Multi-dimensional joint scheduling method for optimizing resource utilization rate of cluster computing network

Abstract

The invention provides a multi-dimensional joint scheduling method for optimizing the resource utilization rate of a cluster computing network, and the method comprises the steps: building a resource fragmentation measurement index system through collecting various types of resource data in a distributed cluster; dividing a training and pushing task into four stages, modeling communication behaviors of each stage, and quantifying and analyzing communication characteristics of each stage; analyzing the communication time axis of the operation in the scheduling window, identifying potential bandwidth conflicts, and adjusting the starting time of the communication stage when the bandwidth conflicts are detected in combination with the task dependency relationship; evaluating the resource adaptability between the jobs in a structural mode of compatibility scoring; constructing an optimal job-node mapping strategy, constructing a feedback error function, and continuously optimizing the resource fragmentation measurement index system; according to the method, the overall distribution condition of cluster resources can be effectively evaluated, GPU fragments can be reduced, higher adaptability and scheduling efficiency are achieved, and the utilization rate of computing network resources can be improved in a large-scale complex cluster environment.

Description

Technical Field

[0001] The present invention relates to a multi-dimensional joint scheduling method for optimizing the utilization efficiency of computing network resources, belonging to the technical field of computer resource scheduling and cluster optimization. Background Art

[0002] As model scale continues to expand, distributed clusters have become critical infrastructure for high-performance computing. Cluster scheduling, as a core component of resource management, directly impacts task execution efficiency and system resource utilization. Furthermore, cluster operations have multidimensional resource requirements, encompassing compute, memory, and network bandwidth. Therefore, how to comprehensively consider compute, memory, and network variations during cluster scheduling and build a more refined and dynamic scheduling model has become a key issue in GPU cluster management.

[0003] Traditional scheduling methods, such as load balancing and priority scheduling, primarily focus on computing resource allocation, lacking coordinated optimization of computing network resources. These methods rely solely on single-dimensional decision-making and ignore the coupling between resources. In practical applications, this can lead to a mismatch between task demand and resource supply, resulting in resource conflicts and redundant allocations. Binpack Scheduling attempts to optimize resource distribution by introducing fragmentation metrics and node load assessments. However, it still fails to fully consider resource affinity and compatibility between jobs, resulting in a lack of flexibility and adaptability in scheduling strategies and limited effectiveness in large-scale systems. Summary of the Invention

[0004] The purpose of the present invention is to provide a multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization to solve the problems of computing resource fragmentation and low multi-dimensional resource scheduling efficiency existing in the prior art.

[0005] The technical solution of the present invention is:

[0006] A multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization includes the following steps:

[0007] S1. Collect various resource data in the distributed cluster, store them in a structured form in a resource status table, and build a resource fragmentation measurement indicator system to evaluate the current resource distribution of the cluster;

[0008] S2. Divide the training and inference task into four phases: input data loading, forward propagation, backpropagation, and parameter synchronization. Detailed modeling is performed on the communication behavior of each phase. Combined with the job type, model structure, and historical execution records, the peak and duration of the communication load are recorded to quantify and analyze the communication characteristics of each phase.

[0009] S3: Analyze the communication timeline of jobs within the scheduling window, identify potential bandwidth conflicts, model the total bandwidth demand, and combine task dependencies to trigger a time-shift scheduling strategy when a bandwidth conflict is detected, adjusting the start time of the communication phase.

[0010] S4. Use the affinity graph model to build a job resource compatibility graph, and use a structured compatibility score to evaluate the resource adaptability between jobs to optimize the task scheduling and allocation plan;

[0011] S5. Combine the affinity graph model and the resource fragmentation measurement indicator system to analyze the resource dimension coupling relationship between jobs, determine the node computing network resource constraints, build the optimal job-node mapping strategy, schedule tasks, and collect task execution data during operation to build a feedback error function, provide feedback, and continuously optimize the resource fragmentation measurement indicator system.

