Calculating power task scheduling method and device and related equipment

Abstract

The invention provides a computing power task scheduling method and device and related equipment, and relates to the field of computing power scheduling. According to the technical scheme, performance scores corresponding to N artificial intelligence agents are calculated, so that the N artificial intelligence agents are associated with different task types, and the task scheduling efficiency is improved. Under the condition that a target computing power task needing to be processed is received, the most matched target agent is determined in the N artificial intelligence agents according to the task type corresponding to the target computing power task and the required performance score value, and therefore the target computing power task is scheduled to the target agent to be processed. The resource and model allocation condition is comprehensively considered when the computing power task is processed, and the scheduling effect of the computing power task is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computing power scheduling, in particular to a computing power task scheduling method and device and related equipment. BACKGROUND

[0002] At present, computing network integration is realized in a computing platform. The computing network integration refers to deep integration of computing resources (computing) and network resources (network) to realize more efficient resource utilization and management. However, in related technologies, the scheduling of computing power tasks usually depends on preset rules or static resource mapping strategies, and lacks intelligent evaluation mechanism of task complexity in the scheduling process, resulting in unreasonable matching of models and resources, and thus poor task scheduling effect. SUMMARY

[0003] The present application provides a computing power task scheduling method, device and related equipment, which solves the problem of poor task scheduling effect in related technologies.

[0004] To solve the above problems, the present application is implemented as follows:

[0005] In a first aspect, the present application provides a computing power task scheduling method, applied to a computing platform, the computing platform comprising N artificial intelligence agents, and the method comprising:

[0006] calculating performance scores corresponding to the N artificial intelligence agents to obtain N performance score values, the performance score value being used to represent the performance of the corresponding artificial intelligence agent in processing a computing power task, and N being a positive integer greater than 1;

[0007] determining task types of the N artificial intelligence agents in processing tasks based on the N performance score values to obtain N task type labels, the task type label being used to represent the task type that the corresponding artificial intelligence agent can process;

[0008] in the case of receiving a target computing power task, determining a target agent among the N artificial intelligence agents based on a target performance score value and a target task type, the target performance score value being a performance score value required by the target computing power task, the target task type being a task type corresponding to the target computing power task, and the target agent being an artificial intelligence agent in the N artificial intelligence agents, whose corresponding task type label matches the target task type and whose corresponding performance score value is greater than or equal to the target performance score value;

[0009] scheduling the target computing power task to the target agent for processing to obtain a processing result.

[0010] Optionally, the calculating of the performance scores corresponding to the N artificial intelligence agents to obtain the N performance score values comprises:

[0011] For a first artificial intelligence agent, determine a computing power shock value corresponding to the first artificial intelligence agent, a hidden dark resource value corresponding to the first artificial intelligence agent, and a data temperature gradient corresponding to the first artificial intelligence agent, the first artificial intelligence agent being any one of the N artificial intelligence agents, wherein the computing power shock value represents the coupling relationship between resource utilization and power consumption limitation of the first artificial intelligence agent, the hidden dark resource value represents the number of resources included in the reserve resource pool of the first artificial intelligence agent that can be called in an emergency, and the data temperature gradient represents the dynamic evolution trend of the data access heat distribution in different levels of storage media of the first artificial intelligence agent;

[0012] According to the first weight information, the computing power shock value, the hidden dark resource value and the data temperature gradient are weighted and calculated to obtain a performance score value corresponding to the first artificial intelligence agent.

[0013] Optionally, the task type includes a first task, a second task and a third task, the task complexity of the first task is lower than the task complexity of the second task, and the task complexity of the second task is lower than the task complexity of the third task, and the N task types are determined based on the N performance score values to obtain N task type labels, including:

[0014] Based on the performance score value corresponding to the first artificial intelligence agent, a data temperature gradient slope of the first artificial intelligence agent processing the computing power task is determined, the data temperature gradient slope representing the migration speed and intensity of the first artificial intelligence agent processing hot data in unit time;

[0015] In the case where the data temperature gradient slope is less than or equal to a first preset value, the task type label corresponding to the first artificial intelligence agent is determined as the first task;

[0016] In the case where the data temperature gradient slope is greater than the first preset value and less than a second preset value, the task type label corresponding to the first artificial intelligence agent is determined as the second task;

[0017] In the case where the data temperature gradient slope is greater than or equal to the second preset value, the task type label corresponding to the first artificial intelligence agent is determined as the third task;

[0018] Wherein, the first preset value is less than the second preset value.

[0019] Optionally, in the case that the target computing power task is received, the target agent is determined in the N artificial intelligence agents based on the target performance score value and the target task type, including:

[0020] In the case that the target computing power task is received, M second artificial intelligence agents are determined in the N artificial intelligence agents based on the target performance score value and the target task type, the second artificial intelligence agents being the artificial intelligence agents in the N artificial intelligence agents corresponding to the task type label consistent with the target task type and the performance score value greater than or equal to the target performance score value, and the M being a positive integer less than or equal to N;

[0021] The comprehensive score value corresponding to the M second artificial intelligence agents is calculated, to obtain M comprehensive score values corresponding to the M second artificial intelligence agents in one-to-one correspondence, the comprehensive score value being used to represent the matching degree of the corresponding second artificial intelligence agent and the target computing power task;

[0022] Based on the M comprehensive score values, a target agent is determined in the M second artificial intelligence agents, the target agent being the second artificial intelligence agent with the highest comprehensive score value in the M second artificial intelligence agents.

[0023] Optionally, the calculation of the comprehensive score value corresponding to the M second artificial intelligence agents to obtain M comprehensive score values corresponding to the M second artificial intelligence agents in one-to-one correspondence includes:

[0024] The computing resource score value corresponding to the third artificial intelligence agent, the network resource score value corresponding to the third artificial intelligence agent, and the storage resource score value corresponding to the third artificial intelligence agent are calculated, the third artificial intelligence agent being any one of the M second artificial intelligence agents;

[0025] According to the second weight information, the computing resource score value, the network resource score value, and the storage resource score value are weighted and calculated to obtain the comprehensive score value corresponding to the third artificial intelligence agent.

[0026] Optionally, after the target computing power task is scheduled to the target agent for processing to obtain a processing result, the method further includes:

[0027] The difference between the processing result and an expected processing result of the target computing power task is calculated to obtain a difference value;

[0028] Based on the difference value, the target agent is updated to obtain an updated target agent.

[0029] In a second aspect, an embodiment of the present application provides a computing power task scheduling device, applied to a computing power platform, the computing power platform comprising N artificial intelligence agents, and the device comprising:

[0030] a calculating module configured to calculate performance scores corresponding to the N artificial intelligence agents, to obtain N performance score values, the performance score values being used to represent the performance of the corresponding artificial intelligence agents in processing computing power tasks, and N being a positive integer greater than 1;

[0031] a determining module configured to determine, based on the N performance score values, the types of tasks that the N artificial intelligence agents can process, to obtain N task type labels, the task type labels being used to represent the types of tasks that the corresponding artificial intelligence agents can process;

[0032] a matching module configured to, in a case where a target computing power task is received, determine a target agent from the N artificial intelligence agents based on a target performance score value and a target task type, the target performance score value being a performance score value required by the target computing power task, the target task type being a task type corresponding to the target computing power task, and the target agent being an artificial intelligence agent from the N artificial intelligence agents, for which a corresponding task type label matches the target task type and a corresponding performance score value is greater than or equal to the target performance score value;

[0033] a processing module configured to schedule the target computing power task to the target agent for processing, to obtain a processing result.

