Method, device, equipment, medium and program product for multi-objective computing power scheduling

Abstract

The application discloses a multi-target computing power scheduling method, device, equipment, medium and program product, and relates to the technical field of computing power networks. The scheduling device of the method receives task carrying willingness scores and resource state information of a plurality of resource pool systems, and splits a computing power task sent by a task party to obtain a plurality of sub-computing power tasks; the task carrying willingness scores, the resource state information and the sub-computing power tasks are matched to obtain candidate resource pool systems, and scheduling information is sent to the candidate resource pool systems. The candidate resource pool systems determine predicted load evaluation results according to the scheduling information and real-time running data and send the predicted load evaluation results to the scheduling device; the scheduling device determines a target resource pool system of each sub-computing power task according to the predicted load evaluation results and the sub-computing power tasks, and sends a corresponding sub-computing power task to the target resource pool system, so that the target resource pool system executes the sub-computing power task. Therefore, the effect of computing power scheduling is improved.

Description

Technical Field

[0001] This application belongs to the field of computing power network technology, and in particular relates to a method, apparatus, equipment, medium and program product for multi-objective computing power scheduling. Background Technology

[0002] With the rapid development of artificial intelligence technology, the demand for computing power has exploded. Computing power has become the core productive force driving the development of the digital economy. Building an efficient, intelligent, and green computing power network infrastructure is of great significance for technological progress and industrial upgrading.

[0003] Existing multi-objective computing power scheduling methods often use multi-objective optimization algorithms such as genetic algorithms and particle swarm optimization to perform static scoring and scheduling of resource pool systems. Although they can take into account multiple conflicting objectives at the same time, such decisions are based on the current state of resources and do not consider the impact of scheduling actions on the state of the resources in the future. Furthermore, they cannot adaptively adjust with the actual dynamic changes of tasks and resources. Therefore, the resource scheduling schemes of existing computing power scheduling methods are prone to discrepancies with the resources that can be scheduled in actual operation, resulting in poor computing power scheduling performance. Summary of the Invention

[0004] This application provides a method, apparatus, device, medium, and program product for multi-objective computing power scheduling, in order to solve the problem that existing resource scheduling schemes are prone to discrepancies with the resources that can be scheduled in actual operation, resulting in poor computing power scheduling performance.

[0005] In a first aspect, embodiments of this application provide a method for multi-objective computing power scheduling, applied to a scheduling device, the method comprising: It receives task carrying willingness scores and resource status information of different types of computing power tasks sent by multiple resource pool systems. The task carrying willingness score is determined by the resource pool system based on the real-time operation data and resource operation trend of the resource pool system. Receive computing power tasks sent by the task provider, and split the computing power tasks into multiple sub-computing power tasks; Based on the task carrying willingness scores, resource status information, and multiple sub-computing tasks of multiple resource pool systems, the resource pool systems and sub-computing tasks are matched to obtain the candidate resource pool system corresponding to each sub-computing task; Scheduling information is generated based on the information of each sub-computing task, and the scheduling information of the sub-computing task is sent to the candidate resource pool system corresponding to each sub-computing task. This allows the candidate resource pool system to determine the predicted load evaluation result of the candidate resource pool system after the execution of the sub-computing task based on the scheduling information and the real-time running data of the candidate resource pool system. Receive the predicted load assessment results sent by multiple candidate resource pool systems, determine the target resource pool system for each sub-computing task based on the predicted load assessment results and sub-computing task, and ensure that the comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than the set threshold. Send the corresponding sub-computing task to the target resource pool system for the target resource pool system to execute the sub-computing task.

[0006] Secondly, embodiments of this application provide a method for multi-objective computing power scheduling, applied to a resource pool system, the method comprising: Obtain real-time operating data, historical operating data, and resource status information of the resource pool system; Determine the predicted resource operation trend within a preset time window based on historical operation data; Based on real-time operational data and resource operation trends, determine the task carrying willingness scores for different types of computing power tasks; The scheduling device sends task carrying willingness scores and resource status information of different types of computing power tasks to the scheduling device, so that the scheduling device can match the resource pool system and the sub-computing power task according to the task carrying willingness scores, resource status information and multiple sub-computing power tasks, and obtain the candidate resource pool system corresponding to each sub-computing power task. The system receives scheduling information for sub-computing tasks from the scheduling device and determines the predicted load assessment result of the resource pool system after the execution of the sub-computing tasks based on the scheduling information and real-time running data. The scheduling information is generated by the scheduling device based on the information of each sub-computing task. Send the predicted load assessment results to the scheduling device so that the scheduling device can determine the target resource pool system for each sub-computing task based on the predicted load assessment results and the sub-computing task. The comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than the set threshold. Receive sub-computing tasks sent by the scheduling device and execute the sub-computing tasks.

[0007] Thirdly, embodiments of this application provide a multi-target computing power scheduling apparatus, applied to a scheduling device, the apparatus comprising: The receiving module is used to receive task carrying willingness scores and resource status information of different types of computing power tasks sent by multiple resource pool systems. The task carrying willingness score is determined by the resource pool system based on the real-time operation data and resource operation trend of the resource pool system. The receiving module is also used to receive computing power tasks sent by the task party and to split the computing power tasks into multiple sub-computing power tasks. The matching module is used to match resource pool systems and sub-computing tasks based on task carrying willingness scores, resource status information, and multiple sub-computing tasks of multiple resource pool systems, to obtain candidate resource pool systems corresponding to each sub-computing task. The sending module is used to generate scheduling information based on the information of each sub-computing task and send the scheduling information of the sub-computing task to the candidate resource pool system corresponding to each sub-computing task, so that the candidate resource pool system can determine the predicted load evaluation result of the candidate resource pool system after the execution of the sub-computing task based on the scheduling information and the real-time running data of the candidate resource pool system. The determination module is used to receive the predicted load assessment results sent by multiple candidate resource pool systems, determine the target resource pool system for each sub-computing task based on the predicted load assessment results and sub-computing task, and ensure that the comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than a set threshold. The sending module is also used to send the corresponding sub-computing power tasks to the target resource pool system for the target resource pool system to execute the sub-computing power tasks.

[0008] Fourthly, embodiments of this application provide a multi-target computing power scheduling apparatus applied to a resource pool system, the apparatus comprising: The acquisition module is used to acquire real-time operating data, historical operating data, and resource status information of the resource pool system; The determination module is used to determine the predicted resource operation trend within a preset time window based on historical operation data; The determination module is also used to determine the task carrying willingness score of different types of computing power tasks based on real-time operating data and resource operation trends; The sending module is used to send task carrying willingness scores and resource status information of different types of computing power tasks to the scheduling device, so that the scheduling device can match the resource pool system and the sub-computing power task according to the task carrying willingness scores, resource status information and multiple sub-computing power tasks of multiple resource pool systems, and obtain the candidate resource pool system corresponding to each sub-computing power task. The receiving module is used to receive the scheduling information of the sub-computing tasks sent by the scheduling device, and to determine the predicted load assessment result of the resource pool system after the execution of the sub-computing tasks based on the scheduling information and real-time running data. The scheduling information is generated by the scheduling device based on the information of each sub-computing task. The sending module is also used to send the predicted load assessment results to the scheduling device, so that the scheduling device can determine the target resource pool system for each sub-computing task based on the predicted load assessment results and the sub-computing task. The comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than a set threshold. The receiving module is also used to receive sub-computing tasks sent by the scheduling device and execute the sub-computing tasks.

[0009] Fifthly, embodiments of this application provide a terminal device, the device including: a processor and a memory storing computer program instructions; the processor, when executing the computer program instructions, implements a method for multi-target computing power scheduling as described in the first or second aspect.

[0010] In a sixth aspect, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement a multi-target computing power scheduling method as described in the first or second aspect.

[0011] In a seventh aspect, embodiments of this application provide a computer program product in which instructions, when executed by a processor of an electronic device, cause the electronic device to perform a multi-target computing power scheduling method as described in the first or second aspect.

[0012] This application provides a method, apparatus, device, medium, and program product for multi-objective computing power scheduling. The scheduling device receives task load willingness scores and resource status information for different types of computing power tasks sent by multiple resource pool systems. The task load willingness score is determined by the resource pool system based on real-time operating data and resource operation trends. The task load willingness score is determined based on both real-time operating data and resource operation trends, rather than relying solely on static resource information, thus possessing foresight and able to proactively avoid performance risks caused by short-term increases in resource load. The scheduling device receives computing power tasks sent by the task provider and splits them into multiple sub-computing power tasks. By splitting them into sub-tasks, parallel execution of sub-tasks is supported, significantly improving the overall execution efficiency of the task and shortening the end-to-end completion time. The scheduling device matches resource pool systems and sub-computing power tasks based on the task load willingness scores, resource status information, and multiple sub-computing power tasks of multiple resource pool systems, obtaining candidate resource pool systems corresponding to each sub-computing power task. The candidate range of resource pool systems is narrowed down based on multi-dimensional screening rules, reducing the computational load of subsequent global optimization and improving scheduling decision efficiency. Scheduling information is generated based on the information of each sub-computing task, and this information is sent to the candidate resource pool system corresponding to each sub-computing task. This allows the candidate resource pool system to determine the predicted load assessment result after the execution of the sub-computing task based on the scheduling information and its real-time operating data. The predicted load assessment result is calculated by the resource pool system based on its own real-time operating data, avoiding a disconnect between decision-making and the actual resource status compared to estimation by the scheduling device. The scheduling device receives predicted load assessment results from multiple candidate resource pool systems, determines the target resource pool system for each sub-computing task based on the predicted load assessment results and the sub-computing task, and ensures that the comprehensive evaluation value of multiple target indicators after all target resource pool systems execute the corresponding sub-computing task is less than a set threshold. The device then sends the corresponding sub-computing task to the target resource pool system for execution. Using the multi-objective comprehensive evaluation value threshold as a constraint, the device achieves global optimization of sub-computing task allocation while considering multiple objective indicator dimensions. Therefore, this embodiment improves the matching degree between computing power scheduling and actual operating status, thereby improving the effectiveness of computing power scheduling. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of the structure of the multi-objective computing power scheduling system provided in the embodiments of this application; Figure 2This is a flowchart illustrating the multi-target computing power scheduling method provided in an embodiment of this application; Figure 3 This is a flowchart illustrating the method for determining the target resource pool system provided in an embodiment of this application; Figure 4 This is a flowchart illustrating the method for updating the target resource pool system provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a device for multi-target computing power scheduling applied to scheduling equipment, as provided in an embodiment of this application; Figure 6 This is a schematic diagram of the structure of the device for multi-objective computing power scheduling applied to a resource pool system provided in the embodiments of this application; Figure 7 This is a schematic diagram of the structure of the terminal device provided in the embodiments of this application. Detailed Implementation

