AI task-based detachable electricity calculation cooperative scheduling method and device, and medium

Abstract

The invention provides an AI task-based detachable electricity calculation cooperative scheduling method and device and a medium, and the method comprises the steps: dividing an arrived AI task based on a preset division target, and calculating the resource demand of each sub-task after division, the resource demand comprising a memory demand and a calculation capability utilization rate; based on the current system state and the task queue, determining a splitting strategy and a distribution strategy for the AI task by executing a Markov decision process; and scheduling each sub-task to a corresponding computing server for execution according to the splitting strategy and the distribution strategy. According to the method, through task internal structure disassembly and dynamic matching of green power supply, the carbon emission of the data center is remarkably reduced while the service quality is ensured, the space-time cooperation efficiency of the computing power resource and the power resource is improved, and a feasible technical path is provided for realizing low-carbon and intelligent scheduling of the data center.

Description

AI-based task-splitable computational collaborative scheduling method, equipment, and media Technical Field

[0001] This document relates to the field of computing-powered collaborative technology, and in particular to a computing-powered collaborative scheduling method, device and medium based on AI task decomposability. Background Technology

[0002] With the widespread application of artificial intelligence (AI) technology, large-scale AI computing tasks have become a significant source of energy consumption and carbon emissions for data centers. Traditional scheduling methods face significant limitations in dealing with such loads: optimization methods based on pre-defined models struggle to adapt to the dynamic fluctuations of tasks and the energy market; while existing online learning methods can achieve preliminary carbon awareness through delayed execution, they cannot change the inherent resource requirements and execution duration of tasks, often leading to service quality degradation and queue congestion when facing long-duration AI tasks. The fundamental problem lies in the fact that existing methods merely treat the computing load as a rigid whole on the time axis for translation, failing to address the internal structure of the task to achieve flexible spatiotemporal coordination between computing power demands and green electricity supply.

[0003] In fact, most AI tasks are decomposable at the computational graph level, which provides new possibilities for simultaneously adjusting task execution time and resource consumption, representing a key technological path to overcome the rigidity of current scheduling. However, existing research focuses primarily on performance optimization in high-performance computing, rarely applying its systems to computation-power-coordinated scheduling aimed at carbon emission reduction. Building such models faces three challenges: establishing a formal model that accurately characterizes task decomposability and resource elasticity; designing a collaborative framework that can simultaneously optimize task decomposition granularity, resource allocation, and execution timing; and ensuring that this framework has online learning capabilities to respond to dynamic carbon signals and load changes.

[0004] Therefore, there is an urgent need for an innovative method that can build a new computing-power collaborative scheduling model based on the decomposability of AI tasks. This model can transform macro-level carbon reduction targets into micro-level scheduling instructions through task decomposition and resource reallocation, thereby achieving high-precision matching of computing load and low-carbon periods and fully tapping the carbon reduction potential of data centers in participating in the regulation of new power systems. Summary of the Invention

[0005] This invention provides a computational-electrical collaborative scheduling method, device, and medium based on AI task decomposability, aiming to solve the above-mentioned problems.

[0006] According to an embodiment of the present invention, a computing-powered collaborative scheduling method based on the decomposability of AI tasks is provided, comprising: S1, dividing arriving AI tasks based on a preset partitioning target, and calculating the resource requirements of each sub-task after partitioning, wherein the resource requirements include memory requirements and computing power utilization; S2, determining a partitioning strategy and an allocation strategy for the AI ​​tasks by executing a Markov decision process based on the current system state and task queue; S3, scheduling each sub-task to the corresponding computing server for execution according to the partitioning strategy and the allocation strategy.

[0007] According to an embodiment of the present invention, an electronic device is provided, comprising: a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the above-described AI task-splitable computation-power cooperative scheduling method.

[0008] According to an embodiment of the present invention, a storage medium is provided for storing computer-executable instructions, which, when executed, implement the steps of the above-described AI task-based computer-computer collaborative scheduling method.

