An agent scheduling method and system for a cloud data server environment

Abstract

The application is an agent scheduling method and system for a cloud data server environment, belonging to the technical field of cloud computing and task scheduling. The agent running state information and capability characteristic information are collected to generate agent load parameters and agent capability parameters; the standard resource requirement sequence is generated according to the task description information and the task and computing resource corresponding relationship data, and the target task is split into multiple subtasks to generate subtask characteristic data and target task subtask sequence; in the running process, the rebalancing is executed by comparing the running index with the preset scheduling warning threshold; and the task and computing resource corresponding relationship data and the agent capability parameters are updated according to the execution deviation information after the task is completed. The application can improve the task allocation matching degree in the multi-agent shared cloud computing power scene, and gradually optimize the scheduling performance through feedback learning.

Description

Technical Field

[0001] This invention relates to the field of cloud computing and multi-agent scheduling technology, specifically to an agent scheduling method and system for cloud data server environments. Background Technology

[0002] With the increasing use of multiple intelligent agents in cloud data server environments, multiple intelligent agent instances, such as dialogue agents, retrieval agents, planning agents, and execution agents, can run simultaneously within a single cloud service.

[0003] In existing technologies, task allocation and computing power scheduling for intelligent agents typically suffer from the following problems: Task allocation relies on simple rules, and tasks are often assigned to agent instances by round-robin or fixed routes without fully considering the agent's capability characteristics and current agent load parameters. This results in some agents running under overload for a long time, while other agents have idle resources.

[0004] The lack of computing power constraints at the agent level means that computing power scheduling in cloud data server environments often configures quotas for virtual machines or container instances, rarely setting thresholds for the proportion of computing power for agents at the "agent instance" level. There is also a lack of control over the task parallelism threshold for a single objective task, which can easily lead to a single agent or a single task occupying most of the total available computing power in the cloud.

[0005] The scheduling process lacks a rebalancing mechanism driven by operational metrics. With a large number of tasks continuously entering the system, without monitoring and early warning of operational metrics such as subtask queuing latency, subtask completion time, agent load distribution, and cloud data server resource utilization, the system will remain in a state of high latency and high congestion for a long time when the scheduling strategy is unbalanced.

[0006] The lack of closed-loop feedback between execution results and resource allocation means that in existing solutions, after the task is completed, the actual resource consumption data and the actual completion time are not used to correct the correspondence between the task and computing resources and the agent's capability parameters, resulting in the standard resource requirement sequence of subsequent tasks deviating from the actual resource requirements in the long term. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent agent scheduling method for a cloud data server environment, comprising the following steps: Step 1: Collect the agent's operational status information and capability characteristic information, generate agent load parameters and agent capability parameters, and write the agent identifier, agent load parameters, and agent capability parameters into the agent information table; Step 2: Receive the target task and generate a standard resource requirement sequence based on the task description information; Step 3: Based on the standard resource requirement sequence and task dependencies, the target task is broken down into multiple subtasks to generate a target task subtask sequence; Step 4: Calculate the fit score based on the target task sub-task sequence and the agent capability parameters and agent load parameters in the agent information table. Then, form a candidate agent set for the corresponding sub-task by agents whose fit scores are greater than the preset fit threshold. Step 5: Collect the resource usage status of the cloud data server and generate a description of the available computing power in the cloud; Step 6: Based on the target task subtask sequence, candidate agent set, agent load parameters and cloud available computing power description information, set agent computing power ratio threshold, preemption priority threshold and task parallelism threshold, generate subtask scheduling scheme under threshold constraints, and allocate subtasks and cloud computing power to target agents. Step 7: During the execution of the subtask scheduling scheme, collect the operation indicators and compare them with the preset scheduling warning threshold. When the operation indicators exceed the scheduling warning threshold, the unfinished subtasks are reallocated and the agent's load parameters are updated under the constraints of the task parallelism threshold and the agent's computing power ratio threshold. Step 8: After the target task is completed, collect actual resource consumption data and actual completion time, generate execution deviation information, and update the task-computation resource correspondence data and agent capability parameters when the execution deviation information exceeds the deviation threshold, so as to adjust the subsequent standard resource requirement sequence and various thresholds.

[0008] Furthermore, the process of collecting the agent's operational status information and capability characteristic information, generating agent load parameters and agent capability parameters, and writing the agent identifier, agent load parameters, and agent capability parameters into the agent information table includes: Collect the number of tasks currently being executed by each agent, the length of the task queue to be executed, the processor utilization, the accelerator utilization, and the network round-trip latency. Combine the collected data into running status information and generate agent load parameters based on the running status information. Collect the types of tasks each agent can handle, the maximum size of the input data it supports, the upper limit of the number of concurrent tasks it supports, and the historical average task completion time. Combine the collected data into capability feature information and generate agent capability parameters based on the capability feature information. Write the agent identifier, agent load parameters, and agent capability parameters of each agent into the agent information table.

[0009] Furthermore, the receiving of the target task and the generation of a standard resource requirement sequence based on the task description information includes: The task type, input data size, and real-time level in the task description information are parsed from the call request of the target task. The task type and input data size are matched with the resource configuration records in the task-computing resource correspondence data to generate initial resource requirement data. Adjust the estimated processing time and recommended parallelism in the initial resource requirement data according to the real-time level to generate standard resource requirement data; According to the order of the target task execution stages, the standard resource requirement data corresponding to each execution stage are arranged in sequence to form a standard resource requirement sequence, and the standard resource requirement sequence is stored in the task-computation resource correspondence data.