[0012] Furthermore, step S1 is specifically:

[0013] S11. Collect various resource data in the distributed cluster, including the computing resources of each GPU node and the resource requirements of the job. The computing resources of the GPU node include video memory and the number of CUDA cores, and the memory usage of each node; the resource requirements of the job include computing load and memory usage;

[0014] S12. Storing the structured cluster resource information in a resource status table in a structured form;

[0015] S13. Establish a resource fragmentation measurement indicator system based on the structured cluster resource information.

[0016] Furthermore, in step S13, the resource fragmentation measurement index system uses computing resource fragmentation and memory resource fragmentation to evaluate the current resource fragmentation distribution of the cluster, specifically,

[0017] S131. Calculate resource fragmentation The expression is: , Where, the utilization of CUDA core is : , existing continuity indicators : , is a weight parameter, and ;

[0018] S132, memory resource fragmentation The expression is: , Where, the discreteness of the memory block : , node memory usage balance : , is the standard deviation, is the weight coefficient, and ;

[0019] S133, calculating resource fragmentation and memory resource fragmentation The current resource fragmentation is obtained by comprehensively considering the fragmentation of : , Where, is the weight coefficient, and .

[0020] Furthermore, in step S2, detailed modeling is performed for the communication behavior at each stage, specifically,

[0021] S21. The expression of bandwidth occupancy in each stage is: , Where, Indicates homework In time Bandwidth usage, Indicates homework In the Bandwidth usage of each stage;

[0022] S22. Model the communication behavior at each stage: , Where, For homework In the Bandwidth requirements for each stage, For homework In the The duration of each phase, For homework No. The relative position of each stage in the training cycle; The entire communication model : ;

[0023] Furthermore, in step S3, when identifying potential bandwidth conflicts, it is detected whether there is excessive communication aggregation during the parameter synchronization phase, specifically,

[0024] Key metrics collected include bandwidth usage of each job, number of jobs performing simultaneous parameter synchronization, network traffic distribution, and communication delay changes;

[0025] When the number of jobs performing parameter synchronization at the same time exceeds the set ratio of the total number of jobs, or the bandwidth utilization exceeds the set value, or the network delay increases beyond the set threshold, the aggregation is triggered. calculate: , where is the number of simultaneous synchronization jobs, is the total number of jobs, is the current bandwidth usage, is the total bandwidth;

[0026] In the concentration When it is higher than the set value, there is excessive communication aggregation during the parameter synchronization phase; otherwise, there is no excessive communication aggregation during the parameter synchronization phase.

[0027] Furthermore, in step S3, the total bandwidth demand is modeled and combined with the task dependency. When a bandwidth conflict is detected, the time offset scheduling strategy is triggered to adjust the start time of the communication phase. Specifically,

[0028] S31. The expression of bandwidth occupancy in each stage is: , Where, It's time The sum of bandwidth requirements of all active job stages within is the number of jobs running simultaneously, For homework In the Bandwidth requirements for each stage, It's homework No. The start time of each communication phase, For homework In the The duration of each phase, ( ) is an indicator function, which is 1 when the condition is met and 0 otherwise;

[0029] S32. When a bandwidth conflict is detected, the time offset scheduling strategy is triggered, and the expression for adjusting the start time of the communication phase is: , Where, Indicates the start time of the calculation time interval, Indicates the end time of the calculation time interval, It's time The sum of bandwidth requirements of all active job stages within Indicates the maximum bandwidth capacity of the network, Indicates that only when When , it will take a positive value, otherwise it will be 0; By adjusting all jobs Minimize the above integral at the start time of each stage i; The task dependency relationship is: ,in, It's homework The start time of the i-1th communication phase, For homework The duration of the phase at the i-1th phase.

[0030] Furthermore, step S4 is specifically:

[0031] S41. Use the affinity graph model to build a job resource compatibility graph: , Where, is the affinity graph model, There are n job nodes A collection of are any two job nodes in set V The edge between A collection of is a collection Edge Weight The set of represents the resource coupling degree between jobs; the weight , indicating the job and homework The affinity between them is normalized so that ,in, is a node To Node The communication needs between It's homework its own internal communication needs, Yes, it's homework its own internal communication needs;

[0032] S42. The expression for quantifying the resource fit between tasks in the affinity graph using compatibility scores is: , Where, Is the compatibility score between the job and the node, calculating the resource matching degree ,in, It's homework The amount of computing resources required, is a node Available computing resource margin, max function comparison operation Required computing resources and nodes The available computing resource margin, the larger value is used for normalization; is the memory resource matching degree, is the network resource matching degree, The rating is based on historical data. is the weight coefficient, and .

