computer nodes reduces or eliminates traffic (including data and / or metadata) between computer nodes when processing the request.
[0016] A distributed virtual hypervisor in a software defined server maintains consistency of a data structure across multiple computer nodes to ensure that information in the data structure remains consistent so that a consumer of the data structure is reading the latest information that may have been written to the data structure by a writer. In a first scenario, the writer writing information to the data structure may be a given virtual processor, and the consumer of the data structure may be a virtual I / O processor in the distributed virtual hypervisor (discussed further below). In this first scenario, the given virtual processor may add information of an access request (write request or read request) to the data structure, and the virtual I / O processor may read the information of the access request from the data structure.
[0017] In a second scenario, the information writer of the data structure may be a virtual I / O processor, and the consumer of the data structure may be a given virtual processor. In this second scenario, the virtual I / O processor may add the result of the access request (e.g., an indication of completion of the write request or read data in response to a read request) to the data structure, and the given virtual processor may read the result of the access request from the data structure.
[0018] Figure 11 is a block diagram of an example arrangement of a virtualized distributed system, referred to as a software-defined server 100, according to some examples. The software-defined server 100 includes a cluster of computer nodes 102-1, 102-2, ..., 102-N (N ≥ 2). A computer node is a physical processing platform with various hardware resources. For example, computer node 102-1 includes hardware resources 104-1, which has processing resources 106-1 for executing machine-readable instructions, physical memory 108-1 for storing information (including data and machine-readable instructions), I / O resources 112-1 for performing I / O operations (such as using a shared storage system 116), and cache memory 114-1 for buffering information for faster access.
[0019] The other computer nodes 102-2 to 102-N similarly include corresponding hardware resources 104-2 to 104-N. The hardware resources 104-2 include processing resources 106-2, physical memory 108-2, I / O resources 112-2, and cache memory 114-2. The hardware resources 104-N include processing resources 106-N, physical memory 108-N, I / O resources 112-N, and cache memory 114-N.
[0020] The processing resources may include one or more hardware processors. The hardware processor may include a microprocessor, a core of a multi-core microprocessor, a microcontroller, a programmable integrated circuit, a programmable gate array, or another hardware processing circuit.
[0021] The physical memory may be implemented using one or more memory devices. The memory devices may include any or some combination of the following: dynamic random access memory (DRAM) devices, static random access memory (SRAM) devices, flash memory devices, etc.
[0022] The I / O resources may include I / O interfaces, such as a network adapter or a network interface controller, to allow computer nodes to communicate over a network, such as the communication fabric 118. Each I / O resource 112-i (i=1 to N) of the computer node 102-i provides a local I / O connection to the shared storage system 116 through the communication fabric 118. The communication fabric 118 may be implemented using any type of network, such as an Ethernet network, a Fibre Channel network, a Fibre Channel over Ethernet network, an Internet Small Computer System Interface (iSCSI) network, a non-volatile memory high-speed network over a Fibre Channel network, a Compute Express Link (CXL) network, and the like.
[0023] The cache memory may include one or more memory devices having an access speed higher than the access speed of the main memory (which may be any of 108-1, 108-2, ..., 108-N). The cache memory may include a hierarchy of cache memory devices at different levels. For example, one or more higher-level cache memory devices (e.g., L1 cache memory devices and L2 cache memory devices) may be part of a hardware processor, and one or more lower-level cache memory devices (e.g., L3 cache memory devices) may be external to the hardware processor.
[0024] A "shared" storage system is a storage system that is shared (i.e., accessible) by any computer node in the cluster of computer nodes 102-1 to 102-N via communication structure 118. Shared storage system 116 is implemented using a collection of storage devices 120. As used herein, a "collection" of items may refer to a single item or multiple items. Thus, a collection of storage devices 120 may include a single storage device or multiple storage devices. Examples of storage devices may include any or some combination of the following: disk-based storage devices, solid-state drives, and the like.
[0025] In the software defined server 100, the VM 122 may be executed across multiple computer nodes 102-1 to 102-N, rather than executing one or more VMs in each corresponding computer node 102-i. Figure 1 One VM 122 is shown, but note that there may be at least one other VM capable of running across multiple computer nodes 102-1 through 102-N. Figure 1 The arrangement of is referred to as a software defined server because the computer nodes 102 - 1 through 102 -N collectively form one virtual machine (VM 122 ) that hosts a guest operating system (OS) 124 . The VM 122 also includes a collection of applications 126 that run in the VM 122 .
[0026] The application 126 can run as a guest thread in a computer node. Multiple threads of the same application 126 or different applications 126 can run on multiple computer nodes. In addition, multiple instances 132-1 to 132-N of the hyperkernel 132 run in corresponding computer nodes 102-1 to 102-N; multiple hyperkernel instances 132-1 to 132-N interact with each other through a network structure, using messages passed through Ethernet or another type of interface to jointly provide the functions of the guest OS 124 in the VM 122. The message passing is based on a message passing protocol that supports parallel computing in a distributed computing environment.
[0027] VM 122 also includes guest address mapping information 128, which maps guest virtual memory addresses of guest physical memory address space 130 (accessible by programs in VM 122, which includes a collection of guest OS 124 and application programs 126) to addresses of a hyperkernel address space used by hyperkernel 132. The guest physical address space refers to the location of guest physical memory. From the perspective of guest OS 124, guest physical memory is treated as physical memory. However, guest physical memory is virtual memory (in the hyperkernel address space) provided by hyperkernel 132.
[0028] The hyper-kernel 132 includes hyper-kernel address mapping information 134, which maps addresses in the hyper-kernel address space to the physical address spaces of the computer nodes 102-1 to 102-N. In some examples, the guest address mapping information 128 includes a guest page table, and the hyper-kernel address mapping information 134 includes an extended page table. A page table is a data structure that stores information to map between addresses in different address spaces. The guest page table and the extended page table together form a hierarchical arrangement of page tables for mapping between the guest physical memory address space 130 and the physical address spaces of the computer nodes 102-1 to 102-N. In other examples, the guest address mapping information 128 and the hyper-kernel address mapping information 134 can be implemented using other types of data structures.