[0034] In a third aspect, the present application further provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and capable of running on the processor, and the computer program, when executed by the processor, implements the steps in the method according to the first aspect.

[0035] In a fourth aspect, the present application further provides a computer-readable storage medium, having a computer program stored thereon, and the computer program, when executed by a processor, implements the steps in the method according to the first aspect.

[0036] In a fifth aspect, the present application further provides a computer program product, comprising computer instructions, and the computer instructions, when executed by a processor, implement the steps in the method according to the first aspect.

[0037] The application provides a computing power task scheduling method and device and related equipment, applied to a computing power platform, the computing power platform comprising N artificial intelligence agents, the method comprising: calculating performance scores corresponding to the N artificial intelligence agents to obtain N performance score values, the performance score value being used to represent the performance of the corresponding artificial intelligence agent in processing a computing power task, N being a positive integer greater than 1; determining the task types of the N artificial intelligence agents in processing tasks based on the N performance score values to obtain N task type labels, the task type label being used to represent the task type that the corresponding artificial intelligence agent can process; in the case of receiving a target computing power task, determining a target agent in the N artificial intelligence agents based on a target performance score value and a target task type, the target performance score value being a performance score value required by the target computing power task, the target task type being a task type corresponding to the target computing power task, and the target agent being an artificial intelligence agent in the N artificial intelligence agents, the corresponding task type label of which matches the target task type, and the corresponding performance score value of which is greater than or equal to the target performance score value; scheduling the target computing power task to the target agent for processing to obtain a processing result. The technical solution of the application calculates the performance score values corresponding to the N artificial intelligence agents, thereby associating the N artificial intelligence agents with different task types, in the case of receiving a target computing power task that needs to be processed, determines the most matched target agent in the N artificial intelligence agents according to the task type corresponding to the target computing power task and the required performance score value, thereby scheduling the target computing power task to the target agent for processing, and realizes the allocation of resources and models in processing computing power tasks, effectively improving the scheduling effect of computing power tasks. BRIEF DESCRIPTION OF DRAWINGS

[0038] To make the technical solutions of the application clearer, the following will briefly introduce the drawings needed in the description of the application. Obviously, the drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor.

[0039] Figure 1 A flowchart of a computing power task scheduling method provided by an embodiment of the application;

[0040] Figure 2 A system structure diagram of a computing power platform provided by an embodiment of the application;

[0041] Figure 3 A system framework diagram of a computing power platform provided by an embodiment of the application;

[0042] Figure 4A structural schematic diagram of a computing power task scheduling device provided by an embodiment of the present application is provided.

[0043] Figure 5 A structural schematic diagram of an electronic device provided by an embodiment of the present application is provided. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0045] The terms "first", "second", and the like in the embodiments of the present application are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or device including a series of steps or units does not necessarily have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product, or device. In addition, "and / or" is used in the present application to represent at least one of the connected objects, for example, A and / or B and / or C represents 7 cases including A alone, B alone, C alone, A and B both exist, B and C both exist, A and C both exist, and A, B and C all exist.

[0046] Reference Figure 1 , Figure 1 is a flowchart of a computing power task scheduling method provided by an embodiment of the present application. As Figure 1 indicated, the computing power task scheduling method is applied to a computing power platform, the computing power platform includes N artificial intelligence agents, and the computing power task scheduling method can include the following steps:

[0047] Step 101, calculate performance scores corresponding to the N artificial intelligence agents to obtain N performance score values, the performance score values are used to represent the performance of the corresponding artificial intelligence agent in processing computing power tasks, and N is a positive integer greater than 1.

[0048] In the present embodiment, the computing power platform includes N artificial intelligence agents, the artificial intelligence agent is AIAgent, and specifically, it can be a computing power node, through which a computing power task distributed by the computing power platform can be processed. It should be noted that the N artificial intelligence agents can be heterogeneous computing power nodes, that is, the plurality of computing power nodes include different types of computing power resources, and thus the processing capabilities of the computing power nodes for computing power tasks are different.

[0049] The performance score values of the N artificial intelligence agents are obtained by calculating the performance scores of the N artificial intelligence agents respectively, wherein the performance score value is the performance of the corresponding artificial intelligence agent in processing the computing power task. For example, the performance of the artificial intelligence agent in processing the computing power task can be evaluated in multiple aspects, such as processing speed (e.g., the number of tasks processed per second), resource efficiency (e.g., the computing power, memory or time consumed per task), accuracy (e.g., the success rate of completing a task or the quality of the output result), flexibility and adaptability, etc. Finally, the performance score value of each artificial intelligence agent is obtained.

[0050] In step 102, based on the N performance score values, the task type of the N artificial intelligence agents in processing the task is determined, and N task type labels are obtained, wherein the task type label represents the type of task that the corresponding artificial intelligence agent can process.

[0051] In this embodiment, according to the performance score value of each artificial intelligence agent, the type of task that each artificial intelligence agent is suitable for processing is determined, and a task type label is assigned to each artificial intelligence agent, wherein the task type label can be a low complexity task, a medium complexity task or a high complexity task, etc.

[0052] Specifically, by analyzing the association between the performance score and the historical task, it can be inferred which score value corresponds to which task type, and thus the task type label corresponding to each artificial intelligence agent is determined.

[0053] In step 103, when a target computing power task is received, a target agent is determined among the N artificial intelligence agents based on a target performance score value and a target task type, wherein the target performance score value is the performance score value required by the target computing power task, the target task type is the task type corresponding to the target computing power task, and the target agent is the artificial intelligence agent in the N artificial intelligence agents, whose corresponding task type label matches the target task type and whose corresponding performance score value is greater than or equal to the target performance score value.

[0054] In this embodiment, when the computing power platform receives a target computing power task, the target agent that is most suitable for processing the target computing power task is determined among the N artificial intelligence agents according to the target performance score value and the target task type corresponding to the target computing power task, wherein the target agent is the artificial intelligence agent in the N artificial intelligence agents, whose corresponding task type label matches the target task type and whose corresponding performance score value is greater than or equal to the target performance score value, i.e., the target agent needs to meet two conditions at the same time.

[0055] It needs to be explained that if there is no artificial intelligence agent that meets both conditions among the N artificial intelligence agents, the target computing power task can be selected to be returned or an artificial intelligence agent that meets one condition can be selected to process.

[0056] Step 104, scheduling the target computing power task to the target agent for processing to obtain a processing result.

[0057] In the embodiment, the target computing power task is scheduled to the target agent for processing, so as to obtain a processing result of processing the target computing power task. The processing result can be successful processing or processing failure, which can also include processing logs and intermediate files of processing the target computing power task, etc., which are not limited in the embodiment.

