Cross-heterogeneous computing power scheduling method based on AI intelligent agent

Abstract

The invention provides a cross-heterogeneous computing power scheduling method based on an AI (artificial intelligence) agent, which utilizes the advantage that the AI agent can dynamically generate or adjust task nodes and dependency relationships thereof, adopts a dynamic recursion splitting method to split tasks, and performs task computing property evaluation on subtasks by corresponding agent members to form a computing power demand scheme. And the computing power scheduling agent transmits the computing power demand to the resource adaptation module, forms a computing power available scheme, performs matching adjustment with a computing power demand scheme, and completes task processing. According to the method, the most suitable computing power type is dynamically judged and selected through the intelligent agent, the task processing cost can be controlled, the cross-node communication delay can be reduced, for the tasks in sequence, the continuity between the tasks is ensured, the waiting time is shortened, and higher-quality and more efficient computing power distribution services are provided for users.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a cross-heterogeneous computing power scheduling method based on an AI agent. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, the size and complexity of AI models continue to increase, and various intelligent applications are widely deployed in image processing, natural language understanding, scientific computing, etc., leading to an explosive growth in demand for high-performance computing resources. Under this background, computing power scheduling, as the core link connecting computing resources and AI tasks, its efficiency and flexibility are directly related to the performance and resource utilization of the overall system.

[0003] Currently, the industry's research and practice in computing power scheduling mainly focuses on how to effectively allocate heterogeneous computing resources (such as GPU, CPU, etc.) to meet the differentiated needs of computing power for different AI application scenarios, such as graphics rendering, deep learning training, industrial simulation, etc. The existing scheduling schemes are mostly based on static task characteristics and resource models, and achieve preliminary matching of resources through task queue management, resource reservation or load balancing mechanisms.

[0004] However, the existing technology still has many limitations:

[0005] 1. Traditional static scheduling strategies are designed for specific application scenarios, lack of universality and flexibility, and are difficult to implement dynamic scheduling strategy adjustment, unable to adapt to the dynamics and uncertainty of AI computing tasks (including the randomness of task arrival time, the unpredictability of execution duration and the diversity of resource requirements) ;

[0006] 2. Modern data centers are usually equipped with multiple heterogeneous computing architectures (such as CPU, GPU, FPGA, etc.), and various resources have significant differences in parallelism, memory bandwidth, energy efficiency, etc. The cross-architecture collaborative scheduling faces the problems of resource compatibility difficulty and insufficient scheduling strategy optimization. SUMMARY

[0007] The purpose of the present application is to overcome the defects of the prior art and provide a cross-heterogeneous computing power scheduling method based on an AI agent, which can solve the problems of lack of universality and flexibility in the prior art, poor compatibility of heterogeneous computing power, and insufficient scheduling, etc.

[0008] In order to achieve the above and other purposes, the present application is realized by including the following technical solutions: as a first aspect, the present application proposes a cross-heterogeneous computing power scheduling method based on an AI agent, including the steps of:

[0009] S1, a cross-heterogeneous computing power scheduling platform is constructed, the cross-heterogeneous computing power scheduling platform comprises an application layer, an agent scheduling layer and a computing power resource layer;

[0010] S2, a user inputs a task demand in the application layer;

[0011] S3, a first agent in the agent scheduling layer performs task splitting, and performs task computing property evaluation on a subtask;

[0012] S4, a second agent in the agent scheduling layer forms a computing power demand scheme according to a task computing property evaluation result;

[0013] S5, a third agent in the agent scheduling layer conveys a computing power demand to a resource adaptation module, the resource adaptation module updates a computing power available scheme in real time according to a resource use condition and in combination with a resource computing property evaluation result;

[0014] S6, the third agent matches the computing power demand scheme with the computing power available scheme, if the computing power available scheme covers the computing power demand scheme, the third agent directly calls a corresponding computing power resource; if it cannot be covered, the third agent calls a substitute resource to process a task according to an association relationship between performance indexes of resources in available resources.

[0015] In an embodiment, the application layer comprises a software resource, a management tool, a monitoring tool and a data analysis application.

[0016] In an embodiment, the agent scheduling layer comprises a plurality of agent teams, the agent teams comprising a master agent, the first agent, the second agent and the third agent.