[0033] Furthermore, in step S5, the node computing network resource constraints are determined and the optimal job-node mapping strategy is constructed, specifically,

[0034] The objective function of the optimal job-node mapping strategy is: , in, is the total number of jobs, is the total number of nodes, is a decision variable, when the job Assigned to a node 1 when it is, otherwise 0; is a node The resource fragmentation, is the balancing factor;

[0035] The expression of node computing network resource constraint is:

[0036] Node computing resource constraints:

[0037]

[0038] in, It's homework The available computing resources margin, is a node Available computing resource margin; where is the total number of jobs, is the total number of nodes, Represents the current node;

[0039] Node memory resource constraints:

[0040]

[0041] in, It's homework Available memory resource margin, is a node Available memory resource margin;

[0042] Node network resource constraints:

[0043] .

[0044] in, It's homework Available network resource margin, is a node The available network resource margin.

[0045] Furthermore, in step S5, task execution data including job completion time, resource utilization, and resource conflict information are collected to construct a feedback error function, which is expressed as:

[0046]

[0047]

[0048]

[0049] in, 、 are the computing resource weights of rounds t+1 and t respectively, is the learning rate, is the weight coefficient, which is used to control the optimization step size. It is based on the collected task execution data including the job completion time , resource utilization , resource conflicts The feedback error function constructed, is the memory resource weight of round t+1, It is an adjustment coefficient used to balance the influencing factors.

[0050] The beneficial effects of the present invention are:

[0051] 1. This multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization can effectively evaluate the overall distribution of cluster resources, reduce GPU fragmentation, has stronger adaptability and scheduling efficiency, and can improve computing network resource utilization in large-scale and complex cluster environments.

[0052] 2. The present invention, when processing network resources, analyzes the bandwidth requirements and timing patterns of the communication phase between jobs and dynamically adjusts the start time of the communication phase of the job. While ensuring the dependency between jobs, it can stagger the overlapping windows of the high-load communication phase as much as possible, thereby effectively avoiding network congestion problems, improving the overall operation efficiency of the cluster, and having stronger adaptability and scheduling efficiency.

[0053] 3. This multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization establishes a scheduling decision model based on the multi-dimensional resource portrait of cluster nodes. It selects the optimal node for scheduling jobs through a fragmentation-aware node optimization algorithm and resource distribution. This can effectively reduce GPU fragmentation and effectively improve the overall system throughput and resource utilization.

[0054] 4. Compared with the existing distributed cluster scheduling method, the present invention realizes the joint consideration of different resource types, breaks through the limitations of the traditional "computing priority and passive adaptation of other resources", realizes the global optimization of computing network resources, and effectively improves the overall cluster throughput and task execution efficiency. It is particularly suitable for high-load, multi-task parallel distributed training environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of a multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to an embodiment of the present invention;

[0056] Figure 2 3 is a schematic diagram comparing the bandwidth requirements of each stage of a job before and after adjusting the communication stage according to task resource adaptability in an embodiment, wherein (a) is a schematic diagram of the bandwidth requirements of each stage of a job before adjustment, and (b) is a schematic diagram of the bandwidth requirements of each stage of a job after adjustment;

[0057] Figure 3 3 is a schematic diagram comparing the effects before and after adjusting the communication phase of a job in the event of a bandwidth conflict in an embodiment, wherein (a) is a graph showing the bandwidth usage of the job before adjustment, and (b) is a graph showing the bandwidth usage of the job after adjustment;

[0058] Figure 4 This is a schematic diagram illustrating a multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to an embodiment. DETAILED DESCRIPTION

[0059] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0060] The embodiment provides a multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization, such as Figure 1 and Figure 4 , including the following steps:

[0061] S1. Collect various resource data in the distributed cluster, store them in a structured form in the resource status table, and build a resource fragmentation measurement indicator system to evaluate the current resource distribution of the cluster.

