A multi-tenant FPGA network interface card (NIC) resource management method, FPGA NIC, system, and media

By adopting a multi-level structure of dynamically reconfigurable regions and a resource scheduling method in multi-tenant FPGA network cards, the problem of imbalance between resource utilization and data flow efficiency is solved, achieving efficient resource utilization and improved data flow efficiency.

CN123114708APending Publication Date: 2026-11-17UNIV OF SCI & TECH OF CHINA +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611601951.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-16
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing multi-tenant FPGA network interface card (NIC) resource management methods struggle to balance resource utilization and data flow efficiency, leading to resource idleness and fragmentation, and failing to meet the requirements of high-performance network protocols for low latency and high bandwidth in the data plane.

Method used

The system adopts a multi-level structure of dynamic reconfigurable zones, decomposes tasks to be processed into deployable units, and deploys them by matching target reconfigurable pages according to resource requirements. Communication is achieved through common parent reconfigurable pages, which monitors and schedules tenant resource requirements, thereby reducing resource idleness and fragmentation.

Benefits of technology

It improves the overall working efficiency of multi-tenant FPGA network cards, increases data flow efficiency, and reduces resource idleness and fragmentation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123114708A_ABST
    Figure CN123114708A_ABST
Patent Text Reader

Abstract

This invention discloses a multi-tenant FPGA network interface card (NIC) resource management method and system, belonging to the field of resource scheduling technology. The invention decomposes the tasks to be processed from multiple tenants into deployable units and deploys them to target reconfigurable pages. The communication relationship between these target reconfigurable pages is configured to communicate with each other through a common parent reconfigurable page. This enables the deployment and processing of tasks based on the deployment. During task processing, the resource requirements of tenants are monitored, and resource scheduling is performed based on these requirements. This reduces resource idleness and fragmentation in the multi-tenant FPGA NIC and improves data flow efficiency, ultimately enhancing the overall working efficiency of the multi-tenant FPGA NIC.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of resource scheduling technology, and in particular to a multi-tenant FPGA network interface card (NIC) resource management method, as well as the FPGA NIC, system, and media. Background Technology

[0002] With the continuous improvement of cloud data center network bandwidth, network interface cards (NICs) are gradually becoming important infrastructure for carrying virtualized data paths, transmission protocol processing, load balancing, security policy enforcement, storage access, and application proximity data movement. FPGA-based NICs combine hardware-level packet processing capabilities with strong programmability, making them suitable for deploying various network functions on the NIC side.

[0003] In practice, network interface cards (NICs) are typically not exclusive to a single tenant, but rather shared by multiple tenants or service chains. Different tenants may have different combinations of network functions, throughput targets, latency targets, and load variation patterns. Tenant traffic is often bursty, and network functions may also be adjusted in response to changes in cloud platform policies, service models, or security rules.

[0004] Existing multi-tenant FPGA network interface card (NIC) resource management methods often employ static partitioning-based resource management or fixed service chain-based service orchestration. These methods allocate fixed physical or logical regions or fixed network function service chains to tenants to handle tasks, easily leading to resource idleness and fragmentation within the multi-tenant FPGA NIC, resulting in low overall resource utilization. To overcome the resource fixation defects of static partitioning-based resource management and fixed service chain-based service orchestration methods, some existing solutions introduce interconnection management methods based on centralized switching architectures. These centralized switching structures facilitate data interaction and resource scheduling between tenants or functional modules. However, as the number of tenants increases, centralized switching structures face severe timing convergence difficulties, cabling congestion, and single-point performance bottlenecks. This causes the effective data throughput in multi-tenant FPGA NICs to decrease significantly with increasing tenant numbers, failing to meet the rigid requirements of high-performance network protocols for low latency and high bandwidth in the data plane.

[0005] It is evident that existing multi-tenant FPGA network interface card (NIC) resource management methods struggle to achieve an effective balance between resource utilization and data flow efficiency. In other words, existing multi-tenant FPGA NICs cannot maintain high overall operating efficiency as the number of tenants expands.

[0006] Therefore, a new multi-tenant FPGA network interface card (NIC) resource management method is needed to solve the problems existing in the current multi-tenant FPGA NIC resource management methods. Summary of the Invention

[0007] This invention provides a multi-tenant FPGA network interface card (NIC) resource management method and system, which improves the overall working efficiency of multi-tenant FPGA NICs by achieving an effective balance between resource utilization and data flow efficiency.

[0008] To address the aforementioned technical problems, the first aspect of this invention discloses a multi-tenant FPGA network interface card (NIC) resource management method. The FPGA NIC includes at least a dynamically reconfigurable region, which has a multi-level structure, including an upper-level reconfigurable page and multiple lower-level reconfigurable pages nested within the upper-level reconfigurable page. The method includes: Obtain pending tasks from multiple tenants and decompose the pending tasks into deployable units; For each deployable unit, the matching page granularity corresponding to the deployable unit is obtained according to the resource requirements of the deployable unit; Based on the matching page granularity corresponding to the deployable unit and the resource status information of the FPGA network card, the target reconfigurable page corresponding to the deployable unit is obtained. The target reconfigurable page is used to carry the deployable unit and execute the computing task corresponding to the deployable unit. The communication relationship between the deployable units is configured to communicate with each other through the common parent reconfigurable page of the target reconfigurable page corresponding to the deployable unit, wherein the common parent reconfigurable page is the common parent reconfigurable page corresponding to all the target reconfigurable pages as the lower reconfigurable pages; Based on the communication relationship between the target reconfigurable page corresponding to all the deployable units and all the deployable units, all the tasks to be processed are processed and the resource status information is updated; During the processing of the pending task, the resource requirements of all the tenants are monitored, and resource scheduling is performed on the tenants according to the resource requirements.

[0009] As an optional implementation, in the first aspect of the present invention, the step of monitoring the resource requirements of all tenants and scheduling resources for the tenants according to the resource requirements during the processing of the task to be processed includes: During the processing of the pending task, the resource requirements of all the tenants are monitored; Based on the resource requirements and resource status information of all the tenants, the dominant resource share corresponding to each tenant is obtained; Resources are allocated to the tenant with the lowest dominant resource share among all the tenants. Reclaim the remaining resources from all tenants whose resource requirements are lower than their allocated share of resources.