[0009] The method proposed in this application effectively addresses the issues of scheduling rigidity and insufficient carbon awareness in existing scheduling schemes when facing large-scale, long-duration AI tasks. By intelligently decomposing AI tasks according to the computation graph structure and dynamically scheduling them in conjunction with real-time system status and carbon emission signals, the spatiotemporal matching accuracy of computing load and green power supply can be significantly improved while ensuring service quality. This fully taps the potential of data centers to participate in the regulation of new power systems, achieving a dual optimization of service quality and carbon reduction targets. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 is a flowchart of the AI ​​task-based computational and power-coordinated scheduling method according to an embodiment of the present invention; Figure 2 is a schematic diagram of the AI ​​task splitting according to an embodiment of the present invention. Detailed Implementation

[0012] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification 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 specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0013] According to an embodiment of the present invention, a computational power-coordinated scheduling method based on the decomposability of AI tasks is provided. Figure 1 is a flowchart of the computational power-coordinated scheduling method based on the decomposability of AI tasks according to an embodiment of the present invention. As shown in Figure 1, the computational power-coordinated scheduling method based on the decomposability of AI tasks according to an embodiment of the present invention specifically includes: S1, dividing the arriving AI tasks based on a preset partitioning target, and calculating the resource requirements of each sub-task after partitioning. The resource requirements include: memory requirements and computing power utilization. The AI ​​tasks depend on AI models, among which the Generalized Matrix Multiplication (GEMM) operator is the most basic element and the main contributor to the workload. Consider a batch size of an AI task. Basic GEMM operators It can be represented as ,in and Due to the separability of matrix operations, each batch of GEMM operations can be performed along different dimensions (e.g., or The original matrix is ​​split into a number of smaller submatrix multiplications, and each matrix multiplication can be executed independently on the server. As shown in Figure 2, an AI task involving multiple GEMM operations can be broken down into... There are 10 subtasks. Each subtask is a smaller matrix multiplication, with a batch size of 1. , dimension , and This corresponds to the multiplication between two light quantum matrices. In this case, the larger split size (i.e., ... This means that the task is executed collaboratively by more servers and consumes more computing power, so the workload can be more concentrated in periods with lower carbon emission factors.

[0014] Because renewable energy sources are integrated into the power system, carbon emission factors (CEFs) vary significantly in space and time. This suggests that concentrating workloads in locations and time periods with low CEFs can effectively reduce carbon emissions. By breaking down original tasks into subtasks and distributing them across different servers within a data center, workloads can be concentrated in periods with lower CEFs, thereby reducing carbon emissions. The computational resource requirements of AI tasks can be broken down into smaller chunks, better meeting computational resource constraints on servers and improving service quality. Multiple data centers are geographically distributed; for data centers... There are several types of servers with heterogeneous computing resource capacity (using...) (represented), each type of server collection is represented by This indicates that, at time step... A set of AI tasks that arrived (using) (This indicates that) we need to determine how these tasks are scheduled, i.e., in which data center each task is executed and when it is executed. The partitioning objective is a weighted sum of service quality rewards and carbon emission rewards, specifically represented by the following formula: ;in, Actions to be decided and They are the first Service quality rewards and carbon emission rewards at each time step. and It is a parameter that balances these two types of rewards.

[0015] After the AI ​​task is broken down, the next step is to estimate the computational resource requirements of each subtask under different splitting strategies: the AI ​​task is split into multiple parallel subtasks and they are scheduled to be executed during periods with lower carbon emission factors (CEF).

[0016] Two strategies can be used to break down a task into multiple parallel subtasks: 1) Model splitting: Split into small matrices, i.e. ; or 2) Data splitting: splitting batches Split into sizes Small batches. The computational cost of each subtask. It can be divided equally as follows: Memory is the minimum space allocated to ensure the successful execution of subtasks. It consists of three parts: model parameter memory, optimizer state memory, and activation memory. Model parameter memory is generated by weights and biases. It can be calculated as... ,in It is the number of bits required for the floating-point number format.

