GPU resource dynamic allocation method and system based on load awareness

Abstract

The invention discloses a GPU resource dynamic allocation method and system based on load awareness, and relates to the technical field of data processing. The method comprises the steps of uploading a multi-thread task on a task management platform, executing first task interpretation and mapping, generating a first task atlas layer, executing multi-dimensional GPU resource state mapping, and determining a second resource atlas layer; migrating the first task atlas layer and the second resource atlas layer to a container arrangement module, migrating a task agent to the container arrangement module, executing a task and resource allocation game, and determining a resource allocation strategy; and task-oriented resource allocation scheduling management is carried out on GPU resources. The technical problems of low resource utilization rate and unreasonable allocation caused by lack of load perception and incapability of dynamically adapting to task change in GPU resource allocation in the prior art are solved, and the technical effects of dynamically and accurately allocating the GPU resources according to the real-time load, improving the resource utilization rate and realizing reasonable and efficient task and resource matching are achieved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of data processing, and particularly relates to a GPU resource dynamic allocation method and system based on load sensing. BACKGROUND

[0002] In current high-performance computing and large-scale data processing scenarios, GPUs have become the core resources for accelerating deep learning training, parallel computing and multi-threaded task processing. However, the existing GPU resource allocation methods generally rely on static presetting or coarse-grained scheduling strategies, lack the ability to perceive real-time task loads, and cannot be flexibly adjusted according to task dynamics and resource states, which easily causes part of the GPU resources to be idle or overloaded, resulting in low overall resource utilization and difficulty in guaranteeing task execution efficiency and stability. Especially in the containerization and multi-agent collaborative computing environment, different tasks have different demands for GPU resources, and traditional methods are difficult to achieve fine matching of tasks and resources. SUMMARY

[0003] The application provides a GPU resource dynamic allocation method and system based on load sensing, which solves the technical problems of low resource utilization and unreasonable allocation caused by the lack of load sensing and the inability to dynamically adapt to task changes in the prior art.

[0004] In a first aspect, the application provides a GPU resource dynamic allocation method based on load sensing, which comprises: A multi-threaded task is uploaded to a task management platform, first task interpretation and mapping are performed, a first task atlas layer is generated, multi-dimensional GPU resource state mapping is performed, and a second resource atlas layer is determined, wherein the first task atlas layer and the second resource atlas layer are dynamically updated; the first task atlas layer and the second resource atlas layer are migrated to a container orchestration module according to a container orchestration port, a task agent is migrated to the container orchestration module by performing an internal game of the agent cluster, and a task and resource allocation game is performed to determine a resource allocation strategy, wherein each agent is identified by a unique code based on a payoff matrix and a strategy space; and the GPU resources are managed by task-oriented resource allocation scheduling according to the resource allocation strategy.

[0005] In a second aspect, the application provides a GPU resource dynamic allocation system based on load sensing, which comprises: The atlas generation component: multi-thread tasks are uploaded on the task management platform, first task interpretation and mapping are performed, a first task atlas layer is generated, multi-dimensional GPU resource state mapping is performed, and a second resource atlas layer is determined, wherein the first task atlas layer and the second resource atlas layer are dynamically updated; the allocation game component: according to a container orchestration port, the first task atlas layer and the second resource atlas layer are migrated to a container orchestration module, task agents are migrated to the container orchestration module through internal competition game of the agent cluster, allocation game of tasks and resources is performed, and a resource allocation strategy is determined, wherein each agent is identified by a unique code based on a payoff matrix and a strategy space; the scheduling management component: according to the resource allocation strategy, task-oriented resource allocation scheduling management of the GPU resources is performed.

[0006] The one or more technical solutions provided in the present application have at least the following technical effects or advantages: First, multi-thread tasks are uploaded on the task management platform, first task interpretation and mapping are performed, a first task atlas layer is generated, multi-dimensional GPU resource state mapping is performed, and a second resource atlas layer is determined, wherein the first task atlas layer and the second resource atlas layer are dynamically updated. Then, according to a container orchestration port, the first task atlas layer and the second resource atlas layer are migrated to a container orchestration module, task agents are migrated to the container orchestration module through internal competition game of the agent cluster, allocation game of tasks and resources is performed, and a resource allocation strategy is determined, wherein each agent is identified by a unique code based on a payoff matrix and a strategy space. Finally, according to the resource allocation strategy, task-oriented resource allocation scheduling management of the GPU resources is performed. The technical problem of low resource utilization and unreasonable allocation caused by lack of load awareness and inability to dynamically adapt to task changes in the prior art is solved, and the technical effects of dynamically and accurately allocating GPU resources according to real-time load, improving resource utilization, and realizing reasonable and efficient task and resource matching are achieved. BRIEF DESCRIPTION OF DRAWINGS

[0007] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0008] Figure 1 A flowchart of a GPU resource dynamic allocation method based on load awareness provided by the embodiments of the present application is shown. Figure 2 A system structure diagram of a GPU resource dynamic allocation system based on load awareness provided by the embodiments of the present application is shown.