[0010] As an optional implementation, in the first aspect of the invention, allocating resources to the tenant based on the resource requirements of the tenant with the lowest dominant resource share among all the tenants includes: An allocation set is obtained based on the dominant resource share of all the resources allocated to all the tenants and the upper limit of the resource share corresponding to all the tenants. The allocation set includes all the tenants whose dominant resource share is lower than the upper limit of the corresponding resource share. Resources are allocated to the tenant with the lowest dominant resource share in the allocation set based on their resource requirements.

[0011] As an optional implementation, in the first aspect of the present invention, the resource status information is a residual resource map, wherein: The nodes in the residual resource graph are used to represent the available information of the corresponding target reconfigurable page. The attributes of the nodes include one or more of the following: resource balance, page granularity, page level, parent reconfigurable page, and current load. The edges in the residual resource graph are used to represent the connection relationship or available communication path between the corresponding target reconfigurable pages, and the attributes of the edges include one or more of the following: ingress bandwidth, egress bandwidth, queuing status, arbitration status, and path level.

[0012] As an optional implementation, in the first aspect of the present invention, the tenant's resource requirements include the tenant's page capacity requirements, ingress bandwidth requirements, egress bandwidth requirements, and storage access requirements. The initial resource requirements of the tenant are derived from one or more of the following: compile-time resource profile, tenant declaration, historical deployment results, or service-level objectives.

[0013] As an optional implementation, in a first aspect of the invention, configuring the communication relationship between the deployable units to communicate with each other through a common parent reconfigurable page of the target reconfigurable page corresponding to the deployable unit includes: An interconnection node is set on the parent reconfigurable page of the target reconfigurable page, and the interconnection node is used to enable communication between the target reconfigurable pages; When the target reconfigurable page is located under the same parent reconfigurable page, the deployable units communicate with each other through the interconnection nodes on the same parent reconfigurable page; When the target reconfigurable page is located under different parent reconfigurable pages, the deployable units communicate with each other through the interconnection node on the nearest common parent reconfigurable page of the target reconfigurable page.

[0014] As an optional implementation, in the first aspect of the present invention, the step of acquiring the pending tasks of multiple tenants and decomposing the pending tasks into deployable units includes: Obtain the pending tasks of multiple tenants and the network function directed acyclic graph corresponding to the pending tasks; Based on one or more of the following: the weight of the edge in the directed acyclic graph corresponding to the network function of the task to be processed, the resource usage of the task to be processed, the throughput target of the task to be processed, the cross-page communication cost of the task to be processed, and the service level target of the task to be processed, the task to be processed is decomposed into deployable units.

[0015] A second aspect of this invention discloses a multi-tenant FPGA network interface card (NIC), wherein the FPGA NIC includes at least a dynamically reconfigurable region, the dynamically reconfigurable region having a multi-level structure, including a higher-level reconfigurable page and multiple lower-level reconfigurable pages nested within the higher-level reconfigurable page, and the FPGA NIC further includes a decomposition module, a matching module, a deployment module, a communication module, a running module, and a scheduling module, wherein: The decomposition module is used to obtain the tasks to be processed from multiple tenants and decompose the tasks to be processed into deployable units. The matching module is used to obtain the matching page granularity corresponding to each deployable unit based on the resource requirements of the deployable unit. The deployment module is used to obtain the target reconfigurable page corresponding to the deployable unit based on the matching page granularity corresponding to the deployable unit and the resource status information of the FPGA network card. The target reconfigurable page is used to carry the deployable unit and execute the computing task corresponding to the deployable unit. The communication module is used to configure the communication relationship between the deployable units to communicate with each other through the common parent reconfigurable page of the target reconfigurable page corresponding to the deployable unit, wherein the common parent reconfigurable page is the common parent reconfigurable page corresponding to all the target reconfigurable pages as the lower reconfigurable pages; The running module is used to process all the tasks to be processed and update the resource status information according to the communication relationship between the target reconfigurable page corresponding to all the deployable units and all the deployable units; The scheduling module is used to monitor the resource requirements of all tenants during the processing of the task to be processed, and to perform resource scheduling on the tenants according to the resource requirements.

[0016] As an optional implementation, in a second aspect of the present invention, the scheduling module monitors the resource requirements of all tenants during the processing of the task to be processed, and performs resource scheduling on the tenants according to the resource requirements, including: During the processing of the pending task, the resource requirements of all the tenants are monitored; Based on the resource requirements and resource status information of all the tenants, the dominant resource share corresponding to each tenant is obtained; Resources are allocated to the tenant with the lowest dominant resource share among all the tenants. Reclaim the remaining resources from all tenants whose resource requirements are lower than their allocated share of resources.

[0017] As an optional implementation, in a second aspect of the present invention, the specific operation method by which the scheduling module allocates resources to the tenant with the lowest dominant resource share among all the tenants includes: An allocation set is obtained based on the dominant resource share of all the resources allocated to all the tenants and the upper limit of the resource share corresponding to all the tenants. The allocation set includes all the tenants whose dominant resource share is lower than the upper limit of the corresponding resource share. Resources are allocated to the tenant with the lowest dominant resource share in the allocation set based on their resource requirements.

[0018] As an optional implementation, in a second aspect of the invention, the resource status information is a residual resource map, wherein: The nodes in the residual resource graph are used to represent the available information of the corresponding target reconfigurable page. The attributes of the nodes include one or more of the following: resource balance, page granularity, page level, parent reconfigurable page, and current load. The edges in the residual resource graph are used to represent the connection relationship or available communication path between the corresponding target reconfigurable pages, and the attributes of the edges include one or more of the following: ingress bandwidth, egress bandwidth, queuing status, arbitration status, and path level.

