A computing cluster global load balancing scheduling method based on an EML operator

Abstract

The application discloses a kind of global load balancing scheduling methods of computing cluster based on EML operator, it is related to computing cluster scheduling technical field, including the following steps: constructing unified computing resource pool, virtualization to heterogeneous computing node and collecting CPU, memory, GPU, network IO and storage IO load, generate multidimensional load vector, receive computing task and extract feature, obtain resource requirement specification, multidimensional load vector is mapped to unified node load metric based on EML operator, global load balancing objective function is constructed by limited times nesting EML operation, the EML operator is defined as, solve global optimal task to node distribution mapping, schedule task according to mapping and monitor cluster, when meeting trigger condition, dynamically reschedule.The application models unified multidimensional load with EML operator, realizes global load balancing, and can be applied to government affairs computing, scientific research computing, industrial simulation, AI training and big data analysis and the like scenes.

Description

Technical Field

[0001] This invention relates to the field of computing cluster scheduling technology, and in particular to a global load balancing scheduling method for computing clusters based on EML operators, which can be applied to various industry scenarios such as government computing, university research computing, industrial simulation, AI training, and big data analysis. Background Technology

[0002] Currently, computing clusters need to simultaneously handle computing tasks from multiple industry scenarios. In the context of heterogeneous tasks deployed across multiple scenarios, existing scheduling methods mainly suffer from the following problems:

[0003] First, the scheduling models are not uniform: the load of multi-dimensional resources such as CPU, memory, GPU, and network needs to be modeled and scheduled using different algorithms or rules. There is a lack of a unified mathematical framework to integrate and process the multi-dimensional resource states, resulting in high complexity of the scheduling system.

[0004] Second, it is difficult to achieve globally optimal load balancing: Existing methods mostly use round-robin, greedy or heuristic algorithms to allocate tasks based on local information. They cannot optimize the overall load status of all nodes in the cluster at once, resulting in unbalanced load between nodes and limited utilization of computing resources.

[0005] Third, insufficient dynamic adaptive capability: When the task type, node performance or network status fluctuates, the existing scheduling system needs to rely on manual adjustment of parameters or strategies. It cannot automatically recalculate and converge to a new optimal allocation scheme based on the real-time changes in the global status of the cluster, which affects the stability of the system operation.

[0006] Therefore, in order to solve the problems existing in the prior art, a global load balancing scheduling method for computing clusters based on EML operators is needed. Summary of the Invention

[0007] The present invention aims to at least partially solve one of the technical problems in the above-mentioned technologies.

[0008] To achieve the above objectives, the first aspect of this invention proposes a global load balancing scheduling method for computing clusters based on EML operators, comprising the following steps:

[0009] S1: Construct a unified computing resource pool, virtualize resources of heterogeneous computing nodes (including CPU, GPU, memory, and network resources) within the cluster to form a unified computing unit, and collect CPU utilization, memory utilization, GPU utilization, network I / O load, and storage I / O load of each node in real time to generate a multi-dimensional load vector; S2: Receive computing tasks, extract features and identify the type of each task to obtain the resource requirement specifications corresponding to the task; S3: Based on EML operators, map the multi-dimensional load vector to a unified form of node load metric, and construct a global load balancing objective function through a finite number of nested EML operations; the EML operator is defined as... ,in It is a natural constant. The input variable is taken from the index value in the multidimensional load vector or its normalized value; S4: Taking minimizing the global load balancing objective function as the optimization objective, and combining the task resource requirement specification and the current load status of the node, solve the globally optimal task-to-node allocation mapping; S5: Schedule the task to the corresponding node for execution according to the allocation mapping, and continuously monitor the cluster status. When the preset trigger condition is detected, return to S1 for dynamic rescheduling.

[0010] In addition, the global load balancing scheduling method for computing clusters based on EML operators proposed above according to the present invention may also have the following additional technical features:

[0011] Furthermore, in S3, the construction of the global load balancing objective function through a finite number of nested EML operations includes: for the i-th node in the cluster, selecting the normalized CPU utilization rate. With memory utilization As the input to the first-level EML, the first-level composite index is calculated. ;Will With normalized GPU utilization As input to the second-level EML, the second-level composite index is calculated. ;Will Normalized network I / O load As the input to the third-layer EML, the node comprehensive load index is obtained. The global load balancing objective function F is constructed for all nodes. The maximum value function or variance function.

[0012] Furthermore, S3 also includes: calculating the overall load index of the nodes. Normalized storage I / O load As the input to the fourth-level EML, the extended comprehensive load index is obtained. The global load balancing objective function F is constructed for all nodes. The maximum value function or variance function.