[0009] Explanation of reference signs: map generation component 11, allocation game component 12, scheduling management component 13. DETAILED DESCRIPTION

[0010] The present application provides a load-aware based GPU resource dynamic allocation method and system, which solves the technical problems of low resource utilization and unreasonable allocation caused by lack of load awareness and inability to dynamically adapt to task changes in the prior art.

[0011] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0012] It should be noted that the terms "comprise" and "have" are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server comprising a series of steps or units does not have to be limited to those clearly listed steps or units, but can include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices.

[0013] Embodiment one, as shown in the present application provides a load-aware based GPU resource dynamic allocation method, wherein the method comprises: Figure 1 executing first task interpretation and mapping on the multi-threaded task on the task management platform, generating a first task map layer, executing multi-dimensional GPU resource state mapping, and determining a second resource map layer, wherein the first task map layer and the second resource map layer are dynamically updated.

[0014] ​After uploading the multi-threaded task on the task management platform, the multi-threaded task is interpreted one by one, wherein the interpretation process includes semantic perception processing of each thread task to identify the task type, performing differential relaxation operation to determine the task size, and obtaining the required video memory requirement and task priority of the task through parameter extraction, thereby generating a task information unit; on the basis of the parsed task information unit, corresponding task sub-graphs are sequentially constructed, and multiple task sub-graphs are fused to form a task global graph by sub-graph splicing and dependency relationship mapping operation, a first task graph layer is established, and the first task graph layer is one-to-one mapped with the multi-threaded task to ensure that each thread task has a unique structured representation in the graph; at the same time, the GPU resource monitoring module collects real-time data of GPU runtime, including video memory occupation state, computing core utilization rate, multi-instance GPU (MIG) division situation and energy consumption data, to generate a GPU resource element set; based on the resource element set, resource sub-graphs are constructed for the GPU resources, and resource global graphs are formed by sub-graph splicing, state normalization processing and architecture reconstruction operation to establish a second resource graph layer; during the whole task execution process, the system monitors the task progress and the GPU resource state changes in real time, and performs dynamic updating operation on the first task graph layer and the second resource graph layer respectively to ensure consistency and real-time performance of the task graph and the resource graph.

[0015] Further, with the uploading of the multi-threaded task, the task is pre-inspected with malicious feature detection, and if it is determined to be a malicious task, the malicious task thread is isolated and marked.

[0016] When the task management platform receives the multi-threaded task, the system calls the built-in malicious feature detection module to pre-inspect the uploaded task, specifically, the instruction set, call chain, GPU resource request mode and data interaction behavior of each thread task are parsed and feature extracted to form a task feature vector; the task feature vector is compared with the preset malicious feature library for similarity, and a malicious task determination result is generated combined with the threshold determination rule; when the determination result is a malicious task, the system immediately isolates the thread task from the multi-threaded task set to form an independent task isolation area, and configures unique identification information for the malicious task thread for marking to prevent it from continuing to enter the subsequent task interpretation, task mapping and resource scheduling links; at the same time, the system returns the marking result of the malicious task to the task management platform for security audit and update of subsequent defense strategies, thereby ensuring the safety and reliability of the GPU resource dynamic allocation process.

[0017] Further, the first task interpretation and mapping are performed to generate a first task graph layer, including: Identify the first thread task, perform task semantic perception and differential relaxation processing, and determine task information; according to the task information, determine a task graph by interpreting task elements, wherein the task elements at least include task type, task size, memory requirement and priority; traverse the multi-threaded task, and perform task interpretation and mapping in parallel until N task graphs are determined; splice the one task graph to the N task graphs to form the first task graph layer, and establish the mapping between the first task graph layer and the multi-threaded task.

