A small amount of intelligent computing resource pooling scheduling method, device and medium

Abstract

This invention discloses a method, device, and medium for pooled scheduling of small-scale intelligent computing resources, relating to the field of resource management technology. The method includes: semantic parsing and resource requirement identification of real-time business requests based on an initial resource status list, and organizing task attributes to generate a task profile list; real-time comparison and optimization of candidate scheduling schemes to form a target execution scheme; task distribution and process control based on the target execution scheme and the task profile list, with dynamic adjustments during execution to obtain an execution result package; resource reclamation and scheduling feedback updates of the target execution scheme based on the execution result package, and correction of subsequent scheduling strategies to form the next round of resource status list. This invention improves the success rate and timeliness of single tasks and even mixed task flows by continuously monitoring the task execution process.

Description

Technical Field

[0001] This invention relates to the field of resource management technology, and in particular to a method, device and medium for pooling and scheduling small-scale intelligent computing resources. Background Technology

[0002] With the continuous growth in demand for artificial intelligence and computing power, intelligent computing environments are evolving towards heterogeneity and distribution. To improve resource utilization, resource pooling and unified scheduling technologies have become core supports for intelligent computing centers and cloud computing platforms. In this process, global schedulers for large-scale clusters and resource management frameworks for general-purpose computing have become relatively mature, effectively integrating heterogeneous computing power such as CPUs and GPUs through centralized orchestration and policy optimization.

[0003] In recent years, to adapt to scenarios such as edge computing and R&D for small and medium-sized enterprises, the demand for pooled scheduling of small-scale, lightweight intelligent computing resources has become increasingly prominent, aiming to achieve flexible and efficient task execution under limited resource scale. However, existing mainstream scheduling technologies and frameworks are mostly designed for data center environments with relatively abundant resources and stable business scale. Scheduling strategies and decision models are usually based on macroscopic and periodic resource states for static or semi-static optimization. When applied to small-scale intelligent computing environments with limited total resources, strong node heterogeneity, and real-time changes in business requests, such schedulers struggle to respond quickly and make accurate adjustments to micro-dynamic changes such as node load fluctuations and resource contention conflicts during task execution. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a pooled scheduling method for small-scale intelligent computing resources to solve the problem of unstable task execution leading to a decline in overall scheduling quality in small-scale, highly heterogeneous intelligent computing environments.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, this invention provides a method for pooling and scheduling small-scale intelligent computing resources, comprising: collecting, pooling, mapping, and organizing the states of heterogeneous resources in a small-scale intelligent computing environment to output an initial resource state list; performing semantic parsing and resource requirement identification on real-time business requests based on the initial resource state list, and organizing task attributes to generate a task profile list; performing resource segmentation and adaptation and node carrying matching based on the task profile list and the initial resource state list, and filtering out high-disturbance schemes to output a candidate scheduling scheme list; performing real-time comparison and optimization decision-making on the candidate scheduling scheme list to form a target execution scheme; performing task distribution and operation process control based on the target execution scheme and the task profile list, and implementing dynamic adjustments during execution to obtain an operation result package; performing resource reclamation and scheduling feedback updates on the target execution scheme based on the operation result package, and correcting subsequent scheduling strategies to form the next round of resource state list.

[0007] As a preferred embodiment of the small-scale intelligent computing resource pooling scheduling method described in this invention, the specific steps for collecting, pooling, mapping, and organizing the heterogeneous resources in the small-scale intelligent computing environment, and outputting an initial resource status list are as follows: Resource detection is performed on computing nodes in a small-scale intelligent computing environment to generate raw resource collection data. Different types of resources in the raw resource collection data are uniformly identified to generate resource identification data. Based on resource identification data, heterogeneous resources scattered across different computing nodes are pooled and mapped to construct a resource pool structure under a unified scheduling view, forming resource pool mapping data. Based on the resource pool mapping data, the availability of resources on each computing node is determined and the status is marked to form resource availability status data. The resource availability status data is then merged and organized according to the node dimension, resource dimension, and service dimension to output an initial resource status list.

[0008] As a preferred embodiment of the small-scale intelligent computing resource pooling scheduling method described in this invention, the steps of performing semantic parsing and resource requirement identification on real-time business requests based on an initial resource status list, and then performing task attribute consolidation to generate a task profile list are as follows: Receive real-time business requests and combine them with the initial resource status list to form basic data for task parsing. Based on the basic data for task parsing, perform semantic parsing and resource requirement identification on the real-time business requests to form task type identification data and task resource requirement data. Based on task resource requirement data, task priority is determined and operational constraints are identified for real-time business requests, forming task constraint data. The task type identification data, task resource requirement data, and task constraint data are then processed to generate a task profile list.