[0017] In an embodiment, a dynamic DAG dependency graph is built in each agent, a task dependency relationship is stored through a distributed graph database, the master agent maintains a global dependency graph, and a subgraph change is synchronized in real time through an incremental graph updating algorithm.

[0018] In an embodiment, the computing power resource layer comprises a computing power resource pool and a resource adaptation module, the resource adaptation module is used for calling performance index data of a computing power resource in the computing power resource pool and converting the performance index data into a standard interface of the cross-heterogeneous computing power scheduling platform according to performance index requirements of a task on the computing power resource, and assisting the third agent in performing computing power resource allocation and scheduling.

[0019] In an embodiment, in the step S3, the first agent adopts a dynamic recursive splitting method to perform task splitting.

[0020] In an embodiment, in the step S4, the second intelligent agent forms a computing power demand scheme by comprehensively considering the performance, cost and dynamic factors.

[0021] The one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0022] The AI intelligent agent can dynamically generate or adjust the task nodes and their dependency relationships, which can adapt to dynamic task load, support multi-type heterogeneous computing power resource compatibility, have high flexibility and wide applicability, not only can distribute tasks to different types of computing power, improve task processing efficiency and accuracy, but also can control task processing cost, reduce cross-node communication delay, guarantee the continuity between tasks for tasks with sequence, shorten waiting time, and provide more high-quality and efficient computing power distribution services for users. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of a cross-heterogeneous computing power scheduling method based on an AI intelligent agent is shown. DETAILED DESCRIPTION

[0024] The embodiments of the present application provide a cross-heterogeneous computing power scheduling method based on an AI intelligent agent, which can solve the problems of lack of universality and flexibility of the scheduling method, poor compatibility of heterogeneous computing power, and insufficient scheduling in the prior art.

[0025] Hereinafter, the technical solutions in the embodiments of the present application will be described clearly and completely with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the example embodiments described herein. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application. In addition, it should be noted that, for convenience of description, only the parts related to the present application are shown in the drawings, not all.

[0026] The technical scheme provided by the application has the following general idea: a cross-heterogeneous computing power scheduling method based on an AI agent, which utilizes the advantages of the AI agent in dynamically generating or adjusting task nodes and their dependency relationships, adopts a dynamic recursive splitting method for task splitting, and performs task computing property evaluation on subtasks by corresponding agent members to form a computing power demand scheme, and the computing power scheduling agent conveys the computing power demand to a resource adaptation module and forms a computing power available scheme, which is matched and adjusted with the computing power demand scheme and completes task processing. The application dynamically judges and selects the most suitable computing power type through the agent, can control the task processing cost, reduce the cross-node communication delay, ensures the continuity between tasks for tasks with a sequence, shortens the waiting time, and provides more high-quality and efficient computing power distribution services for users.

[0027] After introducing the basic principle of the application, various non-limiting embodiments of the application will be specifically introduced below in combination with the accompanying drawings.

[0028] As shown in Figure 1 , the application provides a cross-heterogeneous computing power scheduling method based on an AI agent, which specifically includes the following steps:

[0029] S1, a cross-heterogeneous computing power scheduling platform is constructed, and the cross-heterogeneous computing power scheduling platform includes an application layer, an agent scheduling layer, and a computing power resource layer;

[0030] The application layer includes rich application tools and services, such as office software, design software, rendering and production software, management tools, office collaboration tools, application layer protocol control tools, security protection tools, and data analysis applications.

[0031] The agent scheduling layer includes a plurality of agent teams, and the agent team includes a master agent, a first agent (task disassembly and analysis agent), a second agent (computing power demand planning agent), and a third agent (computing power scheduling agent). The agent team can be flexibly formed according to the task demand (such as task category, work stage, etc.) input by the user in the application layer, and the role-based agent team collaboration mechanism is integrated into the architecture to realize computing task analysis and evaluation, computing power allocation and scheduling, etc. In addition, each agent has a dynamic DAG dependency graph built-in, which stores the task dependency relationship through a distributed graph database, the master agent maintains the global dependency graph, and the incremental graph update algorithm is used to synchronize the subgraph changes in real time. Compared with the AI large model for computing power scheduling, dynamic resource matching and globally optimal scheduling decision can be realized.