[0013] Furthermore, in S2, the computing task comes from one or more of the following industry application scenarios: government computing, university scientific research computing, industrial simulation, AI training, and big data analysis. The feature extraction and type recognition include parsing the industry tags, application framework identifiers, and estimated power demand features contained in the task submission, and mapping them into a unified resource demand vector. The unified resource demand vector includes at least one of the following: number of CPU cores, memory capacity, GPU memory requirements, and network bandwidth expectations.

[0014] Furthermore, in S5, the preset triggering condition is: the CPU utilization, GPU utilization, or network IO load of any node changes by more than a preset threshold, or a new task is submitted to the scheduling queue.

[0015] Furthermore, in S1, the resource virtualization uses containerization or virtual machine technology to abstract physical heterogeneous nodes into computing power units with unified attributes, and maintains a global resource status table to record the current available resources and load indicators of each computing power unit.

[0016] Furthermore, the method also includes: after each global optimization solution, caching the intermediate composite index calculation results generated by the nested operation of the EML operator, and using the cached intermediate composite index calculation results as the initial search point of the optimization algorithm when the next dynamic rescheduling is triggered.

[0017] The second aspect of this invention proposes a global load balancing scheduling system for a computing cluster based on EML operators for implementing the first aspect, comprising: a unified resource pool module, used to virtualize resources of heterogeneous computing nodes of multiple types within the cluster to form a unified computing power unit, and to collect the CPU utilization, memory utilization, GPU utilization, network I / O load and storage I / O load of each node in real time to generate a multi-dimensional load vector; a task identification module, used to receive computing tasks, perform feature extraction and type identification on each task, and obtain the resource requirement specifications corresponding to the task; an EML global scheduling engine, configured to execute S3 and S4 of the method as described in any one of claims 1 to 7, that is, to construct a global load balancing objective function based on EML operators and solve for the globally optimal task-to-node allocation mapping; and a scheduling execution and monitoring module, used to schedule tasks to corresponding nodes for execution according to the allocation mapping, and to initiate dynamic rescheduling when a preset trigger condition is detected.

[0018] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the first aspect.

[0019] Compared with the prior art, the beneficial effects of the present invention are:

[0020] 1. By unifying the multi-dimensional load metrics of nodes such as CPU, memory, GPU, and network as inputs to the EML operator, and generating a comprehensive node load index using the same operator framework, the scheduling model is unified, avoiding the system complexity caused by using multiple heterogeneous algorithms for separate modeling.

[0021] 2. By merging the load status of all nodes into a global objective function through a finite number of nested EML operations and solving for the optimal allocation at once, the load of each node can be balanced from a global perspective, reducing load skew and resource fragmentation caused by local algorithms and improving the overall resource utilization of the cluster.

[0022] 3. When the cluster state changes, global rescheduling based on EML operators can be automatically triggered, and convergence to the new optimal allocation scheme can be achieved without manual parameter tuning, maintaining the stability of load balancing and task execution efficiency in dynamic environments. Attached Figure Description

[0023] Figure 1 This is an overall flowchart of a global load balancing scheduling method for computing clusters based on EML operators according to the present invention.

[0024] Figure 2 This is a schematic diagram of the structure of the EML operator nested construction of the node comprehensive load index in a computing cluster global load balancing scheduling method based on EML operators according to the present invention;

[0025] Figure 3 This is a schematic block diagram of a unified computing resource pool and multi-task scheduling method for a computing cluster based on EML operators according to the present invention.

[0026] Figure 4 This is a flowchart illustrating the dynamic rescheduling triggering and execution process of a computing cluster global load balancing scheduling method based on EML operators according to the present invention. Detailed Implementation

[0027] Embodiments of the present invention are described in detail below, with examples of the embodiments illustrated in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0028] The following describes a global load balancing scheduling method for a computing cluster based on EML operators according to an embodiment of the present invention, with reference to the accompanying drawings.

[0029] like Figures 1 to 4As shown, this embodiment provides a global load balancing scheduling method for a computing cluster based on EML operators. This method is applied to a heterogeneous computing cluster that simultaneously supports three types of services: university scientific research computing, AI training, and offline analysis of government big data. The cluster includes several CPU-intensive nodes, GPU nodes, and storage-balanced nodes.

[0030] S1: Construct a unified computing resource pool. Using containerization or virtual machine technology, abstract all physical nodes into computing units with unified attributes and maintain a global resource status table to record the current available resources and load metrics of each computing unit. Monitoring agents deployed on each node collect CPU utilization, memory utilization, GPU utilization, network throughput, and storage I / O load at preset time intervals. After normalization, these data form a multi-dimensional load vector and are reported to the central scheduler.

[0031] S2: Multi-scenario task identification and requirement profiling. The scheduler receives user-submitted tasks and automatically parses task metadata. For example, task A is tagged with the industry label "AI training" and the application framework identifier "TensorFlow," and the parsed requirements are 16GB GPU memory, 8 CPU cores, and 32GB RAM. Task B is tagged with the industry label "government affairs," and is a data-de-identified batch processing task, emphasizing CPU and memory bandwidth requirements, with no GPU requirement. The task identification module maps these heterogeneous requirements into a standardized resource requirement vector. ,in Indicates the number of CPU cores. Indicates memory capacity. Indicates GPU memory requirements. This indicates the expected network bandwidth.