[0018] In the execution of the first task interpretation and mapping, firstly, the uploaded first thread task is identified in the task management platform, the thread task is subjected to semantic perception processing, its task instruction set and context information are analyzed, so as to determine the logical semantics and execution intention of the task; based on the semantic analysis result, differential relaxation processing is performed to extract the key parameters of the thread task under the condition of multi-threaded concurrency, and the task information including task type, task size, memory requirement and task priority is obtained. Based on the task information, the task element interpretation module is called, the dependency relationship and execution constraint relationship between different elements are mapped to a task graph, so as to realize the structured expression of the thread task. With the continuous traversal of the multi-threaded task, the system performs the above semantic perception, differential relaxation, element interpretation and graph generation steps on each thread task to generate N task graphs in parallel; after generating multiple task graphs, the system performs graph splicing and normalized mapping processing, fuses one to N task graphs, forms the first task graph layer representing the overall task topology structure, and establishes a unique mapping relationship for each thread task in the first task graph layer, so as to ensure that the subsequent GPU resource allocation and scheduling can be accurately matched and regulated according to the task characteristics.

[0019] Further, multi-dimensional GPU resource state mapping is performed to determine the second resource graph layer, including: Determine the GPU resource elements, wherein the resource elements at least include memory state, computing core, MIG division and energy consumption data; call the real-time state of the GPU resource, perform resource architecture reconstruction based on the GPU resource elements, determine the second resource graph layer, and establish the mapping between the second resource graph layer and the GPU resource.

[0020] In the execution of the multi-dimensional GPU resource state mapping to determine the second resource atlas layer, first, the target GPU node is analyzed for resource elements, and the resource elements at least include memory state, computing core utilization, multi-instance GPU (MIG) division information, and energy consumption data and other key indicators; after obtaining the resource elements, the GPU running state data is real-time retrieved through the resource monitoring module to form a resource state vector set containing memory occupancy, core parallelism, MIG instance allocation, and power consumption curve; the resource modeling module is called to perform architecture reconstruction processing on the resource state vector set, including normalization, vectorization, and topological operation on resource elements of different dimensions, and the dependency relationship and constraint conditions between elements are established through relationship mapping, so as to generate a resource subgraph that can represent the overall running state of the GPU; on this basis, multiple resource subgraphs are spliced and expanded to form a complete second resource atlas layer, and a one-to-one mapping relationship between the second resource atlas layer and the actual GPU hardware node is established to ensure that the resource atlas can accurately reflect the real-time running condition of the GPU resource.

[0021] According to the container orchestration port, the first task atlas layer and the second resource atlas layer are migrated to the container orchestration module, the agent cluster is subjected to an internal competition game, the task agent is migrated to the container orchestration module, the task and resource allocation game is executed, and the resource allocation strategy is determined, wherein each agent is identified with a unique code based on the payoff matrix and the strategy space.

[0022] According to the container orchestration port, the system migrates the first task atlas layer and the second resource atlas layer to the container orchestration module, and realizes dynamic loading and synchronous updating of atlas data through a container scheduling interface during the migration process, so as to ensure consistency of the task side and the resource side model under the container environment.

[0023] The system calls the constructed agent cluster in the container orchestration module, introduces an internal competition game mechanism for the matching problem of tasks and resources, that is, multiple agents compete for the matching of task graph and resource graph based on their own strategies without relying on external intervention; wherein each agent makes autonomous decisions based on its built-in payoff matrix and strategy space, the payoff matrix is used to quantify the benefits of task completion, resource occupancy and energy consumption, and the strategy space is used to define the allocation strategies that the agent can take in different benefit situations; the system encodes and converts the payoff matrix and strategy space to configure a unique code identifier for each agent to ensure its uniqueness and traceability in the game process. After completing the internal competition game among agents, the system migrates the task agent to the task allocation channel in the container orchestration module to perform the allocation game of tasks and resources, and finally converges to the optimal resource allocation strategy under global constraints, which is issued to the task management platform and GPU resource pool as a scheduling basis to realize task-oriented dynamic allocation and scheduling control of GPU resources.

[0024] Further, before the internal competition game of the agent cluster, the construction of the agent cluster includes: With resource allocation orientation, the allocation logic is mined, wherein the allocation logic and the resource allocation orientation are one-to-one corresponding; according to the allocation logic, the agent cluster is constructed by performing lightweight supervised training, wherein the agent cluster is built-in in the task management platform and performs agent migration and scheduling based on the container scheduling port.