[0058] The technical scheme of the application, by calculating the performance score values corresponding to the N artificial intelligence agents, associates the N artificial intelligence agents with different task types. When receiving a target computing power task that needs to be processed, the most matched target agent is determined among the N artificial intelligence agents according to the task type and the required performance score value corresponding to the target computing power task, so as to schedule the target computing power task to the target agent for processing. The allocation of resources and models is comprehensively considered when processing the computing power task, and the scheduling effect of the computing power task is effectively improved.

[0059] In some possible implementation manners, optionally, the performance scores corresponding to the N artificial intelligence agents are calculated to obtain N performance score values, including:

[0060] For the first artificial intelligence agent, the computing power shock value corresponding to the first artificial intelligence agent, the hidden dark resource value corresponding to the first artificial intelligence agent, and the data temperature gradient corresponding to the first artificial intelligence agent are determined. The first artificial intelligence agent is any artificial intelligence agent in the N artificial intelligence agents. The computing power shock value is used to represent the coupling relationship between resource utilization and power consumption limitation of the first artificial intelligence agent. The hidden dark resource value is used to represent the number of resources included in the reserve resource pool that can be called in an emergency of the first artificial intelligence agent. The data temperature gradient is used to represent the dynamic evolution trend of data access heat distribution in different levels of storage media of the first artificial intelligence agent.

[0061] According to the first weight information, the computing power shock value, the hidden dark resource value, and the data temperature gradient are weighted and calculated to obtain the performance score value corresponding to the first artificial intelligence agent.

[0062] In the embodiment, the first artificial intelligence agent is any one of the N artificial intelligence agents, and for each first artificial intelligence agent, the corresponding three-dimensional data is calculated, specifically including:

[0063] The computing power shock value is used to describe the coupling relationship between the fragmentation degree of the GPU resources in the artificial intelligence agent node in the space layer and the availability under the power consumption limitation. The calculation basis is:

[0064] The computing power shock value = GPU inter-chip fragmentation rate / chip computing power dark silicon rate

[0065] Among them, the GPU inter-chip fragmentation rate represents the discontinuity degree of the GPU memory idle distribution in unit time, and the computing power dark silicon refers to the part of the resources in the chip that cannot continuously run at the nominal frequency due to the power wall limitation, which is equivalent to the implicit computing power loss of “logically existing and physically limited”.

[0066] The hidden dark resource value: Some machines reserve a part of the reserve resource pool that can only be called in emergency, including part of the CPU threads and GPU computing units / storage units / network connection points.

[0067] Such resources do not participate in regular resource scheduling and are in a “invisible” state at ordinary times;

[0068] Its availability is only activated under certain conditions (such as high-priority tasks, master artificial intelligence agent triggering scheduling instructions).

[0069] Data temperature gradient: used to describe the dynamic evolution trend of the data access heat distribution in different levels of storage media. This index reflects the moving frequency of hot data between cache, main memory and remote storage; the non-uniformity degree of access frequency between different “computing devices”.

[0070] In some embodiments, as shown in Table 1, Table 1 is a performance type table of different dimensions corresponding to the first artificial intelligence agent.

[0071] Table 1 Field Name Type Description gent_ID String Node unique identification Timestamp DateTime Status reporting timestamp Health_Status Enum Node health status marker Storage_Shake_Index Float Reflects GPU fragmentation and the proportion of computing dark silicon (the higher the more unstable) Hidden_Reserve_Flag Boolean Whether there is callable hidden resource (True means available) Data_Temp_Gradient Float Data temperature gradient, thermal data distribution difference

[0072] After obtaining the computing power shock value, the hidden dark resource value and the data temperature gradient, the computing power shock value, the hidden dark resource value and the data temperature gradient are weighted calculated according to the preset first weight information, so as to obtain the performance score value corresponding to the first artificial intelligence agent. It needs to be explained that the first weight information can be adaptively set according to the actual situation, which is not specifically limited in the embodiment.

[0073] Optionally, the task types include a first task, a second task and a third task, the first task has a lower task complexity than the second task, and the second task has a lower task complexity than the third task, and determining the task types of the N artificial intelligence agents based on the N performance score values to obtain N task type labels includes:

[0074] Based on the performance score value corresponding to the first artificial intelligence agent, a data temperature gradient slope of the first artificial intelligence agent processing the computing power task is determined, and the data temperature gradient slope is used to represent the migration speed of the first artificial intelligence agent processing hot data in unit time and the intensity of processing hot data.

[0075] In a case where the data temperature gradient slope is less than or equal to a first preset value, the task type label corresponding to the first artificial intelligence agent is determined as the first task.

[0076] In a case where the data temperature gradient slope is greater than the first preset value and less than a second preset value, the task type label corresponding to the first artificial intelligence agent is determined as the second task.

[0077] In a case where the data temperature gradient slope is greater than or equal to the second preset value, the task type label corresponding to the first artificial intelligence agent is determined as the third task.

[0078] Wherein, the first preset value is less than the second preset value.

[0079] In this embodiment, the task types include the first task, the second task and the third task, the first task has a lower task complexity than the second task, and the second task has a lower task complexity than the third task, and exemplarily, the first task has a small data amount and a short inference path, and is suitable for being processed by an edge node quickly; the second task has a medium resource requirement, and can be deployed on a medium artificial intelligence agent node with a standard GPU; the third task needs a large model for processing, and has a relatively loose delay requirement.

[0080] The third task needs an artificial intelligence agent with high performance to process, and the first task can be processed by an artificial intelligence agent with low performance.

[0081] In this embodiment, according to the performance score value corresponding to each first artificial intelligence agent, the data temperature gradient slope of the first artificial intelligence agent processing computing power task is determined, wherein the data temperature gradient slope (denoted as ∇T_d) measures the speed and intensity of hot data migration between different computing layers per unit time. Its essence reflects the requirements of the task for data locality optimization, cache dependency and remote access tolerance. It needs to be explained that the data temperature gradient slope in this embodiment can be calculated according to the data temperature gradient included in the performance score value, that is, the slope of the data temperature gradient, that is, the dynamic change trend of the hotness distribution as an important parameter to measure the task access characteristics, and combined with the required data access mode of the task, the task is divided into three categories: first task (lightweight), second task (medium complexity) and third task (high complexity).

[0082] In this embodiment, the first preset value is less than the second preset value, and the second preset value is less than the third preset value. Specifically, the first preset value is 0.2 and the second preset value is 0.6. Specifically, the first task (lightweight): gradient slope range: ∇T_d<0.2; feature description: data hotness is concentrated, and high-frequency access data is mostly resident in local cache layer; second task (medium complexity): gradient slope range: 0.2<∇T_d<0.6; feature description: data cold and hot level distribution is relatively balanced, and there is certain main memory and remote IO migration; third task (high complexity): gradient slope range: ∇T_d>0.6; feature description: data migration is frequent, remote access is intensive, and hotness changes dramatically.