[0010] Furthermore, the step of splitting the target task into multiple subtasks based on the standard resource requirement sequence and task dependencies to generate a target task subtask sequence includes: Based on the standard resource requirement data and task dependencies of each execution stage in the standard resource requirement sequence, the target task is divided into sequential execution subtasks and parallel execution subtasks, resulting in multiple subtasks. Generate subtask resource requirement data for each subtask, combine the subtask resource requirement data with the real-time level to determine the subtask priority, and allocate the target task deadline to multiple subtask deadlines according to task dependencies. The subtask resource requirements, subtask priorities, and subtask deadlines are combined into subtask feature data, and the subtask feature data are arranged according to subtask priorities and task dependencies to generate a target task subtask sequence.

[0011] Furthermore, the process of calculating a fit score based on the target task sub-task sequence and the agent capability parameters and agent load parameters in the agent information table, and forming a candidate agent set for the corresponding sub-task with fit scores greater than a preset fit threshold, includes: For each subtask in the target task subtask sequence, the task type and subtask resource requirement data in the subtask feature data are compared with the agent capability parameters in the agent information table. When the agent capability parameters meet the processing conditions corresponding to the subtask resource requirement data, the corresponding agent is selected as a candidate. The suitability score is calculated based on the agent capability parameters, subtask resource requirements, and agent load parameters of the candidate. When the suitability score is greater than the preset suitability threshold, the candidate is added to the candidate agent set of the corresponding subtask. When the suitability score is less than or equal to the preset suitability threshold, the candidate is not added to the candidate agent set. When the candidate agent set is empty, the corresponding subtask is marked as an abnormal subtask and the abnormal mark is recorded.

[0012] Furthermore, the process of collecting the resource usage status of the cloud data server and generating a description of available computing power in the cloud includes: Available processor resources, available accelerator resources, and available memory resources are collected from each cloud data server in the cloud data server environment at preset time intervals. The cloud data server identifier is combined with the available processor resources, available accelerator resources, and available memory resources to form a cloud available computing power record. All cloud available computing power records are summarized into cloud available computing power description information, which reflects the total cloud available computing power.

[0013] Furthermore, the step of setting thresholds for agent computing power ratio, preemption priority, and task parallelism based on the target task subtask sequence, candidate agent set, agent load parameters, and cloud-available computing power description information, and generating a subtask scheduling scheme under threshold constraints to allocate subtasks and cloud computing power to the target agent includes: Based on the relationship between the overall resource requirements of the target task in the standard resource requirement sequence and the total available computing power in the cloud computing power description information, a computing power ratio threshold is configured for each agent. When generating a subtask scheduling scheme, when the ratio of the computing power currently occupied by the agent to the total available computing power in the cloud reaches the computing power ratio threshold of the agent, the allocation of subtasks to the agent is stopped. Set a preemption priority threshold based on the distribution of subtask priority differences in the target task subtask sequence. When the available computing power in the cloud is insufficient to simultaneously meet the resource requirements of two subtasks, computing power that has been allocated to the low-priority subtask but has not yet been used will be allocated to the high-priority subtask when the subtask priority difference between the two subtasks is greater than the preemption priority threshold. Set a task parallelism threshold based on the relationship between the overall resource requirements of the target task in the standard resource requirement sequence and the total available computing power in the cloud computing power description information. When the number of subtasks in the same target task in the execution state reaches the task parallelism threshold, pause the allocation of new subtasks to the target task. Under the conditions of meeting the thresholds for the proportion of computing power of the agent, the preemption priority threshold, and the task parallelism threshold, a target agent is selected from the corresponding candidate agent set for each subtask in the target task subtask sequence, and cloud computing power is allocated to the target agent to generate a subtask scheduling scheme.

[0014] Furthermore, the process of collecting operational metrics during the execution of the subtask scheduling scheme, comparing these metrics with a preset scheduling warning threshold, and reallocating unfinished subtasks and updating agent load parameters under the constraints of task parallelism threshold and agent computing power ratio threshold when the operational metrics exceed the scheduling warning threshold includes: Collect subtask queuing delay, subtask completion time, agent load distribution, and cloud data server resource utilization at preset time intervals, and combine the collected data into operational metrics; Based on the standard resource demand sequence and subtask scheduling scheme, the target subtask queuing delay, target subtask completion time, target agent load distribution and target cloud data server resource utilization are calculated, and a preset scheduling warning threshold is set. The operating indicators are then compared with the preset scheduling warning threshold. When the subtask queuing delay corresponding to the operation index or the subtask completion time corresponding to the operation index exceeds the preset scheduling warning threshold, select an agent whose agent load parameter is less than the preset load threshold and has not reached the agent computing power ratio threshold from the candidate agent set of the corresponding subtask, and migrate the unfinished subtask to the agent. When the corresponding operating indicators of agent load distribution or cloud data server resource utilization exceed the preset scheduling warning threshold, the execution order of subtasks is adjusted while maintaining the task parallelism threshold constraint. Subtasks with priority higher than the preset priority threshold and resource requirement data lower than the preset resource requirement threshold are executed first. After completing the subtask migration and subtask execution order adjustment, the agent load parameters are updated.