[0025] Specifically, taking resource allocation orientation as the core target, the historical task and GPU resource scheduling data are mined for features, the matching relationship between different task types and resource elements is analyzed, and the quantifiable allocation logic is extracted, wherein the allocation logic and the resource allocation orientation maintain a one-to-one correspondence relationship, which is used to describe the optimal task and resource allocation mode under different load conditions; based on the allocation logic, lightweight supervised training is performed, and during the training process, a training set is constructed using labeled samples, and the strategy selection model of the agent is continuously optimized through iterative learning, thereby obtaining multiple agents with autonomous learning and game ability, and combining them to form an agent cluster. The agent cluster is built-in in the task management platform, encapsulated in a modular way, and migrated and scheduled across modules through the container scheduling port, so that the agent can be flexibly called and deployed in different container orchestration scenarios, thereby supporting the subsequent internal competition game and resource allocation strategy generation.

[0026] Further, the construction of the container orchestration module includes: For the agent cluster, define the payoff matrix and strategy space of the agent, and perform encoding conversion to determine the encoding cluster, wherein each agent corresponds to a unique code; based on the internal competition game of the encoding cluster as the first-order training target, and based on the task resource interaction game allocation of the agent as the second-order training target, the container orchestration module is dynamically trained to convergence.

[0027] Specifically, for the agent cluster, define the payoff matrix and strategy space for each agent, wherein the payoff matrix is used to quantify the agent's performance in multiple dimensions such as task execution efficiency, GPU resource utilization, and energy consumption control, and the strategy space is used to limit the agent's action set in different income situations; after setting the payoff matrix and strategy space, the system encodes and converts the strategy parameters of the agent to generate a set of encoding clusters, and ensures that each agent has a unique corresponding code identifier to ensure that it can be accurately identified and tracked in the subsequent game process. Subsequently, the internal competition game based on the encoding cluster is taken as the first-order training target, and the competitive resource allocation process of multiple agents in the container orchestration module is simulated to optimize the collaborative decision-making ability between agents; further, the task resource interaction game allocation based on the agent is taken as the second-order training target, and the dynamic interaction process of real tasks and resources is simulated to continuously correct the adaptability and stability of the agent strategy. Through the step-by-step optimization of the first-order and second-order training targets, the container orchestration module gradually converges in the dynamic iteration process, and finally forms a stable model with adaptive scheduling capability of the task graph layer and the resource graph layer, providing reliable support for subsequent GPU resource allocation and task scheduling.

[0028] Further, the first task graph layer and the second resource graph layer are migrated to the container orchestration module, and the agent cluster is subjected to internal competition game, including: Through the container orchestration port, the first task graph layer and the second resource graph layer are migrated to the container orchestration module of the task management platform; through the container orchestration port, the first-order game allocation of the agent arrangement is used to determine the task agent code and perform agent migration, wherein the game between agents is used as the arrangement method; the mapping relationship between the task agent code, the first task graph layer and the second task graph layer is established to determine the allocation system.

[0029] Before migrating the first task graph layer and the second resource graph layer to the container orchestration module and performing the internal competition game of the agent cluster, first, through the container orchestration port, the first task graph layer and the second resource graph layer generated in the task management platform are synchronously loaded to the container orchestration module, and the real-time updating and consistency of the graph data are maintained during the migration process. Subsequently, based on the container orchestration port, the agent orchestration process is started, the constructed agent cluster is called, the first-order game allocation is taken as the orchestration target, the competitive matching operation between agents is performed, so as to determine the task agent code corresponding to the task, and the task agent with a unique code is migrated to the container orchestration module. After completing the agent migration, the mapping association between the task agent code, the first task graph layer and the second resource graph layer is further established, the corresponding relationship among the task, the resource and the agent is formed, and a complete allocation system is constructed.

[0030] Further, after determining the allocation system, the allocation game of the task and the resource is performed, including: For the allocation system, the container orchestration module performs the second-order allocation game of the resource and the task under the system layer mapping according to the agent, determines the resource allocation strategy, and performs the task-oriented scheduling and control of the GPU resource through the resource allocation scheduling of the task unit or the GPU node unit.