[0083] As shown in Table 2, Table 2 is a mapping relationship between the number of multiple different Agnet and the computing power shock value, hidden dark resources, and task type in this embodiment:

[0084] Table 2 Task Type Computing shock value Whether to enable hidden dark resource Recommended artificial intelligence agent number First task <0.3 (low) No L1, L2 Second task 0.3~0.6 (moderate) Dynamically determined according to load M1, M2, L2 Third task >0.6 (high) Yes (priority) H1, H2, M2

[0085] As shown in Table 3, the system pre-deploys the following representative expert models and binds them to artificial intelligence agent nodes with corresponding computing capabilities:

[0086] Table 3 Artificial intelligence agent number Model category Use category Model parameter size L1 DeepSeek-V2Lite Text classification, keyword extraction 1.2B L2 MobileNetV3 Image classification, target detection 5M–15M M1 LLaMA-2 Semantic matching, simple text generation, code generation 7B M2 ViTBase Image semantic understanding 85M H1 LLaMA-2 Open domain question answering, content generation 13B H2 DeepSeek-V3 Multi-turn dialogue, structured generation 21B

[0087] In this embodiment, the performance score value corresponding to the artificial intelligence agent is used to associate each artificial intelligence agent with the task type, so as to facilitate the matching of computing power tasks and artificial intelligence agents, and improve the resource utilization rate.

[0088] Optionally, in the case where the target computing power task is received, the target agent is determined in the N artificial intelligence agents based on the target performance score value and the target task type, including:

[0089] In the case of receiving a target computing power task, M second artificial intelligence agents are determined in the N artificial intelligence agents based on a target performance score value and a target task type, the second artificial intelligence agents being the artificial intelligence agents in the N artificial intelligence agents corresponding to the task type label meeting the target task type and the performance score value being greater than or equal to the target performance score value, the M being a positive integer less than or equal to N;

[0090] A comprehensive score value corresponding to the M second artificial intelligence agents is calculated, to obtain M comprehensive score values corresponding to the M second artificial intelligence agents in one-to-one correspondence, the comprehensive score value being used to represent the matching degree of the corresponding second artificial intelligence agent and the target computing power task;

[0091] Based on the M comprehensive score values, a target agent is determined in the M second artificial intelligence agents, the target agent being the second artificial intelligence agent in the M second artificial intelligence agents corresponding to the highest comprehensive score value.

[0092] In the embodiment, M second artificial intelligence agents are selected from N artificial intelligence agents, which need to meet two conditions: their task type labels must meet the target task type received, and their performance score values must be greater than or equal to the target performance score value. For the M second artificial intelligence agents selected, the comprehensive score values of the second artificial intelligence agents need to be calculated. For example, the performance score, the task processing efficiency, the historical performance on processing similar tasks and the like can be calculated. Thus, according to the comprehensive score value corresponding to each second artificial intelligence agent, the second artificial intelligence agent with the highest score value is selected as the target agent.

[0093] As Figure 2 shown in the figure, in the embodiment, a parallel decision system composed of a plurality of function-oriented expert artificial intelligence agents can be set, and a core scheduling strategy generation module of cross-domain resource intelligent scheduling is constructed. The module coordinates the optimization targets of a plurality of target functions through a message exchange mechanism supported by the MCP protocol, and realizes the final scheduling strategy fusion and task distribution by a coordination artificial intelligence agent.

[0094] The system divides the key optimization targets in the scheduling process into the following three categories, and independently deploys function-specific expert artificial intelligence agents for each category of targets, as shown in Table 4:

[0095] Table 4 Expert artificial intelligence agent category Core responsibility description End intelligent agent Responsible for managing all inference task path input data acquisition, preprocessing, task number binding, output result reorganization and feedback transmission. HPC collaborative intelligent agent Deployed in cloud-native capable clusters or edge cloud collaborative nodes, dedicated to scheduling high complexity model inference tasks and mixed load computing scenarios. CI / CD linkage intelligent agent Embedded in the CI (Continuous Integration) and CD (Continuous Deployment) link in the model development life cycle, sensing new model versions, patch upgrades and configuration changes.

[0096] All expert artificial intelligence agents are deployed in the resource scheduling layer, maintain MCP connection with the Master artificial intelligence agent, and realize the collaborative decision of "parallel computing and local optimization" through sharing resource state, task profile and local scheduling suggestion. After task scheduling is triggered, the Master artificial intelligence agent sends a "decision request package" to each expert artificial intelligence agent through the MCP task broadcast interface, which contains: task ID and emergency level; target time delay and energy consumption threshold; current selectable node resource picture; model version requirement and constraint rule. After receiving the request, the expert artificial intelligence agent immediately executes the strategy evaluation model locally to generate a set of "recommended scheduling score table" for the candidate nodes, and feeds back to the coordinating artificial intelligence agent through MCP. The output structure of each artificial intelligence agent is uniform, as follows: After receiving the score table of all expert artificial intelligence agents, the coordinating artificial intelligence agent executes a unified strategy fusion function to generate a final scheduling optimization list by comprehensively considering delay, energy consumption and load balancing targets. The expert artificial intelligence agent outputs the following node scores respectively: resource temperature elasticity score: , iterative evolution sensitivity score: and scheduling consistency score: The coordinating artificial intelligence agent supports the "fusion" strategy, and selects the most robust node based on the minimum deviation function:

[0097]

[0098] Wherein represents the mean of the current score group, and the node with the overall score closest to the balance state is selected. The coordinating artificial intelligence agent selects Top-K nodes as scheduling candidates (default K=2~3), preferentially schedules the task to the node with the highest score, and caches the alternative node to the task fault tolerance path. After scheduling is completed, the entire score table, fusion parameters and final scheduling result are returned to the Master artificial intelligence agent through MCP, and enter the task execution process. If scheduling fails (node rejection, predicted abnormality), the Master artificial intelligence agent t will automatically enable suboptimal path rescheduling without reevaluation.

[0099] The embodiment not only considers the performance of each artificial intelligence agent and the type of task, but also ensures the objectivity and reliability of selection through comprehensive scoring.

[0100] Optionally, the calculating the comprehensive score value corresponding to the M second artificial intelligence agents comprises:

[0101] calculate a computing resource score value corresponding to the third artificial intelligence agent, a network resource score value corresponding to the third artificial intelligence agent, and a storage resource score value corresponding to the third artificial intelligence agent, the third artificial intelligence agent being any one of the M second artificial intelligence agents;

[0102] According to the second weight information, the computing resource score value, the network resource score value, and the storage resource score value are weighted and calculated to obtain a comprehensive score value corresponding to the third artificial intelligence agent.

[0103] In this embodiment, when calculating the comprehensive score value corresponding to each second artificial intelligence agent, the computing resource score value, the network resource score value, and the storage resource score value corresponding to any one second artificial intelligence agent, i.e., the third artificial intelligence agent, are considered, so that the computing resource score value, the network resource score value, and the storage resource score value are weighted and calculated according to the preset second weight information to obtain the comprehensive score value corresponding to the third artificial intelligence agent.