[0017] The optimizer state includes the gradient and the corresponding momentum of different orders, which requires... Several times the space. Therefore, the optimizer state memory can be calculated as... ,in It depends on the optimizer type (e.g., when using the SGD optimizer, Activations refer to the tensors created during forward propagation that are necessary for gradient computation. Both activations and their gradients require memory allocation, which can be computed as follows: .

[0018] Therefore, the memory requirements of subtasks It can be written as: ;in and The variations differ under different splitting strategies and can be obtained directly from the original task.

[0019] computing power utilization This refers to the utilization rate of floating-point operations per second when executing a task, which is mainly affected by computing intensity. The impact of computational intensity is defined as the ratio of floating-point operations to memory operations. Typically, and The relationship between them can be represented by a roofline model: This means along with Increase linearly until achieve ,after Stay .

[0020] To evaluate computational power utilization, we need to calculate the computational intensity of subtasks under different splitting strategies. Let... Indicates the first The memory operands in the operator include those for... and Read operations and The write operations generated a total of This involves memory operations. Therefore, the computational intensity under different partitioning strategies can be calculated as follows: It can be seen that the splitting strategy reduces computational intensity. However, when the splitting strategy reduces computational intensity, the task's computation speed may not increase exponentially as expected. Although subtasks can be processed separately on multiple servers, a complete run still requires some information exchange. More specifically, for both splitting strategies, the model parameters, weight gradients, and optimizer state of each subtask need to communicate to update parameters. This brings [a certain cost] to each subtask. This incurs a communication burden. Furthermore, activations and their corresponding gradients also need to be passed between servers, which generates additional overhead. The communication burden is thus calculated as follows: ;in and It varies depending on the splitting strategy and can be obtained directly from the original AI task.

[0021] S2. Based on the current system state and task queue, a Markov Decision Process (MDP) is executed to determine the splitting and allocation strategies for the AI ​​tasks. After modeling the resource requirements of AI tasks under different splitting strategies, a splitting and allocation strategy is determined for all arriving AI tasks to optimize carbon emissions and QoS. The allocation problem is defined as a Markov Decision Process (MDP) and can be solved using a Data Relationship Management (DRL) framework.

[0022] Define the state space as It contains the first Server resource utilization at each time step Task queue and carbon emission factors .

[0023] Server resource utilization is the status of servers across all locations and all types, and can be represented as: ;in , indicating the first The time step, located at the time step The location belongs to the first The first type The utilization rates of computing power, memory, and communication bandwidth of the servers.

[0024] Task queue status yes The state of each observed task. Due to the randomness of AI task arrival rates, the number of AI tasks that need to be scheduled at each time step. It is dynamic. To ensure consistency in the state space dimension, the first P tasks in the waiting sequence are observed, and their information is used to construct the task queue state, which serves as one of the inputs to the DRL framework. If the number of tasks in the waiting sequence is less than... The task queue will be filled with tasks where all elements are set to 0. Therefore, the task queue... It can be represented as: Each task's characteristics comprise six components (i.e., computational cost, two types of memory requirements, computational intensity, and two types of communication burden), and therefore can be modeled as follows: Based on the resource requirements of the tasks that have already been completed in the above steps, the action space of the DRL framework is defined as a combination of splitting and allocation strategies for a task in Tt at each time step, which can be expressed as: ;in .

[0025] At each time step , will from Choose one action .

[0026] if This means that the agent does not schedule any tasks at this moment, and all tasks will be blocked until the next moment; if This means the task It will be scheduled, and the batch size of all GEMM operators within it. and The dimensions will be split into and All subtasks will be scheduled to the data center. middle On the server of this type.

[0027] Since there are multiple tasks in the task queue at each time step, the action space could become excessively large if all actions were considered for decision-making. Furthermore, scheduling decisions are constrained by server resources; tasks must be scheduled to servers that can provide sufficient resources. Therefore, a masking scheme is employed to reduce the action space size, checking the feasibility of all scheduled actions and setting infeasible actions to a value that is not valid. .