[0062] S11. Collect various resource data in the distributed cluster, including the computing resources of each GPU node and the resource requirements of the job. The computing resources of the GPU node include video memory and the number of CUDA cores, and the memory usage of each node; the resource requirements of the job include computing load and memory usage;

[0063] S12. Storing the structured cluster resource information in a resource status table in a structured form;

[0064] S13. Establish a resource fragmentation measurement indicator system based on the structured cluster resource information.

[0065] In step S13, the resource fragmentation measurement index system uses computing resource fragmentation and memory resource fragmentation to evaluate the current resource fragmentation distribution of the cluster, specifically,

[0066] S131. Calculate resource fragmentation The expression is:

[0067]

[0068] Where, the utilization of CUDA core is : , existing continuity indicators : , is a weight parameter, and ;

[0069] In step S131, the resource fragmentation is calculated using Evaluate whether the memory space is scattered and occupied to prevent task deployment from being affected by memory fragmentation.

[0070] S132, memory resource fragmentation The expression is:

[0071]

[0072] Where, the discreteness of the memory block : , node memory usage balance : , is the standard deviation, is the weight coefficient, and .

[0073] In step S132, the memory resource fragmentation is used Evaluate whether the memory space is scattered and occupied to prevent task deployment from being affected by memory fragmentation.

[0074] S133, calculating resource fragmentation and memory resource fragmentation The current resource fragmentation is obtained by comprehensively considering the fragmentation of :

[0075]

[0076] Where, is the weight coefficient, and .

[0077] In step S1, a scheduling decision model is established based on the multi-dimensional resource portrait of the cluster nodes. The optimal node is selected for scheduling jobs through the fragmentation-aware node optimization algorithm and resource distribution. This can effectively reduce GPU fragmentation and effectively improve the overall system throughput and resource utilization.

[0078] S2. Divide the training and inference task into four phases: input data loading, forward propagation, backpropagation, and parameter synchronization. Detailed modeling is performed on the communication behavior of each phase. Combined with the job type, model structure, and historical execution records, the peak and duration of the communication load are recorded to quantify and analyze the communication characteristics of each phase.

[0079] In step S2, detailed modeling is performed for the communication behavior at each stage, specifically,

[0080] S21. The expression of bandwidth occupancy in each stage is:

[0081]

[0082] Where, Indicates homework In time Bandwidth usage, Indicates homework In the Bandwidth usage of each stage;

[0083] S22. Model the communication behavior at each stage:

[0084]

[0085] Where, For homework In the Bandwidth requirements for each stage, For homework In the The duration of each phase, For homework No. The relative position of each stage in the training cycle; The entire communication model : .

[0086] In step S2, the input data loading phase (Data Loading) mainly reads training data from storage, with relatively low bandwidth usage but potentially affected by I / O bottlenecks; the forward propagation phase (Forward Propagation) is a computational phase with low communication requirements; the backward propagation phase (Backward Propagation) involves gradient calculation and generation of updated parameters, which increases communication requirements; the parameter synchronization phase (such as the AllReduce operation) has the highest communication requirements of the four phases and is very likely to cause network congestion.

[0087] S3. Analyze the communication timeline of jobs within the scheduling window, identify potential bandwidth conflicts, model the total bandwidth demand, and combine task dependencies. When bandwidth conflicts are detected, trigger the time offset scheduling strategy and adjust the start time of the communication phase.

[0088] In step S3, when identifying potential bandwidth conflicts, it is detected whether there is excessive communication aggregation during the parameter synchronization phase, specifically,

[0089] Key metrics collected include bandwidth usage of each job, number of jobs performing simultaneous parameter synchronization, network traffic distribution, and communication delay changes;

[0090] When the number of jobs performing parameter synchronization at the same time exceeds the set ratio of the total number of jobs, or the bandwidth utilization exceeds the set value, or the network delay increases beyond the set threshold, the aggregation is triggered. calculate: , where is the number of simultaneous synchronization jobs, is the total number of jobs, is the current bandwidth usage, is the total bandwidth;

[0091] In the concentration When it is higher than the set value, there is excessive communication aggregation during the parameter synchronization phase; otherwise, there is no excessive communication aggregation during the parameter synchronization phase.