[0009] As a preferred embodiment of the small-scale intelligent computing resource pooling scheduling method of the present invention, the steps of performing resource partitioning and adaptation and node bearing matching based on the task profile list and the initial resource status list, and filtering out high-disturbance schemes to output a candidate scheduling scheme list are as follows: Each task in the task profile list is read one by one to obtain the resource information of the node to be matched, and the available resource status of each computing node is extracted by combining the initial resource status list to form the basic data for resource matching. Based on the basic resource matching data, the resource requirements corresponding to each task are segmented and adapted to form resource segmentation data. Based on the resource segmentation data, the node carrying capacity of each computing node is matched to form node matching data. Based on the node matching data, disturbances are identified for each carrier scheme, generating scheme disturbance judgment data. This data is then filtered and processed to output a list of candidate scheduling schemes.

[0010] As a preferred embodiment of the small-scale intelligent computing resource pooling scheduling method described in this invention, the specific steps for real-time comparison and optimization of the candidate scheduling scheme list to form the target execution scheme are as follows: The response time, resource adaptability and operational stability of each candidate scheduling scheme in the candidate scheduling scheme list are compared to form scheme comparison data. Based on the scheme comparison data, the candidate scheduling schemes are sorted, filtered, and prioritized to form scheme decision data. Based on the scheme decision data, the top-ranked candidate scheduling schemes are selected and combined to form the target execution scheme.

[0011] As a preferred embodiment of the small-scale intelligent computing resource pooling scheduling method described in this invention, the steps of issuing tasks and controlling the execution process based on the target execution scheme and the task profile list, and dynamically adjusting during execution to obtain the execution result package, are as follows: According to the target execution plan, each task in the task profile list is sent to the corresponding carrier node, and the execution association between the task and the carrier node is established to form task sending data; Based on the task distribution data, the task execution process is monitored to generate operational monitoring data; The task execution process is dynamically adjusted based on the operation monitoring data to form adjustment execution data. The task distribution data, operation monitoring data, and adjustment execution data are packaged together to obtain the operation result package.

[0012] As a preferred embodiment of the small-scale intelligent computing resource pooling scheduling method of the present invention, the specific steps of dynamically adjusting the task execution process based on operation monitoring data to form adjusted execution data are as follows: Based on operational monitoring data, identify the node load change status, resource consumption change status, task queuing change status, and runtime change status during task execution to form adjustment judgment data; Based on the adjustment judgment data, the computing power share, video memory share, and storage access share in the target execution plan are increased or decreased to form resource adjustment data; Based on the resource adjustment data, the carrier nodes in the target execution plan are switched, adjusted and updated to form path adjustment data. The resource allocation adjustment results in the resource adjustment data are recorded in correspondence with the carrier node switching results and task execution path update results in the path adjustment data to form adjustment execution data.

[0013] As a preferred embodiment of the small-scale intelligent computing resource pooling scheduling method described in this invention, the specific steps of performing resource reclamation and scheduling feedback updates on the target execution plan based on the running result package, and correcting subsequent scheduling strategies to form the resource status list for the next round are as follows: Based on the execution result package, the corresponding carrier nodes and resource allocation relationships of the target execution plan are released to form resource recovery data. The task completion status, resource occupation change status and adjustment execution status during the execution of this round of tasks are fed back and organized to form scheduling feedback data. Based on the scheduling feedback data, the resource allocation order, node selection order, and task scheduling order are corrected to form policy correction data. The resource reclamation data, scheduling feedback data, and policy correction data are then merged to form the next round of resource status list.

[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements any step of the small-scale intelligent computing resource pooling scheduling method as described in the first aspect of the present invention.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any step of the small-scale intelligent computing resource pooling scheduling method described in the first aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by continuously monitoring the task execution process and generating adjustment judgment data in real time based on multi-dimensional operation monitoring data, it can effectively cope with the resource contention and node performance fluctuations common in cloud computing environments, improve the success rate and timeliness of single tasks and even mixed task flows, and ensure that high-quality and highly stable computing service output can still be maintained in resource-constrained cloud-edge collaborative small-scale intelligent computing environments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a small-scale intelligent computing resource pooling scheduling method.

[0019] Figure 2 This diagram illustrates the heterogeneous resource acquisition, pooling mapping, and state management.

[0020] Figure 3 This diagram illustrates the real-time business request parsing and task profile generation.

[0021] Figure 4 A schematic diagram illustrating resource allocation and adaptation, node capacity matching, and candidate solution generation.

[0022] Figure 5 A comparison chart showing the cumulative distribution of actual task runtime under different scheduling schemes.