[0031] Specifically, the container orchestration module takes the allocation system as the constraint condition, calls the agent cluster, and performs the second-order allocation game under the system layer mapping according to the mapping relationship between the task agent code and the first task graph layer and the second resource graph layer. The second-order allocation game gradually converges and generates a globally optimal resource allocation strategy by simulating the dynamic competition of the task for the GPU memory, computing core and energy consumption and the like, and combining the profit matrix and strategy space of each agent. After obtaining the resource allocation strategy, the strategy is further decomposed and allocated: when taking the task unit as the scheduling object, the system performs resource allocation according to the memory demand and execution priority of the single task; when taking the GPU node unit as the scheduling object, the system performs allocation adjustment according to the available resource state and task mapping of the single GPU node. Finally, the system issues the decomposed allocation result to the GPU resource pool, implements task-oriented resource scheduling and control, so as to ensure that the GPU resource can realize efficient, reasonable and dynamic allocation in a multi-task scenario.

[0032] According to the resource allocation strategy, the GPU resource is managed in a task-oriented resource allocation scheduling.

[0033] Specifically, the resource allocation strategy output by the container orchestration module is issued to the scheduling unit of the task management platform, the task allocation information in the resource allocation strategy and the GPU node resource state are analyzed, and corresponding scheduling instructions are generated; then, according to the scheduling instructions, the target task is allocated to the matched GPU node, and according to the memory requirement of the task, the core occupancy rate, the MIG division mode and the energy consumption constraint condition, the allocation proportion and the use priority of the GPU resource are dynamically adjusted; during the task running process, the system continuously monitors the state change of the GPU node, and according to the real-time load condition, the resource allocation strategy is secondarily corrected and scheduling optimized, so as to ensure the continuity and rationality of resource utilization in the task execution process; finally, the task demand-oriented dynamic allocation and unified control of GPU resources are realized, so as to ensure the task completion efficiency while improving the overall GPU resource utilization and the stability of the computing environment.

[0034] Further, the real-time state of the GPU resource is used to determine the GPU node failure; if the first GPU node is determined as a failure node, it is marked as a checkpoint and the internal task migration of the first GPU node is performed.

[0035] During the task execution process, the real-time state of the GPU resource is continuously monitored and determined to realize the GPU node failure detection. Specifically, the resource monitoring module collects the memory occupancy rate, the computing core running frequency, the energy consumption data and the response time delay and other running indexes of each GPU node, and compares them with the preset health threshold range, when it is detected that the running index of a certain GPU node continuously exceeds the threshold or appears an unrecoverable abnormal state, the GPU node is determined as a failure node. If the first GPU node is determined as a failure node, the system immediately marks the node as a checkpoint, and triggers the internal task migration mechanism in the task management platform to migrate the task running on the GPU node to other available nodes in the same GPU resource pool, and the mapping update of the task graph layer and the resource graph layer is maintained during the migration process, so as to ensure the continuity of the task execution and the integrity of the data processing.

[0036] In summary, the embodiments of the present application have at least the following technical effects: Firstly, a multi-thread task is uploaded to a task management platform, first task interpretation and mapping are performed, a first task graph layer is generated, multi-dimensional GPU resource state mapping is performed, and a second resource graph layer is determined, wherein the first task graph layer and the second resource graph layer are dynamically updated. Then, according to a container orchestration port, the first task graph layer and the second resource graph layer are migrated to a container orchestration module, through internal competition game of an agent cluster, a task agent is migrated to the container orchestration module, a task and resource allocation game is performed, and a resource allocation strategy is determined, wherein each agent has a unique code based on a payoff matrix and a strategy space. Finally, according to the resource allocation strategy, task-oriented resource allocation scheduling management is performed on the GPU resources. The technical problems of low resource utilization and unreasonable allocation caused by lack of load awareness and inability to dynamically adapt to task changes in the prior art are solved, and the technical effects of dynamically and accurately allocating GPU resources according to real-time load, improving resource utilization, and realizing reasonable and efficient task and resource matching are achieved.

[0037] In the embodiment two, based on the same inventive concept as the GPU resource dynamic allocation method based on load awareness in the foregoing embodiment one, as shown in the embodiment two, Figure 2 As shown in the embodiment two, the application provides a GPU resource dynamic allocation system based on load awareness, wherein the system comprises: The graph generation component 11: uploading a multi-thread task to a task management platform, performing first task interpretation and mapping, generating a first task graph layer, performing multi-dimensional GPU resource state mapping, and determining a second resource graph layer, wherein the first task graph layer and the second resource graph layer are dynamically updated; the allocation game component 12: according to a container orchestration port, migrating the first task graph layer and the second resource graph layer to a container orchestration module, through internal competition game of an agent cluster, migrating a task agent to the container orchestration module, performing a task and resource allocation game, and determining a resource allocation strategy, wherein each agent has a unique code based on a payoff matrix and a strategy space; the scheduling management component 13: according to the resource allocation strategy, performing task-oriented resource allocation scheduling management on the GPU resources.