[0015] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0016] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0017] Existing computing resource scheduling schemes can generally be divided into two typical modes: The first is the centralized decision-making mode, where the scheduling center uniformly formulates the scheduling strategy. This method makes decisions quickly, but lacks flexibility and cannot respond promptly to changes in demand or resource status, making it particularly unsuitable for scenarios where the scheduling center and resource providers are different stakeholders. The second is the decentralized negotiation mode, where task requesters negotiate independently with multiple resource pools, and each resource pool makes decisions based on its local status. This mode lacks a global perspective, easily leading to uneven resource utilization, idle local resources, and increased task waiting times, thereby reducing overall resource utilization and task execution efficiency.

[0018] To address the problems of existing technologies, this application provides a method, apparatus, device, medium, and program product for multi-objective computing power scheduling. The method involves a scheduling device receiving task load willingness scores and resource status information for different types of computing power tasks from multiple resource pool systems. The task load willingness score is determined by the resource pool system based on real-time operating data and resource operation trends. This score is determined based on both real-time operating data and resource operation trends, rather than relying solely on static resource information, thus possessing foresight and proactively mitigating performance risks caused by short-term increases in resource load. The scheduling device receives computing power tasks from task providers and splits them into multiple sub-computing power tasks. This splitting into sub-tasks supports parallel execution, significantly improving overall task execution efficiency and shortening end-to-end completion time. The scheduling device matches resource pool systems with sub-computing power tasks based on the task load willingness scores, resource status information, and multiple sub-computing power tasks, obtaining candidate resource pool systems for each sub-computing power task. Multi-dimensional filtering rules narrow down the candidate range of resource pool systems, reducing the computational load of subsequent global optimization and improving scheduling decision efficiency. Scheduling information is generated based on the information of each sub-computing task, and this information is sent to the candidate resource pool system corresponding to each sub-computing task. This allows the candidate resource pool system to determine the predicted load assessment result after the execution of the sub-computing task based on the scheduling information and its real-time operating data. The predicted load assessment result is calculated by the resource pool system based on its own real-time operating data, avoiding a disconnect between decision-making and the actual resource status compared to estimation by the scheduling device. The scheduling device receives predicted load assessment results from multiple candidate resource pool systems, determines the target resource pool system for each sub-computing task based on the predicted load assessment results and the sub-computing task, and ensures that the comprehensive evaluation value of multiple target indicators after all target resource pool systems execute the corresponding sub-computing task is less than a set threshold. The device then sends the corresponding sub-computing task to the target resource pool system for execution. Using the multi-objective comprehensive evaluation value threshold as a constraint, the device achieves global optimization of sub-computing task allocation while considering multiple objective indicator dimensions. Therefore, this embodiment improves the matching degree between computing power scheduling and actual operating status, thereby improving the effectiveness of computing power scheduling.

[0019] The following section first introduces the system for multi-objective computing power scheduling, which is applied to the method of multi-objective computing power scheduling provided in the embodiments of this application.

[0020] Figure 1 A schematic diagram of the structure of a multi-objective computing power scheduling system 100 provided in an embodiment of this application is shown. Figure 1As shown, the system may include: a scheduling device 101, a resource pool system 102, and a task-side device 103. The scheduling device 101, as the core hub of the entire system, undertakes the key functions of overall decision-making, resource matching, and dynamic optimization, achieving adaptive adjustment during runtime. The resource pool system 102, as the provider of computing power and the physical carrier for executing computing tasks, may contain multiple heterogeneous resource pools. The task-side device 103 is the initiator of computing power demands. These three components collaborate to form a closed-loop system of demand, scheduling, and computing power coordination, achieving efficient, dynamic, and globally optimal scheduling of computing tasks.

[0021] The method for multi-objective computing power scheduling provided in the embodiments of this application is described below.

[0022] Figure 2 A flowchart illustrating a multi-objective computing power scheduling method provided in an embodiment of this application is shown. Figure 2 As shown, the method may include the following steps: S201 to S205.

[0023] S201, the resource pool system obtains real-time operating data, historical operating data and resource status information of the resource pool system.

[0024] Among them, real-time operation data is the current operation indicator data of the resource pool; historical operation data is the operation record data of the resource pool over a period of time in the past; and resource status information is the basic resource configuration and availability data of the resource pool.

[0025] In some embodiments, the resource pool system can collect its own real-time operating data in real time through data acquisition tools; at the same time, it can retrieve past operating records from the database as historical operating data; and then integrate its own total resource configuration, idle resource quantity and other resource information to obtain resource status information.

[0026] In some embodiments, each resource pool system periodically performs a comprehensive self-check of its own status and reports the results according to a preset cycle agreed upon with the scheduling center. During the self-check, the resource pool system collects real-time operational data, including but not limited to the following indicators, through a data acquisition agent: resource load performance indicators: various performance indicators, including resource idle capacity, CPU utilization, GPU utilization, network inbound and outbound traffic at key nodes, resource fragmentation level, etc.; business load indicators: including the current number of tasks, average task load, task queue length, average task waiting time, task execution stability data, etc.; infrastructure load indicators: including average rack temperature, rack power load rate, and overall data center cooling efficiency coefficient.

[0027] This application embodiment enables the resource pool system to fully grasp its own operating status by acquiring multi-dimensional and full-cycle data, avoiding subsequent decision-making biases caused by incomplete data, and providing data support for accurately assessing task carrying capacity.

[0028] S202, the resource pool system determines the predicted resource operation trend within a preset time window based on historical operation data.

[0029] The preset time window is a pre-defined time range used to predict resource trends; the resource operation trend is the direction and magnitude of change of the resource pool's operation indicators within the preset time window.

[0030] In some embodiments, the resource pool system can select an appropriate prediction model based on the acquired historical operating data, and combine the fluctuation patterns and periodic characteristics of the historical data to predict the resource operating trend within a preset time window, and output the predicted resource operating trend.

[0031] This application's embodiments break through the limitations of traditional scheduling that relies solely on current static data. By predicting future resource states, the task load assessment of the resource pool becomes forward-looking, enabling the avoidance of performance risks caused by excessive future loads and improving the predictability of scheduling decisions.

[0032] In some embodiments, different resource pool systems can be divided according to regions or node clusters, and different models can be used to predict the resource operation trends of resource pool systems in different regions.

[0033] It supports refined regional scheduling, which can allocate tasks to regions with more stable trends and improve the balance of resource utilization.

[0034] S203, the resource pool system determines the task carrying willingness score of different types of computing power tasks based on real-time operation data and resource operation trends.

[0035] The different types of computing power tasks are the task categories. The task carrying willingness score is a quantitative score given by the resource pool system for different types of tasks, based on their real-time status and future trends.

[0036] In some embodiments, when calculating the load impact factor, the weight of energy consumption indicators can be added to prioritize reducing the willingness to carry high-energy-consuming tasks and guide tasks to be allocated to the low-energy-consuming resource pool.

[0037] In some embodiments, the basic adaptability score can be adjusted according to the priority of the task. The basic score for high-priority tasks can be appropriately increased to ensure resource matching for core tasks and improve the execution efficiency of key businesses.

[0038] This application embodiment quantifies the willingness to carry different types of computing power tasks by combining real-time operating data and resource operation trends, thereby achieving accurate and dynamic evaluation of the resource pool's carrying capacity. It takes into account both the current resource load and future resource trend changes, providing an objective and scientific quantitative basis for task and resource matching in the scheduling center, and effectively improving the rationality and accuracy of task allocation.

[0039] S204, the resource pool system sends task carrying willingness scores and resource status information for different types of computing power tasks to the scheduling equipment.

[0040] Among them, the scheduling equipment is the core equipment responsible for coordinating the allocation of computing power tasks.

[0041] In some embodiments, the resource pool system encapsulates the calculated carrying willingness scores for different task types with the acquired resource status information and sends them to the scheduling device through a preset communication protocol.

[0042] This application embodiment sends the computing power task carrying willingness scores and resource status information of different types of resource pools to the scheduling device, enabling the scheduling device to grasp the carrying capacity and resource status of each resource pool in real time and comprehensively. This provides key data support for subsequent task splitting and resource matching, and ensures the smooth progress of the computing power scheduling process.