[0015] Furthermore, the process of collecting actual resource consumption data and actual completion time after the target task is completed, generating execution deviation information, and updating the task-computational resource correspondence data and agent capability parameters when the execution deviation information exceeds the deviation threshold, is used to adjust the subsequent standard resource requirement sequence and various thresholds, including: After the target task is completed, the processor usage time, accelerator usage time, and actual completion time are collected for each subtask in the target task subtask sequence. The collected data are combined into actual execution data, and actual resource consumption data is generated based on the actual execution data. The actual resource consumption data is compared with the standard resource requirement data of the corresponding sub-task, and the actual completion time is compared with the expected processing time of the corresponding sub-task. Based on the comparison results, execution deviation information is generated. When the deviation information exceeds the deviation threshold, the resource configuration record corresponding to the task description information in the task-computation resource correspondence data is adjusted to generate updated standard resource requirement data. The historical average task completion time in the agent capability parameters of the corresponding agent is also updated so that the standard resource requirement sequence generated in step two for subsequent target tasks, as well as the agent computing power ratio threshold, task parallelism threshold, and preset scheduling warning threshold used in steps six and seven, are closer to the actual execution situation.

[0016] An agent scheduling system for a cloud data server environment, employing an agent scheduling method for a cloud data server environment, includes: an agent information management module, a task resource baseline module, a subtask management module, a candidate agent screening module, a cloud computing power monitoring module, a scheduling decision module, a scheduling monitoring and rebalancing module, a result feedback and data update module, and a data processing module; the agent information management module, task resource baseline module, subtask management module, candidate agent screening module, cloud computing power monitoring module, scheduling decision module, scheduling monitoring and rebalancing module, result feedback and data update module, and data processing module are respectively connected to the data processing module; The intelligent agent information management module is used to collect operating status information and capability characteristic information, generate intelligent agent load parameters and intelligent agent capability parameters, and write them into the intelligent agent information table. The task resource baseline module is used to receive the target task, obtain task description information, and generate a standard resource requirement sequence. The subtask management module is used to split the target task and generate a target task subtask sequence based on the standard resource requirement sequence and task dependency relationship; The candidate agent screening module is used to calculate the fit score and generate a candidate agent set based on the target task subtask sequence and the agent capability parameters and agent load parameters in the agent information table. The cloud computing power monitoring module is used to collect the resource usage status of the cloud data server and generate cloud available computing power description information; The scheduling decision module is used to set the intelligent agent computing power ratio threshold, preemption priority threshold and task parallelism threshold and generate a subtask scheduling scheme based on the target task subtask sequence, candidate intelligent agent set, intelligent agent load parameters and cloud available computing power description information. The scheduling monitoring and rebalancing module is used to collect operating indicators, compare the operating indicators with preset scheduling warning thresholds, and reallocate unfinished sub-tasks and update the agent load parameters when the operating indicators exceed the scheduling warning thresholds. The result feedback and data update module is used to collect actual resource consumption data and actual completion time, generate execution deviation information, and update the task-computation resource correspondence data and agent capability parameters.

[0017] The beneficial effects of this invention are: By jointly modeling and calculating the suitability score of agent load parameters and agent capability parameters, subtasks can be assigned to agents that are more suitable for capability matching and load status, thereby improving task execution efficiency.

[0018] By setting thresholds for the proportion of computing power in intelligent agents and for the parallelism of tasks, resource consumption is restricted at both the agent granularity and the target task granularity, thereby achieving a constrained allocation of the total available computing power in the cloud and reducing the degree of competition for computing power.

[0019] By using operational metrics such as subtask queuing latency, subtask completion time, agent load distribution, and cloud data server resource utilization, along with preset scheduling warning thresholds, online monitoring and rebalancing of the scheduling process can be achieved, thereby shortening the duration of high-latency states.

[0020] By updating the task-computation resource correspondence data and agent capability parameters through deviation information, the standard resource demand sequence gradually approaches the actual demand during long-term operation, achieving a sustainable and optimized scheduling effect. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an intelligent agent scheduling method for a cloud data server environment. Figure 2 This is a schematic diagram of the process for collecting and storing information from intelligent agents. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0023] The features and performance of the present invention will be further described in detail below with reference to embodiments.

[0024] Example 1 like Figure 1 As shown, an agent scheduling method for a cloud data server environment includes the following steps: Step 1: Collect the agent's operational status information and capability characteristic information, generate agent load parameters and agent capability parameters, and write the agent identifier, agent load parameters, and agent capability parameters into the agent information table; Step 2: Receive the target task and generate a standard resource requirement sequence based on the task description information; Step 3: Based on the standard resource requirement sequence and task dependencies, the target task is broken down into multiple subtasks to generate a target task subtask sequence; Step 4: Calculate the fit score based on the target task sub-task sequence and the agent capability parameters and agent load parameters in the agent information table. Then, form a candidate agent set for the corresponding sub-task by agents whose fit scores are greater than the preset fit threshold. Step 5: Collect the resource usage status of the cloud data server and generate a description of the available computing power in the cloud; Step 6: Based on the target task subtask sequence, candidate agent set, agent load parameters and cloud available computing power description information, set agent computing power ratio threshold, preemption priority threshold and task parallelism threshold, generate subtask scheduling scheme under threshold constraints, and allocate subtasks and cloud computing power to target agents. Step 7: During the execution of the subtask scheduling scheme, collect the operation indicators and compare them with the preset scheduling warning threshold. When the operation indicators exceed the scheduling warning threshold, the unfinished subtasks are reallocated and the agent's load parameters are updated under the constraints of the task parallelism threshold and the agent's computing power ratio threshold. Step 8: After the target task is completed, collect actual resource consumption data and actual completion time, generate execution deviation information, and update the task-computation resource correspondence data and agent capability parameters when the execution deviation information exceeds the deviation threshold, so as to adjust the subsequent standard resource requirement sequence and various thresholds.

