Multi-node-oriented computing power server cluster resource management system
By constructing a directed acyclic graph of subtasks and a multi-objective optimization scheduling algorithm, combined with resource monitoring and a visual interactive interface, the problems of resource waste and disordered execution order in computing cluster management are solved, achieving efficient resource utilization and orderly task execution, and improving the intelligence level of cluster management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-20
- Publication Date
- 2026-04-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing computing cluster management systems lack fine-grained processing in resource scheduling, leading to cluster load imbalance, resource waste, chaotic execution order, untimely identification of abnormal nodes, and a lack of convenient interaction methods, which affects operational efficiency and management intelligence.
A multi-objective optimization scheduling algorithm is used to construct a directed acyclic graph of subtasks. Combined with resource data and dependencies, a visual interactive interface is used to submit tasks and display results. Real-time monitoring and maintenance checklist generation are implemented to eliminate abnormal nodes and ensure orderly task execution and efficient resource utilization.
It achieves efficient utilization of cluster resources, orderly execution of tasks, and timely handling of abnormal responses, improving the flexibility and stability of cluster management and solving the problems of load imbalance, resource waste, and low interaction efficiency in existing technologies.
Smart Images
Figure CN121785799A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computing power cluster management technology, specifically to a computing power server cluster resource management system for multiple nodes. Background Technology
[0002] With the rapid development of big data, artificial intelligence, and high-performance computing, the demand for computing power is exploding. Multi-node computing server clusters have become the core infrastructure supporting the processing of various complex tasks. Currently, the cluster scale is constantly expanding, and the types of tasks are becoming increasingly diverse. Different tasks have significantly different requirements for computing server resources, which places higher demands on the dynamic management of cluster resources and the efficient scheduling of tasks. Although existing computing cluster management systems can achieve basic resource acquisition and task allocation, most of them adopt a single-dimensional scheduling strategy, lacking fine-grained processing of task dependencies and balanced consideration of the overall resource load of the cluster, making it difficult to adapt to the task execution needs in complex scenarios.
[0003] However, existing technologies still have many shortcomings: on the one hand, resource scheduling algorithms often focus on a single optimization objective, either pursuing task execution speed or simply emphasizing resource utilization, leading to frequent problems such as cluster load imbalance, resource waste, and task blocking; on the other hand, abnormal node identification mechanisms are not accurate enough, often triggering alarms only after node resources are completely exhausted, and the handling of subtask dependencies lacks a systematic method, easily resulting in chaotic execution order. Furthermore, there is a lack of convenient visual interaction methods, making manual adjustment of task logic and tracing execution results inefficient. These problems severely restrict the operational efficiency and intelligent management level of computing clusters, urgently requiring a cluster resource management system that can achieve efficient resource utilization, orderly task execution, and timely response to anomalies. Summary of the Invention
[0004] The purpose of this invention is to provide a resource management system for multi-node computing server clusters to solve the problems mentioned in the background art.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A resource management system for multi-node computing server clusters includes computing server nodes, a control center, and an operation and maintenance monitoring platform. The control center includes a resource perception unit, a task decomposition unit, a subtask scheduling unit, a cross-node collaboration unit, and a task result feedback unit. The control center is communicatively connected to an operation and maintenance monitoring platform and at least one computing server node. The operation and maintenance monitoring platform includes a task submission unit, a cluster resource status monitoring unit, and an operation and maintenance unit. The computing server cluster consists of at least one computing server node. The resource sensing unit acquires resource data from each computing server node; The task decomposition unit decomposes the processing task into several sub-tasks and constructs a directed acyclic graph of the sub-tasks. The subtask scheduling unit allocates each subtask to the optimal computing server node through a multi-objective optimization scheduling algorithm based on the subtask directed acyclic graph and the resource data of each computing server node, thereby obtaining a subtask processing logic list. The cross-node collaborative unit controls the computing server nodes to execute each subtask in an orderly manner according to the subtask processing logic list, and records the subtask results of each subtask. The task result feedback unit aggregates the sub-task results of each sub-task to form a sequence of processing task calculation results; The task submission unit provides a visual interactive interface for handling task submission, manual adjustment of the directed acyclic graph of subtasks, and display of task calculation result sequences. The cluster resource status monitoring unit monitors the resource data of each computing server node, identifies and generates a list of abnormal computing server nodes. The operation and maintenance unit generates an operation and maintenance list based on the list of abnormal computing power server nodes, and distributes the operation and maintenance list to the operation and maintenance personnel for offline operation and maintenance.