[0023] Figure 6 This is a graph showing the changes in the feedback effect of multiple rounds of scheduling. Detailed Implementation

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0027] Reference Figures 1-6 As one embodiment of the present invention, this embodiment provides a method for pooling and scheduling small-scale intelligent computing resources, including the following steps: S1. Collect, pool, map, and organize the heterogeneous resources in a small-scale intelligent computing environment, and output an initial resource status list.

[0028] S1.1. Perform resource detection on computing nodes in a small-scale intelligent computing environment to form raw resource collection data, and perform unified identification processing on different types of resources in the raw resource collection data to form resource identification data.

[0029] Specifically, the computing nodes continuously read the running records corresponding to the status of central processing resources, accelerated computing resources, memory resources, video memory resources, local storage resources, network links, and running services, and summarize the collection results of each computing node in terms of resource usage, idle changes, connectivity status, and running status in chronological order to form raw resource collection data. The raw resource data is classified and organized according to the location and resource type of the computing node. Then, each resource record in the raw resource data is written with the corresponding computing node identifier, resource type identifier, resource location identifier and status identifier. The status of central processing resources, accelerated computing resources, memory resources, video memory resources, local storage resources, network link status and running service status on the same computing node are mapped to the same identifier system to form resource identifier data.

[0030] S1.2. Based on resource identification data, perform pooling mapping on heterogeneous resources scattered on different computing nodes, construct a resource pool structure under a unified scheduling view, and form resource pool mapping data.

[0031] Specifically, resource identification data scattered across different computing nodes are categorized and aggregated according to the same resource type identifier, and each resource record is associated with and retained with the corresponding computing node identifier and resource location identifier. This completes the pooling mapping of heterogeneous resources scattered across different computing nodes, resulting in resource pool mapping results.

[0032] The pooling mapping results are arranged and processed according to the organization method under the unified scheduling view. The status of central processing resources, accelerated computing resources, memory resources, video memory resources, local storage resources, network link status, and running service status are respectively mapped to the resource category layer, computing node location layer, and resource status layer, and the positions are fixed to form resource pool mapping data.

[0033] S1.3. Based on the resource pool mapping data, determine the availability and status of resources on each computing node to form resource availability status data. Then, merge and organize the resource availability status data according to the node dimension, resource dimension, and service dimension to output the initial resource status list.

[0034] Specifically, based on the central processing resource status, accelerated computing resource status, memory resource status, video memory resource status, local storage resource status, network link status, and running service status of each computing node in the resource pool mapping data, and according to the resource occupancy, connectivity, and running status, the availability of resources on each computing node is judged. Resources that can participate in subsequent scheduling are marked as available, and resources that do not participate in subsequent scheduling are marked as unavailable, thus forming resource availability status data. The resource availability status data is centrally organized according to the node dimension, the resource status corresponding to the same computing node is classified and organized according to the resource dimension, and the relationship between the running service status and the corresponding resource status is organized according to the service dimension, and an initial resource status list is output.

[0035] S2. Based on the initial resource status list, perform semantic parsing and resource requirement identification on real-time business requests, and organize task attributes to generate a task profile list.

[0036] S2.1. Receive real-time business requests and combine them with the initial resource status list to form basic task parsing data. Based on the basic task parsing data, perform semantic parsing and resource requirement identification on the real-time business requests to form task type identification data and task resource requirement data.

[0037] Specifically, it receives the request content, request source, processing target, output requirements and time limit requirements from real-time business requests, and reads the available resource status and running service status of each computing node from the initial resource status list. It then organizes the real-time business request content with the available resource status in the initial resource status list to form the basic data for task parsing. Based on the task parsing basic data, semantic parsing is performed on the text content, processing objectives, and output requirements in real-time business requests to identify the task category and processing direction corresponding to the real-time business requests, forming task type identification data. Combined with the task type identification data, resource requirements such as computing power consumption, video memory consumption, storage access requirements, network transmission requirements, and execution time requirements of real-time business requests are identified to form task resource requirement data.

[0038] S2.2. Based on the task resource requirement data, determine the task priority and identify the operation constraints of real-time business requests to form task constraint data. Then, organize and process the task type identification data, task resource requirement data and task constraint data to generate a task profile list.

[0039] Specifically, based on the execution time requirement, computing power consumption requirement, video memory consumption requirement, storage access requirement, and network transmission requirement in the task resource requirement data, the task priority and operation constraint identification are performed for each real-time business request. Real-time business requests with short processing time limits and short execution times are identified as high task priority, real-time business requests with moderate processing time limits and stable resource consumption are identified as medium task priority, and real-time business requests with lenient processing time limits and long execution times are identified as low task priority. At the same time, the resource consumption constraints, execution time constraints, and concurrent operation constraints corresponding to real-time business requests are identified by combining computing power consumption requirements, video memory consumption requirements, storage access requirements, and network transmission requirements, forming task constraint data. The task type identification data, task resource requirement data, and task constraint data are organized accordingly, and task category, resource requirement, task priority, and operation constraints are written for each individual real-time business request to generate a task profile list.