[0032] The computing power resource layer includes a computing power resource pool and a resource adaptation module. The computing power resource pool includes various computing power resources such as GPUs, CPUs, Nvidia, and Ascend. The resource adaptation module is used to call the performance index data of the computing power resources in the computing power resource pool according to the performance index requirements of the tasks on the computing power resources, and convert the performance index data into a standard interface of a cross-heterogeneous computing power scheduling platform, to assist the third intelligent agent in allocating and scheduling the computing power resources.

[0033] Since the advantages and characteristics of different types of computing power resources are different, and the computing power performance provided by different suppliers is also different, a resource adaptation module needs to be added to identify the performance index data of various computing power resources, to solve the adaptation problem between different computing power resources. Specifically, a corresponding resource adaptation module can be added for each type of computing power resource. Each resource adaptation module converts the performance index (including hardware performance index such as the number of computing resources, the number of network resources, the number of storage resources, etc.; service demand index such as input data size, computing resource demand, output data size, computing energy efficiency and cost; resource state index such as real-time resource state, resource identification and classification, etc.) of different types of computing power resources into a standard interface that can be understood by the scheduling platform. For different types of tasks, the first intelligent agent judges the key performance index of the task after the task is input, and uses these key performance indexes as the computing power scheduling judgment standard. For example, AI model training and inference tasks have requirements for the computing power resource performance indexes such as computing accuracy, memory bandwidth, video memory capacity, and delay; real-time rendering tasks have requirements for the computing power resource performance indexes such as FLOPS (floating point operation per second) and video memory bandwidth; video editing tasks have requirements for transmission, network, and storage, etc. Therefore, the resource adaptation module standardizes the performance index data of the computing power resources, and assists the third intelligent agent in allocating and scheduling the computing power resources, which can solve the compatibility problem. At the same time, the performance index data of different computing power resources called by the resource adaptation module can also provide a basis for the computing power scheduling alternative scheme, so that when the preferred computing power scheduling scheme is in short supply of computing power resources, different types and amounts of computing power resources with the same effect can be provided to ensure that the alternative scheme can achieve the same task processing effect.

[0034] S2, a user inputs a task demand in an application layer;

[0035] S3, the first intelligent agent in the intelligent agent scheduling layer adopts a dynamic recursive splitting method to split the task, and performs task computing property evaluation on the subtasks;

[0036] The task computing property evaluation refers to: based on the accumulated resource usage data of historical task processing, reusable evaluation indexes are formed to assist the second intelligent agent to quickly judge the resource performance. Specifically, the task analysis intelligent agent performs task analysis, determines the required software and hardware resources for task processing, learns the historical computing power resource usage of similar tasks, judges which key indicators of computing power resources the task requires, the type and amount of computing power used by historical tasks, and the effect of the computing power scheduling scheme adopted (transmission time, processing time, computing performance, result satisfaction, etc.), to provide a basis for the second intelligent agent to form a computing power demand scheme.

[0037] S4, the second intelligent agent forms a computing power demand scheme according to the task computing property evaluation result, comprehensively considering the performance, cost and dynamic factors;

[0038] The performance dimension: taking the delivery period (processing time + transmission time + waiting time) as the core, measures the efficiency of resources completing tasks. The cost dimension: compares the prices of different resources, and judges the cost rationality combined with the performance difference (such as selecting low price when the performance is similar, and selecting low price when the performance is slightly lower but the price difference is more than 30%). Dynamic factor: considering the tidal characteristics of computing power resources (load fluctuation over time), avoiding high-load resources during peak hours, and optimizing the timeliness of computing power scheduling.

[0039] S5, the third intelligent agent transmits the computing power demand to the resource adaptation module, and the resource adaptation module updates the computing power available scheme in real time according to the resource usage and combined with the resource computing property evaluation result;

[0040] Specifically, through regular systematic evaluation and iterative optimization of resource performance, the actual performance of resources (such as processing time, transmission time, waiting time, etc., which together constitute the delivery period), price and dynamic characteristics (such as tidal computing power) are dynamically monitored in real time, the actual processing capacity, efficiency and other characteristic indexes of resources are evaluated, and the basis for subsequent computing power resource selection and alternative scheme is provided, and finally the rational utilization and maximum cost performance of resources are realized.