[0029] The physical address space of computer nodes 102 - 1 through 102 -N includes physical addresses that refer to locations in physical memory 108 - 1 through 108 -N, cache memory 114 - 1 through 114 -N, and shared storage system 116 .
[0030] Hyper-kernel 132 includes software that acts as a hypervisor in each computer node. Figure 1 As shown, hyperkernel instances 132-1 to 132-N run in corresponding computer nodes 102-1 to 102-N. Each hyperkernel instance 132-i (i=1 to N) simulates the hardware resources 104-i of the corresponding computer node 102-i and allocates the simulated hardware resources for use by VM 122. Hyperkernel 132 (acting as a distributed hypervisor) also manages the execution of VM 122, including creating, tearing down or modifying VM 122. If present in software-defined server 100, hyperkernel 132 can also manage the execution of other VMs.
[0031] Hyper-kernel instances 132 - 1 through 132 -N are able to communicate with each other through a message passing interface 138 , which may be a network interface such as Ethernet or another type of interface.
[0032] Each hyperkernel instance 132-i can observe the software-defined server 100 running in real time and optimize the resources of the corresponding computer node 102-i to match the requirements of the software-defined server 100 during operation. The hyperkernel instances 132-1 to 132-N federate (aggregate) the hardware resources 104-1 to 104-N of the computer nodes 102-1 to 102-N and present the federated set to the guest OS 124.
[0033] The emulation of processing resources 106-1 through 106-N by hyperkernel instances 132-1 through 132-N provides virtual processors (or equivalently, virtual central processing units or VCPUs) in VM 122. The VCPUs may execute on processing resources of the corresponding computer nodes. Figure 1 In the example of , VCPUs 140 - 11 and 140 - 12 run in compute node 102 - 1 , VCPU 140 - 21 runs in compute node 102 - 2 , and VCPUs 140 - N1 , 140 - N2 , and 140 - N3 run in compute node 102 -N .
[0034] Each VCPU can execute a thread of an application 126. The thread of an application 126 running on a VCPU is called a "guest thread". The guest threads of (multiple) application 126 can be scheduled by the guest OS 124 on the corresponding VCPU. From the perspective of the guest OS 124, each VCPU is a physical CPU, and the scheduler of the guest OS 124 can run the corresponding guest thread of the application 126 on the physical CPU. The guest thread running on the VCPU can issue a request to perform I / O operations, such as to access data of virtual storage. Virtual storage is a simulated instance of the shared storage system 116.
[0035] The VCPUs on computer node 102-i are bound to physical processors in processing resources 106-i of computer node 102-i. Binding a VCPU to a physical processor refers to presenting a simulated instance of a physical processor to guest OS 124 as a VCPU. A VCPU may be migrated from a source computer node to a destination computer node. If a VCPU is migrated, the VCPU will be bound to a physical processor that is part of the processing resources of the destination computer node.
[0036] A virtual I / O processor (VIOP) is also executed in each computer node 102-1 to 102-N. In some examples, each VIOP is implemented as part of a corresponding hypercore instance. For example, hypercore instance 132-1 includes VIOP 136-1, hypercore instance 132-2 includes VIOP 136-2, and hypercore instance 132-N includes VIOP 136-N.
[0037] In order to access virtual storage, the VCPU interacts with the corresponding VIOP on the computer node on which the VCPU runs. In some examples, a virtual I / O interface is provided between each VCPU and the VIOP. As an example, the virtual I / O interface between the VCPU and the VIOP (or more generally, between the VM 122 and the hyper-kernel 132) is a VirtIO interface. The VirtIO interface is a standardized interface that allows a VM to access virtual I / O devices, such as virtual storage devices.
[0038] The VirtIO interface uses descriptors that include information describing the corresponding I / O tasks to be executed. The I / O tasks can be requested by the VCPU, or more specifically, by a guest thread executing on the VCPU. In some examples, the VirtIO interface includes a descriptor queue corresponding to the corresponding I / O tasks to be executed. The chain of descriptors represents a collection of I / O tasks requested by the I / O request. As an example, the I / O request can specify that data is written to a corresponding storage location of the shared storage system 116. The storage locations may be discontinuous, so multiple I / O tasks are executed to write data to discontinuous storage locations.
[0039] In some examples, the VirtIO interface may include multiple queues between the VCPU and the VIOP. For different I / O requests issued by the corresponding VCPU, descriptors may be added to the multiple queues.
[0040] like Figure 1 As shown in FIG. 1 , the dotted line between the VCPU and the VIOP represents the corresponding VirtIO interface. Figure 1 In the example of , VirtIO interface 144-11 is provided between VCPU 140-11 and VIOP 136-1, and VirtIO interface 144-12 is provided between VCPU 140-12 and VIOP 136-1. In addition, VirtIO interface 144-21 is provided between VCPU 140-21 and VIOP 136-2, and VirtIO interfaces 144-N1, 144-N2, and 144-N3 are provided between corresponding VCPUs 140-N1, 140-N2, and 140-N3 and VIOP 136-N.
[0041] Each VirtIO interface includes a corresponding queue to allow interaction between a corresponding pair of VCPUs and VIOPs. In this example, it is assumed that there is one queue per VirtIO interface. More generally, a VirtIO interface may include a collection of queues (a single queue or multiple queues). The queues of a VirtIO interface are part of the guest physical address space 130. Figure 1In the example of FIG. 1 , queues 146 - 11 , 146 - 12 , . . . , 146 -N3 are part of corresponding VirtIO interfaces 144 - 11 , 144 - 12 , . . . , 144 -N3 .
[0042] Each queue includes a collection of pages (single page or multiple pages), where "page" refers to a unit of data of a given size. The VCPU can add a descriptor of an I / O task to one or more pages of the queue, and the VIOP can add information about the result of the I / O request to one or more pages of the queue. The VCPU can also retrieve read data from one or more pages of the queue.