[0092] In step S3, when identifying potential bandwidth conflicts, focus on whether there is excessive communication aggregation during the parameter synchronization phase. Monitoring probes are mainly deployed on each computing node and network device, and key indicators including bandwidth usage of each job, the number of jobs performing parameter synchronization at the same time, network traffic distribution, and communication delay changes are continuously collected through a sliding time window (1-5 seconds). These data will be aggregated to the central monitoring system in real time to form a dynamic view of the global network communication status. On this basis, a hierarchical response mechanism is set according to historical data: when the number of jobs performing parameter synchronization at the same time exceeds 60% of the total number of jobs, or the bandwidth utilization exceeds 75%, or the network delay increases by more than 2 times the normal value, the aggregation level is triggered. Calculate the degree of aggregation When the load exceeds the set value, the time offset scheduling strategy is triggered. By delaying the parameter synchronization of some jobs, prioritizing bandwidth resources for critical jobs, and distributing communication traffic across multiple network paths, the startup time of the job communication phase is dynamically adjusted. This effectively avoids overlapping high-load communication phases, achieves a reasonable allocation of network resources, and improves overall system efficiency.

[0093] In step S3, the total bandwidth demand is modeled and combined with the task dependencies. When a bandwidth conflict is detected, the start time of the communication phase is adjusted. Specifically,

[0094] S31. The expression of bandwidth occupancy in each stage is:

[0095]

[0096] Where, It's time The sum of bandwidth requirements of all active job stages within is the number of jobs running simultaneously, For homework In the Bandwidth requirements for each stage, It's homework No. The start time of each communication phase, For homework In the The duration of each phase, ( ) is an indicator function, which is 1 when the condition is met and 0 otherwise;

[0097] S32. When a bandwidth conflict is detected, the time offset scheduling strategy is triggered, and the expression for adjusting the start time of the communication phase is:

[0098]

[0099] Where, Indicates the start time of the calculation time interval, Indicates the end time of the calculation time interval, It's time The sum of bandwidth requirements of all active job stages within Indicates the maximum bandwidth capacity of the network, Indicates that only when When , it will take a positive value, otherwise it will be 0; By adjusting all jobs Minimize the above integral at the start time of each stage i;

[0100] The task dependency relationship is: ,in, It's homework The start time of the i-1th communication phase, For homework The duration of the phase at the i-1th phase.

[0101] In step S3, due to the fact that the communication behavior of jobs is repeated periodically during the model training process, and their bandwidth occupancy patterns are different, when the parameters of multiple jobs are synchronized at the same time, bandwidth contention will occur, significantly reducing the efficiency of job execution. To alleviate this problem, this method adopts a communication time offset scheduling strategy to dynamically adjust the start time of the job communication phase according to the communication requirements and timing pattern of the job, avoiding the concurrent operation of high-load communication tasks. Figure 2 (a) is the bandwidth requirement diagram for each stage of the job before adjustment. Figure 2 (b) is the bandwidth requirement diagram of each stage of the job after adjustment. Figure 2 As can be seen from (a) and (b), after adjustment, it is possible to effectively avoid high-communication demand phases between tasks and avoid overlapping high-load communication phases. High-load communication tasks are those that generate a large amount of data exchange, communicate frequently, and are highly sensitive to network bandwidth and communication delays during execution. In the processing of network resources, the present invention analyzes the bandwidth demand and timing pattern of the communication phase between jobs and dynamically adjusts the start time of the communication phase of the job. Figure 3 (a) is the operation bandwidth usage curve before adjustment. Figure 3 (b) is the adjusted operation bandwidth usage curve, which is Figure 3 As can be seen from (a) and (b), after adjustment, while ensuring the dependencies between jobs, the overlapping windows of the high-load communication phase can be staggered as much as possible, thereby effectively avoiding network congestion problems, improving the overall operation efficiency of the cluster, and having stronger adaptability and scheduling efficiency.