[0041] The resource computing property evaluation refers to: taking the theoretical indexes of resources at the time of factory shipment as the initial benchmark for performance evaluation; and actually measuring the resource performance in the actual use process, and verifying the landing situation of the theoretical indexes according to the measured data (such as processing speed, stability, etc.).

[0042] S6, the third intelligent agent matches the computing power demand scheme with the computing power available scheme, if the computing power available scheme covers the computing power demand scheme, the corresponding computing power resource is directly called by the third intelligent agent; if it cannot be covered, the alternative resources are selected according to the correlation between the performance indexes of the resources in the available resources for task processing.

[0043] For example, the processing time of resource A is 3s, and the processing time of resource B is 6s. When resource A is short and resource B is available, two resource Bs are used to replace one resource A for processing to achieve the same effect.

[0044] In summary, the AI agent of the present application can dynamically generate or adjust the task nodes and their dependency relationships. The method of dynamic recursive splitting is used for task splitting. The corresponding agent members evaluate the task computing properties of the subtasks and form the computing power demand scheme. The computing power scheduling agent conveys the computing power demand to the resource adaptation module and forms the computing power available scheme. The computing power available scheme is matched and adjusted with the computing power demand scheme and the task processing is completed. Not only can the tasks be distributed to different types of computing power, improving the task processing efficiency and accuracy, but also the most suitable computing power type can be selected by the agent to control the task processing cost, reduce the cross-node communication delay, ensure the continuity between tasks for tasks with a sequence, shorten the waiting time, and provide more high-quality and efficient computing power distribution services for users.

[0045] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the same technology of the present application, the present application also intends to include these modifications and variations.

Claims

1. A cross-heterogeneous computing power scheduling method based on AI intelligent agents, characterized in that, Including the following steps: S1. Construct a cross-heterogeneous computing power scheduling platform, which includes an application layer, an agent scheduling layer, and a computing power resource layer. S2. The user inputs task requirements at the application layer; S3. The first agent in the agent scheduling layer performs task splitting and evaluates the task computation properties of the sub-tasks. S4. The second agent in the agent scheduling layer forms a computing power requirement scheme based on the evaluation results of the task computing nature. S5. In the intelligent agent scheduling layer, the third intelligent agent conveys the computing power requirement to the resource adaptation module. The resource adaptation module updates the available computing power scheme in real time based on the resource usage and the evaluation results of the resource computing properties. S6. The third intelligent agent matches the computing power demand scheme with the computing power available scheme. If the computing power available scheme covers the computing power demand scheme, the third intelligent agent directly calls the corresponding computing power resources. If they cannot cover each other, the third intelligent agent calls alternative resources from the available resources according to the correlation between the performance indicators of the resources to process the task.

2. The cross-heterogeneous computing power scheduling method based on AI agents according to claim 1, characterized in that, The application layer includes software resources, management tools, monitoring tools, and data analysis applications.

3. The cross-heterogeneous computing power scheduling method based on AI agents according to claim 1, characterized in that, The agent scheduling layer includes multiple agent teams, which include a master agent, the first agent, the second agent, and the third agent.

4. The cross-heterogeneous computing power scheduling method based on AI agents according to claim 3, characterized in that, Each agent has a built-in dynamic DAG dependency graph, and the task dependencies are stored in a distributed graph database. The master agent maintains the global dependency graph and synchronizes subgraph changes in real time through an incremental graph update algorithm.

5. The cross-heterogeneous computing power scheduling method based on AI agents according to claim 1, characterized in that, The computing power resource layer includes a computing power resource pool and a resource adaptation module. The resource adaptation module is used to retrieve the performance index data of the computing power resources in the computing power resource pool according to the performance index requirements of the task for computing power resources and convert it into the standard interface of the cross-heterogeneous computing power scheduling platform to assist the third intelligent agent in the allocation and scheduling of computing power resources.

6. The cross-heterogeneous computing power scheduling method based on AI agents according to claim 1, characterized in that, In step S3, the first agent uses a dynamic recursive splitting method to split the task.

7. The cross-heterogeneous computing power scheduling method based on AI agents according to claim 1, characterized in that, In step S4, the second intelligent agent formulates a computing power requirement scheme by comprehensively considering three dimensions: performance, cost, and dynamic factors.

Images