[0028] The decisions made by the agent will lead to different state transitions. Let... , and These represent changes in computing power, memory, and communication bandwidth, respectively. If in the... The action at each time step is ,So , and All values ​​are equal to 0. Otherwise, the server's state will undergo corresponding changes as the resource requirements for each subtask are calculated according to the steps above.

[0029] During the execution of each subtask, the required resources will be continuously occupied. These resources will only be released after all subtasks have been completed. When a task... When scheduled, its execution time is: The reward function guides the training process of the DRL-based agent, thereby optimizing the overall QoS and carbon emissions throughout the entire cycle. Therefore, the reward at each time step... Including QoS rewards and carbon emission rewards , can be represented as: QoS reflects task latency. Low-carbon task allocation strategies tend to postpone tasks to periods with lower carbon emission factors, which may result in longer task delays and thus reduce overall service quality. QoS rewards are defined as the difference between the expected completion time and the actual completion time of a task. Therefore, at each time step, The acceleration gained by the tasks to be assigned, minus the latency of tasks still blocked in the queue, can be calculated as follows: Carbon emission rewards Defined as the sum of carbon emissions from all servers. The carbon emissions of each server can be assessed by its computing power utilization and computing efficiency. and carbon emission factors Multiplying them together gives the result. Therefore, It can be calculated as follows: A DQN-based agent is proposed, which includes a policy network for approximating future rewards. (Depend on (parameterization), and a target network designed to improve DRL stability. (Depend on (Parameterization). At each time step The agent first observes the current state. Then, the total future reward for all actions can be estimated as follows: The intelligent agent will choose As the first The state changes accordingly with each time step of the action. To guide network training, the loss function is generated as follows: ;in By providing instant rewards It is obtained by adding it to the future reward of the next time step, and the calculation method is as follows: ;in This is a discount factor that measures the importance of future rewards. Then, it can be minimized using the loss function. Update .

[0030] S3. Based on the splitting and allocation strategies, schedule each subtask to the corresponding computing server for execution.

[0031] Based on the splitting and allocation strategies determined in step S2, the scheduling system generates a specific set of scheduling instructions. This set of instructions specifies in detail the target execution location (i.e., the specific data center and server), execution timing requirements, and required resource quotas for each subtask. The system then issues these instructions to the corresponding computing resource management components.

[0032] Upon receiving the scheduling instruction, the resource management component first verifies whether the real-time resource availability of the target server meets the requirements of the subtask, including computing power, memory space, and communication bandwidth. If the verification is successful, the component will reserve the necessary resource blocks for each subtask on its target server and lock these resource blocks until the task is completed to prevent resource conflicts.

[0033] After successful resource reservation, the system will initiate the loading and initialization process of subtasks. The code snippets, model parameter slices, and data processing logic corresponding to each subtask will be transmitted and deployed to the designated server. For training tasks involving parameter synchronization, the system will also establish communication links between subtasks and determine the timing and frequency of parameter synchronization.

[0034] After loading their subtasks, each server begins executing its assigned computational tasks in parallel. During execution, the system continuously monitors the progress of each subtask, resource consumption, and the health of the server nodes. Simultaneously, the system collects key runtime metrics, such as actual computation time, peak memory usage, and network traffic. This data is fed back to update the system status and provides a basis for subsequent scheduling decisions.

[0035] Once a subtask has completed all its computations, it reports its completion status to the scheduling system and releases the computing resources and memory it occupied. After confirming that all subtasks of a given AI task have been successfully completed, the scheduling system triggers the subsequent result aggregation process. For example, for a distributed training task, the system will coordinate the synchronization and aggregation of gradients across servers and update the global model parameters.

[0036] Through the above-mentioned refined scheduling and execution process, this invention not only realizes the transformation of macroscopic computing and power coordination goals into microscopic server-level instructions, but also ensures the reliable and efficient operation of large-scale AI tasks in complex heterogeneous environments by introducing monitoring, feedback and fault tolerance mechanisms during execution, ultimately achieving the comprehensive optimization goal of service quality and carbon emission reduction.