[0032] S3: Construction of the global load balancing objective function based on EML operators, refer to Figure 2 In this embodiment, the EML operator is defined as follows: , All values ​​are normalized to the interval (0, 1).

[0033] For the i-th node in the cluster, perform nested operations sequentially:

[0034] First layer: Normalized CPU utilization With memory utilization As input, calculate the first-level composite index. ;

[0035] Second layer: With normalized GPU utilization As input, the second-level composite index is calculated. ;

[0036] Third layer: Normalized network I / O load As input, the node comprehensive load index is obtained. .

[0037] This comprehensive load index integrates four dimensions of load information: CPU, memory, GPU, and network. Nodes with heavier loads... The larger the value.

[0038] Assuming the cluster has N nodes, the global load balancing objective function is constructed for all nodes. variance: ,in For all nodes The mean can also be constructed as a maximum value function. .

[0039] In another embodiment, if storage I / O load needs to be further considered, a fourth nested operation is performed: the node's overall load index is calculated. Normalized storage I / O load As the input to the fourth-level EML, the extended comprehensive load index is obtained. And the global load balancing objective function F is constructed for all nodes. The maximum value function or variance function.

[0040] S4: Solving the global optimal allocation mapping involves matching the set of tasks to be scheduled with the currently available resources of the nodes. The optimization objective is to minimize the global load balancing objective function F. The gradient descent method is used to solve the allocation variables. During the iterative optimization process, the exponential operation in the EML operator amplifies the differences in node load values, so that the gradient information contains a significant distinction between heavily loaded and lightly loaded nodes. The logarithmic operation compresses the range of input values, suppressing the numerical oscillation of lightly loaded nodes during the optimization process, thereby converging the optimization process to the globally load-balanced task-to-node allocation mapping.

[0041] S5: Task scheduling and dynamic rescheduling, see reference Figure 3 and Figure 4 The central scheduler distributes tasks to target nodes for execution based on the obtained allocation mapping. The system continuously monitors the load status of each node. When any of the following trigger conditions are detected, it returns to S1 to initiate a new round of dynamic rescheduling: the CPU utilization, GPU utilization, or network IO load of any node changes by more than a preset threshold, or a new task is submitted to the scheduling queue.

[0042] Furthermore, after each global optimization solution is completed, the scheduler will perform intermediate composite indices generated by nested EML operator operations. , The calculation results are cached. When dynamic rescheduling is triggered again, the cached intermediate composite index calculation results are used as the initial search point of the optimization algorithm, so that the optimization process converges near the existing solution and reduces the online computing overhead.

[0043] Reference Figure 3 The unified resource pool module performs resource virtualization and status collection for various types of heterogeneous computing nodes within the cluster. The task identification module receives computing tasks from various industry scenarios such as government computing, university scientific research computing, industrial simulation, AI training, and big data analysis, and performs feature extraction. The EML global scheduling engine executes the construction of the global load balancing objective function and the solution of the optimal allocation mapping. The scheduling execution and monitoring module is responsible for task distribution and cluster status monitoring.

[0044] Reference Figure 3 This embodiment provides a global load balancing scheduling system for computing clusters based on EML operators, including: a unified resource pool module, a task identification module, an EML global scheduling engine, and a scheduling execution and monitoring module.

[0045] The unified resource pool module is used to virtualize resources of various types of heterogeneous computing nodes in the cluster to form a unified computing power unit. It also collects the CPU utilization, memory utilization, GPU utilization, network I / O load and storage I / O load of each node in real time to generate a multi-dimensional load vector. Resource virtualization is implemented using containerization or virtual machine technology, and a global resource status table is maintained to record the current available resources and load indicators of each computing power unit.

[0046] The task identification module is used to receive computing tasks, extract features and identify the type of each task, and obtain the resource requirement specifications corresponding to the task. The computing tasks come from one or more industry application scenarios such as government computing, university scientific research computing, industrial simulation, AI training, and big data analysis. Feature extraction includes parsing the industry tags, application framework identifiers and estimated power requirement features contained in the task submission.

[0047] The EML global scheduling engine is configured to execute S3 and S4 in the method, that is, to construct a global load balancing objective function based on EML operators, and to optimize by minimizing the objective function, and to solve the globally optimal task-to-node allocation mapping by combining task resource requirements and the current load status of nodes.