[0040] S3. Based on the task profile list and the initial resource status list, perform resource partitioning and adaptation and node carrying matching, filter out high-disturbance schemes, and output a list of candidate scheduling schemes.

[0041] S3.1. Read each task in the task profile list one by one to obtain the resource information of the node to be matched, and extract the available resource status of each computing node in combination with the initial resource status list to form the basic data for resource matching.

[0042] Specifically, the task category, resource requirements, task priority, and running constraints corresponding to each task in the task profile list are extracted and expanded according to the reading order of individual tasks; at the same time, the central processing resource status, accelerated computing resource status, memory resource status, video memory resource status, local storage resource status, network link status, and running service status corresponding to each computing node are extracted from the initial resource status list, and resource records in the available state are filtered to obtain the resource information of the node to be matched. The resource requirements, task priorities, and operational constraints corresponding to each task are matched one-to-one with the resource information of the nodes to be matched for each computing node, forming the basic data for resource matching.

[0043] S3.2. Based on the resource matching basic data, the resource requirements corresponding to each task are segmented and adapted to form resource segmentation data. Based on the resource segmentation data, node load matching is performed on each computing node to form node matching data.

[0044] Specifically, for computing power usage requirements, the computing power consumption range corresponding to the task execution process is divided into an allocable computing power share. For video memory usage requirements, the video memory usage range corresponding to the task execution process is divided into an allocable video memory share. For storage access requirements and network transmission requirements, the data access intensity corresponding to the task execution process is divided into an allocable access share. The overall resource requirements corresponding to a single task are converted into resource allocation results that can be allocated among different computing nodes, forming resource allocation data. The computing power share, video memory share, storage access share, and network transmission share corresponding to each task are compared with the available resource status of each computing node. The computing nodes that can accommodate the corresponding resource allocation results and meet the task priority and running constraints are reserved as bearable nodes. The bearing relationship between each task and the corresponding bearable node is recorded to form node matching data.

[0045] S3.3. Based on the node matching data, identify disturbances in each carrying scheme, form scheme disturbance judgment data, filter and process the scheme disturbance judgment data, and output a list of candidate scheduling schemes.

[0046] Specifically, for a single bearer scheme, the changes in resource usage, task queuing, and running status of the computing nodes corresponding to the bearer scheme are read. Combined with the task priority and running constraints in the task profile list, the fluctuation of computing power share, fluctuation of video memory share, increase in task waiting time, and degree of restriction on yielding of high-priority tasks are determined after the current task is accessed. Based on the fluctuation of computing power share, the fluctuation of video memory share, the increase of task waiting time, and the degree of restriction on high-priority tasks, the scheme disturbance judgment value corresponding to each bearing scheme is calculated to obtain the disturbance identification result and form scheme disturbance judgment data. When filtering the data for determining the disturbance of the scheme, the carrying scheme with high disturbance level is removed, and the carrying scheme with low disturbance level and meeting the task priority and operation constraints is retained, and a list of candidate scheduling schemes is output.

[0047] It should be noted that the calculation of the first... The expression for the disturbance identification result corresponding to each bearing scheme is: ; in, Indicates the first The disturbance judgment value corresponding to each load-bearing scheme; Indicates the first Fluctuation in computing power share corresponding to each carrying scheme; Indicates the first The fluctuation of the video memory share corresponding to each carrying scheme; Indicates the first The task waiting time increment corresponding to each bearer scheme; Indicates the first The degree of restriction on high-priority tasks corresponding to each bearer scheme; This represents the index variable for the carrying scheme.

[0048] S4. Compare and optimize the candidate scheduling scheme list in real time to form the target execution scheme.

[0049] S4.1. Compare the response time, resource adaptability, and operational stability of each candidate scheduling scheme in the candidate scheduling scheme list to form scheme comparison data.

[0050] Specifically, the response time of each candidate scheduling scheme is compared. The task waiting time, resource switching time and task execution time corresponding to each candidate scheduling scheme are read and compared in order of time to obtain the response speed results of each candidate scheduling scheme. Resource matching comparison is performed on each candidate scheduling scheme. The computing power share, video memory share, storage access share and network transmission share corresponding to each candidate scheduling scheme are read and compared with the resource requirements and operational constraints in the task profile list to obtain the resource matching results corresponding to each candidate scheduling scheme. The operational stability of each candidate scheduling scheme is compared. The changes in resource consumption, task queuing, and operational status of the corresponding bearer nodes for each candidate scheduling scheme are read. The changes before and after the current task is connected are compared to obtain the operational fluctuation results of each candidate scheduling scheme. The response speed results, resource matching results, and operational fluctuation results are written into the corresponding candidate scheduling schemes to form scheme comparison data.