[0037] By employing the embodiments of the present invention, the following beneficial effects are achieved: The method proposed in this application can effectively solve the problems of insufficient scheduling rigidity and carbon awareness capability in existing scheduling schemes when facing large-scale, long-duration AI tasks. By intelligently decomposing AI tasks according to the computation graph structure and dynamically scheduling them in conjunction with real-time system status and carbon emission signals, the spatiotemporal matching accuracy of computing load and green power supply can be significantly improved while ensuring service quality. This fully taps the potential of data centers to participate in the regulation of new power systems, achieving a dual optimization of service quality and carbon emission reduction targets.

[0038] According to an embodiment of the present invention, an electronic device is provided, comprising: a processor; and a memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the above-described AI task-splitable computation-power cooperative scheduling method.

[0039] According to an embodiment of the present invention, a storage medium is provided for storing computer-executable instructions, which, when executed, implement the steps of the above-described AI task-based computer-computer collaborative scheduling method.

[0040] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A computational-powered collaborative scheduling method based on AI task decomposability, characterized in that... include: S1. Divide the arriving AI tasks based on a preset partitioning target, and calculate the resource requirements of each sub-task after partitioning. The resource requirements include memory requirements and computing power utilization. S2. Based on the current system state and task queue, determine the splitting strategy and allocation strategy for the AI ​​tasks by executing a Markov decision process. S3. According to the splitting strategy and allocation strategy, schedule each sub-task to the corresponding computing server for execution.

2. The method according to claim 1, characterized in that, The target for this allocation is a weighted sum of incentives for optimizing service quality and carbon emission, specifically expressed by the following formula: ;in, Actions to be decided and They are the first Service quality rewards and carbon emission rewards at each time step. and It is a parameter that balances these two types of rewards.

3. The method according to claim 1, characterized in that, The memory requirements include: model parameter memory, optimizer state memory, and activation memory; the model parameter memory is generated by weights and biases, and is expressed as follows: ;in, This refers to the number of bits required in floating-point format; the optimizer state includes the gradient and the corresponding momentum of different orders, expressed as: Where m represents the subtask Divide into m smaller matrices. Parameters related to the optimizer type; the active memory is the tensor created during forward propagation that is necessary for gradient computation, represented as: Where B is the batch size. The output of the GEMM operation represents the first... A result vector.

4. The method according to claim 1, characterized in that, The computing power utilization rate is obtained using the following formula: ;in Indicates the calculated strength. This is the preset calculation intensity threshold.

5. The method according to claim 1, characterized in that, The Markov decision process is executed online by a pre-trained reinforcement learning agent, including: at each decision time step, the agent observes the current system state; based on the observed state, it evaluates the total future reward of all possible actions using a pre-trained policy network; and selects the action with the highest reward as the scheduling decision for the current time step based on the evaluation results; the feasibility of the possible actions is determined by checking whether the task resource requirements match the available resources of the target server.

6. The method according to claim 5, characterized in that, The current system status includes: server resource utilization status, which represents the real-time utilization of computing power, memory and communication bandwidth of each data center and type of server; task queue status, including resource requirement information of the previous task in the waiting scheduling sequence; and carbon emission factor information, which represents the carbon emission intensity of each data center at each time step.

7. The method according to claim 5, characterized in that, The optional action is defined as a scheduling decision for a task in the current task queue, expressed as: Among them, if If this is true, it means the agent will not schedule any tasks at this moment, and all tasks will be blocked until the next time step; if This indicates the task. It will be scheduled, and the batch size of all GEMM operators within it. and The dimensions will be split into and All subtasks will be scheduled to the data center. middle On the server of this type.

8. The method according to claim 1, characterized in that, The reward function guides the training process of the agent, and the reward function is expressed as follows: ;in, The acceleration is gained from the tasks to be assigned. Incentives for carbon emissions.

9. An electronic device, comprising: processor; And a memory configured to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the AI ​​task-decomposable computational-electrical collaborative scheduling method as described in any one of claims 1-8.

10. A storage medium for storing computer-executable instructions, which, when executed, implement the steps of the AI ​​task-splitable computer-computer collaborative scheduling method as described in any one of claims 1-8.

Images