[0025] Specifically, the cloud data server environment includes multiple cloud data servers, each providing processor resources, accelerator resources, and memory resources. Multiple intelligent agents run in the cloud data server environment, with each agent performing one or more task types.

[0026] The intelligent agent scheduling system is deployed in a cloud data server environment and interacts with intelligent agents and the cloud data server via a network. The intelligent agent scheduling system includes: The module includes: agent information management module; task resource baseline module; subtask management module; candidate agent screening module; cloud computing power monitoring module; scheduling decision module; scheduling monitoring and rebalancing module; and result feedback and data update module.

[0027] like Figure 2 As shown, the intelligent agent information management module performs the following operations periodically: 1) Collect the number of tasks currently being executed by each agent, the length of the task queue to be executed, the processor utilization, the accelerator utilization, and the network round-trip latency, and use the above data as running status information.

[0028] 2) Generate agent load parameters based on operational status information. The generation process allows configuring preset weights for each indicator, transforming them to a uniform numerical range, and then weighting and summing them according to their weights to obtain a single value as the agent load parameter. The larger this value, the higher the agent's current load.

[0029] 3) Collect the types of tasks that each agent can handle, the maximum size of the input data it supports, the upper limit of the number of concurrent tasks it supports, and the historical average task completion time, and use the above data as capability feature information.

[0030] 4) Generate agent capability parameters based on capability feature information. Agent capability parameters can be described in a structured manner, including the set of task types, maximum input data size, upper limit of the number of concurrent tasks, and historical average task completion time.

[0031] 5) Write the agent identifier, agent load parameters, and agent capability parameters as a record into the agent information table for use in steps four, six, and seven.

[0032] In this embodiment, the weights and numerical ranges are preset by the operation and maintenance personnel in the configuration file. The intelligent agent information management module reads the configuration and applies it, and the specific values ​​are not limited.

[0033] The task resource baseline module performs the following operations upon receiving a target task: 1) Parse the task description information from the target task call request. The task description information includes the task type, input data size, and real-time level.

[0034] 2) Based on the task type and the size of the input data, locate the resource configuration record in the task-computing resource mapping data. The resource configuration record should include at least the processor requirement, accelerator requirement, estimated processing time, and recommended parallelism.

[0035] 3) Use resource configuration records as initial resource requirement data. Adjust the estimated processing time and recommended parallelism based on the real-time level: shorten the estimated processing time and increase the recommended parallelism when the real-time level is high, and extend the estimated processing time and decrease the recommended parallelism when the real-time level is low.

[0036] 4) Use the adjusted data as the standard resource requirement data. If the target task contains multiple execution phases, generate standard resource requirement data separately for each execution phase.

[0037] 5) Arrange the standard resource requirement data of each execution stage in sequence to form a standard resource requirement sequence, and store the standard resource requirement sequence in the storage structure to which the task-computation resource correspondence data belongs, for subsequent steps to call.

[0038] The standard resource requirement sequence serves as a unified resource baseline for subsequent subtask decomposition, threshold setting, and operation monitoring.

[0039] The subtask management module performs the following operations based on the standard resource requirement sequence and task dependencies: 1) Based on the standard resource requirement data of each execution stage in the standard resource requirement sequence, and combined with the task dependency relationship, the target task is divided into sequential execution subtasks and parallel execution subtasks, resulting in multiple subtasks.

[0040] 2) Use the standard resource requirement data corresponding to each subtask as the subtask resource requirement data.

[0041] 3) Determine the priority of subtasks based on their real-time requirements and task dependencies. The higher the real-time requirement or the more critical the subtask's location on a critical dependency path, the higher its priority.

[0042] 4) Based on the target task's deadline, allocate the deadlines to each subtask along the task dependencies to obtain the deadlines for each subtask. For subtasks with successor subtasks, the deadline can be determined by subtracting the estimated processing time of the subtask from its deadline.

[0043] 5) Combine the subtask resource requirements, subtask priority, and subtask deadline to form subtask feature data.

[0044] 6) Arrange the subtask feature data from high to low priority, and within the same subtask priority, arrange them in order of task dependency to obtain the target task subtask sequence.

[0045] The target task subtask sequence will be used in steps four, six, and seven.

[0046] The candidate agent selection module performs the following operations for each subtask in the target task subtask sequence: 1) Read the agent capability parameters and agent load parameters from the agent information table.

[0047] 2) Compare the task type and resource requirement data in the subtask feature data with the agent's capability parameters. If the agent's capability parameters include the corresponding task type and the maximum input data size is not less than the subtask input data size, then the agent is considered a candidate.