[0043] The guest physical address space 130 also includes other data structures, including VirtIO metadata 148-11, 148-12, ..., 140-N3 for corresponding VirtIO interfaces 144-11, 144-12, ..., 144-N3. Each VirtIO metadata includes information associated with the corresponding VirtIO interface. For example, the VirtIO metadata may include information indicating which computer node the queue (or page of the queue) is located on.
[0044] The data structure of the guest physical address space 130 can be stored in the virtual memory of VM 122, wherein the virtual memory is a simulated instance of one or more physical memories in the physical memory 108-1, 108-2, ..., 108-N in the computer node 102-1, 102-2, ..., 102-N. Thus, for example, the page of the queue of the VirtIO interface can be stored in the virtual memory as the simulated instance of the physical memory. The data of the queue page can physically reside on the physical memory of a specific computer node. For example, the (multiple) pages of the queue 146-11 can be stored in the virtual memory, which is the simulated instance of the physical memory 108-1 in the computer node 102-1. At a later point in time, the (multiple) pages of the queue 146-11 may be migrated from the computer node 102-1 to another computer node, such as due to the VCPU 140-11 migrating to another computer node and requesting the access of the (multiple) pages of the queue 146-11.
[0045] Each hypercore instance 132-i includes a corresponding arbitrator (ARB), a consistency engine (COH), and a migration engine (MIG). The arbitrator, consistency engine, and migration engine may be implemented using machine-readable instructions of the hypercore instance 132-i. Figure 1In the example of , hyperkernel instance 132-1 includes arbitrator instance 150-1, consistency engine instance 152-1, and migration engine instance 154-1; hyperkernel instance 132-2 includes arbitrator instance 150-2, consistency engine instance 152-2, and migration engine instance 154-2; and hyperkernel instance 132-N includes arbitrator instance 150-N, consistency engine instance 152-N, and migration engine instance 154-N. Collectively, arbitrator instances 150-1 to 150-N form an arbitrator (or more specifically, a distributed arbitrator), consistency engine instances 152-1 to 152-N form a consistency engine (or more specifically, a distributed consistency engine), and migration engine instances 154-1 to 154-N form a migration engine (or more specifically, a distributed migration engine).
[0046] The arbitrator is responsible for ensuring the fairness of access to data pages (including queue pages) by various entities in the cluster of computer nodes 102-1 to 102-N. For example, in a scenario where multiple entities are requesting access to queue pages at the same time, the arbitrator can set an upper limit on how long the consumer of the queue page waits to access the queue page. For example, the VCPU and VIOP can concurrently attempt to access the queue page. In order to ensure fairness to both the VCPU and the VIOP, the arbitrator can ensure that the VCPU and the VIOP will not wait for more than the upper limit delay time for access to the requested queue page. If the consumer has been waiting for the queue page for too long (that is, the delay of the access of the queue page is close to the upper limit delay time), the arbitrator can give a higher priority to the request for accessing the queue page of the consumer. In a scenario where the queue page must be migrated back and forth between multiple computer nodes, the possibility of delayed access to the queue page is greater. An example of the migration of the queue page is further provided below.
[0047] In other examples where multiple VCPUs are assigned to one queue, the arbitrator may also arbitrate between the multiple VCPUs to ensure fairness, such as by setting an upper limit latency time for access requested by the VCPUs.
[0048] The consistency engine is used to maintain the consistency of data in each page (including queue pages) due to competition for pages by multiple entities. Maintaining consistency ensures that any read of data in a page (including queue pages) is the correct version of the data in the page.
[0049] The migration engine manages the migration of VCPUs and pages (including queue pages) between computer nodes. As discussed further below, the migration engine includes a machine learning module that applies a machine learning process to decide whether VCPUs and pages are to be moved, and if so, to which computer nodes.
[0050] The arbitrator instances 150-1, 150-2, ..., 150-N can interact with each other through the message passing interface 138 to perform arbitration tasks, the consistency engine instances 152-1, 152-2, ..., 152-N can interact with each other through the message passing interface 138 to maintain the consistency of data pages, and the migration engine instances 154-1, 154-2, ..., 154-N can interact with each other through the message passing interface 138 to manage the migration of VCPUs and pages.
[0051] Figure 2 1 is a block diagram of an example arrangement including computer nodes 102-1 and 102-2 to illustrate an example in which a VCPU (VCPU A) has issued a request to add a descriptor associated with an I / O access to a queue page (Queue Page A) that is located on a different computer node than the computer node on which VCPU A is running. Figure 2 In the example of , VCPU A is executing in computer node 102 - 1 , and queue page A is stored at computer node 102 - 2 . Figure 2 Also shown are VIOP 136-1 executing in computer node 102-1 and VIOP 136-2 executing in computer node 102-2.
[0052] Note that initially VCPU A may have been running on computer node 102-2, but for some reason VCPU A was migrated from computer node 102-2 to computer node 102-1. An example of a reason for migrating VCPU A between computer nodes is that the VCPU may be running inefficiently on computer node 102-2 and there is a destination computer node (102-1) with less load. In such a scenario, the migration engine may decide to migrate VCPU A from computer node 102-2 to computer node 102-1.
[0053] like Figure 2 As further shown, VM 122 includes a guest page table 250. Guest page table 250 is Figure 1 An example of guest address mapping information 128.
[0054] The hyperkernel 132 includes a hyperkernel page table, a migration engine, a consistency engine, and an arbitrator. The hyperkernel page table (also called an extended page table) is Figure 1 An example of hyper-kernel address mapping information 134.
[0055] like Figure 2As shown, an instance of the hyperkernel page table 202 resides in each computer node (hyperkernel page table instance 202-1 in computer node 102-1, and hyperkernel page table instance 202-2 in computer node 102-2). In addition, the migration engine includes migration engine instances 154-1 and 154-2 in corresponding computer nodes 102-1 and 102-2, the consistency engine includes consistency engine instances 152-1 and 152-2 in corresponding computer nodes 102-1 and 102-2, and the arbitrator includes arbitrator instances 150-1 and 150-2 in corresponding computer nodes 102-1 and 102-2.