[0102] S4. Use the affinity graph model to build a job resource compatibility graph, and evaluate the resource adaptability between jobs in a structured way using compatibility scores to optimize the task scheduling and allocation plan.

[0103] S41. Use the affinity graph model to build a job resource compatibility graph:

[0104]

[0105] Where, is the affinity graph model, There are n job nodes A collection of are any two job nodes in set V The edge between A collection of is a collection Edge Weight The set of represents the resource coupling between jobs; the weight , indicating the job and homework The affinity between them is normalized so that ,in, is a node To Node The communication needs between It's homework its own internal communication needs, Yes, it's homework its own internal communication needs;

[0106] S42. The expression for quantifying the resource fit between tasks in the affinity graph using compatibility scores is:

[0107]

[0108] Where, Is the compatibility score between the job and the node, calculating the resource matching degree ,in, It's homework The amount of computing resources required, is a node Available computing resource margin, max function comparison operation Required computing resources and nodes The available computing resource margin is used for normalization, which measures the relative mismatch between computing demand and available capacity. is the memory resource matching degree, is the network resource matching degree, The rating is based on historical data. is the weight coefficient, and .

[0109] In step S4, due to the varying requirements of different jobs for computing, memory, and communication resources, the present invention discovered that some jobs can coexist with each other in a complementary manner. Therefore, the present invention introduces an affinity graph model to construct a job-resource compatibility graph, which allows for a structured assessment of resource compatibility between jobs and optimizes task allocation.

[0110] S5. Combine the affinity graph model and the resource fragmentation measurement indicator system to analyze the resource dimension coupling relationship between jobs, determine the node computing network resource constraints, build the optimal job-node mapping strategy, schedule tasks, and collect task execution data during operation to build a feedback error function, provide feedback, and continuously optimize the resource fragmentation measurement indicator system.

[0111] In step S5, the node computing network resource constraints are determined and the optimal job-node mapping strategy is constructed. Specifically,

[0112] The objective function of the optimal job-node mapping strategy is:

[0113]

[0114] in, is the total number of jobs, is the total number of nodes, is a decision variable, when the job Assigned to a node 1 when it is, otherwise 0; is a node The resource fragmentation, is the balancing factor;

[0115] The expression of node computing network resource constraint is:

[0116] Node computing resource constraints:

[0117]

[0118] in, It's homework The available computing resources margin, is a node Available computing resource margin; where is the total number of jobs, is the total number of nodes, Represents the current node;

[0119] Node memory resource constraints:

[0120]

[0121] in, It's homework Available memory resource margin, is a node Available memory resource margin;

[0122] Node network resource constraints:

[0123] .

[0124] in, It's homework Available network resource margin, is a node The available network resource margin.

[0125] In step S5, the task execution data including job completion time, resource utilization, and resource conflict information are collected to construct a feedback error function, which is expressed as: , , , in, 、 are the computing resource weights of rounds t+1 and t respectively, is the learning rate, is the weight coefficient, which is used to control the optimization step size. It is based on the collected task execution data including the job completion time , resource utilization , resource conflicts The feedback error function constructed, is the memory resource weight of round t+1, It is an adjustment coefficient used to balance the influencing factors.

[0126] This multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization can effectively evaluate the overall distribution of cluster resources, reduce GPU fragmentation, has stronger adaptability and scheduling efficiency, and can improve computing network resource utilization in large-scale complex cluster environments.

[0127] This multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization constructs a joint decision-making framework, and combines fragmentation-aware node optimization strategies and real-time resource distribution characteristics to achieve efficient job scheduling. This method can reduce resource fragmentation while accurately matching GPU resource requirements. In terms of network resource optimization, it innovatively introduces a dynamic orchestration mechanism for communication timing. By analyzing the bandwidth requirements and timing dependencies of the inter-job communication phase, it intelligently adjusts the phase offset of the key communication window, and reduces the risk of network peak conflicts while ensuring the logical order of tasks. The present invention breaks through the limitations of the single resource dimension of the traditional scheduling system, adopts a collaborative optimization strategy of computing power and network, and significantly enhances the throughput capacity of the cluster in high-concurrency training scenarios through multi-dimensional resource modeling and global optimization scheduling. It is particularly suitable for the complex resource scheduling requirements of large-scale AI training environments.