[0104] Specifically, to ensure efficient coordination of tasks among computing power, network, and storage heterogeneous resources, a sequential scheduling mechanism is adopted in this step to divide the entire resource matching process into three stages: “computing node selection→network path evaluation→storage resource binding”, local score calculation is performed respectively, and finally fused into global scheduling priority. The entire process is driven by the coordination artificial intelligence agent to execute the scheduling strategy, and the resource calling and instruction issuing are realized through the MCP protocol.

[0105] First, in the computing resource stage: preferentially selecting nodes with model capabilities and moderate loads

[0106] In the first stage, the coordination artificial intelligence agent preferentially selects nodes with corresponding inference capabilities (such as supporting M1 / M2 / H1 / H2 models) according to the model requirements and complexity of the task.

[0107] The computing power adaptation score of the candidate node is calculated, and the indicators include the remaining computing power of the node, whether it has the corresponding model instance, the current load pressure, etc., so that the scheduling system can optimize the node under the principle of “enough but not waste”. For each candidate computing node A, the execution adaptation score is calculated according to the task and the ability of the intelligent agent:

[0108]

[0109] wherein is the available computing power (TFLOPS or model parameter scale support) of the node; is the model matching flag, which is 1 if the artificial intelligence agent has the required model capability, and 0 otherwise; The current processing queue length of the node; for example, for large model inference tasks (dialog generation), the scheduling system preferentially selects artificial intelligence agents (such as H1 / H2 nodes) with high parameter model loading capabilities, and for structured data classification or text keyword extraction tasks, the light artificial intelligence agent (such as L1 / L2) can be scheduled to improve the overall resource utilization.

[0110] Secondly, in the network resource stage: ensure that the task data flow transfer delay is controllable and the bandwidth is stable; after task matching, it needs to be transmitted to the target node for execution, so network delay and bandwidth become the key points of the second stage decision. The coordination artificial intelligence agent calculates the data path topology of each candidate node based on the RTT (Round-Trip Time) and available bandwidth information collected in the resource state heartbeat. For the network path corresponding to the computing node, the system evaluates its structure score under the current topology state:

[0111]

[0112] Among them End-to-end delay of the task from the entrance to the target node; Current bandwidth availability; the scoring model in this stage ensures that the selected path meets the lower limit of the task requirements in terms of throughput capacity and delay control.

[0113] Finally, in the storage resource stage: guarantee the cache ability of the intermediate state and model weight in the task

[0114] Some tasks (such as multi-stage inference, model distillation, and edge cache acceleration) have explicit requirements for storage space, so the system performs adaptive evaluation on the local or adjacent cache resources of the candidate nodes in the third stage.

[0115]

[0116] Among them Current free cache space; Whether to support task persistent storage (tasks that require result feedback or logs need to be enabled); the scoring in this stage is dominated by cache space and functional support, avoiding task rollback or migration due to intermediate state write failure.

[0117] In other embodiments, to enhance the stability of the scoring ranking and the reliability of the scheduling result, a double-sequence-intersection ranking strategy can be used to rank the computing resource score value, the network resource score value, and the storage resource score value. Specifically, the three scoring dimensions are sorted in descending order to obtain the following three sorted lists: the storage sorting table: , the network sorting table: And the computing sorting table: .

[0118] The matching screening in order and reverse order is performed on each ranking table respectively: forward (Top-down) selection: the top three (or Top-K) are taken to form a set; reverse (Bottom-up) selection: the top three of the lower ranking are taken to form a set; and the overlapping nodes are screened.

[0119] The cross-position consistency screening is performed to find the node set in which the same node has the same ranking position in the forward and reverse ranking tables in three dimensions:

[0120]

[0121] These nodes are the preferred objects in the three-dimensional stability score intersection, indicating that they show consistency and stability in the system resource state change.

[0122] Fine-grained fine-tuning and ranking confirmation The node introduces a scheduling history index for fine-tuning, including the average task queuing time, scheduling interruption history, and MCP packet response RTT. According to these indexes, the ranking result is slightly adjusted to obtain the final task scheduling priority sequence.

[0123] In this embodiment, the computing resource score value, the network resource score value, and the storage resource score value corresponding to each artificial intelligence agent are calculated to obtain a comprehensive score value corresponding to each artificial intelligence agent, and thus it is determined whether each artificial intelligence agent is matched with a target computing power task according to the comprehensive score value, thereby effectively improving the scheduling effect of the computing power task.

[0124] Optionally, after the target computing power task is scheduled to the target agent for processing to obtain a processing result, the method further includes:

[0125] The difference between the processing result and the expected processing result of the target computing power task is calculated to obtain a difference value;

[0126] The target agent is updated based on the difference value to obtain an updated target agent.

[0127] In this embodiment, the expected processing result is an ideal processing result of processing the target computing power task, and the target agent is updated according to the difference value between the expected processing result and the processing result to obtain an updated target agent, thereby realizing self-adaptation of the artificial intelligence agent.

[0128] Specifically, to improve long-term scheduling performance, the system deploys a lightweight "local feedback processing module" in each artificial intelligence agent node to monitor, evaluate and learn key running indicators generated during current task execution. Mainly includes: each artificial intelligence agent dynamically updates the local reward function Rt according to the task completion quality, which is used to fine-tune and optimize the current model parameters (scheduling estimation model): the reward function result will be used to fine-tune the "execution preference" of the current artificial intelligence agent in task scheduling, and the node task priority selection strategy is updated through a lightweight learner, so that it can select more suitable task types in the future.

[0129] As Figure 3 shown, Figure 3 The system framework diagram in this embodiment, this embodiment constructs a task complexity recognition mechanism, introduces a multi-expert artificial intelligence agent division and cooperation strategy, and integrates MCP communication protocol, resource portrait standard, scheduling reliable receipt chain and other core components, effectively solves the problems of resource waste, path instability, scheduling rigidity and other problems in current heterogeneous computing resource scheduling. The system supports task grading and shunting for large AI model inference tasks, and adapts to "light-medium-heavy" models and node systems. In terms of resource cost, the system supports scheduling backtracking at the level of "task-resource-model" triplets, can automatically repair abnormal paths and shield inefficient nodes, combined with feedback learning mechanism can realize, significantly reduce the operation and energy cost in the era of large models.

[0130] Specifically, the application proposes a computing power shock value, hidden dark resources, and a three-dimensional data of data temperature gradient for constructing a heterogeneous artificial intelligence agent resource-aware constraint condition for the MCP protocol. Different types of artificial intelligence agent nodes can automatically complete registration and resource state reporting, and generate a structured "resource portrait package" uniformly, which contains key indicators such as storage shock value, hidden dark resources, and data temperature gradient, and comprehensively reflects the availability and load characteristics of the node. Based on the relationship between the computing power shock value, hidden dark resources, and data temperature gradient three-dimensional data, according to the data temperature gradient slope, the type of task is determined, and according to the computing power shock value, hidden dark resources, and task type, the numbers of multiple different artificial intelligence agents are associated. Calculate the computing power adaptation score of the candidate node, the indicators cover the node's remaining computing power, whether it has the corresponding model instance, and the current load pressure, so that the scheduling system can optimize the node under the principle of "enough but not waste". For each candidate computing node A, calculate its execution adaptation score according to the task and the ability of the intelligent agent: network resource stage: ensure that the task data flow transfer delay is controllable and the bandwidth is stable; after task matching, it needs to be transmitted to the target node for execution, so network delay and bandwidth become the key points of the second stage decision. The coordination artificial intelligence agent collects the RTT (Round-Trip Time) and available bandwidth information based on the resource state heartbeat, and calculates the data path topology of each candidate node. For the network path corresponding to the computing node, the system evaluates its structure score in the current topology state: storage resource stage: guarantee the cache ability of the task intermediate state and model weight.