[0043] In some embodiments, the resource pool can push only the bearer willingness score related to the currently scheduled task to the scheduling device according to the task type, thereby reducing the amount of data transmission, improving communication efficiency, and reducing the parsing pressure on the scheduling device.

[0044] S205, the task-side device sends a computing power task to the scheduling device.

[0045] In this context, the task-side equipment refers to the terminal or server that submits the computing power request. The computing power task refers to the specific computing requirements submitted by the task-side provider.

[0046] In some embodiments, the task-oriented device encapsulates its own computing power requirement information in a standardized format to obtain a computing power task, and sends the computing power task to the scheduling device.

[0047] This application embodiment sends computing power tasks to the scheduling device through the task-side device, providing a clear objective for the entire computing power scheduling process, laying the foundation for subsequent intelligent task splitting and precise resource matching, and ensuring the pertinence and effectiveness of computing power scheduling.

[0048] S206, the scheduling device receives the computing power task sent by the task party and splits the computing power task into multiple sub-computing power tasks.

[0049] Sub-computing tasks are small, independently executable tasks obtained by splitting the original computing task according to certain rules. The task splitting rules are the splitting criteria set for different task types.

[0050] In some embodiments, after the scheduling device receives the computing power task sent by the task party, it first matches the corresponding splitting rule according to the task type of the computing power task, and then splits the computing power task according to the splitting rule; finally, it splits the computing power task into multiple sub-computing power tasks according to preset conditions.

[0051] In some embodiments, when splitting tasks, the scheduling device can combine the resource types of each resource pool to split out sub-tasks that are adapted to different resource types, enabling the sub-tasks to accurately match the hardware capabilities of the resource pool system and improve task execution efficiency.

[0052] In some embodiments, the scheduling device can dynamically adjust the size of subtasks based on the amount of idle resources in the current resource pool, which can flexibly adapt to the load of the resource pool and avoid increased scheduling overhead caused by subtasks being too small.

[0053] This application embodiment intelligently splits the computing power task sent by the task party into multiple sub-computing power tasks by scheduling equipment, which can transform a complex single task into a cluster of sub-tasks that can be executed in parallel, giving full play to the distributed computing power advantages of multiple resource pools and shortening the overall execution time of the task.

[0054] S207, the scheduling device receives task carrying willingness scores and resource status information of different types of computing power tasks, and matches resource pool systems and sub-computing power tasks according to the task carrying willingness scores, resource status information and multiple sub-computing power tasks to obtain candidate resource pool systems corresponding to each sub-computing power task.

[0055] Among them, the candidate resource pool system is a resource pool system that has the ability to carry the corresponding sub-computing power tasks.

[0056] In some embodiments, the scheduling device aggregates the bearer willingness scores and resource status information sent by all resource pools, and then, for each sub-computing task, filters resource pools according to preset matching rules, that is, filters resource pools whose resource types and specifications meet the requirements of the sub-task, and then filters resource pools whose bearer willingness scores reach the threshold, and finally forms a candidate resource pool list for each sub-task.

[0057] In some embodiments, the historical task execution success rate metric of the resource pool can be added during matching to prioritize the selection of resource pools with high success rates, such as resource pools with success rates higher than a set threshold, thereby improving the stability of subtask execution and reducing the risk of task failure.

[0058] In some embodiments, for subtasks with local computing requirements, resource pools in the same region can be prioritized to reduce data transmission latency, improve task response speed, and meet low-latency service requirements.

[0059] This application embodiment combines task carrying willingness scores, resource status information, and multiple sub-computing tasks with a scheduling device to match and select candidate resource pools for each sub-task. This not only narrows the scope of subsequent global optimization decisions and reduces computational overhead, but also ensures the compatibility between candidate resource pools and sub-tasks through multi-dimensional matching rules, laying the foundation for selecting the optimal resource pool and improving the overall efficiency of computing power scheduling.

[0060] S208, the scheduling device sends the scheduling information of the sub-computing task to the candidate resource pool system corresponding to each sub-computing task.

[0061] In some embodiments, the scheduling device encapsulates information such as the subtask ID, resource requirements, SLA requirements, and data path into scheduling information for each sub-computing task, and then sends it to all candidate resource pool systems corresponding to the subtask, notifying the candidate resource pools to evaluate the feasibility of undertaking the subtask.

[0062] This application embodiment realizes the synchronization of task information between the scheduling device and the candidate resource pool, enabling the candidate resource pool to clearly understand the specific requirements of the sub-tasks, providing a basis for subsequent predictive load assessment, and ensuring the pertinence of the assessment process.

[0063] In some embodiments, the priority of subtasks can be added to the scheduling information, with high-priority subtasks being evaluated first, guiding the candidate resource pool to prioritize the evaluation of high-priority subtasks and ensuring the efficiency of core task progress.

[0064] In some embodiments, scheduling information fields adapted to the resource type can be sent to different types of candidate resource pools. For example, only GPU and related requirements are sent to the GPU resource pool, reducing interference from irrelevant information and improving the evaluation efficiency of candidate resource pools.

[0065] S209, the resource pool system receives the scheduling information of the sub-computing task sent by the scheduling device, and determines the predicted load assessment result of the resource pool system after the execution of the sub-computing task based on the scheduling information and real-time running data.

[0066] Among them, the predicted load assessment result is a quantitative assessment data of the future load changes of the resource pool after simulating the execution of sub-tasks.

[0067] In some embodiments, after receiving the subtask scheduling information sent by the scheduling device, the candidate resource pool system combines its own real-time operating data and simulates the execution of the subtask through a load prediction model, evaluates the changes in indicators such as resource load, power consumption, and temperature after execution, and finally generates a predicted load evaluation result.

[0068] This application embodiment determines the predicted load assessment result after the execution of subtasks by using a resource pool system based on scheduling information and its own real-time operating data. By utilizing the resource pool's accurate understanding of its own operating status, it replaces the remote estimation of the scheduling equipment, effectively avoiding assessment deviations caused by information asymmetry, improving the accuracy and reliability of load assessment, and providing high-quality data references for the subsequent global optimization decisions of the scheduling equipment.

[0069] S210, the resource pool system sends the predicted load assessment results to the scheduling equipment.

[0070] In some embodiments, the candidate resource pool system encapsulates the generated predicted load assessment results in the format required by the scheduling device and sends them to the scheduling device through a preset communication protocol, ensuring that the scheduling device can directly parse the load assessment data of each candidate resource pool for the same subtask.

[0071] This application embodiment sends the predicted load assessment results after the execution of subtasks to the scheduling device through the resource pool system, realizing the efficient feedback of candidate resource pool assessment data to the scheduling device. This allows the scheduling device to comprehensively summarize the load impact assessment information of each resource pool on the same subtask, providing rich and accurate data support for subsequent multi-objective global optimization decisions and ensuring the scientific and comprehensive nature of scheduling decisions.

[0072] In some embodiments, a timestamp can be added to the predicted load assessment results to mark the time point of the assessment, so that the scheduling device can distinguish the assessment results at different times and avoid the interference of outdated data caused by changes in resource status with decision-making.

[0073] S211, the scheduling device receives the predicted load assessment results and determines the target resource pool system for each sub-computing task based on the predicted load assessment results and the sub-computing task. The comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than the set threshold.

[0074] The target resource pool system is the optimal resource pool selected from the candidate resource pool that meets the comprehensive evaluation requirements of multiple objective indicators after executing sub-tasks. The multiple objective indicators are the core dimensions used by the scheduling equipment for evaluation. The set threshold is a pre-defined upper limit for the comprehensive evaluation value.

[0075] In some embodiments, the scheduling device aggregates the predicted load assessment results of all candidate resource pools, constructs a multi-objective optimization model for each sub-computing task, calculates the comprehensive evaluation value of each candidate resource pool after executing the sub-task, then filters out resource pools whose comprehensive evaluation value is less than a set threshold, and finally determines the target resource pool for each sub-task.

[0076] This application embodiment uses a multi-objective optimization model to select target resource pools, which can take into account the needs of multiple dimensions such as performance, cost, and energy consumption, and achieve global optimal scheduling. At the same time, threshold constraints ensure the load safety of the resource pool and avoid overload operation.

[0077] S212, the scheduling device sends the corresponding sub-computing power task to the target resource pool system.

[0078] In some embodiments, the scheduling device encapsulates the corresponding execution instructions, data location, SLA requirements, and other information into a task issuance instruction for each sub-computing task, and sends it to the target resource pool system of the sub-task to notify the target resource pool to start the sub-task execution process.

[0079] This application embodiment sends corresponding sub-computing tasks to the target resource pool system through scheduling equipment, realizing the implementation of computing power scheduling decisions from theory to actual execution. It ensures that sub-tasks can be accurately distributed to the optimal target resource pool. At the same time, through clear execution instructions and constraints, it ensures that sub-tasks can be executed efficiently and stably.

[0080] S213, the resource pool system receives and executes the sub-computing tasks sent by the scheduling device.

[0081] This embodiment acquires real-time, historical, and status data through a resource pool system and predicts resource trends. It combines dual-dimensional data to determine the task carrying willingness score and feeds it back to the scheduling device. Simultaneously, the task party submits a computing power task. The scheduling device intelligently splits the task into sub-tasks and matches candidate resource pools based on the resource pool's willingness score and status information. Then, it performs multi-objective global optimization based on the predicted load evaluation results of the candidate resource pools, determines the target resource pool, and issues the sub-tasks for execution. The entire process breaks through the limitations of traditional static scheduling, realizes dynamic adaptation between task requirements and resource status, and takes into account the global optimization of multiple objective dimensions such as performance, cost, and energy consumption, thereby improving the efficiency and accuracy of computing power resource scheduling.