[0128] This multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization establishes a scheduling decision model based on the multi-dimensional resource portrait of cluster nodes. It selects the optimal node for scheduling jobs through a fragmentation-aware node optimization algorithm and resource distribution. This can effectively reduce GPU fragmentation while effectively improving the overall system throughput and resource utilization.

[0129] Compared with existing distributed cluster scheduling methods, the present invention realizes the joint consideration of different resource types, breaks through the limitations of the traditional "computing priority and passive adaptation of other resources", realizes the global optimization of computing network resources, and effectively improves the overall cluster throughput and task execution efficiency. It is particularly suitable for high-load, multi-task parallel distributed training environments.

[0130] The above description is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiment. Any equivalent modifications or changes made by ordinary technicians in this field based on the contents disclosed in the present invention should be included in the protection scope recorded in the claims.

Claims

1. A multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization, characterized by: The following steps are included: S1. Collect various resource data in the distributed cluster, store them in a structured form in a resource status table, and build a resource fragmentation measurement indicator system to evaluate the current resource distribution of the cluster; S2. Divide the training and inference task into four phases: input data loading, forward propagation, backpropagation, and parameter synchronization. Detailed modeling is performed on the communication behavior of each phase. Combined with the job type, model structure, and historical execution records, the peak and duration of the communication load are recorded to quantify and analyze the communication characteristics of each phase. S3: Analyze the communication timeline of jobs within the scheduling window, identify potential bandwidth conflicts, model the total bandwidth demand, and combine task dependencies to trigger a time-shift scheduling strategy when a bandwidth conflict is detected, adjusting the start time of the communication phase. S4. Use the affinity graph model to build a job resource compatibility graph, and use a structured compatibility score to evaluate the resource adaptability between jobs to optimize the task scheduling and allocation plan; S5. Combine the affinity graph model and the resource fragmentation measurement indicator system to analyze the resource dimension coupling relationship between jobs, determine the node computing network resource constraints, build the optimal job-node mapping strategy, schedule tasks, and collect task execution data during operation to build a feedback error function, provide feedback, and continuously optimize the resource fragmentation measurement indicator system.

2. The multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to claim 1, characterized in that: Step S1 is specifically: S11. Collect various resource data in the distributed cluster, including the computing resources of each GPU node and the resource requirements of the job. The computing resources of the GPU node include video memory and the number of CUDA cores, and the memory usage of each node; the resource requirements of the job include computing load and memory usage; S12. Storing the structured cluster resource information in a resource status table in a structured form; S13. Establish a resource fragmentation measurement indicator system based on the structured cluster resource information.

3. The multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to claim 1, characterized in that: In step S13, the resource fragmentation measurement index system uses computing resource fragmentation and memory resource fragmentation to evaluate the current resource fragmentation distribution of the cluster, specifically, S131. Calculate resource fragmentation The expression is: ， Where, the utilization of CUDA core is : , existing continuity indicators : , is a weight parameter, and ; S132, memory resource fragmentation The expression is: ， Where, the discreteness of the memory block : , node memory usage balance : , is the standard deviation, is the weight coefficient, and ; S133, calculating resource fragmentation and memory resource fragmentation The current resource fragmentation is obtained by comprehensively considering the fragmentation of : ，， Where, is the weight coefficient, and .

4. The multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to any one of claims 1 to 3, characterized in that: In step S2, detailed modeling is performed for the communication behavior at each stage, specifically, S21. The expression of bandwidth occupancy in each stage is: ， Where, Indicates homework In time Bandwidth usage, Indicates homework In the Bandwidth usage of each stage; S22. Model the communication behavior at each stage: ， Where, For homework In the Bandwidth requirements for each stage, For homework In the The duration of each phase, For homework No. The relative position of each stage in the training cycle; The entire communication model : .