[0051] It should be noted that the calculation of the first... The expression for the resource matching result corresponding to each candidate scheduling scheme is: ; in, Indicates the first Resource matching results corresponding to each candidate scheduling scheme; Indicates the first The computing power share corresponding to each candidate scheduling scheme; Indicates the first The computing power requirements for each task; Indicates the first The memory share corresponding to each candidate scheduling scheme; Indicates the first The video memory usage requirements for each task; Indicates the first Storage access share corresponding to each candidate scheduling scheme; Indicates the first The storage access requirements corresponding to each task; Indicates the first The network transmission share corresponding to each candidate scheduling scheme; Indicates the first The network transmission requirements corresponding to each task; This represents the index variable for candidate scheduling schemes.

[0052] S4.2. Based on the scheme comparison data, sort, filter and prioritize each candidate scheduling scheme to form scheme decision data. Based on the scheme decision data, filter and combine the top-ranked candidate scheduling schemes to form the target execution scheme.

[0053] Specifically, each candidate scheduling scheme is ranked according to its response speed, resource matching, and operational fluctuation, and candidate scheduling schemes with poor response time, low resource matching, and large operational fluctuation are eliminated. Prioritize the retained candidate scheduling schemes, focusing on reading the task priority, task waiting time, resource consumption intensity, and operation constraints corresponding to the retained candidate scheduling schemes. For candidate scheduling schemes with high task priority and short task waiting time, give them a higher execution priority. For candidate scheduling schemes with the same task priority, continue to compare resource consumption intensity and operation constraints, and rank the candidate scheduling scheme with more stable resource consumption and better compliance with operation constraints first, so as to obtain the priority decision results corresponding to each candidate scheduling scheme and form the scheme decision data. Based on the decision data, the candidate scheduling schemes that rank highly are selected and combined. The carrier nodes, resource allocation relationships and task execution paths that rank highly in the priority decision results are combined accordingly to form the target execution scheme.

[0054] S5. Based on the target execution plan and the task profile list, tasks are assigned and the execution process is controlled, and dynamic adjustments are made during execution to obtain the execution result package.

[0055] S5.1. Based on the target execution plan, distribute each task in the task profile list to the corresponding carrier node, and establish the execution association between the task and the carrier node to form task distribution data.

[0056] Specifically, according to the target execution plan, each task in the task profile list is distributed to the corresponding carrier node. The task category, resource requirements, task priority, and operational constraints in the task profile list are bound to the carrier nodes, resource allocation relationships, and task execution paths in the target execution plan, either by assigning a single task to a single carrier node or by assigning a single task to multiple carrier nodes. When establishing the execution association between tasks and carrier nodes, the carrier node information, resource allocation information, and task execution path information corresponding to each task are written into the same task distribution record. At the same time, each carrier node can be reversed to correspond to the task content and execution order it undertakes, forming the execution association between tasks and carrier nodes, and generating task distribution data.

[0057] S5.2. Monitor the task execution process based on the task distribution data to generate operational monitoring data.

[0058] Specifically, based on the task distribution data, the system continuously tracks each task that has been distributed to the corresponding bearer node, reads the changes in the waiting order, waiting time, and release status of the preceding tasks on the bearer node, and continuously records the changes in the execution time of each task from the time of distribution to the current time, as well as the changes in resource usage and operational busyness of each bearer node during task execution. The resource occupancy status, task queuing status, runtime status, and node load status are written into the monitoring records of each task to form operational monitoring data.

[0059] S5.3. Dynamically adjust the task execution process based on the operation monitoring data to form adjustment execution data. Package the task distribution data, operation monitoring data and adjustment execution data to obtain the operation result package.

[0060] S5.3.1. Based on the operation monitoring data, identify the node load change status, resource usage change status, task queuing change status, and runtime change status during task execution, and form adjustment judgment data.

[0061] Specifically, based on the operation monitoring data, the continuous monitoring records of each task during the execution process are read and compared with the corresponding time before and after the task execution. This identifies the increase, decrease or fluctuation of the node load corresponding to the carrying node, the increase, decrease or fluctuation of resource consumption corresponding to each task, the extension, shortening or backlog of task queues corresponding to each task, and the acceleration, slowdown or overrun of runtime corresponding to each task, thus forming adjustment judgment data.