[0056] Each migration engine instance 154-i includes a machine learning module 204-i, which includes an access frequency probability function 206-i and a Markov model 208-i. The Markov model 208-i applies a Markov decision process, which is a discrete-time stochastic process for decision making. In other examples, other models can be used by the machine learning module 204-i. As part of learning based on training data or previous operations of the machine learning module 204-i, the parameters of the access frequency probability function 206-i and the Markov model 208-i can be dynamically adjusted.
[0057] like Figure 3 As further shown, to initiate an I / O access (read access or write access), VCPU A (at 302) issues a request to update queue page A (210). Since queue page A is not resident at computer node 102-1, an access conflict occurs because the requested resource is not available in computer node 102-1, and in response, VM 122 (at 304) issues an interrupt that is intercepted by hyperkernel 132. In response to the interrupt, since queue page A is not resident in computer node 102-1, a hyperkernel page table fault is issued by hyperkernel page table instance 202-1 at computer node 102-1 (at 306).
[0058] In response to the hyperkernel page table failure, the machine learning module instance 204-1 in the computer node 102-1 determines (at 308) whether VCPU A is to be migrated from the computer node 102-1 to another computer node, in this case the computer node 102-2 containing the queue page A.
[0059] Inputs considered by the Markov model 208-1 in the machine learning module instance 204-1 include the VCPU node affinity information 212-1 stored at the computer node 102-1, and whether VCPU A has issued a request to perform an I / O access (a read or write of virtual storage). Other inputs may also be considered by the Markov model 208-1 to determine whether VCPU A should be migrated.
[0060] The VCPU node affinity information 212-1 includes information indicating the affinity of each VCPU to each computer node in the cluster of computer nodes. The affinity of a VCPU to a computer node can be based on various factors, such as (1) the frequency with which the VCPU is accessing the resources (e.g., pages) of the computer node, (2) how efficiently the VCPU is running in the computer node, (3) whether the VCPU is in the middle of executing an instruction set in the computer node, and the like. If the VCPU accesses the resources of the computer node at a higher frequency, the affinity between the VCPU and the computer node can be increased. If the VCPU is running inefficiently in the computer node, the affinity between the VCPU and the computer node can be reduced. On the other hand, if the VCPU is running efficiently in the computer node, the affinity between the VCPU and the computer node can be increased. If the VCPU is in the middle of executing an instruction set in the computer node, the affinity between the VCPU and the computer node can be increased (migrating the VCPU while the VCPU is executing an instruction set will be inefficient because the execution context will have to be moved along with any cached data).
[0061] Another input considered by Markov model 208-1 is whether VCPU A has issued a request to perform an I / O access (a read or write of virtual storage). If VCPU A has issued a request to perform an I / O access, then based on the goal of preferring local processing of I / O accesses to data in shared storage system 116 (as opposed to forwarding I / O access requests to another computer node for processing), it will be weighted more heavily in favor of retaining VCPU A at the current node.
[0062] The affinity between a VCPU and a computer node can be expressed using an affinity metric adjusted by the hypercore 132 based on various factors. The VCPU node affinity information 212-1 is part of a global page map, which is a data structure used to store metadata for pages including queue pages for the VirtIO interface. An instance of the global page map is stored in each computer node. Figure 2A global page map instance 210-1 in computer node 102-1 and a global page map instance 210-2 in computer node 102-2 are shown. Global page map instance 210-2 includes VCPU node affinity information 212-2. Hypercore 132 may maintain global page map instances 210-1 and 210-2 in sync with each other (i.e., to ensure that global page map instances remain the same when they are modified).
[0063] Each global page map (GPM) instance 210-1 or 210-2 also includes corresponding VIOP-page affinity information 214-1 or 214-2. The VIOP-page affinity information indicates the affinity between a VIOP and a page (such as a queue page). If the VIOP is accessing (writing or reading) a page, the VIOP page affinity is set to a first value. On the other hand, if the VIOP is not accessing the page, the VIOP-page affinity is set to a second value different from the first value. If the corresponding VIOP-page affinity indicating the affinity between a given VIOP and a given page (e.g., a queue page) is set to a first value, this indicates a condition that the given page should not be moved. However, once the given VIOP has completed its access to the given page, the corresponding VIOP-page affinity is reset to a second value to allow migration of the given page.
[0064] If the Markov model 208-1 of the machine learning module 204-1 outputs a decision that VCPU A is to be migrated, then the migration engine 154-1 migrates (at 310) VCPU A to the computer node 102-2.
[0065] On the other hand, if the Markov model 208-1 of the machine learning module 204-1 outputs a decision that VCPU A should not be migrated from the computer node 102-1, the migration engine instance 154-1 sends a message to another computer node (at 312) for queue page A. Metadata (e.g., Figure 1 1) may be stored at computer node 102-1 indicating where queue page A was last known to be located. Assuming the metadata indicates that queue page A is located at computer node 102-2, migration engine 154-1 sends a message for queue page A to computer node 102-2.
[0066] The message contains information to be used at computer node 102-2 to decide whether to move queue page A from computer node 102-2 to computer node 102-1. The information contained in the message may include a write error (indicating that the message was sent because VCPU A attempted to write to queue page A but queue page A was not available at computer node 102-1), an identification of VCPU A, and other information.
[0067] In response to receiving the message at computer node 102-2, machine learning module 204-2 in migration engine instance 154-2 determines (at 314) whether queue page A is to be migrated to computer node 102-1 based on the information in the message and information related to queue page A at computer node 102-2. A factor considered by Markov model 208-2 in machine learning module instance 204-2 in deciding whether to migrate queue page A is contention for queue page A. Contention for queue page A may be indicated by VIOP-page affinity information 214-2 at computer node 102-2. If VIOP 136-2 is accessing queue page A, the VIOP-page affinity for VIOP 136-2 to queue page A is set to a first value (as described above) to prevent migration of queue page A as long as VIOP 136-2 is actively accessing queue page A. If another entity (such as another VCPU or another entity) is accessing queue page A, there may also be contention for queue page A. In the example where the queue is assigned to only one VCPU, it is unlikely that another VCPU will access queue page A. However, in other examples where the queue may be assigned to multiple VCPUs, then there may be more than one VCPU that may compete for access to queue page A.