[0048] 3) For each candidate, calculate the fit score based on the following factors: Does the task type match? The relationship between subtask resource requirements and the maximum input data size and the upper limit of concurrent tasks; the degree of similarity between the historical average task completion time and the expected processing time of the subtask; The relationship between agent load parameters and preset load thresholds: when agent load parameters exceed the preset load threshold, the fit score decreases; when agent load parameters are below the preset load threshold, the fit score increases.

[0049] 4) The above factors are weighted and summed according to preset weights to obtain the fit score.

[0050] 5) When the fit score is greater than the preset fit threshold, the candidate object is written into the candidate agent set of the corresponding subtask. When the fit score is less than or equal to the preset fit threshold, it is not written into the candidate agent set.

[0051] 6) When the candidate agent set is empty, mark the corresponding subtask as an abnormal subtask and record the abnormal mark for operation and maintenance personnel or upper management programs to handle.

[0052] The cloud computing power monitoring module performs the following operations at preset time intervals: 1) Collect the number of available processor cores, available accelerator computing power, and available memory capacity from each cloud data server.

[0053] 2) Combine the cloud data server identifier with the above resource data to form a record of available computing power in the cloud.

[0054] 3) All available computing power records in the cloud are summarized to form a description of available computing power in the cloud, including the total number of available processor cores in the cloud, the total available computing power of accelerators in the cloud, and the total available memory capacity in the cloud.

[0055] The available computing power description information in the cloud is used to set the threshold for the proportion of computing power of intelligent agents, the threshold for task parallelism, and to determine whether there is insufficient computing power.

[0056] The scheduling decision module performs the following operations based on the target task subtask sequence, candidate agent set, agent load parameters, and cloud-available computing power description information: 1) Setting the threshold for the proportion of computing power of intelligent agents Calculate the total available computing power in the cloud based on the description of available computing power in the cloud. According to the service policy, configure a maximum percentage limit for each agent, for example, specifying the maximum percentage of total available computing power a given agent can use. Record this percentage as the agent's computing power percentage threshold.

[0057] 2) Preemption priority threshold setting Assign integer priority levels to subtasks. Analyze the distribution of priority differences between different subtasks in the target task's subtask sequence.

[0058] Based on the business's tolerance for preemption, an integer value is selected as the preemption priority threshold. When the priority difference between two subtasks is greater than the preemption priority threshold, the higher-priority subtask is allowed to preempt the computing power pre-allocated to the lower-priority subtask.

[0059] 3) Setting the task parallelism threshold Estimate the reasonable parallelism range of the target task based on the overall resource requirements of the target task in the standard resource requirement sequence. Determine an integer value as the task parallelism threshold by combining the total available computing power in the cloud and the number of other tasks. During execution, control the number of subtasks of the same target task in the execution state to ensure that it does not exceed the task parallelism threshold.

[0060] 4) Subtask scheduling scheme generation Traverse the subtasks in the order of the target task's subtask sequence. Select the target agent from the corresponding candidate agent set whose agent load parameters are low and whose current computing power usage ratio to the total available computing power in the cloud does not reach the agent computing power ratio threshold.

[0061] The resource requirements of subtasks are mapped to specific processor resources, accelerator resources, and memory resources, and corresponding resources are reserved for the cloud data server corresponding to the target intelligent agent in the cloud available computing power description information.

[0062] When the available computing power in the cloud is insufficient to meet the resource requirements of multiple subtasks, if the priority difference between the subtasks is greater than the preemption priority threshold, the unused reserved computing power of the low-priority subtasks will be released and allocated to the high-priority subtasks.

[0063] When the number of subtasks in the same target task that are in the execution state reaches the task parallelism threshold, the allocation of new subtasks to the target task is paused until the existing subtasks are completed.

[0064] After completing the above processing, the scheduling decision module generates a subtask scheduling scheme and sends the subtask scheduling scheme to the corresponding intelligent agent.

[0065] The scheduling, monitoring, and rebalancing module collects operational metrics at preset time intervals. These operational metrics include at least: Subtask queuing delay: This is obtained by the difference between the time a subtask enters the queue and the time it starts executing. It can be calculated as an average or a maximum value.

[0066] Subtask completion time: This is obtained by the difference between the start time and the completion time of the subtask. It can be calculated as an average or a maximum value.

[0067] Agent load distribution: Composed of agent load parameters of each agent, it can analyze the maximum value, minimum value and difference.

[0068] Cloud data server resource utilization: This is composed of the processor utilization, accelerator utilization, and memory utilization of each cloud data server, and its average and maximum values ​​can be calculated.

[0069] The scheduling monitoring and rebalancing module calculates the target subtask queuing delay, target subtask completion time, target agent load distribution, and target cloud data server resource utilization based on the standard resource demand sequence and subtask scheduling scheme, and sets preset scheduling warning thresholds for each indicator. The preset scheduling warning thresholds are used to limit the maximum allowable deviation range.

[0070] When operational metrics exceed preset scheduling warning thresholds, the scheduling monitoring and rebalancing module performs the following operations: 1) When the queuing delay or completion time of a subtask exceeds the corresponding preset scheduling warning threshold, select an agent whose load parameter is less than the preset load threshold and whose current computing power ratio has not reached the agent computing power ratio threshold from the candidate agent set of the corresponding subtask, and migrate the subtask that has not yet started execution to that agent.