[0062] S5.3.2. Based on the adjustment judgment data, the computing power share, video memory share, and storage access share in the target execution plan are increased or decreased to form resource adjustment data.

[0063] Specifically, adjustments are made based on the node load change status, resource usage change status, task queuing change status, and runtime change status identified in the judgment data. The resource allocation relationships in the target execution plan are adjusted accordingly. When the node load increases, the task queue lengthens, or the runtime increases more rapidly, the computing power share, video memory share, and storage access share corresponding to the current task are increased. When the node load decreases, resource usage decreases, or the task queue lengthens, the computing power share, video memory share, and storage access share corresponding to the current task are decreased. When resource usage or node load fluctuates, the computing power share, video memory share, and storage access share are increased or decreased synchronously according to the actual changes in usage during the execution of the current task, forming resource adjustment data.

[0064] S5.3.3. Based on the resource adjustment data, the carrier nodes in the target execution plan are switched, adjusted and updated to form path adjustment data. The resource allocation adjustment results in the resource adjustment data are recorded in correspondence with the carrier node switching results and task execution path update results in the path adjustment data to form adjustment execution data.

[0065] Specifically, the process reads the computing power share adjustment results, video memory share adjustment results, and storage access share adjustment results corresponding to the current task from the resource adjustment data. Then, the bearer nodes that still cannot meet the execution requirements of the current task after resource adjustment are removed from the target execution plan. The bearer nodes that match the resource adjustment results are selected from the remaining bearer nodes corresponding to the target execution plan to take over the current task execution position, thus completing the bearer node switching adjustment. During the synchronization update, the resource allocation relationship and task execution path corresponding to the switch of the carrier node are rewritten to the corresponding content after the switch, so that the carrier node, resource allocation relationship and task execution path corresponding to the current task are consistent, forming path adjustment data; then, based on the resource adjustment data and path adjustment data, the resource adjustment result, carrier node switch result and task execution path update result corresponding to the current task are written into the adjustment execution record to generate adjustment execution data.

[0066] It should be noted that, Figure 5 The cumulative distribution of actual task runtime under different scheduling schemes is shown. The method group of the present invention represents the complete process of scheduling using the small-volume intelligent computing resource pooling scheduling method of the present invention. The "No Dynamic Adjustment" group means that after completing the initial resource status list construction, task profile list generation, candidate scheduling scheme list screening, target execution scheme formation, and task issuance, each task is executed only according to the initially determined bearing nodes, resource allocation relationships, and task execution paths. During task execution, although resource occupancy status, task queuing status, runtime status, and node load status can still be recorded, adjustment judgment data is no longer formed based on operation monitoring data, nor is the resource allocation relationship in the target execution scheme readjusted. The computing power share, video memory share, and storage access share of the current task are not increased or decreased, and bearing node switching adjustments and path synchronization updates are not performed. The group that does not screen out high-disturbance schemes means that after completing the initial resource status list construction, task profile list generation, resource partitioning and adaptation, and node carrying matching, although a set of carrying schemes is still formed and subsequent target execution scheme decisions can continue, the group does not identify disturbances in the resource occupancy changes, task queuing changes, and running status changes corresponding to each carrying scheme, does not generate scheme disturbance judgment data, or even if relevant data is generated, it does not remove carrying schemes with high disturbance levels based on this data, but allows high-disturbance schemes and low-disturbance schemes to enter the subsequent comparison, sorting, and selection process together; Depend on Figure 5 As can be seen, the cumulative distribution curve corresponding to the method group of the present invention is located to the left of the other two groups of curves, indicating that under the same proportion of task samples, the actual running time of the task group of the present invention is shorter. The present invention continuously monitors the task execution process and generates adjustment judgment data based on the operation monitoring data, dynamically adjusts the resource allocation relationship and carrying nodes in the target execution plan, and filters out high disturbance schemes in the pre-scheduling stage, which can effectively shorten the actual running time of the task and improve the timeliness of task execution.

[0067] S6. Based on the execution result package, perform resource reclamation and scheduling feedback updates on the target execution plan, revise subsequent scheduling strategies, and form the resource status list for the next round.

[0068] S6.1. Based on the execution result package, release the carrier nodes and resource allocation relationships corresponding to the target execution plan to form resource reclamation data, and provide feedback and organize the task completion status, resource occupation change status and adjustment execution status during the current task execution process to form scheduling feedback data.