[0068] If there is no contention for queue page A, migration engine 154-2 migrates queue page A to computer node 102-1 (at 316). Once queue page A is migrated to computer node 102-1, VCPU A may add a descriptor for the I / O access to queue page A. VIOP 136-1 at computer node 102-1 may read the descriptor from queue page A and initiate an I / O access.
[0069] However, if there is contention for queue page A, migration engine 154-2 sends a rejection indication (e.g., a rejection message, a rejection information element, etc.) to computer node 102-1 (at 318). The rejection indication is forwarded by hyperkernel instance 132-1 at computer node 102-1 to VCPU A. The rejection indication may include a retry indicator to indicate to VCPU A that its request for access to queue page A has been rejected, but that VCPU A should retry the request at a later point in time, such as after a certain timeout interval.
[0070] The rejection indication with a retry is to allow contention for queue page A to be resolved, such as due to VOIP 136-2 or another entity completing access to queue page A. When VCPU A later retries the request for queue page A and contention no longer exists, migration engine instance 154-2 can migrate queue page A to computer node 102-1.
[0071] In other examples, rather than rejecting the request with a retry indicator, migration engine instance 154-2 may simply send a busy indication back to computer node 102-1 to notify VCPU A that there is currently an ongoing access to queue page A, and migration engine 154-1 will send queue page A to computer node 102-1 when the access is complete.
[0072] Assuming that machine learning module 204-2 determines that queue page A is to be migrated from computer node 102-2 to computer node 102-1, machine learning module 204-2 also determines what state queue page A should be set to when migrating to computer node 102-1. Access frequency probability function instance 206-2, which is part of machine learning module 204-2, generates a value indicating the frequency of reads or writes to queue page A.
[0073] As an example, the value generated by the access frequency probability function instance 206-2 can indicate the frequency of reading queue page A. The value can be compared with the frequent read threshold to determine whether queue page A is a read-dominated page. If the value exceeds the frequent read threshold, it indicates that queue page A is a read-dominated page. If the value does not exceed the frequent read threshold, it indicates that queue page A is not a read-dominated page.
[0074] As another example, the value generated by the access frequency probability function instance 206-2 can indicate the frequency of writes to queue page A. The value can be compared with the frequent write threshold to determine whether queue page A is a read-dominated page. If the value is less than the frequent write threshold, it indicates that queue page A is a read-dominated page. If the value exceeds the frequent write threshold, it indicates that queue page A is not a read-dominated page.
[0075] Since queue page A is likely to be written at a relatively high frequency (i.e., queue page A is unlikely to be a read-dominated page), the machine learning module instance 206-2 can set (or maintain) queue page A in a readable and writable state. In such an example, the migration engine 154-2 migrates queue page A in a readable and writable state, which means that queue page A can be read or written after being migrated to the computer node 102-1.
[0076] Other possible states of the page may include a read-only state or an execute-only state. If the access frequency probability function instance 206-2 indicates that queue page A is a read-dominated page, the machine learning module instance 206-2 may set queue page A to a read-only state. In such an example, the migration engine 154-2 migrates queue page A in a read-only state, which means that queue page A can only be read after being migrated to the computer node 102-1.
[0077] The access frequency probability function (including multiple access frequency probability function instances 206-1 and 206-2) is also used by the consistency engine to decide whether a queue page is to be replicated. Because the queue page may not be a read-dominated page, the consistency engine can prevent the replication of the queue page. In this way, the consistency engine ensures that there is only a single copy of the queue page, so that the consistency engine does not have to perform invalidation of queue page copies as part of the consistency protocol.
[0078] Figure 4 is a flow diagram of an example process according to some examples. VCPU A in computer node 102-1 issues (at 402) a write request to add a descriptor to queue page A (which is initially located at computer node 102-2). Subsequently, VOIP 136-2 in computer node 102-2 issues (at 404) a read request to access queue page A. Coherence engine 152, which is part of hyperkernel 132, maintains the consistency of the contents of queue page A by (1) causing any modified copies of queue page A to be flushed to shared storage system 116, and (2) ensuring that VIOP 136-2 reads the latest version of the contents of queue page A updated by VCPU A. Coherence engine 152 includes Figure 4 In the example, consistency engine instances 152-1 and 152-2.
[0079] In response to the write request from VCPU A and the read request from VIOP 136-2, consistency engine 152 issues a temporary block on writes by performing a permission change (at 406). The permission change may refer to changing the permissions associated with queue page A to prohibit VCPU A from writing to queue page A. The permissions may be part of the metadata associated with queue page A.
[0080] Coherence engine 152 triggers (at 408) a flush of any modified copies of queue page A (which may be stored in cache memory). Coherence engine 152 may cause computer node 102-2 to flush any modified copies of queue page A. Coherence engine 152 may also cause other computer nodes, if any, to flush any modified copies of queue page A. For example, the flush may be a cache line (or another cache portion).
[0081] After the flush has been completed, the hyperkernel 132 (or more specifically, the migration engine in the hyperkernel) migrates (at 410) queue page A from computer node 102-2 to computer node 102-1. The hyperkernel 132 also changes the permissions for VCPU A to be able to write to queue page A (to remove the temporary block issued at 406). At this point, VCPU A in computer node 102-1 is able to write to queue page A (at 412). After queue page A has been written, the migration engine migrates (at 414) queue page A to computer node 102-2, at which point VIOP 136-2 can read queue page A (at 416).