[0082] In some embodiments, the scheduling device splits the computing power task into multiple sub-computing power tasks, which may include: The scheduling device splits the computing power tasks according to the preset splitting rules of the task types, and obtains a directed acyclic graph including sub-task nodes and the dependencies between sub-task nodes. Among them, the preset splitting rules are the splitting standards pre-set for different types of computing power tasks; the directed acyclic graph is a graphical structure used to describe the subtask nodes and the dependencies between nodes. Nodes represent subtasks, and directed edges represent the execution order of subtasks, and there are no circular dependencies. If the subtask corresponding to the subtask node meets the preset conditions, the scheduling device splits the subtask so that none of the subtask nodes in the directed acyclic graph meet the preset conditions, resulting in multiple sub-computing tasks.

[0083] This application embodiment first constructs a directed acyclic graph containing subtask nodes and dependencies based on preset splitting rules for task types using a scheduling device. Then, iteratively splits subtask nodes that meet preset conditions until all nodes no longer meet the preset conditions. This achieves refined splitting of computing power tasks from macro to micro levels, ensuring both the orderliness of subtask execution logic and the adaptability of the final sub-computing power tasks to the resource pool capabilities. It fully leverages the parallel computing advantages of multiple resource pools, effectively improving the execution efficiency and resource utilization of computing power tasks, while avoiding the subjectivity and inefficiency of manual splitting.

[0084] In some embodiments, the scheduling device splits the computing task according to a preset splitting rule for the task type, resulting in a directed acyclic graph including sub-task nodes and dependencies between sub-task nodes, which may include: The scheduling center parses received task requests and transforms them into an internally standardized data structure containing information such as SLA, cost budget, and business requirements. This includes, but is not limited to: Task Identifier: A globally unique identifier (UUID) for the task; Business Requirements: Clearly defining the task type (e.g., training, inference), model type (e.g., Transformer model, CNN model), and data scale (dataset size and location); Service Level Agreement (SLA) Requirements: Required performance metrics for the task (e.g., TPS ≥ 5000), task execution latency requirements, task completion deadline, and fault recovery time; Cost Budget Constraints: The budget range for task execution and the billing model (e.g., billing by card time, billing by fixed duration and specifications, etc.); Task Preferences: Specific preference conditions for the task, such as resource type restrictions (GPU / NPU), deployment location preferences, and isolation level requirements (dedicated instance, shared instance, etc.).

[0085] Based on common task types in intelligent computing scenarios, the scheduling center pre-organizes a structured task type knowledge base, clarifies the standard splitting rules for different task types, and uses a fine-tuned large model to match the task description information with the task types in the knowledge base. It automatically determines the task type and selects the corresponding standard splitting rules, outputs the task sub-task nodes and the data dependencies between nodes, and forms a structured task DAG graph.

[0086] In some embodiments, the task type may include a large model training task, a large model inference task, and a data preprocessing task. The preset splitting rules for large model training may include data parallel splitting, tensor parallel splitting, or pipeline parallel splitting. The preset splitting rules for large model inference tasks may include batch processing parallel splitting. The data preprocessing task may include data sharding parallel splitting.

[0087] In some embodiments, large-scale model training tasks refer to computational tasks involving training large-scale neural network models, requiring massive computing resources and data support. Data parallelism involves splitting the training dataset into multiple data shards, with different computing nodes simultaneously using the same model parameters to train different data shards, and finally aggregating gradients to update the model. Tensor parallelism involves splitting the tensor parameters of the large model across multiple computing nodes, with each node storing only a subset of parameters and performing the corresponding computations; this is suitable for ultra-large-scale parameter models. Pipeline parallelism involves splitting the model's network layers into multiple stages, with different computing nodes responsible for different stages of network layer computation, executed sequentially in a pipeline, improving computational continuity.

[0088] This application's embodiments employ differentiated splitting rules for different features in large model training tasks, maximizing adaptability to task requirements and resource capabilities. Data parallelism fully utilizes massive amounts of data to improve training speed; tensor parallelism solves the problem that a single node cannot handle ultra-large parameter models; pipeline parallelism optimizes the computation process of multi-layer models and reduces node waiting time. Each of these three methods has its own focus, achieving efficient splitting of large model training tasks.

[0089] In some embodiments, large model inference tasks refer to tasks that use a trained large model to perform calculations such as prediction and generation on input data, requiring low latency and high throughput. Batch processing parallelism involves packaging multiple inference requests into a batch and distributing them to multiple computing nodes for parallel processing, or processing multiple requests within a batch in parallel on a single node, thereby improving inference throughput.

[0090] This application's embodiments employ a batch processing parallel approach to split large model inference tasks, effectively improving inference throughput and reducing the average computational cost per request. Simultaneously, by setting a reasonable batch size, a balance can be struck between throughput and latency, satisfying both the high-efficiency processing requirements of offline tasks and the low-latency requirements of real-time tasks, thereby improving the resource utilization efficiency of inference tasks.

[0091] In some embodiments, data preprocessing tasks refer to computational tasks performed on raw data before model training or inference, such as cleaning, transformation, normalization, and feature extraction. Data sharding in parallel processing involves dividing massive amounts of raw data into multiple shards based on data volume or characteristics, and distributing them to multiple computing nodes for parallel processing, thus shortening the overall preprocessing time.

[0092] This application's embodiments employ a data sharding parallel approach to decompose data preprocessing tasks. This fully leverages the parallel computing capabilities of multiple nodes, breaking down massive data preprocessing tasks into multiple sub-tasks for parallel execution, significantly reducing the overall preprocessing time. Furthermore, targeted splitting based on data type and operation type improves the accuracy of preprocessing operations, providing high-quality data support for subsequent model training or inference tasks.

[0093] In some embodiments, the preset conditions may include: the computing resource requirements of the subtask exceed the preset computing resource limit of a single node, the end-to-end execution time of the subtask exceeds the preset time limit, or the business characteristics of the subtask can be split and executed in parallel.

[0094] This application uses three conditions—that the subtask's computing resource requirements exceed the preset upper limit of a single node's computing resources, that the end-to-end execution time exceeds a preset time limit, and that the business characteristics are divisible and can be executed in parallel—as the criteria for splitting computing power tasks. This comprehensively covers the three core requirements of resource adaptation, time limit compliance, and efficiency improvement in task execution. It ensures the feasibility of subtasks in terms of resources and time, fully explores the parallelization potential of tasks, realizes the distributed and efficient utilization of computing power resources, and significantly improves the scientific nature of overall computing power scheduling and the efficiency and stability of task execution.

[0095] In some embodiments, for tasks that need to be split, based on the task's DAG graph, a comprehensive evaluation is conducted on communication overhead between subtasks, heterogeneity of resource requirements, differences in task node computation types, and overall task SLA constraints to determine the optimal splitting strategy and splitting location, achieving efficient task execution and optimized resource allocation. Specific factors considered in the comprehensive evaluation include: The principle of minimizing communication overhead: Based on the data communication volume between subtask nodes in the task DAG, edge weights are constructed. The minimum cut algorithm in graph theory is used to find the task splitting point with the minimum cross-node data communication volume, thereby reducing the bandwidth requirements and latency of cross-node communication. Resource demand heterogeneity principle: For adjacent subtask nodes in the task DAG, analyze the resource demand type of each node, and calculate the Euclidean distance between the resource demand profiles of different nodes to split nodes with significantly different resource demand types into different resource pools as much as possible, thereby optimizing the overall utilization efficiency of the resource pool. Task node computation type difference analysis: When making task splitting decisions, the differences in computation types of different nodes in the task DAG are comprehensively considered, such as computation-intensive (e.g., model training), communication-intensive (e.g., parallel model computing), storage and I / O-intensive (e.g., data processing, model saving) computation characteristics, to ensure that the split subtasks can efficiently match the computing capabilities of the corresponding resource pools. Task end-to-end SLA constraint analysis: Based on the overall end-to-end SLA requirements of the task, the task execution time of each potential splitting scheme is estimated. The execution efficiency and latency of the task are predicted by the computational load and resource matching degree of the task nodes, ensuring that the final splitting scheme meets the task's SLA execution time limit requirements.

[0096] In some embodiments, such as Figure 3 As shown, the system receives predicted load assessment results from multiple candidate resource pool systems, and determines the target resource pool system for each sub-computing task based on the predicted load assessment results and the sub-computing task. This may include steps S301 to S304.

[0097] S301, the scheduling device obtains multiple candidate resource allocation schemes, and each sub-computing task of each candidate resource allocation scheme corresponds to a candidate resource pool system.

[0098] Among them, the candidate resource allocation scheme is the complete allocation combination formed after all sub-computing tasks are matched with the candidate resource pool.

[0099] In some embodiments, the scheduling device aggregates the candidate resource pool lists of all sub-computing tasks, and matches a candidate resource pool to each sub-computing task through a combination and arrangement method, generating multiple sets of non-overlapping candidate resource allocation schemes. Each scheme must cover all sub-computing tasks, and each sub-task corresponds to only one candidate resource pool.

[0100] This application embodiment matches candidate resource pools for all sub-computing tasks and generates multiple complete candidate resource allocation schemes by scheduling equipment, which fully covers the matching possibilities between sub-tasks and resource pools, providing a complete scheme basis for subsequent multi-objective comprehensive evaluation and global optimal decision-making, and avoiding the problem of local optima caused by scheme omissions.