[0071] 2) When the intelligent agent load distribution or cloud data server resource utilization index exceeds the corresponding preset scheduling warning threshold, under the condition of maintaining the task parallelism threshold constraint, the execution order of the currently unexecuted subtasks and the subtasks to be scheduled is adjusted, and the subtasks with the priority higher than the preset priority threshold and the resource requirement data of the subtasks are less than the preset resource requirement threshold are scheduled in advance.

[0072] 3) Update the agent load parameters in the agent information table based on the results of subtask migration and execution order adjustment.

[0073] The aforementioned rebalancing mechanism can reduce long-term load imbalance and long queuing times.

[0074] The results feedback and data update module performs the following operations upon completion of the target task: 1) For each subtask in the target task subtask sequence, collect the processor usage time, accelerator usage time, and actual completion time. Combine the collected data into actual execution data and obtain the actual resource consumption data from the actual execution data.

[0075] 2) Compare the actual resource consumption data with the standard resource requirement data of the corresponding sub-task, and compare the actual completion time with the expected processing time of the corresponding sub-task to obtain the resource consumption difference and the time difference.

[0076] 3) Record the difference in resource consumption and the difference in execution time as execution deviation information. Execution deviation information is used to explain the magnitude and direction of the difference between the standard resource requirement data and the actual execution result.

[0077] 4) When the deviation information exceeds the deviation threshold, adjust the resource configuration records corresponding to the task description information in the task-computation resource mapping data to generate updated standard resource requirement data. Adjustment methods may include increasing or decreasing processor requirements, increasing or decreasing accelerator requirements, increasing or decreasing expected processing time, and adjusting the recommended parallelism.

[0078] 5) At the same time, based on the difference between the actual completion time and the historical average task completion time, update the historical average task completion time in the corresponding agent's capability parameters.

[0079] Through the above update operations, the standard resource requirement sequence generated in step two for subsequent target tasks, as well as the agent computing power ratio threshold, task parallelism threshold, and preset scheduling warning threshold used in steps six and seven, are made closer to the actual execution situation.

[0080] Example 2 An agent scheduling system for a cloud data server environment, employing an agent scheduling method for a cloud data server environment, includes: an agent information management module, a task resource baseline module, a subtask management module, a candidate agent screening module, a cloud computing power monitoring module, a scheduling decision module, a scheduling monitoring and rebalancing module, a result feedback and data update module, and a data processing module; the agent information management module, task resource baseline module, subtask management module, candidate agent screening module, cloud computing power monitoring module, scheduling decision module, scheduling monitoring and rebalancing module, result feedback and data update module, and data processing module are respectively connected to the data processing module; The intelligent agent information management module is used to collect operating status information and capability characteristic information, generate intelligent agent load parameters and intelligent agent capability parameters, and write them into the intelligent agent information table. The task resource baseline module is used to receive the target task, obtain task description information, and generate a standard resource requirement sequence. The subtask management module is used to split the target task and generate a target task subtask sequence based on the standard resource requirement sequence and task dependency relationship; The candidate agent screening module is used to calculate the fit score and generate a candidate agent set based on the target task subtask sequence and the agent capability parameters and agent load parameters in the agent information table. The cloud computing power monitoring module is used to collect the resource usage status of the cloud data server and generate cloud available computing power description information; The scheduling decision module is used to set the intelligent agent computing power ratio threshold, preemption priority threshold and task parallelism threshold and generate a subtask scheduling scheme based on the target task subtask sequence, candidate intelligent agent set, intelligent agent load parameters and cloud available computing power description information. The scheduling monitoring and rebalancing module is used to collect operating indicators, compare the operating indicators with preset scheduling warning thresholds, and reallocate unfinished sub-tasks and update the agent load parameters when the operating indicators exceed the scheduling warning thresholds. The result feedback and data update module is used to collect actual resource consumption data and actual completion time, generate execution deviation information, and update the task-computation resource correspondence data and agent capability parameters.

[0081] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for intelligent agent scheduling in a cloud data server environment, characterized in that, Includes the following steps: Step 1: Collect the agent's operational status information and capability characteristic information, generate agent load parameters and agent capability parameters, and write the agent identifier, agent load parameters, and agent capability parameters into the agent information table; Step 2: Receive the target task and generate a standard resource requirement sequence based on the task description information; Step 3: Based on the standard resource requirement sequence and task dependencies, the target task is broken down into multiple subtasks to generate a target task subtask sequence; Step 4: Calculate the fit score based on the target task sub-task sequence and the agent capability parameters and agent load parameters in the agent information table. Then, form a candidate agent set for the corresponding sub-task by agents whose fit scores are greater than the preset fit threshold. Step 5: Collect the resource usage status of the cloud data server and generate a description of the available computing power in the cloud; Step 6: Based on the target task subtask sequence, candidate agent set, agent load parameters and cloud available computing power description information, set agent computing power ratio threshold, preemption priority threshold and task parallelism threshold, generate subtask scheduling scheme under threshold constraints, and allocate subtasks and cloud computing power to target agents. Step 7: During the execution of the subtask scheduling scheme, collect the operation indicators and compare them with the preset scheduling warning threshold. When the operation indicators exceed the scheduling warning threshold, the unfinished subtasks are reallocated and the agent's load parameters are updated under the constraints of the task parallelism threshold and the agent's computing power ratio threshold. Step 8: After the target task is completed, collect actual resource consumption data and actual completion time, generate execution deviation information, and update the task-computation resource correspondence data and agent capability parameters when the execution deviation information exceeds the deviation threshold, so as to adjust the subsequent standard resource requirement sequence and various thresholds.