[0019] As an optional implementation, in a second aspect of the present invention, the tenant's resource requirements include the tenant's page capacity requirements, ingress bandwidth requirements, egress bandwidth requirements, and storage access requirements. The initial resource requirements of a tenant are derived from one or more of the following: compile-time resource profile, tenant declaration, historical deployment results, or service-level objectives.

[0020] As an optional implementation, in a second aspect of the invention, the specific operation method by which the communication module configures the communication relationship between the deployable units to communicate with each other through the common parent reconfigurable page of the target reconfigurable page corresponding to the deployable unit includes: An interconnection node is set on the parent reconfigurable page of the target reconfigurable page, and the interconnection node is used to enable communication between the target reconfigurable pages; When the target reconfigurable page is located under the same parent reconfigurable page, the deployable units communicate with each other through the interconnection nodes on the same parent reconfigurable page; When the target reconfigurable page is located under different parent reconfigurable pages, the deployable units communicate with each other through the interconnection node on the nearest common parent reconfigurable page of the target reconfigurable page.

[0021] As an optional implementation, in a second aspect of the invention, the specific operation method by which the decomposition module obtains pending tasks from multiple tenants and decomposes the pending tasks into deployable units includes: Obtain the pending tasks of multiple tenants and the network function directed acyclic graph corresponding to the pending tasks; Based on one or more of the following: the weight of the edge in the directed acyclic graph corresponding to the network function of the task to be processed, the resource usage of the task to be processed, the throughput target of the task to be processed, the cross-page communication cost of the task to be processed, and the service level target of the task to be processed, the task to be processed is decomposed into deployable units.

[0022] The third aspect of this invention discloses a multi-tenant FPGA network interface card (NIC) system, the system comprising: a memory storing executable program code; a processor coupled to the memory; the processor calling the executable program code in the memory to execute the multi-tenant FPGA NIC resource management method disclosed in the first aspect of this invention.

[0023] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when executed by a processor, implement the multi-tenant FPGA network interface resource management method disclosed in the first aspect of the present invention.

[0024] Implementing this invention has the following beneficial effects: This invention provides a multi-tenant FPGA network interface card (NIC) resource management method. It decomposes the tasks to be processed from multiple tenants into deployable units and deploys them to target reconfigurable pages. The communication relationship between these target reconfigurable pages is configured to communicate with each other through a common parent reconfigurable page. This enables the deployment and processing of tasks based on the deployment. During task processing, the resource requirements of tenants are monitored, and resource scheduling is performed based on these requirements. This reduces resource idleness and fragmentation in the multi-tenant FPGA NIC and improves data flow efficiency, ultimately enhancing the overall working efficiency of the multi-tenant FPGA NIC. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart illustrating a multi-tenant FPGA network interface card resource management method disclosed in an embodiment of the present invention; Figure 2 This is a schematic diagram of a process for resource scheduling for tenants disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a multi-tenant FPGA network interface resource management system disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of another multi-tenant FPGA network interface resource management system disclosed in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a multi-tenant FPGA network card disclosed in an embodiment of the present invention. Detailed Implementation

[0027] To better understand and implement this invention, the technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0028] It should be noted that, unless otherwise expressly specified and limited, the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0029] Example 1 Existing multi-tenant FPGA network interface card (NIC) resource management methods often adopt static or rigid allocation modes, which can easily lead to resource idleness and fragmentation in multi-tenant FPGA NICs, as well as a significant decrease in effective data throughput as the number of tenants increases. This results in insufficient resource utilization and internal data flow efficiency in multi-tenant FPGA NICs, ultimately affecting the overall working efficiency of multi-tenant FPGA NICs.

[0030] Therefore, this invention discloses a new multi-tenant FPGA network interface card (NIC) resource management method to solve the problems existing in the above-mentioned existing multi-tenant FPGA NIC resource management methods.

[0031] Please see Figure 1 The method may include: Step 110: Obtain the pending tasks from multiple tenants and decompose the pending tasks into deployable units.

[0032] The tasks to be processed include, but are not limited to, data plane forwarding tasks, security protocol encryption / decryption / unloading tasks, storage protocol remote direct memory access processing tasks, and custom message parsing and reassembly tasks. As long as the task can be abstracted into having a clear input data stream, output data stream, and independently schedulable processing logic, this embodiment does not limit the specific form of the tasks to be processed.

[0033] The deployable unit can be a single independent subtask obtained by partitioning the network function directed acyclic graph of the task to be processed, or a multi-node combined subtask with dependent edges obtained by partitioning the directed acyclic graph, or an independent scheduling granule obtained by partitioning the execution priority level of the task to be processed, or a deployment slice obtained by dynamically dividing the physical distribution of the currently idle hardware resources of the target FPGA; as long as the deployable unit has independent data input receiving capability and processing result output capability, and the execution order between the deployable unit and other deployable units can be determined by the logical dependency relationship inside the task to be processed, the preset execution priority or the external scheduling strategy, this embodiment does not limit the specific form of the deployable unit.

[0034] Step 120: For each deployable unit, obtain the matching page granularity corresponding to the deployable unit based on the resource requirements of the deployable unit.

[0035] Here, page granularity refers to the size of the resource block corresponding to the multi-level partitioning of the dynamically reconfigurable region. The specific page size corresponding to the page granularity can be a preset fixed value or a variable value configured according to application requirements, as long as the page size can serve as the basic unit of measurement for the dynamically reconfigurable region in multi-level partitioning. Those skilled in the art can flexibly configure it according to the actual hardware resources and application scenarios, and this embodiment does not impose any restrictions.

[0036] Step 130: Based on the matching page granularity corresponding to the deployable unit and the resource status information of the FPGA network card, obtain the target reconfigurable page corresponding to the deployable unit.

[0037] The resource status information can be a resource residual map, a resource occupancy bitmap, a resource allocation table, a resource utilization vector, a list of free resource blocks, or a resource usage heatmap; it can even be a combination of the above, such as a comprehensive index structure formed by overlaying the residual map and the occupancy bitmap. In short, as long as the information can directly or indirectly reflect at least one of the current usage, remaining capacity, occupancy distribution, or fragmentation degree of various resources in the FPGA network card, it can be used as the resource status information described in this application. Those skilled in the art can flexibly select or customize the above representation forms according to different actual scheduling strategies, resource management granularity, and hardware monitoring methods; this embodiment does not impose any limitations.