[0101] S302, the scheduling device determines the comprehensive evaluation value of multiple target indicators for each candidate resource allocation scheme based on the predicted load evaluation results of the corresponding sub-computing power tasks executed by each candidate resource pool system.

[0102] In some embodiments, the scheduling device can assign weights to each target indicator, calculate the individual target indicator values ​​of each subtask corresponding to the resource pool in each candidate scheme based on the predicted load evaluation results of the candidate resource pool, and then calculate the comprehensive evaluation value of the scheme through a weighted summation formula.

[0103] This application embodiment calculates a comprehensive evaluation value by weighting multiple objective indicators, transforming scattered performance, cost, energy consumption and other indicators into a quantifiable single score, realizing a unified evaluation of objectives in different dimensions, and making the comparison of candidate solutions more intuitive and objective.

[0104] In some embodiments, the weights of indicators can be dynamically adjusted according to the needs of different business scenarios. This makes the evaluation criteria more aligned with actual business requirements and improves the scenario adaptability of the scheduling scheme.

[0105] S303, the scheduling device selects the candidate resource allocation scheme with the smallest comprehensive evaluation value among all candidate resource allocation schemes as the target resource allocation scheme.

[0106] In some embodiments, the scheduling device sorts the comprehensive evaluation values ​​of all candidate resource allocation schemes and selects the scheme with the smallest comprehensive evaluation value as the target resource allocation scheme. This scheme is the optimal scheme among all feasible schemes in terms of target indicators such as cost, energy consumption, performance, and resource balance.

[0107] This application embodiment ensures the global optimality of scheduling decisions by selecting the scheme with the smallest comprehensive evaluation value, taking into account the balance of multiple objective dimensions such as cost, energy consumption, and performance, and avoiding the situation where one objective is optimal while other objectives are compromised.

[0108] S304, The scheduling device determines the candidate resource pool system corresponding to each sub-computing task in the target resource allocation scheme as the target resource pool system for the sub-computing task.

[0109] This application embodiment first generates a set of possible candidate solutions covering all sub-tasks and candidate resource pools by scheduling equipment, then calculates the comprehensive evaluation value of each solution based on multi-objective indicators, selects the solution with the smallest comprehensive evaluation value as the target solution, and finally clarifies the target resource pool corresponding to each sub-task. This achieves global multi-objective optimal decision-making for computing power scheduling, which not only fully covers potential matching possibilities, but also ensures the objectivity and optimality of the solution through quantitative evaluation, taking into account the needs of multiple objective indicators, and effectively improving the utilization efficiency of computing power resources and the stability of task execution.

[0110] In some embodiments, the scheduling device selects the candidate resource allocation scheme with the smallest comprehensive evaluation value among all candidate resource allocation schemes as the target resource allocation scheme, which may include: Define the multi-objective optimization function as follows: in, The economic cost of using resource pool j for subtask i. The power consumption cost of using resource pool j for subtask i reflects the energy efficiency of the resource scheme; the lower the better. The heat load trend of the resource pool j scheme for subtask i affects the cost. The performance cost of using resource pool j for subtask i. The degree of uneven distribution of overall computing power utilization caused by using resource pool j scheme for subtask i. This is an incentive indicator for the use of domestically produced resources. It has a positive incentive effect when selecting domestically produced chips or domestic resource pools. Weighting factor. This is used to adjust the importance of different indicators, and to adjust them according to the overall optimization goal.

[0111] Multi-objective optimization decisions are made by matching resource solutions in a step-by-step manner based on predefined task priorities. Specifically, during decision-making, a comprehensive cost evaluation (including performance, economic cost, energy consumption, thermal load, resource balance, and domestic content indicators) is performed on candidate resource solutions according to the multi-objective optimization objective function. A multi-objective genetic algorithm is then used to quickly determine the optimal resource allocation scheme for that priority level of task. Once the task at that priority level is determined, the scheduling center updates the load prediction status of the resource pool and uses this status as the base load status for optimizing the next priority level of task. By executing this process step-by-step, the resource schemes for all tasks are ultimately determined, completing the overall global optimization decision.

[0112] In some embodiments, employing a multi-objective genetic algorithm to quickly determine the optimal resource allocation scheme for this level of task may include: A group of initial individuals is randomly or heuristically selected from the global combinatorial decision space, with each individual representing a defined multi-task resource combination. The overall fitness of each combination is evaluated according to a defined multi-objective optimization function, while the performance of the combination is assessed across various objective dimensions to ensure the diversity and effectiveness of Pareto front solutions.

[0113] Based on the fitness scores of the combination schemes, combinations with higher fitness are selected from the current population to enter the next generation, ensuring the continued preservation of high-quality combinations. An exchange is then performed between two selected parent combinations, randomly selecting a subset of tasks and exchanging the resource schemes for these tasks to generate entirely new offspring combinations with mixed characteristics.

[0114] In the generated progeny combinations, the resource configuration of a task is modified randomly or based on certain heuristic rules, thereby introducing new combination variations, enhancing population diversity, and preventing the algorithm from getting trapped in local optima too early. The selection, crossover, and mutation operations are repeated. After each iteration, the fitness of all combinations is re-evaluated, and iteration continues until any of the following termination conditions are met: reaching the preset maximum number of iterations (e.g., 50-100 generations); the Pareto front solution set is stable, with no significant changes in consecutive iterations (i.e., algorithm convergence); finally, the algorithm outputs a set (5-10) of Pareto optimal combinations, which provide different optimal trade-offs among performance, cost, energy consumption, and other multi-dimensional objectives.

[0115] After obtaining multiple Pareto optimal combinations, a single optimal combination is selected using clearly defined objective rules. By assigning explicit secondary weights to each objective of the multiple Pareto optimal combinations, a comprehensive score is calculated, and the single optimal combination is determined after ranking. Once the overall combination optimization decision for a batch of tasks is completed and a single optimal combination has been selected, the scheduling center updates the resource load status of each resource pool according to this combination, and uses the updated resource pool status as the initial load status basis for the optimization decision of the next batch of tasks. This optimization process is executed batch by batch to continuously ensure the overall global optimization effect of task scheduling and resource allocation.

[0116] In some embodiments, after the scheduling center quickly finds the overall optimal resource allocation combination scheme based on the above optimization algorithm, it generates a contract (i.e., scheduling information) containing the corresponding negotiation results and terms for each successful "task-resource scheme". This contract becomes effective after being digitally signed by three parties. According to the resource scheme in the contract, the scheduling center automatically generates clear resource scheduling execution instructions and issues them to the corresponding resource pool to complete the specific task deployment and operation.

[0117] In this embodiment, the task requester clarifies its specific computing power requirements through a dynamic contract, including business metrics, service level agreements (SLAs), budget ranges, and task preferences. The resource pool provider proactively offers corresponding resource solutions based on its current resource status and clarifies the resource's adaptability, including the types of tasks the resource is suitable for and its load capacity. The scheduling device monitors the task execution status and resource load level in real time based on the task requirements and resource supply conditions clearly agreed upon by both parties in the contract. It also dynamically coordinates and adjusts the matching relationship between tasks and resources, taking into account both the local needs of individual tasks and the optimization goals of global computing power resources. This effectively addresses real-time changes in task requirements and dynamic fluctuations in resource status, achieving long-term stability and continuous improvement in overall resource utilization efficiency.

[0118] In some embodiments, the core fields of a ternary dynamic contract are shown in Table 1.

[0119] Table 1. Fields of the Ternary Dynamic Contract In some embodiments, such as Figure 4 As shown, the method may further include: S401 to S405.

[0120] S401, the scheduling device obtains the execution status data of each target resource pool system after executing the corresponding sub-computing task in real time.

[0121] Among them, the execution status data of the sub-computing task is the real-time monitoring data generated by the sub-task during its operation in the target resource pool.

[0122] In some embodiments, the scheduling device can collect core status data in real time during the execution of sub-computing tasks through a monitoring agent deployed in the target resource pool. This data may include task performance data, resource load data, and infrastructure data, and the collected data is transmitted to the scheduling device in a standardized format and stored in a database for subsequent analysis.

[0123] In some embodiments, the monitoring metrics for execution status data include, but are not limited to: task performance metrics monitoring: real-time monitoring of key performance metrics such as task throughput (TPS), response latency, and real-time progress percentage; resource load metrics monitoring: real-time monitoring of the utilization of CPU, GPU, and NPU resources used by task execution; monitoring of infrastructure load such as changes in overall power consumption of the resource pool and temperature trends in the data center generated by task execution; and monitoring of the overall stability of task execution in the resource pool, such as abnormal fluctuations or sudden events.

[0124] This application embodiment enables scheduling equipment to promptly grasp the actual situation of task operation and resource load by acquiring execution status data in real time, providing data support for discovering execution deviations and triggering optimization actions.

[0125] S402, the scheduling device determines the deviation value based on the execution status data of the target resource pool system and the preset execution indicators of the corresponding sub-computing power tasks executed by the target resource pool system. When the deviation value exceeds the set threshold, the device updates the information of the corresponding sub-computing power tasks based on the deviation value, and determines the new candidate resource pool system corresponding to each sub-computing power task based on the task carrying willingness score, resource status information, and multiple updated sub-computing power tasks of multiple resource pool systems.

[0126] The preset execution metric is the subtask execution target value agreed upon before scheduling. The deviation value is the difference between the execution status data and the preset execution metric. The set threshold is the critical value for determining whether the deviation needs to trigger optimization.