[0069] Specifically, based on the execution result package, the task distribution record, resource occupancy status, and adjusted execution results corresponding to this round of tasks are read item by item, and the corresponding locations are determined according to the carrier nodes and resource allocation relationships in the target execution plan; the tasks that have been completed are removed from the task execution relationship of the corresponding carrier nodes, and the computing power share, video memory share, and storage access share that have been allocated to the corresponding tasks are released from the resource allocation relationship in the target execution plan, so that the computing power share, video memory share, and storage access share that have been occupied are restored to a resource state that can participate in subsequent scheduling, forming resource reclamation data; The task completion status, resource usage change status, and adjustment execution status in the execution result package are organized according to individual tasks to form scheduling feedback data.

[0070] S6.2. Based on the scheduling feedback data, the resource allocation order, node selection order, and task scheduling order are corrected to form strategy correction data. The resource reclamation data, scheduling feedback data, and strategy correction data are merged to form the resource status list for the next round.

[0071] Specifically, based on the task completion status, resource usage change status, and adjustment execution status of each task in the scheduling feedback data, the resource allocation order, node selection order, and task scheduling order are corrected. For bearer nodes with stable resource usage changes and normal task completion, the corresponding resource allocation order is moved forward; for bearer nodes with large resource usage fluctuations or multiple adjustment executions, the corresponding resource allocation order is moved backward. For bearer nodes with high task completion efficiency and few adjustments, the corresponding node selection order is moved forward; for bearer nodes with significant task queuing backlog or frequent bearer switching, the corresponding node selection order is moved backward. For tasks with high priority and good completion time, the corresponding task scheduling order is moved forward; for tasks with long execution time and consistently high resource usage, the corresponding task scheduling order is moved backward, thus forming strategy correction data. Based on resource recovery data, scheduling feedback data, and strategy correction data, the available resource status after release, the task execution status after feedback, and the corrected scheduling order are written into the same round of status record to form the resource status list for the next round.

[0072] It should be noted that, as Figure 6 The diagram shows the changes in feedback effects of the method group of the present invention during multi-round scheduling, where one curve represents the on-time completion rate and the other curve represents the average task waiting time.

[0073] The method group of this invention describes the complete process of scheduling using the small-scale intelligent computing resource pooling scheduling method described in this invention. This involves filtering out high-perturbation schemes after the candidate scheduling scheme list is formed, dynamically adjusting based on operational monitoring data during task execution, and generating scheduling feedback data based on the execution result package after each round of tasks to correct the resource allocation order, node selection order, and task scheduling order. As the scheduling rounds progress, the on-time completion rate remains at a high level, and the average task waiting time shows a decreasing or stable optimization trend. This indicates that this invention can continuously correct subsequent scheduling strategies by updating scheduling feedback during multiple rounds of scheduling, making the resource status list of the next round more consistent with the resource and task status changes in the current small-scale intelligent computing environment, thus demonstrating good adaptive scheduling capabilities and stable output capabilities.

[0074] This embodiment also provides a computer device applicable to the small-scale intelligent computing resource pooling scheduling method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the small-scale intelligent computing resource pooling scheduling method proposed in the above embodiment.

[0075] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0076] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for pooling and scheduling small-scale intelligent computing resources as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0077] In summary, this invention, by continuously monitoring the task execution process and generating adjustment judgment data in real time based on multi-dimensional operational monitoring data, can effectively address common resource contention and node performance fluctuations in cloud computing environments, improve the success rate and timeliness of single tasks and even mixed task flows, and ensure that high-quality and highly stable computing service output can still be maintained in resource-constrained cloud-edge collaborative small-scale intelligent computing environments.

[0078] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for pooling and scheduling small-scale intelligent computing resources, characterized in that, include: Step S1: Collect, pool, map, and organize the heterogeneous resources in the small-scale intelligent computing environment, and output the initial resource status list; Step S2: Based on the initial resource status list, perform semantic parsing and resource requirement identification on real-time business requests, and organize task attributes to generate a task profile list; Step S3: Based on the task profile list and the initial resource status list, perform resource partitioning and adaptation and node carrying matching, filter out high-disturbance schemes, and output a candidate scheduling scheme list; Step S4: Compare and optimize the candidate scheduling scheme list in real time to form the target execution scheme; Step S5: Based on the target execution plan and the task profile list, tasks are assigned and the execution process is controlled, and dynamic adjustments are made during execution to obtain the execution result package; Step S6: Based on the execution result package, perform resource reclamation and scheduling feedback updates on the target execution plan, revise the subsequent scheduling strategy, and form the resource status list for the next round.

2. The method for pooling and scheduling small-scale intelligent computing resources as described in claim 1, characterized in that, The specific steps for collecting, pooling, mapping, and organizing the heterogeneous resources in a small-scale intelligent computing environment, and outputting an initial resource status list, are as follows: Resource detection is performed on computing nodes in a small-scale intelligent computing environment to generate raw resource collection data. Different types of resources in the raw resource collection data are uniformly identified to generate resource identification data. Based on resource identification data, heterogeneous resources scattered across different computing nodes are pooled and mapped to construct a resource pool structure under a unified scheduling view, forming resource pool mapping data. Based on the resource pool mapping data, the availability of resources on each computing node is determined and the status is marked to form resource availability status data. The resource availability status data is then merged and organized according to the node dimension, resource dimension, and service dimension to output an initial resource status list.