[0038] Step 140: Configure the communication relationship between deployable units to communicate with each other through the common parent reconfigurable page of the target reconfigurable page corresponding to the deployable unit.

[0039] The common upper-level reconfigurable page can be either the common upper-level reconfigurable page corresponding to the target reconfigurable page when it is a lower-level reconfigurable page, or a node reconfigurable page selected from the topology that is at a higher level than all target reconfigurable pages. This node reconfigurable page is configured specifically for data exchange between target reconfigurable pages. Furthermore, the number of common upper-level reconfigurable pages can be one or more; when there are multiple, the common upper-level reconfigurable pages can be redundant with each other or collaboratively share the communication load between target reconfigurable pages. As long as the common upper-level reconfigurable page can connect all target reconfigurable pages and realize the relay exchange of data between target reconfigurable pages, this embodiment does not limit the specific implementation method of the common upper-level reconfigurable page.

[0040] In this embodiment of the invention, by configuring the communication relationship between deployable units to communicate with each other through the common upper-level reconfigurable page of the target reconfigurable page corresponding to the deployable unit, the latency of data transmission between deployable units and the data conflict when multiple deployable units communicate are reduced, the frequency of congestion is reduced, and the data flow efficiency in multi-tenant FPGA network cards is ultimately improved.

[0041] Step 150: Based on the communication relationship between the target reconfigurable page corresponding to all deployable units and all deployable units, process all pending tasks and update resource status information.

[0042] In this embodiment of the invention, updating resource status information reflects the current usage of various resources in the multi-tenant FPGA network card in a timely manner. This is used to match available target reconfigurable pages for each deployable unit in step 130 and to select available resources for scheduling in the subsequent process of scheduling resources for tenants.

[0043] Step 160: During the processing of pending tasks, monitor the resource requirements of all tenants and schedule resources for tenants according to the resource requirements.

[0044] In this embodiment of the invention, during the process of a multi-tenant FPGA network card processing a pending task, the resource requirements of all tenants in the multi-tenant FPGA network card are monitored, and various resources in the FPGA network card are allocated to the tenants according to their resource requirements and idle resources are reclaimed. This ensures that each tenant has sufficient resources to process the pending task and reduces the phenomenon of idle resources in the multi-tenant FPGA network card.

[0045] As can be seen, the embodiments of the present invention reduce resource fragmentation in multi-tenant FPGA network cards by decomposing the tasks to be processed by multiple tenants into deployable units and deploying them to target reconfigurable pages. Furthermore, the communication relationship between target reconfigurable pages is configured to communicate with each other through a common parent reconfigurable page, thereby reducing data transmission latency between deployable units and data conflicts when multiple deployable units communicate, reducing the frequency of congestion, and improving data flow efficiency in multi-tenant FPGA network cards. In addition, during the processing of tasks to be processed, the resource requirements of tenants are monitored and resources are scheduled according to those requirements, allocating the necessary resources to each tenant and reducing resource idleness in multi-tenant FPGA network cards, ultimately improving the overall working efficiency of multi-tenant FPGA network cards.

[0046] However, this invention further reveals that during network interface card (NIC) resource management, tenants may occupy shared resources for extended periods, affecting the task processing of other tenants. Furthermore, allocating resources to tenants based on only a single dimension can easily lead to an overabundance of one type of resource while a shortage of another. All of these problems reduce the overall efficiency of multi-tenant FPGA NICs. Therefore, based on the above findings, this invention makes the following optimizations: In an optional embodiment, such as Figure 2 As shown, step 160 may include: Step 161: During the processing of pending tasks, monitor the resource requirements of all tenants.

[0047] Step 162: Based on the resource requirements and resource status information of all tenants, obtain the dominant resource share corresponding to each tenant.

[0048] The dominant resource share can be the proportion of the tenant's dominant resource demand to the total available resources, the proportion of the total allocated resources, or a relative value normalized based on the resource occupancy rate of other tenants. As long as it can represent the degree of occupancy of the tenant's most scarce resource, this embodiment does not limit it.

[0049] Step 163: Allocate resources to the tenant with the lowest dominant resource share among all tenants.

[0050] Step 164: Reclaim the remaining resources from all tenants whose resource requirements are lower than their allocated share of resources.

[0051] As can be seen, in this optional embodiment, by monitoring the tenant's resource requirements and obtaining the tenant's dominant resource share, and allocating resources to the tenant and reclaiming the remaining resources based on the dominant resource share, it is ensured that each tenant in the multi-tenant FPGA network card can be allocated the required resources and the idle resource phenomenon in the multi-tenant FPGA network card is reduced, thereby further improving the overall working efficiency of the multi-tenant FPGA network card.

[0052] However, embodiments of the present invention further discover that during network interface card (NIC) resource management, a certain tenant may occupy too many resources, thereby causing a shortage of overall resources for other tenants. Therefore, embodiments of the present invention make the following optimizations based on the above problems: In another alternative embodiment, step 163 may include: Step 163a: Obtain the allocation set based on the dominant resource share of all tenants and the upper limit of the resource share corresponding to all tenants.

[0053] The allocation set can be an array of identifiers of tenants whose dominant resource share is lower than the corresponding resource share limit, or it can be a list of structures containing each tenant's identifier and the type and quantity of resources required in the FPGA network card. It can also be a selected status flag represented in bit vector form, with each bit corresponding to a tenant. As long as it can clearly indicate the tenants whose dominant resource share is lower than the corresponding resource share limit and their related information, this embodiment does not impose any limitations.

[0054] Step 163b: Allocate resources to the tenant with the lowest dominant resource share in the allocation set, based on the tenant's resource requirements.