[0131] The technical scheme of the application associates N artificial intelligence agents with different task types by calculating the performance score values corresponding to the N artificial intelligence agents, and in the case of receiving a target computing power task that needs to be processed, determines the most matched target agent in the N artificial intelligence agents according to the task type and the required performance score value corresponding to the target computing power task, thereby scheduling the target computing power task to the target agent for processing, which comprehensively considers the allocation of resources and models when processing computing power tasks, and effectively improves the scheduling effect of computing power tasks.

[0132] Referring to Figure 4 , Figure 4 is a structural diagram of a computing power task scheduling device provided by an embodiment of the application. As Figure 4 shown, the computing power task scheduling device 400 includes:

[0133] A computing module 410 is configured to calculate performance scores corresponding to the N artificial intelligence agents to obtain N performance score values, wherein the performance score values are used to represent the performance of the corresponding artificial intelligence agents in processing computing power tasks, and N is a positive integer greater than 1.

[0134] The determining module 420 is configured to determine, based on the N performance score values, a task type of the N artificial intelligence agents in processing a task, to obtain N task type labels, the task type label being used to represent a type of task that the corresponding artificial intelligence agent can process.

[0135] The matching module 430 is configured to, in a case where a target computing power task is received, determine a target agent from the N artificial intelligence agents based on a target performance score value and a target task type, the target performance score value being a performance score value required by the target computing power task, the target task type being a type of task corresponding to the target computing power task, and the target agent being an artificial intelligence agent in the N artificial intelligence agents, for which a corresponding task type label matches the target task type and a corresponding performance score value is greater than or equal to the target performance score value.

[0136] The processing module 440 is configured to schedule the target computing power task to the target agent for processing, to obtain a processing result.

[0137] Optionally, the calculating module 410 includes:

[0138] The first determining sub-module is configured to determine, for a first artificial intelligence agent, a computing power fluctuation value corresponding to the first artificial intelligence agent, a hidden dark resource value corresponding to the first artificial intelligence agent, and a data temperature gradient corresponding to the first artificial intelligence agent, the first artificial intelligence agent being any one of the N artificial intelligence agents, wherein the computing power fluctuation value is used to represent a coupling relationship between resource utilization and power consumption limitation of the first artificial intelligence agent, the hidden dark resource value is used to represent a number of resources included in a reserve resource pool of the first artificial intelligence agent that can be called in an emergency, and the data temperature gradient is used to represent a dynamic evolution trend of data access heat distribution in different levels of storage media of the first artificial intelligence agent.

[0139] The first calculating sub-module is configured to perform weighted calculation on the computing power fluctuation value, the hidden dark resource value, and the data temperature gradient according to first weight information, to obtain a performance score value corresponding to the first artificial intelligence agent.

[0140] Optionally, the task type includes a first task, a second task, and a third task, a task complexity of the first task is lower than a task complexity of the second task, and the task complexity of the second task is lower than a task complexity of the third task, and the determining module 420 includes:

[0141] The first determining unit is configured to determine, based on the performance score value corresponding to the first artificial intelligence agent, a data temperature gradient slope of the first artificial intelligence agent in processing the computing power task, the data temperature gradient slope being used to represent a migration speed of the first artificial intelligence agent in processing hot data in a unit of time and an intensity of the first artificial intelligence agent in processing hot data.

[0142] The second determining unit is configured to determine, in a case where the data temperature gradient slope is less than or equal to a first preset value, a task type label corresponding to the first artificial intelligence agent as the first task.

[0143] The third determining unit is configured to determine, in a case where the data temperature gradient slope is greater than the first preset value and less than a second preset value, the task type label corresponding to the first artificial intelligence agent as the second task.

[0144] The fourth determining unit is configured to determine, in a case where the data temperature gradient slope is greater than or equal to the second preset value, the task type label corresponding to the first artificial intelligence agent as the third task.

[0145] The first preset value is less than the second preset value.

[0146] Optionally, the matching module 430 comprises:

[0147] The second determining sub-module is configured to determine, in a case where a target computing power task is received, M second artificial intelligence agents from the N artificial intelligence agents based on a target performance score value and a target task type, the second artificial intelligence agent being an artificial intelligence agent corresponding to the task type label in the N artificial intelligence agents that is consistent with the target task type and has a performance score value greater than or equal to the target performance score value, and the M being a positive integer less than or equal to N.

[0148] The second calculating sub-module is configured to calculate a comprehensive score value corresponding to the M second artificial intelligence agents, to obtain M comprehensive score values corresponding to the M second artificial intelligence agents in one-to-one correspondence, the comprehensive score value being used to represent a matching degree of the corresponding second artificial intelligence agent and the target computing power task.

[0149] The third determining sub-module is configured to determine, based on the M comprehensive score values, a target agent from the M second artificial intelligence agents, the target agent being a second artificial intelligence agent corresponding to the highest comprehensive score value from the M second artificial intelligence agents.

[0150] Optionally, the third determining sub-module comprises:

[0151] The first calculation unit is configured to calculate a calculation resource score value corresponding to a third artificial intelligence agent, a network resource score value corresponding to the third artificial intelligence agent, and a storage resource score value corresponding to the third artificial intelligence agent, the third artificial intelligence agent being any one of the M second artificial intelligence agents;

[0152] The second calculation unit is configured to perform weighted calculation on the calculation resource score value, the network resource score value, and the storage resource score value according to second weight information, to obtain a comprehensive score value corresponding to the third artificial intelligence agent.

[0153] Optionally, the technical solution further comprises:

[0154] The difference calculation module is configured to calculate a difference between the processing result and an expected processing result of the target computing power task, to obtain a difference value;

[0155] The updating module is configured to update the target agent based on the difference value, to obtain an updated target agent.

[0156] The technical solution of the present application calculates performance score values corresponding to N artificial intelligence agents, thereby associating the N artificial intelligence agents with different task types. When a target computing power task that needs to be processed is received, the most matched target agent is determined from the N artificial intelligence agents according to a task type corresponding to the target computing power task and a required performance score value, thereby scheduling the target computing power task to the target agent for processing. The allocation of resources and models is comprehensively considered when processing the computing power task, and the scheduling effect of the computing power task is effectively improved.

[0157] The present application also provides an electronic device. Please refer to Figure 5 The electronic device can include a processor 501, a memory 502, and a program 5021 stored in the memory 502 and executable on the processor 501.