3. The method for pooling and scheduling small-scale intelligent computing resources as described in claim 2, characterized in that, The steps for semantic parsing and resource requirement identification of real-time business requests based on the initial resource status list, and for task attribute organization to generate a task profile list are as follows: Receive real-time business requests and combine them with the initial resource status list to form basic data for task parsing. Based on the basic data for task parsing, perform semantic parsing and resource requirement identification on the real-time business requests to form task type identification data and task resource requirement data. Based on task resource requirement data, task priority is determined and operational constraints are identified for real-time business requests, forming task constraint data. The task type identification data, task resource requirement data, and task constraint data are then processed to generate a task profile list.

4. The method for pooling and scheduling small-scale intelligent computing resources as described in claim 3, characterized in that, The steps for resource allocation and adaptation and node capacity matching based on the task profile list and the initial resource status list, and for filtering out high-disturbance schemes to output a candidate scheduling scheme list are as follows: Each task in the task profile list is read one by one to obtain the resource information of the node to be matched, and the available resource status of each computing node is extracted by combining the initial resource status list to form the basic data for resource matching. Based on the basic resource matching data, the resource requirements corresponding to each task are segmented and adapted to form resource segmentation data. Based on the resource segmentation data, the node carrying capacity of each computing node is matched to form node matching data. Based on the node matching data, disturbances are identified for each carrying scheme, forming scheme disturbance judgment data. The scheme disturbance judgment data is then filtered and processed to output a list of candidate scheduling schemes.

5. The method for pooling and scheduling small-scale intelligent computing resources as described in claim 4, characterized in that, The specific steps for real-time comparison and optimization of the candidate scheduling scheme list to form the target execution scheme are as follows: The response time, resource adaptability and operational stability of each candidate scheduling scheme in the candidate scheduling scheme list are compared to form scheme comparison data. Based on the scheme comparison data, the candidate scheduling schemes are sorted, filtered, and prioritized to form scheme decision data. Based on the scheme decision data, the top-ranked candidate scheduling schemes are selected and combined to form the target execution scheme.

6. The method for pooling and scheduling small-scale intelligent computing resources as described in claim 5, characterized in that, The process of task assignment and execution control based on the target execution plan and task profile list, with dynamic adjustments made during execution, yields an execution result package. The specific steps are as follows: According to the target execution plan, each task in the task profile list is sent to the corresponding carrier node, and the execution association between the task and the carrier node is established to form task sending data; Based on the task distribution data, the task execution process is monitored to generate operational monitoring data; The task execution process is dynamically adjusted based on the operation monitoring data to form adjustment execution data. The task distribution data, operation monitoring data, and adjustment execution data are packaged together to obtain the operation result package.

7. The method for pooling and scheduling small-scale intelligent computing resources as described in claim 6, characterized in that, The specific steps for dynamically adjusting the task execution process based on operational monitoring data to generate adjusted execution data are as follows: Based on operational monitoring data, identify the node load change status, resource consumption change status, task queuing change status, and runtime change status during task execution to form adjustment judgment data; Based on the adjustment judgment data, the computing power share, video memory share, and storage access share in the target execution plan are increased or decreased to form resource adjustment data; Based on the resource adjustment data, the carrier nodes in the target execution plan are switched, adjusted and updated to form path adjustment data. The resource allocation adjustment results in the resource adjustment data are recorded in correspondence with the carrier node switching results and task execution path update results in the path adjustment data to form adjustment execution data.

8. The method for pooling and scheduling small-scale intelligent computing resources as described in claim 6, characterized in that, The steps for reclaiming resources and updating scheduling feedback based on the execution result package, correcting subsequent scheduling strategies, and forming the resource status list for the next round are as follows: Based on the execution result package, the corresponding carrier nodes and resource allocation relationships of the target execution plan are released to form resource recovery data. The task completion status, resource occupation change status and adjustment execution status during this round of task execution are fed back and organized to form scheduling feedback data. Based on the scheduling feedback data, the resource allocation order, node selection order, and task scheduling order are corrected to form policy correction data. The resource reclamation data, scheduling feedback data, and policy correction data are then merged to form the resource status list for the next round.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the small-scale intelligent computing resource pooling scheduling method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the small-scale intelligent computing resource pooling scheduling method according to any one of claims 1 to 8.

Images