[0055] As can be seen, in this optional embodiment, by filtering out all tenants whose dominant resource share is lower than the corresponding resource share limit, an allocation set is obtained, and resources are allocated only to tenants in the allocation set, thereby limiting a tenant from occupying too many resources and solving the problem of overall resource shortage for other tenants caused by a tenant occupying too many resources.

[0056] In yet another optional embodiment, the resource status information can be a residual resource graph, wherein: The nodes in the residual resource graph are used to represent the available information of the corresponding target reconfigurable page. The attributes of the nodes can include resource balance, page granularity, page level, parent reconfigurable page, and current load. Edges in the residual resource graph represent the connections or available communication paths between corresponding target reconfigurable pages. Edge attributes can include ingress bandwidth, egress bandwidth, queuing status, arbitration status, and path level.

[0057] In yet another alternative embodiment, a tenant’s resource requirements may include the tenant’s page capacity requirements, ingress bandwidth requirements, egress bandwidth requirements, and storage access requirements. A tenant’s initial resource requirements can come from compile-time resource profiles, tenant declarations, historical deployment results, or service-level goals.

[0058] In yet another alternative embodiment, step 140 may include: Step 141: Set up an interconnection node on the parent reconfigurable page of the target reconfigurable page.

[0059] Step 142a: When the target reconfigurable page is located under the same parent reconfigurable page, the deployable units communicate with each other through the interconnection nodes on the same parent reconfigurable page.

[0060] Step 142b: When the target reconfigurable page is located under different parent reconfigurable pages, the deployable units communicate with each other through the interconnection node on the nearest common parent reconfigurable page of the target reconfigurable page.

[0061] In yet another alternative embodiment, step 110 may include: Step 111: Obtain the pending tasks for multiple tenants and the corresponding directed acyclic graph of network functions for the pending tasks.

[0062] Step 112: Based on one or more of the following: the weight of the edge in the directed acyclic graph corresponding to the network function of the task to be processed, the resource usage of the task to be processed, the throughput target of the task to be processed, the cross-page communication cost of the task to be processed, and the service level target of the task to be processed, decompose the task to be processed into deployable units.

[0063] Example 2 Based on the same concept, this embodiment discloses another multi-tenant FPGA network interface card (NIC) resource management method. This method decomposes the tasks to be processed from multiple tenants into deployable units and deploys them to target reconfigurable pages. The communication relationship between these target reconfigurable pages is configured to communicate with each other through a common parent reconfigurable page. This enables the deployment of tasks to be processed and their processing based on the deployment. During the processing of tasks, the resource requirements of tenants are monitored, and resource scheduling is performed based on these requirements. This reduces resource idleness and fragmentation in the multi-tenant FPGA NIC and improves data flow efficiency, ultimately enhancing the overall working efficiency of the multi-tenant FPGA NIC. The method specifically includes: Configure multi-tenant FPGA network interface cards: Multi-tenant FPGA network interface cards (NICs) may include a static shell region, a dynamically reconfigurable region, an interconnect module, and a resource management module, wherein: Static shells can be used to carry network ports, MAC subsystems, host interfaces, storage control interfaces, and runtime control logic; The dynamic reconfigurable area is divided into multiple basic pages. A basic page is the smallest independent reconfigurable resource unit in the dynamic reconfigurable area. Basic pages can be combined with each other to form composite pages, which are used to provide larger logic, on-chip memory, digital signal processing and interface resources. All reconfigurable pages in the dynamic reconfigurable area are composed of basic pages and / or composite pages. Configure the dynamic reconfigurable area as a multi-level structure, including a parent reconfigurable page and multiple child reconfigurable pages nested within the parent reconfigurable page. The parent and child reconfigurable pages are relative concepts of adjacent levels. When a child reconfigurable page is the current level, it can be nested as the parent reconfigurable page of the next level and continue to nest the next level of child reconfigurable pages. Interconnection nodes are configured on reconfigurable pages through interconnection modules. Interconnection nodes may include uplinks, downlinks, peer page streaming switching units, routing tables, output arbitrators, and queue buffers. Interconnection nodes can be used for communication between reconfigurable pages and between reconfigurable pages and network ports. Interconnection modules can also map data packets from different tenants, different destinations, or different service levels to different virtual channels. When a certain output direction is congested, packets from other output directions can still be forwarded through the corresponding virtual channels. The resource management module may include a monitoring module, a decision-making module, and an execution module. During the processing of tasks to be processed, the monitoring module monitors the resource requirements of all tenants and updates the residual resource graph in the current multi-tenant FPGA. The nodes in the residual resource graph are used to represent the availability information of the corresponding target reconfigurable page. The decision-making module is used to formulate a resource allocation plan and hand it over to the execution module for execution.

[0064] Process pending tasks and perform resource scheduling: Obtain the tenant's pending tasks and the corresponding directed acyclic graph of network functions through the static shell region.

[0065] Based on the weights of the edges in the directed acyclic graph corresponding to the network function of the task to be processed, the resource usage of the task to be processed, the throughput target of the task to be processed, the cross-page communication cost of the task to be processed, and the service level target of the task to be processed, the task to be processed is decomposed into one or more deployable units.

[0066] For each deployable unit, the matching page granularity corresponding to each deployable unit is determined based on the resource requirements of each deployable unit.

[0067] Based on the matching page granularity and resource residual graph corresponding to each deployable unit, select the minimum available page granularity that can meet the logic, flip-flops, block memory, digital signal processor and bandwidth requirements of each deployable unit.

[0068] The communication relationship between the target reconfigurable pages corresponding to the deployable units is configured by the interconnect module to allow communication between them through the interconnect nodes on the common parent reconfigurable page of the target reconfigurable pages.

[0069] The task to be processed is carried out through the communication relationship between the target reconfigurable page corresponding to the aforementioned deployable unit and all deployable units.

[0070] During the processing of pending tasks, the monitoring module maintains and updates the residual resource graph. Nodes in the residual resource graph represent the available information of the corresponding target reconfigurable page. The attributes of a node may include resource availability, page granularity, page level, parent reconfigurable page, and current load. Edges in the residual resource graph represent the connection relationship or available communication path between the corresponding target reconfigurable pages. The attributes of an edge may include ingress bandwidth, egress bandwidth, queuing status, arbitration status, and path level.