[0158] The program 5021, when executed by the processor 501, can implement any step in the corresponding method embodiment: Figure 1

[0159] The N performance scores corresponding to the N artificial intelligence agents are calculated to obtain N performance score values, the performance score values being used to represent the performance of the corresponding artificial intelligence agent in processing the computing power task, the N being a positive integer greater than 1;

[0160] Based on the N performance score values, the task types of the N artificial intelligence agents are determined to obtain N task type labels, the task type labels being used to represent the task types that the corresponding artificial intelligence agent can process; ​

[0161] In the case of receiving a target computing power task, a target agent is determined among the N artificial intelligence agents based on a target performance score value and a target task type, the target performance score value is a performance score value required by the target computing power task, the target task type is a task type corresponding to the target computing power task, and the target agent is an artificial intelligence agent in the N artificial intelligence agents, whose corresponding task type label matches the target task type and whose corresponding performance score value is greater than or equal to the target performance score value;

[0162] The target computing power task is scheduled to the target agent for processing to obtain a processing result.

[0163] Optionally, the N performance scores corresponding to the N artificial intelligence agents are calculated to obtain N performance score values, including:

[0164] For a first artificial intelligence agent, a computing power shock value corresponding to the first artificial intelligence agent, a hidden dark resource value corresponding to the first artificial intelligence agent, and a data temperature gradient corresponding to the first artificial intelligence agent are determined, the first artificial intelligence agent being any one of the N artificial intelligence agents, wherein the computing power shock value is used to represent a coupling relationship between resource utilization and power consumption limitation of the first artificial intelligence agent, the hidden dark resource value is used to represent a resource quantity included in a reserve resource pool that can be called in an emergency situation of the first artificial intelligence agent, and the data temperature gradient is used to represent a dynamic evolution trend of data access heat distribution in different levels of storage media of the first artificial intelligence agent.

[0165] According to first weight information, the computing power shock value, the hidden dark resource value, and the data temperature gradient are weighted and calculated to obtain a performance score value corresponding to the first artificial intelligence agent.

[0166] Optionally, the task type includes a first task, a second task, and a third task, a task complexity degree of the first task is lower than a task complexity degree of the second task, and the task complexity degree of the second task is lower than a task complexity degree of the third task, and the N performance score values are used to determine a task type of a task processed by the N artificial intelligence agents to obtain N task type labels, including:

[0167] Based on the performance score value corresponding to the first artificial intelligence agent, a data temperature gradient slope of the first artificial intelligence agent processing a computing power task is determined, the data temperature gradient slope being used to represent a migration speed of processing hot data and an intensity of processing hot data of the first artificial intelligence agent within a unit time.

[0168] In a case where the data temperature gradient slope is less than or equal to a first preset value, a task type label corresponding to the first artificial intelligence agent is determined as the first task.

[0169] In a case where the data temperature gradient slope is greater than the first preset value and less than a second preset value, the task type label corresponding to the first artificial intelligence agent is determined as the second task.

[0170] In a case where the data temperature gradient slope is greater than or equal to the second preset value, the task type label corresponding to the first artificial intelligence agent is determined as the third task.

[0171] The first preset value is less than the second preset value.

[0172] Optionally, in a case where the target computing power task is received, a target agent is determined from the N artificial intelligence agents based on a target performance score value and a target task type, including:

[0173] In a case where the target computing power task is received, M second artificial intelligence agents are determined from the N artificial intelligence agents based on a target performance score value and a target task type, the second artificial intelligence agents being artificial intelligence agents in the N artificial intelligence agents that have a task type label corresponding to the target task type and a performance score value greater than or equal to the target performance score value, and the M being a positive integer less than or equal to N.

[0174] A comprehensive score value corresponding to each of the M second artificial intelligence agents is calculated, to obtain M comprehensive score values corresponding one-to-one to the M second artificial intelligence agents, the comprehensive score value being used to represent a matching degree of the corresponding second artificial intelligence agent to the target computing power task.

[0175] Based on the M comprehensive score values, a target agent is determined from the M second artificial intelligence agents, the target agent being a second artificial intelligence agent in the M second artificial intelligence agents that has the highest comprehensive score value.

[0176] Optionally, the calculation of the comprehensive score value corresponding to each of the M second artificial intelligence agents includes:

[0177] A computing resource score value corresponding to a third artificial intelligence agent, a network resource score value corresponding to the third artificial intelligence agent, and a storage resource score value corresponding to the third artificial intelligence agent are calculated, the third artificial intelligence agent being any one of the M second artificial intelligence agents.

[0178] According to the second weight information, the computing resource score value, the network resource score value and the storage resource score value are weighted and calculated to obtain a comprehensive score value corresponding to the third artificial intelligence agent.

[0179] Optionally, after the target computing power task is dispatched to the target agent for processing to obtain a processing result, the method further includes:

[0180] A difference between the processing result and an expected processing result of the target computing power task is calculated to obtain a difference value.

[0181] The target agent is updated based on the difference value to obtain an updated target agent.

[0182] In a second aspect, an embodiment of the present application provides a computing power task scheduling device applied to a computing power platform, the computing power platform including N artificial intelligence agents, and the device including:

[0183] A computing module is configured to calculate performance scores corresponding to the N artificial intelligence agents to obtain N performance score values, the performance score values being used to represent the performance of the corresponding artificial intelligence agents in processing computing power tasks, and the N being a positive integer greater than 1.

[0184] A determining module is configured to determine, based on the N performance score values, the task types of the N artificial intelligence agents in processing tasks to obtain N task type labels, the task type labels being used to represent the task types that the corresponding artificial intelligence agents can process.

[0185] A matching module is configured to, when a target computing power task is received, determine a target agent in the N artificial intelligence agents based on a target performance score value and a target task type, the target performance score value being a performance score value required by the target computing power task, the target task type being a task type corresponding to the target computing power task, and the target agent being an artificial intelligence agent in the N artificial intelligence agents, for which a corresponding task type label matches the target task type and a corresponding performance score value is greater than or equal to the target performance score value.

[0186] A processing module is configured to dispatch the target computing power task to the target agent for processing to obtain a processing result.

[0187] The technical solution of the present application associates N artificial intelligence agents with different task types by calculating the performance score values corresponding to the N artificial intelligence agents, and in the case that a target computing power task needing to be processed is received, determines the most matched target agent from the N artificial intelligence agents according to the task type and the required performance score value corresponding to the target computing power task, so as to schedule the target computing power task to the target agent for processing, thereby realizing the allocation of resources and models in processing the computing power task, and effectively improving the scheduling effect of the computing power task.

[0188] The embodiment of the present application also provides a computer readable storage medium, and a computer program is stored on the computer readable storage medium. The computer program is executed by a processor to realize each process of the scheduling embodiment of the computing power task and achieve the same technical effect. To avoid repetition, details are not described herein. The computer readable storage medium includes a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, and the like.