[0127] In some embodiments, the scheduling device first compares the collected execution status data with preset execution indicators and calculates the degree of deviation of each indicator; then it determines whether the deviation value exceeds the set threshold. If it does, the task execution is determined to be abnormal; then it updates the sub-computing task information, such as resource requirements and execution priority, according to the cause of the deviation; finally, it calls back the task carrying willingness score and resource status information of the resource pool and selects a new candidate resource pool system according to the initial matching rules.

[0128] In some embodiments, the deviation indicators analyzed may include: task performance commitment indicator analysis and comparison: automatically analyzing the deviation between the actual task performance (TPS, response latency, task completion time) and the contractually agreed benchmark value, and triggering real-time performance indicator anomaly warnings based on a preset performance indicator deviation threshold (e.g., ±5%); resource load commitment indicator analysis and comparison: automatically comparing the actual resource utilization rate of the resource pool (e.g., CPU / GPU load, power consumption, thermal load) with the load prediction indicators specified in the contract, and identifying potential risks of resource load exceeding the commitment range in real-time based on a preset load change deviation threshold (e.g., ±10%).

[0129] The embodiments of this application realize the adaptive adjustment capability of the scheduling system. By judging the deviation value, execution anomalies can be detected in a timely manner to prevent tasks from continuing to run in a state that deviates from the target. At the same time, the sub-task information is updated and the candidate resource pool is re-selected based on the cause of the deviation to ensure the targeting of optimization actions and improve the flexibility and effectiveness of scheduling decisions.

[0130] S403, the scheduling device generates a new scheduling protocol based on the information of each sub-computing task, and sends the new scheduling protocol of the sub-computing task to the new candidate resource pool system corresponding to each sub-computing task.

[0131] The new scheduling protocol consists of scheduling instructions generated based on the updated sub-computing task information. The new candidate resource pool system comprises selected resource pools capable of supporting the updated sub-tasks.

[0132] This application embodiment generates a new scheduling protocol based on the updated sub-computing power task information through a scheduling device and sends it to a new candidate resource pool system. This achieves accurate transmission of optimization requirements, enabling the candidate resource pool to perform load assessment based on the latest task requirements, thus ensuring the pertinence and effectiveness of subsequent optimization decisions.

[0133] S404, the scheduling device receives new predicted load assessment results from multiple new candidate resource pool systems, and determines a new target resource pool system for each sub-computing task based on the new predicted load assessment results and the sub-computing task.

[0134] In some embodiments, after receiving the new scheduling protocol, the new candidate resource pool system, in conjunction with its own real-time resource status, evaluates the impact of executing the updated subtasks using a load prediction model, generates new predicted load evaluation results, and feeds them back to the scheduling device. The scheduling device summarizes the evaluation results of all new candidate resource pools and determines the comprehensive evaluation value for each resource pool, selecting the resource pool with the smallest comprehensive evaluation value as the new target resource pool system.

[0135] S405, the scheduling device sends the corresponding sub-computing task to the new target resource pool system for the new target resource pool system to execute the sub-computing task.

[0136] This application embodiment obtains the subtask execution status data of the target resource pool in real time through the scheduling device, calculates the deviation value between the execution status and the preset index, updates the subtask information and re-selects new candidate resource pools when the deviation exceeds the threshold, generates a new scheduling protocol and obtains new load assessment results, and finally determines the new target resource pool and issues the task for execution. This realizes dynamic optimization of the entire process of computing power scheduling, breaks the limitations of traditional static scheduling, and can correct execution deviations in a timely manner and adapt to dynamic changes in resources.

[0137] In some embodiments, the scheduling center can also continuously perform adaptive learning and optimization. This involves automatically tracking and recording all historical scheduling data, including successful resource scheduling schemes, actual execution performance data, and various anomalies or deviations, to accumulate continuous data. Regular evaluation and analysis of the overall scheduling performance are then conducted to gradually optimize the decision-making mechanism. The specific process includes: From a holistic perspective, the scheduling center selects several clear and practically meaningful evaluation indicators to periodically assess the actual effectiveness of resource scheduling. These evaluation indicators include, but are not limited to: Overall resource utilization efficiency: average CPU, GPU, and NPU utilization rates (target: ≥70%); resource pool idle rate or inefficient utilization rate (target: ≤10%); Overall task service quality (SLA compliance): overall task SLA compliance rate (target: ≥95%); overall task average response latency (target: ≤100ms); Economic cost control: deviation rate between resource operating costs and expected costs (target: within ±5%); Energy consumption optimization effect: deviation between overall actual energy consumption and expected target energy consumption (target: within ±5%); Resource load balancing: standard deviation or coefficient of variation of load for each resource pool (target: standard deviation ≤10%); Effectiveness of domestic resource utilization: percentage of domestic resources used (target: ≥50%).

[0138] The dispatch center conducts automatic analysis of the overall effectiveness of resource scheduling in each evaluation cycle (e.g., weekly or monthly), collecting data on task execution, resource utilization efficiency, cost and energy consumption, and resource balance. It employs a comprehensive approach, combining longitudinal trend assessment (comparing trends with historical data) and horizontal target gap assessment (analyzing gaps with clearly defined expected targets), to provide a clear quantitative evaluation of the overall resource scheduling effectiveness. The center compares the trends of each evaluation indicator (e.g., resource utilization rate, SLA compliance rate, energy consumption indicators) with those of the previous cycle or historical data. This trend comparison clearly identifies the continuous improvement or degradation of the overall resource scheduling effectiveness. Expected target values ​​are clearly set for each indicator (e.g., resource utilization rate ≥70%, SLA compliance rate ≥95%), and the system automatically compares the deviations between the actual monitored indicator values ​​and the defined expected targets for the current cycle. Based on the degree of deviation, the center analyzes the actual achievement of the overall scheduling effect and identifies indicators with significant gaps.

[0139] The system automatically generates a clear overall optimization goal evaluation report, analyzing the optimization results and obvious shortcomings of the current decision-making rules. Based on the problems and analysis conclusions identified in the overall evaluation report, the system automatically or manually optimizes and adjusts the decision-making rules: adjusting the weights of various indicators in the multi-objective optimization objective function (such as performance weight, cost weight, energy consumption weight, and resource balance weight), including adjusting the specific trigger thresholds of proactive optimization and risk triggering rules; optimizing the weight coefficients of the resource pool task carrying willingness scoring model; and optimizing the task priority weight setting rules to better match actual business needs.

[0140] In some embodiments, the resource pool system determines task load willingness scores for different types of computing power tasks based on real-time operational data and resource operation trends, which may include: The resource pool system determines the real-time load impact factor based on real-time operating data and preset sensitivity weights for different types of computing power tasks. The preset sensitivity weights are the impact weights set for various operating indicators for different types of computing power tasks, reflecting the sensitivity of the indicators to task load. The real-time load impact factor is a quantitative value calculated by combining real-time operating data and sensitivity weights, reflecting the degree of impact of the current load on task load. The resource pool system uses resource operation trends and real-time load impact factors to determine the comprehensive load impact factor; the comprehensive load impact factor is a quantitative value that integrates real-time load impact factors and resource operation trends, which reflects both the current load status and the impact of future trends. The resource pool system determines the task carrying willingness score for different types of computing power tasks based on the preset basic adaptability score and comprehensive load impact factor. The preset basic adaptability score is the initial adaptability score set for different types of tasks based on the hardware characteristics of the resource pool and historical carrying data.

[0141] This application embodiment first calculates the real-time load impact factor by combining real-time operating data and the sensitivity weights of different task types in the resource pool system, then determines the comprehensive load impact factor by integrating resource operation trends, and finally obtains the task carrying willingness score based on the basic adaptability score and the comprehensive load impact factor. This achieves a comprehensive quantitative evaluation of the task carrying capacity of the resource pool from three dimensions: real-time status, future trends, and inherent adaptability. It avoids the short-sightedness of static evaluation and takes into account the characteristic differences of different task types, making the scoring results accurate, objective, and forward-looking. This provides a high-quality decision-making basis for the global optimization and matching of the scheduling center, and significantly improves the adaptability and overall efficiency of computing power scheduling.

[0142] In some embodiments, the formula for calculating the load impact factor can be: in, Weights for major indicator categories; The sensitivity weight of the indicator in a specific task type; These are the normalized values ​​for each indicator.

[0143] In some embodiments, the load impact factor is adjusted based on future trend forecasts, and the calculation formula can be: in, The load impact factor is calculated based on the current real-time monitoring indicators; The trend direction factor for the i-th predictive indicator is set to positive (+1) for an upward trend, 0 for a stable trend, and negative (-1) for a downward trend. Adjust the weight of the i-th predictive indicator for the trend of a specific task type. Map the above influencing factors to load adjustment factors (between [0,1]). When the load impact is large, the adjustment factor is small, and the willingness decreases; when the load impact is small, the adjustment factor is large, and the willingness increases.

[0144] In some embodiments, the final task load willingness score is calculated by combining the basic adaptability score and the load adjustment factor. The calculation formula can be: in, Score the willingness to undertake the task. Based on the basic fitness score, This is the load adjustment factor.

[0145] In some embodiments, basic adaptability scores for various typical tasks can be preset based on the resource pool's own resource characteristics and historical performance. As shown in Table 2, these may include high-load tasks, high-SLA tasks, and non-sensitive tasks. The weights shown in Table 2 are only for illustrating the implementation examples of this application and do not constitute a limitation of this application. In actual implementation, they can be flexibly adjusted according to specific scenarios.

[0146] Table 2 Task-Based Adaptability Score Table In some embodiments, weights of index categories are defined for different target indicators, as shown in Table 3. The weights shown in Table 3 are only used to illustrate the implementation examples of this application and do not constitute a limitation on the scope of the claims of this application. They can be flexibly adjusted according to specific scenarios in actual implementation.