[0048] The scheduling execution and monitoring module is used to schedule tasks to the corresponding nodes for execution based on the allocation mapping output by the EML global scheduling engine, and continuously monitor the cluster status. When the CPU utilization, GPU utilization or network IO load of any node exceeds the preset threshold, or when a new task is submitted to the scheduling queue, the EML global scheduling engine is triggered to initiate dynamic rescheduling.

[0049] The specific implementation of each module is consistent with the description of the corresponding steps in the above method embodiments.

[0050] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the inventive concept of the present invention and the contents of the specification and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

[0051] In the description of this specification, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0052] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0053] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A global load balancing scheduling method for computing clusters based on EML operators, characterized in that, Includes the following steps: S1: Construct a unified computing resource pool, virtualize resources of heterogeneous computing nodes of various types, including CPU, GPU, memory and network resources, within the cluster to form a unified computing unit, and collect the CPU utilization, memory utilization, GPU utilization, network IO load and storage IO load of each node in real time to generate a multi-dimensional load vector. S2: Receive computing tasks, perform feature extraction and type identification for each task, and obtain the resource requirement specifications corresponding to the task; S3: Based on EML operators, the multidimensional load vector is mapped to a unified form of node load metric, and a global load balancing objective function is constructed through a finite number of nested EML operations; The EML operator is defined as follows: ,in It is a natural constant. The input variables are taken from the index values ​​in the multidimensional load vector or their normalized values; S4: Taking minimizing the global load balancing objective function as the optimization objective, and combining the task resource requirements and the current load status of the nodes, solve for the globally optimal task-to-node allocation mapping; S5: Schedule the task to the corresponding node for execution according to the allocation mapping, and continuously monitor the cluster status. When the preset trigger condition is detected, return to S1 for dynamic rescheduling.

2. The global load balancing scheduling method for a computing cluster based on EML operators according to claim 1, characterized in that, In S3, the construction of the global load balancing objective function through a finite number of nested EML operations includes: For the i-th node in the cluster, select the normalized CPU utilization. With memory utilization As the input to the first-level EML, the first-level composite index is calculated. ; Will With normalized GPU utilization As input to the second-layer EML, calculate the second-layer composite index. ; Will Normalized network I / O load As the input to the third-layer EML, the node comprehensive load index is obtained. ; The global load balancing objective function F is constructed for all nodes. The maximum value function or variance function.

3. The global load balancing scheduling method for a computing cluster based on EML operators according to claim 2, characterized in that, S3 also includes: The node's overall load index Normalized storage I / O load As the input to the fourth-level EML, the extended comprehensive load index is obtained. ; The global load balancing objective function F is constructed for all nodes. The maximum value function or variance function.

4. The global load balancing scheduling method for a computing cluster based on EML operators according to claim 1, characterized in that, In S2, the computing task comes from one or more of the following industry application scenarios: government computing, university scientific research computing, industrial simulation, AI training, and big data analysis. The feature extraction and type recognition include parsing the industry tags, application framework identifiers, and estimated power demand features contained in the task submission, and mapping them into a unified resource demand vector. The unified resource demand vector includes at least one of the following: number of CPU cores, memory capacity, GPU memory requirements, and network bandwidth expectations.

5. The global load balancing scheduling method for a computing cluster based on EML operators according to claim 1, characterized in that, In S5, the preset triggering conditions are: the CPU utilization, GPU utilization, or network IO load of any node changes by more than a preset threshold, or a new task is submitted to the scheduling queue.

6. The global load balancing scheduling method for a computing cluster based on EML operators according to claim 1, characterized in that, In S1, the resource virtualization uses containerization or virtual machine technology to abstract physical heterogeneous nodes into computing power units with unified attributes, and maintains a global resource status table to record the current available resources and load indicators of each computing power unit.

7. A global load balancing scheduling method for a computing cluster based on EML operators according to any one of claims 1 to 6, characterized in that, The method further includes: after each global optimization solution, caching the intermediate composite index calculation results generated by the nested operation of the EML operator, and using the cached intermediate composite index calculation results as the initial search point of the optimization algorithm when the next dynamic rescheduling is triggered.

8. A global load balancing scheduling system for computing clusters based on EML operators, characterized in that, include: The unified resource pool module is used to virtualize resources of various types of heterogeneous computing nodes in the cluster, forming a unified computing power unit, and to collect the CPU utilization, memory utilization, GPU utilization, network I / O load and storage I / O load of each node in real time to generate a multi-dimensional load vector. The task identification module is used to receive computing tasks, extract features and identify the type of each task, and obtain the resource requirement specifications corresponding to the task. The EML global scheduling engine is configured to execute S3 and S4 in the method described in any one of claims 1 to 7, namely, to construct a global load balancing objective function based on EML operators and solve for the globally optimal task-to-node allocation mapping. The scheduling execution and monitoring module is used to schedule tasks to corresponding nodes for execution according to the allocation mapping, and to initiate dynamic rescheduling when a preset trigger condition is detected.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1 to 7.

Images