2. The intelligent agent scheduling method for a cloud data server environment according to claim 1, characterized in that, The process of collecting the agent's operational status information and capability characteristic information, generating agent load parameters and agent capability parameters, and writing the agent identifier, agent load parameters, and agent capability parameters into the agent information table includes: Collect the number of tasks currently being executed by each agent, the length of the task queue to be executed, the processor utilization, the accelerator utilization, and the network round-trip latency. Combine the collected data into running status information and generate agent load parameters based on the running status information. Collect the types of tasks each agent can handle, the maximum size of the input data it supports, the upper limit of the number of concurrent tasks it supports, and the historical average task completion time. Combine the collected data into capability feature information and generate agent capability parameters based on the capability feature information. Write the agent identifier, agent load parameters, and agent capability parameters of each agent into the agent information table.

3. The intelligent agent scheduling method for a cloud data server environment according to claim 1, characterized in that, The process of receiving the target task and generating a standard resource requirement sequence based on the task description information includes: The task type, input data size, and real-time level in the task description information are parsed from the call request of the target task. The task type and input data size are matched with the resource configuration records in the task-computing resource correspondence data to generate initial resource requirement data. Adjust the estimated processing time and recommended parallelism in the initial resource requirement data according to the real-time level to generate standard resource requirement data; According to the order of the target task execution stages, the standard resource requirement data corresponding to each execution stage are arranged in sequence to form a standard resource requirement sequence, and the standard resource requirement sequence is stored in the task-computation resource correspondence data.

4. The intelligent agent scheduling method for a cloud data server environment according to claim 1, characterized in that, The process of splitting the target task into multiple subtasks based on the standard resource requirement sequence and task dependencies, and generating a target task subtask sequence, includes: Based on the standard resource requirement data and task dependencies of each execution stage in the standard resource requirement sequence, the target task is divided into sequential execution subtasks and parallel execution subtasks, resulting in multiple subtasks. Generate subtask resource requirement data for each subtask, combine the subtask resource requirement data with the real-time level to determine the subtask priority, and allocate the target task deadline to multiple subtask deadlines according to task dependencies. The resource requirements, priority, and deadline of each subtask are combined into subtask feature data. The subtask feature data are then arranged according to subtask priority and task dependency to generate a sequence of subtasks for the target task.

5. The intelligent agent scheduling method for a cloud data server environment according to claim 1, characterized in that, The process of calculating a fit score based on the target task sub-task sequence and agent capability and load parameters in the agent information table, and forming a candidate agent set for the corresponding sub-task with fit scores greater than a preset fit threshold, includes: For each subtask in the target task subtask sequence, the task type and subtask resource requirement data in the subtask feature data are compared with the agent capability parameters in the agent information table. When the agent capability parameters meet the processing conditions corresponding to the subtask resource requirement data, the corresponding agent is selected as a candidate. The suitability score is calculated based on the agent capability parameters, subtask resource requirement data and agent load parameters of the candidate. When the suitability score is greater than the preset suitability threshold, the candidate is written into the candidate agent set of the corresponding subtask. When the suitability score is less than or equal to the preset suitability threshold, the candidate is not written into the candidate agent set. When the candidate agent set is empty, the corresponding subtask is marked as an abnormal subtask and the abnormal mark is recorded.

6. The intelligent agent scheduling method for a cloud data server environment according to claim 1, characterized in that, The process of collecting the resource usage status of cloud data servers and generating information describing the available computing power in the cloud includes: Available processor resources, available accelerator resources, and available memory resources are collected from each cloud data server in the cloud data server environment at preset time intervals. The cloud data server identifier is combined with the available processor resources, available accelerator resources, and available memory resources to form a record of available computing power in the cloud. All available computing power records in the cloud are summarized into cloud available computing power description information, which reflects the total available computing power in the cloud.

7. The agent scheduling method for a cloud data server environment according to claim 1, characterized in that, The process of setting thresholds for agent computing power ratio, preemption priority, and task parallelism based on the target task subtask sequence, candidate agent set, agent load parameters, and cloud-available computing power description information, generating a subtask scheduling scheme under threshold constraints, and allocating subtasks and cloud computing power to the target agent includes: Based on the relationship between the overall resource requirements of the target task in the standard resource requirement sequence and the total available computing power in the cloud computing power description information, a computing power ratio threshold is configured for each agent. When generating a subtask scheduling scheme, when the ratio of the computing power currently occupied by the agent to the total available computing power in the cloud reaches the computing power ratio threshold of the agent, the allocation of subtasks to the agent is stopped. Set a preemption priority threshold based on the distribution of subtask priority differences in the target task subtask sequence. When the available computing power in the cloud is insufficient to simultaneously meet the resource requirements of two subtasks, computing power that has been allocated to the low-priority subtask but has not yet been used will be allocated to the high-priority subtask when the subtask priority difference between the two subtasks is greater than the preemption priority threshold. Set a task parallelism threshold based on the relationship between the overall resource requirements of the target task in the standard resource requirement sequence and the total available computing power in the cloud computing power description information. When the number of subtasks in the same target task in the execution state reaches the task parallelism threshold, pause the allocation of new subtasks to the target task. Under the conditions of meeting the thresholds for the proportion of computing power of the agent, the preemption priority threshold, and the task parallelism threshold, a target agent is selected from the corresponding candidate agent set for each subtask in the target task subtask sequence, and cloud computing power is allocated to the target agent to generate a subtask scheduling scheme.