[0071] During the processing of pending tasks, the monitoring module also monitors the resource requirements of all tenants. These requirements can include page capacity requirements, ingress bandwidth requirements, egress bandwidth requirements, and storage access requirements. The initial resource requirements of a tenant can come from compile-time resource profiles, tenant declarations, historical deployment results, or service-level targets.

[0072] It should be noted that the above-mentioned compile-time resource profiles, tenant declarations, historical deployment results, or service-level targets are merely examples of preferred methods for obtaining resource requirements. Those skilled in the art will understand that tenants' initial resource requirements can also be obtained through other methods, such as dynamic prediction based on machine learning or runtime adaptive parameter adjustment. This invention does not limit the specific source of resource requirements; any technical solution that uses the acquisition of tenant resource requirements and performs resource management accordingly falls within the protection scope of this invention.

[0073] The decision-making module calculates the dominant resource share for each tenant based on the resource requirements and residual resource graph of all tenants.

[0074] The dominant resource share can be the proportion of the tenant's dominant resource demand to the total available resources, the proportion of the total allocated resources, or a relative value normalized based on the resource occupancy rate of other tenants. As long as it can represent the degree of occupancy of the tenant's most scarce resource, this embodiment does not limit it.

[0075] The allocation set is obtained by dividing the dominant resource share among all the resources allocated to each tenant by the upper limit of the resource share corresponding to each tenant.

[0076] The allocation set can be an array of identifiers of tenants whose dominant resource share is lower than the corresponding resource share limit, or it can be a list of structures containing each tenant's identifier and the type and quantity of resources required in the FPGA network card. It can also be a selected status flag represented in bit vector form, with each bit corresponding to a tenant. As long as it can clearly indicate the tenants whose dominant resource share is lower than the corresponding resource share limit and their related information, this embodiment does not impose any limitations.

[0077] The decision-making module formulates a resource allocation plan based on the dominant resource share of all tenants in the allocation set, prioritizing the allocation of resources to the tenant with the lowest dominant resource share in the allocation set and formulating a remaining resource recovery plan for tenants whose resource demand is lower than the share of resources allocated to them. It is evident that through the coordinated efforts of the monitoring, decision-making, and execution modules within the resource management module, tenant resource scheduling in multi-tenant FPGA network cards can be achieved, reducing resource idleness and fragmentation. Furthermore, the interconnect module configures communication between reconfigurable pages to communicate through interconnect nodes on a common upper-level reconfigurable page, and maps data packets from different tenants, destinations, or service levels to different virtual channels. This reduces data transmission latency and congestion during data transmission, improving the data flow efficiency of multi-tenant FPGA network cards and ultimately enhancing their overall operational efficiency.

[0078] Example 3 This invention discloses a multi-tenant FPGA network interface card system, such as... Figure 3 As shown, this multi-tenant FPGA network interface card (NIC) system includes a server host, an FPGA NIC, and a cloud network, wherein: The FPGA network card is installed at the network interface of the server host. The control configuration information on the server host side and the inbound packets on the cloud network side enter the dynamic reconfigurable area through the static shell area of ​​the FPGA network card. The network function instances of different tenants are loaded into the corresponding reconfigurable pages. Interconnection nodes are set on the reconfigurable pages. Different reconfigurable pages communicate with each other, between reconfigurable pages and network ports, and between reconfigurable pages and host ports through the interconnection nodes.

[0079] The FPGA network interface card (NIC) includes a static shell area and a dynamic reconfigurable area. The static shell area includes a host interface, a network interface, and a runtime resource management module. The host interface is used to receive control configurations from the server host side, and the network interface is used to receive inbound packets from the cloud network side. The runtime resource management module includes a monitoring module, a decision-making module, and an execution module. During the processing of the task to be processed, the monitoring module monitors the resource requirements of all tenants and updates the residual resource graph in the current multi-tenant FPGA. The decision-making module formulates a resource allocation plan based on the resource requirements of all tenants and the residual resource graph monitored by the monitoring module, according to the dominant resource share of all tenants, and then submits it to the execution module for execution. The dynamic reconfigurable area has a multi-level structure, including a higher-level reconfigurable page and multiple lower-level reconfigurable pages nested in the higher-level reconfigurable page. The higher-level reconfigurable page and the lower-level reconfigurable page are relative concepts of adjacent levels. When a lower-level reconfigurable page is the current level, it can continue to nest as the higher-level reconfigurable page of the next level. The specific process by which the decision-making module formulates a resource allocation plan is as follows: the decision-making module generates a demand vector di for each tenant based on the resource demands of all tenants monitored by the monitoring module.<diP, diI, diE, diM> Where diP represents page capacity requirement, diI represents ingress bandwidth requirement, diE represents egress bandwidth requirement, and diM represents storage access requirement. The initial requirements come from compile-time resource profiles, tenant declarations, historical deployment results, or service-level targets. The decision module calculates the proportion of each tenant's required resources to all resources of that type in the FPGA network card based on the requirement vector di corresponding to each tenant and the total amount of resources in the FPGA network card corresponding to the residual resource graph. The maximum value of each type of resource proportion for each tenant is taken as the dominant resource share of that tenant. Resources are allocated to the tenant with the lowest dominant resource share among all tenants whose dominant resource share does not exceed the preset upper limit of the tenant's resource share. The remaining resources of all tenants who have not used up the allocated resources are reclaimed to reduce the idle resources in the FPGA network card.