[0082] Figure 5 is how an indicator is associated with a logical storage volume 502 to indicate that the data of the logical storage volume should be globally accessible from any computer node in the cluster of computer nodes 500 (such as Figure 1 1 to 102-N) in the example of a block diagram accessible by a logical storage volume. A "logical storage volume" refers to a logical storage structure for storing data, such as a logical unit (LU) identified by a logical unit number (LUN). The data of the logical storage volume 502 will be stored in the shared storage system 116.
[0083] A user at a management console 504 (eg, which may be a computer) assigns a global label 506 to a logical storage volume 502. The management console 504 may communicate with the cluster 500 of computer nodes over a network.
[0084] The "global tag" includes an indicator (e.g., a predefined value, a flag, etc.) that the logical storage volume 502 will be available from any computer node of the cluster of computer nodes. The global tag 506 is added to the configuration information 508 stored by the hyperkernel 132 on the cluster of computer nodes 500. The global tag 506 in the configuration information 508 indicates to the hyperkernel 132 that the data of the logical storage volume 502 will be accessed in parallel (such as across multiple nodes of the cluster of computer nodes) for requests initiated or received by the corresponding computer node. Figure 1 1 to 112-N in the corresponding computer nodes shown in FIG), and each request should be processed locally by the receiving computer node to avoid sending the I / O task of the request across multiple computer nodes. Based on the global tag 506, the hyper core 132 creates a local path from each computer node to the logical storage volume 502 stored in the shared storage system 116.
[0085] Figure 6 is a block diagram of a distributed system 600 according to some examples. Examples of distributed system 600 are shown in FIG. Figure 1 The distributed system 600 includes a plurality of computer nodes 602 , with the computer nodes 620 including corresponding physical I / O interfaces 604 to network attached storage systems 606 .
[0086] Distributed system 600 includes a distributed virtualization hypervisor 608 that executes across multiple computer nodes 602. An example of distributed virtualization hypervisor 608 is Figure 1 The hyperkernel 132 shown in FIG.
[0087] The distributed system 600 also includes a virtual I / O interface 610 between the distributed hypervisor 608 and a guest OS 612 of a VM 614. The VM 614 executes across multiple computer nodes 602. In some examples, the virtual I / O interface 610 includes a VirtIO interface, or more specifically, multiple VirtIO interfaces between corresponding virtual processors and VIOPs.
[0088] The distributed system 600 also includes a plurality of virtual processors 616 executing in the VM 614. The plurality of virtual processors 616 will access the network attached storage system 606 in parallel through the virtual I / O interface 610 using the corresponding physical I / O interfaces. A first virtual processor in a first computer node in the plurality of computer nodes 602 will submit a request to access data in the network attached storage system 606 using a data structure 618 in the virtual I / O interface 610. An example of the data structure 618 is a VirtIO queue. The first virtual processor will also use the data structure 618 to receive the result of the request. The access to the data from the network attached storage system 606 that satisfies the request is performed locally at the first computer node using the physical I / O interface of the first computer node.
[0089] The distributed hypervisor 608 also maintains consistency of the data structure 618 across the plurality of computer nodes 602 .
[0090] In some examples, the coherency maintained by the distributed virtual hypervisor 608 controls selective access by the data structure 618 of different entities including the first virtual processor and the virtual I / O processor in the first computer node.
[0091] In some examples, the first virtual processor will add a representation of the request (eg, a descriptor) to data structure 618 , and the virtual I / O processor will add the results of the request to data structure 618 .
[0092] In some examples, the distributed virtual hypervisor 608 will arbitrate requests for the data structure 618 from different entities and, as part of the arbitration, set an upper limit on the latency to be experienced by the first virtual processor when accessing the data structure 618 .
[0093] In some examples, the consistency maintained by distributed virtual hypervisor 608 forces a single copy for data structure 618 to exist in distributed system 600 for the first virtual processor.
[0094] In some examples, multiple data structures are assigned to respective virtual processors 616. The respective virtual processors 616 submit requests to access data in the network attached storage system 606 using the multiple data structures.
[0095] In some examples, data structure 618 is assigned only to the first virtual processor.
[0096] In some examples, data structure 618 is assigned to the first virtual processor and at least one other virtual processor. Coherency maintained by distributed virtual hypervisor 608 controls access to data structure 618 by the first virtual processor and at least one other virtual processor.
[0097] In some examples, in response to the request from the first virtual processor, the distributed virtual hypervisor 608 determines that the data structure 618 is in the second computer node, and in response to determining that the data structure 618 is in the second computer node, the distributed virtual hypervisor 608 migrates the data structure 618 from the second computer node to the first computer node.
[0098] In some examples, before moving the data structure 618 from the second computer node to the first computer node, the distributed virtual hypervisor 608 determines whether the data structure 618 is requested at the second computer node to update the data structure 618 with a result for another request. In response to determining that the data structure is requested at the second computer node, the distributed virtual hypervisor 608 delays moving the data structure 618 from the second computer node to the first computer node.
[0099] In some examples, the delay is based on the distributed virtual hypervisor 608 sending a rejection indication in response to the request, the rejection indication being used to cause the first virtual processor to retry the request at a later time.
[0100] In some examples, in response to the request from the first virtual processor, the distributed virtual hypervisor 608 applies a machine learning process to determine whether to move the data structure 618 from the second computer node to the first computer node. The decision of the machine learning process is based on an indication that an error occurred at the first computer node in response to the request from the first virtual processor and on information related to the use of the data structure at the second computer node.
[0101] In some examples, the decision of the machine learning process is also based on the affinity between the first virtual processor and each computer node in the plurality of computer nodes 602 .
[0102] In some examples, different virtual processors in different computer nodes 602 will access data in the network attached storage system 606 in parallel through the respective local I / O connections of the different computer nodes 602 .
[0103] In some examples, the distributed hypervisor 608 detects global tags (e.g., Figure 5 In response to detecting the global tag, the distributed hypervisor 608 performs local processing of the request at a corresponding computer node of the plurality of computer nodes 602 .
[0104] Figure 7 is a block diagram of a non-transitory machine-readable or computer-readable storage medium 700 storing machine-readable instructions that, when executed, cause a distributed system to perform various tasks. The machine-readable instructions include VM execution instructions 702 for executing a VM across multiple computer nodes that are connected to a network attached storage system through corresponding physical I / O interfaces.