[0147] Table 3 Target Indicator Weights In some embodiments, sensitivity weights for various types of loads are defined for different task types to reflect the degree of influence of the load on the intention, as shown in Table 4. The weights shown in Table 4 are only used to illustrate the implementation examples of this application and do not constitute a limitation of this application. They can be flexibly adjusted according to specific scenarios in actual implementation.

[0148] Table 4. Weighting of Load Indicators for Sensitivity In some embodiments, for different task types, sensitivity weights for trend prediction are defined as shown in Table 5, reflecting the degree of influence of the corresponding load change trend on intention. The weights shown in Table 5 are only used to illustrate the implementation examples of this application and do not constitute a limitation on the scope of the claims of this application. In actual implementation, they can be flexibly adjusted according to specific scenarios.

[0149] Table 5 Sensitivity Weights for Trend Forecasting In some embodiments, the calculated preference score ranges from [0,1], as shown in Table 6. A higher score indicates that the resource pool has a higher preference for this type of task in the current and future period.

[0150] Table 6 Task Bearing Willingness Score Figure 5This application illustrates a multi-target computing power scheduling apparatus 500, which is applied to a scheduling device and may include: The receiving module 501 is used to receive task carrying willingness scores and resource status information of different types of computing power tasks sent by multiple resource pool systems. The task carrying willingness score is determined by the resource pool system based on the real-time operation data and resource operation trend of the resource pool system. The receiving module 501 is also used to receive computing power tasks sent by the task party and to split the computing power tasks into multiple sub-computing power tasks. The matching module 502 is used to match resource pool systems and sub-computing tasks based on the task carrying willingness scores, resource status information and multiple sub-computing tasks of multiple resource pool systems, to obtain the candidate resource pool system corresponding to each sub-computing task. The sending module 503 is used to generate scheduling information based on the information of each sub-computing task and send the scheduling information of the sub-computing task to the candidate resource pool system corresponding to each sub-computing task, so that the candidate resource pool system can determine the predicted load evaluation result of the candidate resource pool system after the execution of the sub-computing task based on the scheduling information and the real-time running data of the candidate resource pool system. The determination module 504 is used to receive the predicted load assessment results sent by multiple candidate resource pool systems, determine the target resource pool system for each sub-computing task based on the predicted load assessment results and sub-computing task, and ensure that the comprehensive evaluation value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than a set threshold. The sending module 503 is also used to send the corresponding sub-computing task to the target resource pool system for the target resource pool system to execute the sub-computing task.

[0151] In some embodiments, the multi-target computing power scheduling apparatus 500 may further include: The splitting module is used to split computing tasks according to preset splitting rules for the task type of computing tasks, and obtain a directed acyclic graph including sub-task nodes and the dependencies between sub-task nodes. The splitting module is also used to split the subtasks when the subtasks corresponding to the subtask nodes meet the preset conditions, so that the subtask nodes in the directed acyclic graph do not meet the preset conditions, resulting in multiple sub-computing tasks.

[0152] In some embodiments, the splitting module is further configured to split the computing task according to a preset splitting rule for the task type of the computing task, to obtain a directed acyclic graph including sub-task nodes and dependencies between sub-task nodes. The task types include large model training tasks, large model inference tasks, and data preprocessing tasks. The preset splitting rule for large model training includes data parallel splitting, tensor parallel splitting, or pipeline parallel splitting. The preset splitting rule for large model inference tasks includes batch processing parallel splitting. The data preprocessing task includes data sharding parallel splitting.

[0153] In some embodiments, the splitting module is further configured to split the subtask if the subtask corresponding to the subtask node meets preset conditions. The preset conditions include: the computing resource requirement of the subtask exceeds the preset computing resource limit of a single node, the end-to-end execution time of the subtask exceeds the preset time limit, or the business characteristics of the subtask can be split and can be executed in parallel.

[0154] In some embodiments, the multi-target computing power scheduling apparatus 500 may further include: The acquisition module is used to acquire multiple candidate resource allocation schemes. Each sub-computing task of each candidate resource allocation scheme corresponds to a candidate resource pool system. The determination module 504 is also used to determine the comprehensive evaluation value of multiple target indicators for each candidate resource allocation scheme based on the predicted load evaluation results of the corresponding sub-computing power tasks executed by each candidate resource pool system. The selection module is used to select the candidate resource allocation scheme with the smallest comprehensive evaluation value among all candidate resource allocation schemes as the target resource allocation scheme. The determination module 504 is also used to determine the candidate resource pool system corresponding to each sub-computing power task in the target resource allocation scheme as the target resource pool system of the sub-computing power task.

[0155] In some embodiments, the acquisition module is further configured to acquire in real time the execution status data of each target resource pool system after executing the corresponding sub-computing power task; The determination module 504 is also used to determine the deviation value based on the execution status data of the target resource pool system and the preset execution indicators of the corresponding sub-computing power tasks executed by the target resource pool system. When the deviation value exceeds the set threshold, the information of the corresponding sub-computing power tasks is updated based on the deviation value. The new candidate resource pool system corresponding to each sub-computing power task is determined based on the task carrying willingness score, resource status information and multiple updated sub-computing power tasks of multiple resource pool systems. The sending module 503 is also used to generate a new scheduling protocol based on the information of each sub-computing task, and send the new scheduling protocol of the sub-computing task to the new candidate resource pool system corresponding to each sub-computing task. The determination module 504 is also used to receive new predicted load assessment results sent by multiple new candidate resource pool systems, and determine a new target resource pool system for each sub-computing task based on the new predicted load assessment results and the sub-computing task. The sending module 503 is also used to send the corresponding sub-computing task to the new target resource pool system for the new target resource pool system to execute the sub-computing task.

[0156] Figure 5 The various modules in the device shown can achieve Figure 2 The various steps involved, and the corresponding technical effects achieved, will not be elaborated upon here for the sake of brevity.

[0157] Figure 6 This application illustrates a multi-target computing power scheduling apparatus 600, applied to a resource pool system. The apparatus may include: The acquisition module 601 is used to acquire real-time operating data, historical operating data and resource status information of the resource pool system; The determination module 602 is used to determine the predicted resource operation trend within a preset time window based on historical operation data; The determination module 602 is also used to determine the task carrying willingness score of different types of computing power tasks based on real-time operating data and resource operation trends; The sending module 603 is used to send task carrying willingness scores and resource status information of different types of computing power tasks to the scheduling device, so that the scheduling device can match the resource pool system and the sub-computing power task according to the task carrying willingness scores, resource status information and multiple sub-computing power tasks of multiple resource pool systems, and obtain the candidate resource pool system corresponding to each sub-computing power task. The receiving module 604 is used to receive the scheduling information of the sub-computing task sent by the scheduling device, and determine the predicted load assessment result of the resource pool system after the execution of the sub-computing task based on the scheduling information and real-time running data. The scheduling information is generated by the scheduling device based on the information of each sub-computing task. The sending module 603 is also used to send the predicted load assessment result to the scheduling device, so that the scheduling device can determine the target resource pool system for each sub-computing task based on the predicted load assessment result and the sub-computing task, and the comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than a set threshold. The receiving module 604 is also used to receive sub-computing tasks sent by the scheduling device and execute the sub-computing tasks.

[0158] In some embodiments, the determining module 602 is further configured to determine the real-time load impact factor based on the real-time running data and the preset index sensitivity weights of different types of computing power tasks. The determination module 602 is also used to determine the comprehensive load impact factor by utilizing resource operation trends and real-time load impact factors; The determination module 602 is also used to determine the task carrying willingness score of different types of computing power tasks based on the preset basic adaptability score and comprehensive load impact factor of different types of computing power tasks.

[0159] Figure 6 The various modules in the device shown can achieve Figure 2 The various steps involved, and the corresponding technical effects achieved, will not be elaborated upon here for the sake of brevity.

[0160] Figure 7 A schematic diagram of the hardware structure of the terminal device provided in an embodiment of this application is shown.

[0161] The terminal device may include a processor 701 and a memory 702 storing computer program instructions.

[0162] Specifically, the processor 701 may include a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0163] Memory 702 may include mass storage for data or instructions. For example, and not limitingly, memory 702 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. In one instance, memory 702 may include removable or non-removable (or fixed) media, or memory 702 may be non-volatile solid-state memory. Memory 702 may be internal or external to the integrated gateway disaster recovery device.

[0164] In one example, memory 702 may include read-only memory (ROM), random access memory (RAM), disk storage media device, optical storage media device, flash memory device, electrical, optical, or other physical / tangible memory storage device. Therefore, typically, memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method of multi-target computing power scheduling according to this disclosure.

[0165] The processor 701 reads and executes computer program instructions stored in the memory 702 to achieve... Figure 1 The multi-objective computing power scheduling method in the illustrated embodiment.

[0166] In one example, the terminal device may also include a communication interface 703 and a bus 704. Wherein, for example... Figure 7 As shown, the processor 701, memory 702, and communication interface 703 are connected through bus 704 and complete communication with each other.

[0167] The communication interface 703 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0168] Bus 704 includes hardware, software, or both, that couples components of an end device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Extended Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), a Hyper Transport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 704 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0169] Furthermore, in conjunction with the multi-target computing power scheduling method in the above embodiments, this application embodiment can provide a computer storage medium for implementation. The computer storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the multi-target computing power scheduling methods in the above embodiments.