[0080] The specific implementation method of the resource allocation scheme determined by the execution module is as follows: Regarding page capacity, if a tenant is in a computationally constrained state for multiple consecutive periods, the system selects available reconfigurable pages based on the resource residual graph to create new replicas. New replicas are preferentially placed under the same parent reconfigurable page as the existing tenant's deployable unit. For ingress bandwidth, adjust the token generation rate, bucket depth, or burst tolerance parameters of the tenant token bucket; For outbound bandwidth, adjust the weights or scheduling quantum of the weighted fair queue; For storage access, adjust the number of credits, request queue weights, or available channel shares in the storage broker; The module performs scaling down and resource reclamation. When tenant load decreases, the module performs scaling down. For page capacity, redundant copies are revoked and the corresponding pages are released; for ingress and egress bandwidth, the token rate or queue weight is reduced; for storage access, unused credits are reclaimed. Scaling down can be configured with a hysteresis threshold or continuous periodicity to avoid frequent reconstructions due to short-term fluctuations.

[0081] The multi-tenant FPGA network card system in this embodiment of the invention also includes a cloud network. The cloud network serves as the external interaction environment for the FPGA network card and includes switches, external networks, and storage and / or service clusters. The FPGA network card communicates with the external network and storage and / or service clusters through the switches to receive inbound messages from the external network or to send data processed by the FPGA network card to the external network or storage and / or service clusters.

[0082] Example 4 Please see Figure 4 , Figure 4 This is a schematic diagram of another multi-tenant FPGA network interface card system disclosed in an embodiment of the present invention. Figure 4 As shown, the device including memory and processor may include: Memory containing executable program code; A processor coupled to memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in any of the multi-tenant FPGA network card resource management methods in Embodiment 1 of the present invention.

[0083] Example 5 Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a multi-tenant FPGA network card disclosed in an embodiment of the present invention. Figure 5As shown, the FPGA network interface card includes a decomposition module, a matching module, a deployment module, a communication module, a running module, and a scheduling module, wherein: The decomposition module is used to obtain the tasks to be processed from multiple tenants and decompose the tasks to be processed into deployable units; The matching module is used to obtain the matching page granularity corresponding to each deployable unit based on the resource requirements of the deployable unit. The deployment module is used to obtain the target reconfigurable page corresponding to the deployable unit based on the matching page granularity corresponding to the deployable unit and the resource status information of the FPGA network card. The target reconfigurable page is used to carry the deployable unit and execute the computing tasks corresponding to the deployable unit. The communication module is used to configure the communication relationship between deployable units to communicate with each other through the common upper-level reconfigurable page of the target reconfigurable page corresponding to the deployable unit. The common upper-level reconfigurable page is the common upper-level reconfigurable page corresponding to all target reconfigurable pages when they are lower-level reconfigurable pages. The runtime module is used to process all pending tasks and update resource status information based on the communication relationship between the target reconfigurable page corresponding to all deployable units and all deployable units; The scheduling module is used to monitor the resource requirements of all tenants during the processing of pending tasks, and to schedule resources for tenants based on the resource requirements.

[0084] In an optional embodiment, the scheduling module monitors the resource requirements of all tenants during the processing of tasks to be processed, and performs resource scheduling for tenants based on these requirements, including: During the processing of pending tasks, monitor the resource requirements of all tenants; Based on the resource requirements and resource status information of all tenants, the dominant resource share of each tenant is obtained. Allocate resources to the tenant with the lowest dominant resource share among all tenants; Reclaim the remaining resources from all tenants whose resource requirements are lower than their allocated share of resources.

[0085] In another optional embodiment, the scheduling module allocates resources to the tenant with the lowest dominant resource share among all tenants, including the following specific operation: Based on the dominant resource share of all tenants and the corresponding resource share limit of all tenants, an allocation set is obtained. The allocation set includes tenants whose dominant resource share is lower than the corresponding resource share limit. Resources are allocated to the tenant whose resource share is the lowest in the allocation centralization.

[0086] In yet another optional embodiment, the resource status information is a residual resource graph, wherein: The nodes in the residual resource graph are used to represent the available information of the corresponding target reconfigurable page. The attributes of the nodes include one or more of the following: resource balance, page granularity, page level, parent reconfigurable page, and current load. Edges in the residual resource graph represent the connections or available communication paths between corresponding target reconfigurable pages. Edge attributes include one or more of the following: ingress bandwidth, egress bandwidth, queuing status, arbitration status, and path level.

[0087] In yet another alternative embodiment, the tenant’s resource requirements include the tenant’s page capacity requirements, ingress bandwidth requirements, egress bandwidth requirements, and storage access requirements. A tenant’s initial resource requirements come from one or more of the following: compile-time resource profile, tenant declaration, historical deployment results, or service-level objectives.

[0088] In yet another optional embodiment, the specific operation method by which the communication module configures the communication relationship between deployable units to communicate with each other through the common parent reconfigurable page of the target reconfigurable page corresponding to the deployable unit includes: Set up an interconnect node on the parent reconfigurable page of the target reconfigurable page. The interconnect node is used to enable communication between target reconfigurable pages. When the target reconfigurable page is located under the same parent reconfigurable page, the deployable units communicate with each other through interconnect nodes on the same parent reconfigurable page; When the target reconfigurable page is located under different parent reconfigurable pages, the deployable units communicate with each other through interconnect nodes on the nearest common parent reconfigurable page of the target reconfigurable page.

[0089] In another optional embodiment, the decomposition module, upon acquiring the pending tasks from multiple tenants, decomposes the pending tasks into deployable units in the following specific operational manner: Obtain the pending tasks for multiple tenants and the corresponding directed acyclic graph of network functions for the pending tasks; Based on one or more of the following factors: the weight of the edge in the directed acyclic graph corresponding to the network function of the task to be processed, the resource usage of the task to be processed, the throughput target of the task to be processed, the cross-page communication cost of the task to be processed, and the service level target of the task to be processed, the task to be processed is decomposed into deployable units.

[0090] The foregoing has provided a detailed description of a multi-tenant FPGA network interface card resource management method and system disclosed in the embodiments of the present invention. Specific embodiments have been used to illustrate the principles and implementation methods of the present invention. However, the above preferred embodiments are not intended to limit the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be determined by the scope defined in the claims.