[0105] The machine-readable instructions include distributed virtual hypervisor execution instructions 704 for executing an instance of a distributed virtual hypervisor in a corresponding computer node in the plurality of computer nodes. The instance of the distributed virtual hypervisor may include Figure 1 The hyperkernel instances 132-1 to 132-N in FIG.
[0106] The machine-readable instructions include virtual I / O interface provisioning instructions 706 for providing a virtual I / O interface between the distributed hypervisor and the guest OS of the VM. The virtual I / O interface includes a queue that can be used by multiple virtual processors in the VM to request parallel access to data in the network attached storage system through corresponding physical I / O interfaces.
[0107] The machine readable instructions include request receiving instructions 708 for receiving, at the distributed virtual hypervisor, a request from a first virtual processor in a first computer node to use a first queue in a virtual I / O interface.
[0108] The machine-readable instructions include queue location determination instructions 710 for determining, using the distributed virtual hypervisor, that the first queue is located at the second computer node. The machine-readable instructions include queue migration instructions 712 for migrating, using the distributed virtual hypervisor, the first queue from the second computer node to the first computer node to allow the first virtual processor to add requested information to the first queue to allow local access to data from the network attached storage system that satisfies the request using a physical I / O interface of the first computer node.
[0109] The machine readable instructions include queue coherency maintaining instructions 714 for maintaining coherency of the first queue across the first computer node and the second computer node utilizing the distributed hypervisor.
[0110] Figure 8 is a flow diagram of a process 800 according to some examples, which may be performed in a distributed system.
[0111] Process 800 includes executing (at 802) a VM across a plurality of computer nodes connected to a network attached storage system through respective physical I / O interfaces.
[0112] Process 800 includes (at 804) executing an instance of a distributed hypervisor in a corresponding computer node of a plurality of computer nodes. Process 800 includes (at 806) providing a virtual I / O interface between the distributed hypervisor and a guest OS of a VM, the virtual I / O interface including a queue usable by a plurality of virtual processors in the VM to request concurrent access to data in a network attached storage system through corresponding physical I / O interfaces.
[0113] Process 800 includes receiving (at 808) at a distributed hypervisor from a first virtual processor in a first computer node a request to use a first queue in a virtual I / O interface. Process 800 includes determining (at 810) by the distributed hypervisor that the first queue is at a second computer node.
[0114] Process 800 includes (at 812) determining, by the distributed hypervisor, using a machine learning process, that a first queue is to be migrated from a second computer node to a first computer node, wherein a decision of the machine learning process is based on information regarding usage of the first queue at the second computer node and an affinity of the first virtual processor to each of the plurality of computer nodes.
[0115] Process 800 includes (at 814) migrating, by the distributed hypervisor, a first queue from a second computer node to the first computer node to allow a first virtual processor to add requested information to the first queue to allow access to data specified by the request from a network attached storage system using a physical I / O interface of the first computer node, wherein access to the data specified by the request is performed locally at the first computer node without involving any other computer node in accessing the data specified by the request.
[0116] Process 800 includes (at 816) maintaining, by the distributed hypervisor, consistency of the first queue across the first computer node and the second computer node.
[0117] Storage media (e.g. Figure 7700 in ) may include any or some combination of the following: a semiconductor memory device (such as a dynamic or static random access memory (DRAM or SRAM), an erasable and programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), and flash memory; a magnetic disk, such as a fixed disk, a floppy disk, and a removable disk; another magnetic medium, including magnetic tape; an optical medium, such as a compact disk (CD) or a digital video disk (DVD); or another type of storage device. Note that the instructions discussed above may be provided on one computer-readable or machine-readable storage medium, or alternatively, may be provided on multiple computer-readable or machine-readable storage media distributed in a large system that may have multiple nodes. Such (multiple) computer-readable or machine-readable storage media are considered part of an article (or article of manufacture). An article or article of manufacture may refer to any single component or multiple components of any manufacture. One or more storage media may be located in the machine that runs the machine-readable instructions, or at a remote site from which the machine-readable instructions may be downloaded over a network for execution.
[0118] In the present disclosure, the use of the terms "a", "an" or "the" is also intended to include the plural forms, unless the context clearly indicates otherwise. In addition, the terms "includes", "including", "comprises", "comprising", "have" or "having" when used in the present disclosure specify the presence of stated elements, but do not exclude the presence or addition of other elements.
[0119] In the foregoing description, many details are set forth to provide an understanding of the subject matter disclosed herein. However, implementations may be practiced without some of these details. Other implementations may include modifications and variations from the details discussed above. The appended claims are intended to cover such modifications and variations.
Claims
1. A distributed system, comprising: a plurality of computer nodes including corresponding physical input / output I / O interfaces to a network attached storage system; a distributed virtualization hypervisor executed across the plurality of computer nodes; a virtual I / O interface between the distributed virtual hypervisor and a guest operating system OS of a virtual machine VM, the VM being executed across the plurality of computer nodes; as well as a plurality of virtual processors to execute in the VMs and to access the network attached storage system through the virtual I / O interfaces in parallel using the corresponding physical I / O interfaces, wherein a first virtual processor in a first computer node of the plurality of computer nodes is to submit requests to access data in the network attached storage system using data structures in the virtual I / O interfaces and to receive results of the requests, and wherein the accessing of data from the network attached storage system to satisfy the requests is performed locally at the first computer node using the physical I / O interfaces of the first computer node, The distributed hypervisor is configured to maintain consistency of the data structure across a plurality of computer nodes in the plurality of computer nodes.
2. The distributed system of claim 1 , wherein the consistency is maintained by the distributed virtual hypervisor, the distributed virtual hypervisor controlling selective access to the data structure by different entities, the different entities comprising the first virtual processor and a virtual I / O processor in the first computer node.