[0170] This application also provides a computer program product, including a computer program, which, when executed, implements any of the multi-target computing power scheduling methods described in the above embodiments.

[0171] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0172] The functional blocks shown in the above block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or text segments used to perform the required tasks. Programs or text segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, read-only memory (ROM), flash memory, erasable read-only memory (EROM), floppy disks, compact disc read-only memory (CD-ROM), optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Text segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0173] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0174] The aspects of this disclosure have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0175] The above are merely specific embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for multi-objective computing power scheduling, characterized in that, Applied to scheduling equipment, the method includes: The system receives task carrying willingness scores and resource status information of different types of computing power tasks sent by multiple resource pool systems. The task carrying willingness scores are determined by the resource pool system based on the real-time operation data and resource operation trends of the resource pool system. Receive computing power tasks sent by the task provider, and split the computing power tasks into multiple sub-computing power tasks; Based on the task carrying willingness scores, resource status information, and multiple sub-computing tasks of multiple resource pool systems, the resource pool systems and sub-computing tasks are matched to obtain the candidate resource pool system corresponding to each sub-computing task; Scheduling information is generated based on the information of each sub-computing task, and the scheduling information of the sub-computing task is sent to the candidate resource pool system corresponding to each sub-computing task, so that the candidate resource pool system can determine the predicted load evaluation result of the candidate resource pool system after the execution of the sub-computing task based on the scheduling information and the real-time running data of the candidate resource pool system. Receive predicted load assessment results from multiple candidate resource pool systems, determine the target resource pool system for each sub-computing task based on the predicted load assessment results and the sub-computing task, and ensure that the comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than a set threshold. Send the corresponding sub-computing task to the target resource pool system for the target resource pool system to execute the sub-computing task.

2. The method for multi-objective computing power scheduling according to claim 1, characterized in that, The process of splitting the computing power task into multiple sub-computing power tasks includes: The computing power task is split according to the preset splitting rules of the task type to obtain a directed acyclic graph including sub-task nodes and the dependencies between sub-task nodes. If the subtask corresponding to the subtask node meets the preset conditions, the subtask is split so that none of the subtask nodes in the directed acyclic graph meet the preset conditions, resulting in multiple sub-computing tasks.

3. The method for multi-objective computing power scheduling according to claim 2, characterized in that, The task types include large model training tasks, large model inference tasks, and data preprocessing tasks. The preset splitting rules for the large model training include data parallel splitting, tensor parallel splitting, or pipeline parallel splitting. The preset splitting rules for the large model inference task include batch processing parallel splitting. The data preprocessing task includes data sharding parallel splitting.

4. The method for multi-objective computing power scheduling according to claim 2, characterized in that, The preset conditions include: the computing resource requirements of the subtask exceed the preset computing resource limit of a single node; the end-to-end execution time of the subtask exceeds the preset time limit; or the business characteristics of the subtask can be split and executed in parallel.

5. The method for multi-objective computing power scheduling according to claim 1, characterized in that, The step of receiving predicted load assessment results from multiple candidate resource pool systems, and determining the target resource pool system for each sub-computing task based on the predicted load assessment results and the sub-computing task, includes: Multiple candidate resource allocation schemes are obtained, and each sub-computing task of each candidate resource allocation scheme corresponds to a candidate resource pool system. Based on the predicted load assessment results of the corresponding sub-computing power tasks executed by each candidate resource pool system, a comprehensive evaluation value of multiple target indicators for each candidate resource allocation scheme is determined. The candidate resource allocation scheme with the smallest comprehensive evaluation value among all candidate resource allocation schemes is selected as the target resource allocation scheme. The candidate resource pool system corresponding to each sub-computing power task in the target resource allocation scheme is determined as the target resource pool system of the sub-computing power task.

6. The method for multi-objective computing power scheduling according to claim 1, characterized in that, The method further includes: Real-time acquisition of execution status data after each target resource pool system executes the corresponding sub-computing power task; The deviation value is determined based on the execution status data of the target resource pool system and the preset execution indicators of the corresponding sub-computing power tasks executed by the target resource pool system. When the deviation value exceeds a set threshold, the information of the corresponding sub-computing power tasks is updated based on the deviation value. A new candidate resource pool system corresponding to each sub-computing power task is determined based on the task carrying willingness score, resource status information, and multiple updated sub-computing power tasks of multiple resource pool systems. A new scheduling protocol is generated based on the information of each sub-computing task, and the new scheduling protocol of the sub-computing task is sent to the new candidate resource pool system corresponding to each sub-computing task. Receive new predicted load assessment results from multiple new candidate resource pool systems, and determine a new target resource pool system for each sub-computing task based on the new predicted load assessment results and the sub-computing task. Send the corresponding sub-computing task to the new target resource pool system so that the new target resource pool system can execute the sub-computing task.

7. A method for multi-objective computing power scheduling, characterized in that, Applied to a resource pool system, the method includes: Obtain real-time operating data, historical operating data, and resource status information of the resource pool system; Determine the predicted resource operation trend within a preset time window based on historical operation data; Based on real-time operational data and resource operation trends, determine the task carrying willingness scores for different types of computing power tasks; The task carrying willingness score and resource status information of the different types of computing power tasks are sent to the scheduling device so that the scheduling device can match the resource pool system and the sub-computing power task according to the task carrying willingness score, resource status information and multiple sub-computing power tasks of multiple resource pool systems, and obtain the candidate resource pool system corresponding to each sub-computing power task. The system receives scheduling information for sub-computing tasks sent by the scheduling device, and determines the predicted load assessment result of the resource pool system after the execution of the sub-computing tasks based on the scheduling information and real-time operating data. The scheduling information is generated by the scheduling device based on the information of each sub-computing task. The predicted load assessment result is sent to the scheduling device so that the scheduling device can determine the target resource pool system for each sub-computing task based on the predicted load assessment result and the sub-computing task. The comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than a set threshold. Receive sub-computing tasks sent by the scheduling device and execute the sub-computing tasks.

8. The method for multi-objective computing power scheduling according to claim 7, characterized in that, The process of determining the task carrying willingness score for different types of computing power tasks based on real-time operating data and resource operation trends includes: The real-time load impact factor is determined based on the real-time operating data and the preset sensitivity weights of different types of computing power tasks. The comprehensive load impact factor is determined using the resource operation trend and real-time load impact factor. The task carrying willingness score for different types of computing power tasks is determined based on the preset basic adaptability score for different types of computing power tasks and the comprehensive load impact factor.

9. A device for multi-objective computing power scheduling, characterized in that, Applied to scheduling equipment, the device includes: The receiving module is used to receive task carrying willingness scores and resource status information of different types of computing power tasks sent by multiple resource pool systems. The task carrying willingness scores are determined by the resource pool system based on the real-time operation data and resource operation trends of the resource pool system. The receiving module is also used to receive computing power tasks sent by the task party, and to split the computing power tasks into multiple sub-computing power tasks. The matching module is used to match resource pool systems and sub-computing tasks based on task carrying willingness scores, resource status information, and multiple sub-computing tasks of multiple resource pool systems, to obtain candidate resource pool systems corresponding to each sub-computing task. The sending module is used to generate scheduling information based on the information of each sub-computing task, and send the scheduling information of the sub-computing task to the candidate resource pool system corresponding to each sub-computing task, so that the candidate resource pool system can determine the predicted load evaluation result of the candidate resource pool system after the execution of the sub-computing task based on the scheduling information and the real-time running data of the candidate resource pool system. The determination module is used to receive the predicted load assessment results sent by multiple candidate resource pool systems, determine the target resource pool system for each sub-computing task based on the predicted load assessment results and the sub-computing task, and the comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than a set threshold. The sending module is also used to send the corresponding sub-computing power task to the target resource pool system for the target resource pool system to execute the sub-computing power task.

10. A device for multi-objective computing power scheduling, characterized in that, The device, applied to a resource pool system, includes: The acquisition module is used to acquire real-time operating data, historical operating data, and resource status information of the resource pool system; The determination module is used to determine the predicted resource operation trend within a preset time window based on historical operation data; The determination module is also used to determine the task carrying willingness score of different types of computing power tasks based on real-time operating data and resource operation trends; The sending module is used to send the task carrying willingness score and resource status information of the different types of computing power tasks to the scheduling device, so that the scheduling device can match the resource pool system and the sub-computing power task according to the task carrying willingness score, resource status information and multiple sub-computing power tasks of multiple resource pool systems, and obtain the candidate resource pool system corresponding to each sub-computing power task. The receiving module is used to receive the scheduling information of the sub-computing power tasks sent by the scheduling device, and determine the predicted load assessment result of the resource pool system after the execution of the sub-computing power tasks based on the scheduling information and real-time running data. The scheduling information is generated by the scheduling device based on the information of each sub-computing power task. The sending module is also used to send the predicted load assessment result to the scheduling device, so that the scheduling device can determine the target resource pool system for each sub-computing task based on the predicted load assessment result and the sub-computing task, and the comprehensive assessment value of multiple target indicators of all target resource pool systems after executing the corresponding sub-computing task is less than a set threshold. The receiving module is also used to receive sub-computing tasks sent by the scheduling device and execute the sub-computing tasks.

11. A terminal device, characterized in that, The device includes: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, it implements the multi-target computing power scheduling method as described in any one of claims 1-8.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, implement the multi-target computing power scheduling method as described in any one of claims 1-8.

13. A computer program product, characterized in that, When the instructions in the computer program product are executed by the processor of the electronic device, the electronic device performs the multi-target computing power scheduling method as described in any one of claims 1-8.

Images