[0189] The embodiment of the present application also provides a computer readable storage medium, and a computer program is stored on the computer readable storage medium. The computer program is executed by a processor to realize each process of the scheduling embodiment of the computing power task and achieve the same technical effect. To avoid repetition, details are not described herein. The computer readable storage medium includes a read-only memory (ROM), a random access memory (RAM), a magnetic disk, an optical disk, and the like.

[0190] It should be noted that in this paper, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "includes a" does not exclude the presence of another identical element in the process, method, article or device including the element.

[0191] From the above description of the embodiments, those skilled in the art can clearly understand that the above embodiment method can be realized by software and necessary general hardware platform, of course, it can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solution of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for making a communication device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) execute the method described in each embodiment of the present application.

[0192] The embodiments of the present application are described above with reference to the accompanying drawings, but the present application is not limited to the above-described specific embodiments, and the above-described specific embodiments are merely illustrative, but not restrictive, and a person of ordinary skill in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims.

Claims

1. A method for scheduling a computing power task, characterized in that, The method is applied to a computing power platform including N artificial intelligence agents, and the method comprises the following steps: calculating performance scores corresponding to the N artificial intelligence agents to obtain N performance score values, wherein the performance score values represent the performance of the corresponding artificial intelligence agents in processing computing power tasks, and N is a positive integer greater than 1; determining the task types of the N artificial intelligence agents in processing tasks based on the N performance score values to obtain N task type labels, wherein the task type labels represent the types of tasks that can be processed by the corresponding artificial intelligence agents; in the case of receiving a target computing power task, determining a target agent among the N artificial intelligence agents based on a target performance score value and a target task type, wherein the target performance score value is a performance score value required by the target computing power task, the target task type is the task type corresponding to the target computing power task, and the target agent is an artificial intelligence agent in the N artificial intelligence agents, which has a corresponding task type label matching the target task type and a corresponding performance score value greater than or equal to the target performance score value; scheduling the target computing power task to the target agent for processing to obtain a processing result.

2. The method of claim 1, wherein, The calculation of the performance scores corresponding to the N artificial intelligence agents to obtain N performance score values comprises the following steps: for a first artificial intelligence agent, determining a computing power shock value corresponding to the first artificial intelligence agent, a hidden dark resource value corresponding to the first artificial intelligence agent, and a data temperature gradient corresponding to the first artificial intelligence agent, wherein the first artificial intelligence agent is any one of the N artificial intelligence agents, the computing power shock value represents the coupling relationship between the resource utilization and the power consumption limitation of the first artificial intelligence agent, the hidden dark resource value represents the number of resources included in the reserve resource pool that can be called by the first artificial intelligence agent in an emergency, and the data temperature gradient represents the dynamic evolution trend of the data access heat distribution of the first artificial intelligence agent in different levels of storage media; performing weighted calculation on the computing power shock value, the hidden dark resource value, and the data temperature gradient according to first weight information to obtain a performance score value corresponding to the first artificial intelligence agent.

3. The method of claim 2, wherein, The task types include a first task, a second task, and a third task, the task complexity of the first task is lower than that of the second task, and the task complexity of the second task is lower than that of the third task, and the determination of the task types of the N artificial intelligence agents in processing tasks based on the N performance score values to obtain N task type labels comprises the following steps: determining a data temperature gradient slope of the first artificial intelligence agent in processing computing power tasks based on the performance score value corresponding to the first artificial intelligence agent, wherein the data temperature gradient slope represents the migration speed and intensity of the first artificial intelligence agent in processing hot data within a unit of time. In a case where the data temperature gradient slope is less than or equal to a first preset value, a task type label corresponding to the first artificial intelligence agent is determined as the first task; In a case where the data temperature gradient slope is greater than the first preset value and less than a second preset value, the task type label corresponding to the first artificial intelligence agent is determined as the second task; In a case where the data temperature gradient slope is greater than or equal to the second preset value, the task type label corresponding to the first artificial intelligence agent is determined as the third task; The first preset value is less than the second preset value.

4. The method of claim 1, wherein, In a case where the target computing power task is received, a target agent is determined from the N artificial intelligence agents based on a target performance score value and a target task type, including: In a case where the target computing power task is received, M second artificial intelligence agents are determined from the N artificial intelligence agents based on a target performance score value and a target task type, the second artificial intelligence agents being artificial intelligence agents corresponding to the target task type and having performance score values greater than or equal to the target performance score value, and M being a positive integer less than or equal to N; A comprehensive score value corresponding to each of the M second artificial intelligence agents is calculated, to obtain M comprehensive score values corresponding to the M second artificial intelligence agents, the comprehensive score value being used to represent a matching degree of the corresponding second artificial intelligence agent and the target computing power task; A target agent is determined from the M second artificial intelligence agents based on the M comprehensive score values, the target agent being a second artificial intelligence agent having the highest comprehensive score value among the M second artificial intelligence agents.

5. The method of claim 4, wherein, The calculation of the comprehensive score value corresponding to each of the M second artificial intelligence agents includes: A computing resource score value corresponding to a third artificial intelligence agent, a network resource score value corresponding to the third artificial intelligence agent, and a storage resource score value corresponding to the third artificial intelligence agent are calculated, the third artificial intelligence agent being any one of the M second artificial intelligence agents; The computing resource score value, the network resource score value, and the storage resource score value are weighted and calculated according to second weight information, to obtain a comprehensive score value corresponding to the third artificial intelligence agent.

6. The method according to any one of claims 1 to 5, characterized in that, After the target computing power task is dispatched to the target agent for processing to obtain a processing result, the method further includes: A difference between the processing result and an expected processing result of the target computing power task is calculated, to obtain a difference value; The target agent is updated based on the difference value, to obtain an updated target agent.

7. A computing power task scheduling apparatus characterized by comprising: The device is applied to a computing power platform, the computing power platform including N artificial intelligence agents, and the device including: A calculation module is configured to calculate performance scores corresponding to the N artificial intelligence agents, to obtain N performance score values, the performance score value being used to represent a performance of an artificial intelligence agent in processing a computing power task, and N being a positive integer greater than 1. determining, based on the N performance score values, a task type of a task processed by the N artificial intelligence agents, to obtain N task type labels, the task type label being used to represent a type of task that the corresponding artificial intelligence agent is capable of processing; matching, in a case where a target computing power task is received, a target agent from the N artificial intelligence agents based on a target performance score value and a target task type, the target performance score value being a performance score value required by the target computing power task, the target task type being a type of task corresponding to the target computing power task, the target agent being an artificial intelligence agent from the N artificial intelligence agents that has a corresponding task type label matching the target task type and a corresponding performance score value greater than or equal to the target performance score value; processing, to schedule the target computing power task to the target agent for processing, to obtain a processing result.

8. An electronic device, comprising: comprising: a processor, a memory, and a program stored on the memory and executable on the processor, the program, when executed by the processor, implementing the steps of the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, a computer program stored on the computer readable storage medium, the computer program, when executed by the processor, implementing the steps of the method of any one of claims 1 to 6.

10. A computer program product, characterised in that, computer instructions, the computer instructions, when executed by the processor, implementing the steps of the method of any one of claims 1 to 6.

Images