Claims

1. A multi-tenant FPGA network interface card (NIC) resource management method, the FPGA NIC comprising at least a dynamic reconfigurable region, the dynamic reconfigurable region having a multi-level structure including a super reconfigurable page and a plurality of sub reconfigurable pages nested in the super reconfigurable page, the method comprising: The method includes: ​ Obtain pending tasks from multiple tenants and decompose the pending tasks into deployable units; For each deployable unit, the matching page granularity corresponding to the deployable unit is obtained according to the resource requirements of the deployable unit; Based on the matching page granularity corresponding to the deployable unit and the resource status information of the FPGA network card, the target reconfigurable page corresponding to the deployable unit is obtained. The target reconfigurable page is used to carry the deployable unit and execute the computing task corresponding to the deployable unit. The communication relationship between the deployable units is configured to communicate with each other through the common parent reconfigurable page of the target reconfigurable page corresponding to the deployable unit, wherein the common parent reconfigurable page is the common parent reconfigurable page corresponding to all the target reconfigurable pages as the lower reconfigurable pages; Based on the communication relationship between the target reconfigurable page corresponding to all the deployable units and all the deployable units, all the tasks to be processed are processed and the resource status information is updated; During the processing of the pending task, the resource requirements of all the tenants are monitored, and resource scheduling is performed on the tenants according to the resource requirements.

2. The multi-tenant FPGA network card resource management method of claim 1, wherein, During the processing of the task to be processed, monitoring the resource requirements of all tenants and scheduling resources for the tenants according to the resource requirements includes: During the processing of the pending task, the resource requirements of all the tenants are monitored; Based on the resource requirements and resource status information of all the tenants, the dominant resource share corresponding to each tenant is obtained; Resources are allocated to the tenant with the lowest dominant resource share among all the tenants. Reclaim the remaining resources from all tenants whose resource requirements are lower than their allocated share of resources.

3. The multi-tenant FPGA network card resource management method of claim 2, wherein, The step of allocating resources to the tenant with the lowest dominant resource share among all tenants includes: An allocation set is obtained based on the dominant resource share of all the resources allocated to all the tenants and the upper limit of the resource share corresponding to all the tenants. The allocation set includes all the tenants whose dominant resource share is lower than the upper limit of the corresponding resource share. Resources are allocated to the tenant with the lowest dominant resource share in the allocation set based on their resource requirements.

4. The multi-tenant FPGA network card resource management method of claim 1, wherein, The resource status information is a residual resource graph, wherein: The nodes in the residual resource graph are used to represent the available information of the corresponding target reconfigurable page. The attributes of the nodes include one or more of the following: resource balance, page granularity, page level, parent reconfigurable page, and current load. The edges in the residual resource graph are used to represent the connection relationship or available communication path between the corresponding target reconfigurable pages, and the attributes of the edges include one or more of the following: ingress bandwidth, egress bandwidth, queuing status, arbitration status, and path level.

5. The multi-tenant FPGA network card resource management method of claim 1, wherein, The tenant's resource requirements include the tenant's page capacity requirements, ingress bandwidth requirements, egress bandwidth requirements, and storage access requirements. The initial resource requirements of a tenant are derived from one or more of the following: compile-time resource profile, tenant declaration, historical deployment results, or service-level objectives.

6. The multi-tenant FPGA network card resource management method of claim 1, wherein, The step of configuring the communication relationship between the deployable units to communicate with each other through the common parent reconfigurable page of the target reconfigurable page corresponding to the deployable unit includes: An interconnection node is set on the parent reconfigurable page of the target reconfigurable page, and the interconnection node is used to enable communication between the target reconfigurable pages; When the target reconfigurable page is located under the same parent reconfigurable page, the deployable units communicate with each other through the interconnection nodes on the same parent reconfigurable page; When the target reconfigurable page is located under different parent reconfigurable pages, the deployable units communicate with each other through the interconnection node on the nearest common parent reconfigurable page of the target reconfigurable page.

7. The multi-tenant FPGA network card resource management method of claim 1, wherein, The step of acquiring pending tasks from multiple tenants and decomposing the pending tasks into deployable units includes: Obtain the pending tasks of multiple tenants and the network function directed acyclic graph corresponding to the pending tasks; Based on one or more of the following: the weight of the edge in the directed acyclic graph corresponding to the network function of the task to be processed, the resource usage of the task to be processed, the throughput target of the task to be processed, the cross-page communication cost of the task to be processed, and the service level target of the task to be processed, the task to be processed is decomposed into deployable units.

8. A multi-tenant FPGA network card, the FPGA network card comprising at least a dynamically reconfigurable region having a multi-level structure including a super-level reconfigurable page and a plurality of sub-level reconfigurable pages nested in the super-level reconfigurable page, characterized in that, The FPGA network interface card also includes: The decomposition module is used to obtain the tasks to be processed from multiple tenants and decompose the tasks to be processed into deployable units. The matching module is used to obtain the matching page granularity corresponding to each deployable unit based on the resource requirements of the deployable unit. The deployment module is used to obtain the target reconfigurable page corresponding to the deployable unit based on the matching page granularity corresponding to the deployable unit and the resource status information of the FPGA network card. The target reconfigurable page is used to carry the deployable unit and execute the computing task corresponding to the deployable unit. A communication module is configured to configure the communication relationship between the deployable units to communicate with each other through the common upper-level reconfigurable page of the target reconfigurable page corresponding to the deployable unit, wherein the common upper-level reconfigurable page is the common upper-level reconfigurable page corresponding to all the target reconfigurable pages as lower-level reconfigurable pages; The running module is used to process all the tasks to be processed and update the resource status information according to the communication relationship between the target reconfigurable page corresponding to all the deployable units and all the deployable units; The scheduling module is used to monitor the resource requirements of all tenants during the processing of the task to be processed, and to schedule resources for the tenants according to the resource requirements.

9. A multi-tenant FPGA network card system, comprising: The system includes: a memory storing executable program code; a processor coupled to the memory; the processor calling the executable program code in the memory to execute the multi-tenant FPGA network interface resource management method as described in any one of claims 1-7.

10. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when executed by a processor, implement the multi-tenant FPGA network interface resource management method as described in any one of claims 1-7.