[0038] Further, the graph generation component 11 is used to perform the following method: A first thread task is identified, task semantic awareness and differential relaxation processing are performed, and task information is determined; according to the task information, a task graph is determined by interpreting task elements, wherein the task elements at least include task type, task size, memory demand, and priority; the multi-thread task is traversed, and task interpretation and mapping are performed in parallel until N task graphs are determined; the one task graph until the N task graphs are spliced to constitute the first task graph layer, and the mapping of the first task graph layer and the multi-thread task is established.

[0039] Further, the graph generation component 11 is configured to perform the following method: determining GPU resource elements, wherein the resource elements at least include video memory status, computing core, MIG division, and energy consumption data; calling real-time status of GPU resources, performing resource architecture reconstruction based on GPU resource elements, determining a second resource graph layer, and establishing mapping between the second resource graph layer and GPU resources.

[0040] Further, the allocation game component 12 is configured to perform the following method: mining allocation logic based on resource allocation guidance, wherein the allocation logic corresponds to the resource allocation guidance one by one; constructing an agent cluster by performing lightweight supervision training according to the allocation logic, wherein the agent cluster is built-in in a task management platform and performs agent migration scheduling based on a container scheduling port.

[0041] Further, the allocation game component 12 is configured to perform the following method: defining a payoff matrix and a strategy space of the agent for the agent cluster, performing encoding conversion to determine an encoding cluster, wherein each agent corresponds to a unique code; taking an internal competition game based on the encoding cluster as a first-order training target and taking a task resource interaction game allocation based on the agent as a second-order training target, and dynamically training the container orchestration module until convergence.

[0042] Further, the allocation game component 12 is configured to perform the following method: migrating the first task graph layer and the second resource graph layer to the container orchestration module of the task management platform through the container orchestration port; determining task agent codes and performing agent migration through the first-order game allocation of agent orchestration through the container orchestration port, wherein the game between agents is used as the orchestration method; establishing mapping association between the task agent codes, the first task graph layer, and the second task graph layer to determine the allocation system.

[0043] Further, the allocation game component 12 is configured to perform the following method: For the allocation system, the container orchestration module performs a second-order allocation game of resources and tasks under the system layer mapping according to the agent to determine a resource allocation strategy; the resource allocation strategy is decomposed and planned down by resource allocation scheduling of a task unit or a GPU node unit to perform task-oriented scheduling and control of GPU resources.

[0044] Further, the graph generation component 11 is configured to perform the following method: With uploading of the multi-threaded task, task pre-inspection is performed by malicious feature detection, and if the task is determined to be malicious, the malicious task thread is isolated and marked.

[0045] Further, the scheduling management component 13 is configured to perform the following method: The real-time state of the GPU resource is used to determine the failure of the GPU node. If the first GPU node is determined as a failure node, the first GPU node is marked as a checkpoint and the internal task migration of the first GPU node is performed.

[0046] It should be noted that the above-mentioned sequence of the embodiments of the present application is only for description, and does not represent the advantages and disadvantages of the embodiments. The above describes a specific embodiment of the present application. The processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are also possible or may be advantageous.

[0047] The above only describes the preferred embodiments of the present application, and does not limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

[0048] The present application is only an exemplary description of the present application, and should be considered as covering any and all modifications, changes, combinations or equivalents within the scope of the present application. Obviously, those skilled in the art can make various modifications and changes to the present application without departing from the scope of the present application. Thus, if these modifications and changes of the present application belong to the scope of the present application and its equivalent technology, the present application intends to include these modifications and changes.

Claims

1. A load-aware based GPU resource dynamic allocation method, characterized in that, The method comprises: On the task management platform, multi-threaded tasks are executed, first task interpretation and mapping are performed, a first task atlas layer is generated, multi-dimensional GPU resource state mapping is performed, and a second resource atlas layer is determined, wherein the first task atlas layer and the second resource atlas layer are dynamically updated; According to the container orchestration port, the first task atlas layer and the second resource atlas layer are migrated to the container orchestration module, the task agent is migrated to the container orchestration module through the internal competition game of the agent cluster, the task and resource allocation game is executed, and the resource allocation strategy is determined, wherein each agent is identified by a unique code based on the payoff matrix and the strategy space; According to the resource allocation strategy, the GPU resources are managed by task-oriented resource allocation scheduling.