8. The intelligent agent scheduling method for a cloud data server environment according to claim 1, characterized in that, The process of collecting operational metrics during the execution of the subtask scheduling scheme, comparing these metrics with a preset scheduling warning threshold, and reallocating unfinished subtasks and updating agent load parameters under the constraints of task parallelism threshold and agent computing power ratio threshold when the operational metrics exceed the scheduling warning threshold includes: Collect subtask queuing delay, subtask completion time, agent load distribution, and cloud data server resource utilization at preset time intervals, and combine the collected data into operational metrics; Based on the standard resource demand sequence and subtask scheduling scheme, the target subtask queuing delay, target subtask completion time, target agent load distribution and target cloud data server resource utilization are calculated, and a preset scheduling warning threshold is set. The operating indicators are then compared with the preset scheduling warning threshold. When the subtask queuing delay corresponding to the operation index or the subtask completion time corresponding to the operation index exceeds the preset scheduling warning threshold, select an agent whose agent load parameter is less than the preset load threshold and has not reached the agent computing power ratio threshold from the candidate agent set of the corresponding subtask, and migrate the unfinished subtask to the agent. When the corresponding operating indicators of agent load distribution or cloud data server resource utilization exceed the preset scheduling warning threshold, the execution order of subtasks is adjusted while maintaining the task parallelism threshold constraint. Subtasks with priority higher than the preset priority threshold and resource requirement data lower than the preset resource requirement threshold are executed first. After completing the subtask migration and subtask execution order adjustment, the agent load parameters are updated.

9. The intelligent agent scheduling method for a cloud data server environment according to claim 1, characterized in that, The process of collecting actual resource consumption data and actual completion time after the target task is completed, generating execution deviation information, and updating the task-computation resource correspondence data and agent capability parameters when the execution deviation information exceeds the deviation threshold, is used to adjust the subsequent standard resource requirement sequence and various thresholds, including: After the target task is completed, the processor usage time, accelerator usage time, and actual completion time are collected for each subtask in the target task subtask sequence. The collected data are combined into actual execution data, and actual resource consumption data is generated based on the actual execution data. The actual resource consumption data is compared with the standard resource requirement data of the corresponding sub-task, and the actual completion time is compared with the expected processing time of the corresponding sub-task. Based on the comparison results, execution deviation information is generated. When the deviation information exceeds the deviation threshold, the resource configuration record corresponding to the task description information in the task-computation resource correspondence data is adjusted to generate updated standard resource requirement data. The historical average task completion time in the agent capability parameters of the corresponding agent is also updated so that the standard resource requirement sequence generated in step two for subsequent target tasks, as well as the agent computing power ratio threshold, task parallelism threshold, and preset scheduling warning threshold used in steps six and seven, are closer to the actual execution situation.

10. An intelligent agent scheduling system for a cloud data server environment, characterized in that, An agent scheduling method for a cloud data server environment according to any one of claims 1-9 includes: an agent information management module, a task resource baseline module, a subtask management module, a candidate agent screening module, a cloud computing power monitoring module, a scheduling decision module, a scheduling monitoring and rebalancing module, a result feedback and data update module, and a data processing module; the agent information management module, task resource baseline module, subtask management module, candidate agent screening module, cloud computing power monitoring module, scheduling decision module, scheduling monitoring and rebalancing module, result feedback and data update module, and data processing module are respectively connected to the data processing module; The intelligent agent information management module is used to collect operating status information and capability characteristic information, generate intelligent agent load parameters and intelligent agent capability parameters, and write them into the intelligent agent information table. The task resource baseline module is used to receive the target task, obtain task description information, and generate a standard resource requirement sequence. The subtask management module is used to split the target task and generate a target task subtask sequence based on the standard resource requirement sequence and task dependency relationship; The candidate agent screening module is used to calculate the fit score and generate a candidate agent set based on the target task subtask sequence and the agent capability parameters and agent load parameters in the agent information table. The cloud computing power monitoring module is used to collect the resource usage status of the cloud data server and generate cloud available computing power description information; The scheduling decision module is used to set the intelligent agent computing power ratio threshold, preemption priority threshold and task parallelism threshold and generate a subtask scheduling scheme based on the target task subtask sequence, candidate intelligent agent set, intelligent agent load parameters and cloud available computing power description information. The scheduling monitoring and rebalancing module is used to collect operating indicators, compare the operating indicators with preset scheduling warning thresholds, and reallocate unfinished sub-tasks and update the agent load parameters when the operating indicators exceed the scheduling warning thresholds. The result feedback and data update module is used to collect actual resource consumption data and actual completion time, generate execution deviation information, and update the task-computation resource correspondence data and agent capability parameters.

Images