5. The multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to any one of claims 1 to 3, characterized in that: In step S3, when identifying potential bandwidth conflicts, it is detected whether there is excessive communication aggregation during the parameter synchronization phase, specifically, Key metrics collected include bandwidth usage of each job, number of jobs performing simultaneous parameter synchronization, network traffic distribution, and communication delay changes; When the number of jobs performing parameter synchronization at the same time exceeds the set ratio of the total number of jobs, or the bandwidth utilization exceeds the set value, or the network delay increases beyond the set threshold, the aggregation is triggered. calculate: , where is the number of simultaneous synchronization jobs, is the total number of jobs, is the current bandwidth usage, is the total bandwidth; In the concentration When it is higher than the set value, there is excessive communication aggregation during the parameter synchronization phase; otherwise, there is no excessive communication aggregation during the parameter synchronization phase.

6. The multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to any one of claims 1 to 3, characterized in that: In step S3, the total bandwidth demand is modeled and combined with the task dependencies. When a bandwidth conflict is detected, the time offset scheduling strategy is triggered to adjust the start time of the communication phase. Specifically, S31. The expression of bandwidth occupancy in each stage is: ， Where, It's time The sum of bandwidth requirements of all active job stages within is the number of jobs running simultaneously, For homework In the Bandwidth requirements for each stage, It's homework No. The start time of each communication phase, For homework In the The duration of each phase, ( ) is an indicator function, which is 1 when the condition is met and 0 otherwise; S32. When a bandwidth conflict is detected, the time offset scheduling strategy is triggered, and the expression for adjusting the start time of the communication phase is: ， Where, Indicates the start time of the calculation time interval, Indicates the end time of the calculation time interval, It's time The sum of bandwidth requirements of all active job stages within Indicates the maximum bandwidth capacity of the network, Indicates that only when When , it will take a positive value, otherwise it will be 0; By adjusting all jobs Minimize the above integral at the start time of each stage i; The task dependency relationship is: ,in, It's homework The start time of the i-1th communication phase, For homework The duration of the phase at the i-1th phase.

7. The multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to any one of claims 1 to 3, characterized in that: Step S4 is specifically: S41. Use the affinity graph model to build a job resource compatibility graph: ， Where, is the affinity graph model, There are n job nodes A collection of are any two job nodes in set V The edge between A collection of is a collection Edge Weight The set of represents the resource coupling degree between jobs; the weight , indicating the job and homework The affinity between them is normalized so that ,in, is a node To Node The communication needs between It's homework its own internal communication needs, Yes, it's homework its own internal communication needs; S42. The expression for quantifying the resource fit between tasks in the affinity graph using compatibility scores is: ， Where, Is the compatibility score between the job and the node, calculating the resource matching degree ,in, It's homework The amount of computing resources required, is a node Available computing resource margin, max function comparison operation Required computing resources and nodes The available computing resource margin, the larger value is used for normalization; is the memory resource matching degree, is the network resource matching degree, The rating is based on historical data. is the weight coefficient, and .

8. The multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to claim 7, characterized in that: In step S5, the node computing network resource constraints are determined and the optimal job-node mapping strategy is constructed. Specifically, The objective function of the optimal job-node mapping strategy is: ， in, is the total number of jobs, is the total number of nodes, is a decision variable, when the job Assigned to a node 1 when it is, otherwise 0; is a node The resource fragmentation, is the balancing factor; The expression of node computing network resource constraint is: Node computing resource constraints: ， in, It's homework The available computing resources margin, is a node Available computing resource margin; where is the total number of jobs, is the total number of nodes, Represents the current node; Node memory resource constraints: ， in, It's homework Available memory resource margin, is a node Available memory resource margin; Node network resource constraints: ， in, It's homework Available network resource margin, is a node The available network resource margin.

9. The multi-dimensional joint scheduling method for optimizing cluster computing network resource utilization according to claim 7, characterized in that: In step S5, the task execution data including job completion time, resource utilization, and resource conflict information are collected to construct a feedback error function, which is expressed as: ， ， ， in, 、 are the computing resource weights of rounds t+1 and t respectively, is the learning rate, is the weight coefficient, which is used to control the optimization step size. It is based on the collected task execution data including the job completion time , resource utilization , resource conflicts The feedback error function constructed, is the memory resource weight of round t+1, It is an adjustment coefficient used to balance the influencing factors.

Images