2. The method of claim 1, wherein, Performing first task interpretation and mapping to generate a first task atlas layer comprises: Identifying first thread tasks, performing task semantic perception and differential relaxation processing, and determining task information; According to the task information, a task atlas is determined by interpreting task elements, wherein the task elements at least include task type, task size, video memory requirement and priority; Iterate through the multi-threaded tasks and perform task interpretation and mapping in parallel until N task atlases are determined; Splice the one task atlas to the N task atlases to form the first task atlas layer, and establish a mapping between the first task atlas layer and the multi-threaded tasks.

3. The method of claim 1, wherein, Performing multi-dimensional GPU resource state mapping to determine a second resource atlas layer comprises: Determining GPU resource elements, wherein the resource elements at least include video memory state, computing core, MIG division and energy consumption data; Accessing the real-time state of the GPU resources, performing resource architecture reconstruction based on the GPU resource elements, determining the second resource atlas layer, and establishing a mapping between the second resource atlas layer and the GPU resources.

4. The method of claim 1, wherein, Before the internal competition game of the agent cluster, the construction of the agent cluster comprises: Mining allocation logic in a resource allocation oriented manner, wherein the allocation logic corresponds to the resource allocation orientation one by one; According to the allocation logic, the agent cluster is constructed by performing lightweight supervised training, wherein the agent cluster is built-in in the task management platform and the agent migration scheduling is performed based on the container scheduling port.

5. The method of claim 4, wherein, The construction of the container orchestration module comprises: For the agent cluster, define the payoff matrix and strategy space of the agent, perform encoding conversion to determine the encoding cluster, wherein each agent corresponds to a unique code; Take the internal competition game based on the encoding cluster as the first-order training target, and take the task resource interaction game allocation based on the agent as the second-order training target, and dynamically train the container orchestration module to convergence.

6. The method of claim 5, wherein, Migrating the first task atlas layer and the second resource atlas layer to the container orchestration module, migrating the task agent to the container orchestration module through the internal competition game of the agent cluster comprises: Through the container orchestration port, the first task atlas layer and the second resource atlas layer are migrated to the container orchestration module of the task management platform; Determine the task agent code and perform agent migration through the container orchestration port, and the first-order game distribution of the agent orchestration, wherein the inter-agent game is used as the orchestration method; Establish the mapping relationship between the task agent code, the first task graph layer, and the second task graph layer to determine the distribution system.

7. The method of claim 6, wherein, After determining the distribution system, perform the allocation game of tasks and resources, including: For the distribution system, the container orchestration module performs the second-order allocation game of resources and tasks under the system layer mapping according to the agent to determine the resource allocation strategy; Decompose and plan the resource allocation strategy through the resource allocation and scheduling of the task unit or the GPU node unit to perform the task-oriented scheduling and control of the GPU resources.

8. The method of claim 1, wherein, With the uploading of the multi-threaded task, perform task pre-inspection through malicious feature inspection, and if the task is determined to be malicious, perform task cluster isolation and marking on the malicious task thread.

9. The method of claim 1, wherein, Determine the GPU node failure through the real-time state of the GPU resources; If the first GPU node is determined to be a failure node, mark it as a checkpoint and perform internal task migration on the first GPU node. 10.A load-aware GPU resource dynamic allocation system, characterized in that, A GPU resource dynamic allocation method based on load sensing, for implementing any one of claims 1-9, the system comprises: A graph generation component: upload multi-threaded tasks to the task management platform, perform first task interpretation and mapping, generate a first task graph layer, perform multi-dimensional GPU resource state mapping, and determine a second resource graph layer, wherein the first task graph layer and the second resource graph layer are dynamically updated; An allocation game component: according to the container orchestration port, migrate the first task graph layer and the second resource graph layer to the container orchestration module, migrate the task agent to the container orchestration module through the internal competition game of the agent cluster, perform the allocation game of tasks and resources, and determine the resource allocation strategy, wherein each agent is identified by a unique code based on the revenue matrix and the strategy space; A scheduling management component: according to the resource allocation strategy, perform task-oriented resource allocation and scheduling management of the GPU resources.

Images