3. The distributed system of claim 2, wherein the first virtual processor is configured to add a representation of the request to the data structure, and the virtual I / O processor is configured to add a result of the request to the data structure.
4. The distributed system of claim 1 , wherein the distributed virtual hypervisor is configured to arbitrate requests for the data structure from different entities and, as part of the arbitration, to set an upper limit on a delay to be experienced by the first virtual processor in accessing the data structure.
5. The distributed system of claim 1, wherein the consistency maintained by the distributed virtual hypervisor forces a single copy of the data structure to exist in the distributed system for the first virtual processor.
6. The distributed system of claim 1, comprising a plurality of data structures assigned to respective ones of the plurality of virtual processors, the respective virtual processors to use the plurality of data structures to submit requests to access data in the network attached storage system.
7. The distributed system of claim 1, wherein the data structure is assigned only to the first virtual processor.
8. The distributed system of claim 1 , wherein the data structure is assigned to the first virtual processor and at least one other virtual processor, and wherein the coherency maintained by the distributed virtual hypervisor controls access to the data structure by the first virtual processor and the at least one other virtual processor.
9. The distributed system of claim 1, wherein the distributed hypervisor is configured to: In response to the request from the first virtual processor, determining that the data structure is in a second computer node; and In response to determining that the data structure is in the second computer node, the data structure is migrated from the second computer node to the first computer node.
10. The distributed system of claim 9, wherein in response to the first virtual processor migrating from the second computer node to the first computer node, the first virtual processor is in the first computer node and the data structure is in the second computer node.
11. The distributed system of claim 9, wherein the distributed hypervisor is configured to: Prior to moving the data structure from the second computer node to the first computer node, determining whether the data structure is requested at the second computer node to update the data structure with a result for another request; and In response to determining that the data structure is requested at the second computer node, moving the data structure from the second computer node to the first computer node is delayed.
12. The distributed system of claim 11, wherein the delay is based on the distributed virtual hypervisor sending a rejection indication in response to the request, the rejection indication being used to cause the first virtual processor to retry the request at a later time.
13. The distributed system of claim 9, wherein the distributed hypervisor is configured to: In response to the request from the first virtual processor, applying a machine learning process to determine whether to move the data structure from the second computer node to the first computer node, wherein a decision of the machine learning process is based on an indication that a failure occurred at the first computer node in response to the request from the first virtual processor and based on information related to usage of the data structure at the second computer node.
14. The distributed system of claim 13, wherein the decision of the machine learning process is further based on an affinity between the first virtual processor and each of the plurality of computer nodes.
15. The distributed system of claim 1, wherein different virtual processors in different computer nodes are used to access data in the network attached storage system in parallel through respective local I / O connections of the different computer nodes.
16. The distributed system of claim 1, wherein the distributed hypervisor is configured to: detecting a global label for a logical storage volume in the network attached storage system; and In response to detecting the global tag, local processing of the request is performed at a corresponding computer node of the plurality of computer nodes.
17. A non-transitory machine-readable storage medium comprising instructions that, when executed, cause a distributed system to: executing a virtual machine VM across a plurality of computer nodes, the plurality of computer nodes being connected to a network attached storage system by respective physical input / output I / O interfaces; executing an instance of a distributed virtualization hypervisor in a corresponding computer node of the plurality of computer nodes; providing a virtual I / O interface between the distributed hypervisor and a guest operating system (OS) of the VM, the virtual I / O interface including a queue usable by a plurality of virtual processors in the VM to request concurrent access to data in the network attached storage system through the corresponding physical I / O interface; receiving, at the distributed virtual hypervisor, a request from a first virtual processor in a first computer node to use a first queue in the virtual I / O interface; determining, using the distributed hypervisor, that the first queue is at a second computer node; migrating, using the distributed virtual hypervisor, the first queue from the second computer node to the first computer node to allow the first virtual processor to add information of the request to the first queue to allow local access to data of the network attached storage system that satisfies the request using the physical I / O interface of the first computer node; as well as The consistency of the first queue is maintained across the first computer node and the second computer node using the distributed hypervisor.
18. The non-transitory machine-readable storage medium of claim 17, wherein the migration of the first queue comprises migrating one or more pages of the first queue from the second computer node to the first computer node.
19. A method of a distributed system, comprising: executing a virtual machine VM across a plurality of computer nodes, the plurality of computer nodes being connected to a network attached storage system by respective physical input / output I / O interfaces; executing an instance of a distributed hypervisor in a corresponding computer node of the plurality of computer nodes; providing a virtual I / O interface between the distributed hypervisor and a guest operating system (OS) of the VM, the virtual I / O interface including a queue usable by a plurality of virtual processors in the VM to request parallel access to data in the network attached storage system through the corresponding physical I / O interface; receiving, at the distributed hypervisor, a request from a first virtual processor in a first computer node to use a first queue in the virtual I / O interface; determining, by the distributed hypervisor, that the first queue is at a second computer node; determining, by the distributed hypervisor using a machine learning process, that the first queue is to be migrated from the second computer node to the first computer node, wherein a decision of the machine learning process is based on information regarding usage of the first queue at the second computer node and an affinity of the first virtual processor to each of the plurality of computer nodes; as well as migrating, by the distributed hypervisor, the first queue from the second computer node to the first computer node to allow the first virtual processor to add information of the request to the first queue to allow access to data specified by the request from the network attached storage system using the physical I / O interface of the first computer node, wherein the access to the data specified by the request is performed locally at the first computer node without involving any other computer node in accessing the data specified by the request; as well as Coherency of the first queue is maintained across the first computer node and the second computer node by the distributed hypervisor.
20. The method of claim 19, wherein the machine learning process is weighted to favor retaining the first virtual processor at the first computer node and migrating the first queue to allow local processing of the request at the first computer node.
Citation Information
Patent Citations
Intraconnection communication method of distributed virtual machine monitoring apparatus
CN101430674A
Hierarchical dynamic scheduling
US11159605B2
Network attached memory using selective resource migration
US11240334B2
Virtual disk management for efficient bin packing across nodes
US20220100551A1