[0006] Preferably, the resource data includes CPU utilization, memory usage, and bandwidth occupancy. The resource sensing unit collects the resource data of each computing server node in real time according to a preset sampling period by deploying a resource collection agent on each computing server node, and transmits the collected resource data back to the resource sensing unit through a communication link established by the TCP / IP protocol.
[0007] Preferably, the method for constructing the directed acyclic graph of the subtasks is as follows: S1. The syntax structure of the processing task submitted by the user is parsed by the LL(1) parser, the execution statements and parameter configurations of the processing task are extracted, and the BERT-based task semantic parsing model is used to mine the logical associations, data flow and execution steps in the execution statements. Based on this, the results are integrated to form the execution logic script of the processing task. The execution logic script includes the task execution process, data input and output rules and logical branch information.
[0008] S2. Based on the execution logic script, the processing task is decomposed into several subtasks according to the principle of functional independence, forming a set of subtasks. And analyze the dependencies between subtasks. If subtasks exist... Execution requires subtasks The output of the subtask is the input, or the subtask's result is the input. Need to be in subtask Once completed and started, the dependency relationships will be recorded. This forms a dependency set. Dependencies include two categories: data dependencies and execution dependencies; the set of subtasks Each subtask Includes subtask identifier And the required computing resources, which include CPU usage requirements, memory usage requirements, bandwidth transmission requirements, and estimated computing time; S3. Using subtasks in the subtask set T as nodes and dependency pairs in the dependency set D as directed edges, construct an initial directed graph. The node set V=T, where each node's attributes encapsulate the subtask identifier. Given the required computing resources, the edge set E=D, and the attribute label of each directed edge indicating the type of the corresponding dependency relationship.
[0009] S4. Use the topological sorting algorithm to process the initial directed graph. Perform topological sort verification, wherein the topological sort verification is as follows: if the initial directed graph If a cycle exists, identify the sub-task set of the cycle, update the dependency set D by splitting the dependencies between subtasks within the sub-task set, and then return to S3 to rebuild the initial directed graph. If the initial directed graph If there are no cycles, then the initial directed graph will be... The subtasks are identified as a directed acyclic graph. The method for identifying the loop is as follows: calculate the initial directed graph. To determine the in-degree of each node, construct an in-degree table and adjacency table for each node. Then, iterate through all nodes, selecting nodes with an in-degree of 0 and adding them to the topology sequence. Remove the directed edges associated with each node and update the in-degree of the corresponding node. If, after traversing all nodes, the length of the topology sequence is less than the total number of subtasks n, then the initial directed graph is considered complete. If a cycle exists, and the length of the topological sequence is equal to n, then the initial directed graph is... No loop road.
[0010] Preferably, the process of obtaining the subtask processing logic list is as follows: S01. Extract the subtask identifiers of all subtasks in the directed acyclic graph. Required computing resources ( and dependency set , D represents a subtask. All corresponding dependencies, and simultaneously collect resource data from all computing server nodes ( Together, they constitute the scheduling input dataset. ∈n}∪ m is the total number of computing server nodes. This is the identifier for the computing server node; S02. Set optimization objectives including minimizing the total execution time of subtasks, balancing the resource load of computing server nodes, and maximizing resource utilization. Construct a multi-objective optimization function based on these objectives. The details are as follows: Minimize the total execution time of subtasks: ; in, Assign decision variables to subtask nodes. Represents the i-th subtask Assigned to the j-th computing server node , Represents the i-th subtask Not assigned to the j-th computing server node , The sum of the estimated execution times of all subtasks on the corresponding allocated computing power server nodes represents the total execution time of the task. Load balancing of computing server nodes: ; in, Let m be the variance of the load on the computing server nodes, and m be the total number of computing server nodes. For the j-th computing server node The sum of the estimated execution times of all subtasks assigned above. The average load of all computing server nodes is represented by the load variance of the computing server nodes, which characterizes the resource load balance of the computing server nodes. Maximize resource utilization: ; in, This represents the overall average resource utilization rate of the cluster. For the i-th subtask CPU usage requirements For the i-th subtask Memory usage requirements For the i-th subtask bandwidth transmission requirements , , The j-th node The total number of CPU cores, total memory, and bandwidth limit of the cluster; the overall average resource utilization rate of the cluster represents the overall utilization efficiency of the cluster resources, i.e., resource utilization rate. The multi-objective optimization function is: ; in, , These are the weighting coefficients for the total execution time of the subtask, the load variance of the computing server nodes, and the resource utilization rate, respectively, satisfying... and , The maximum estimated total execution time for all subtasks allocated to a single computing server node. This represents the maximum possible load variance of all computing server nodes in the computing server cluster, used to... and Unify to the [0,1] interval; S03. Using the set of computational resource requirements and dependencies of subtasks in the scheduling input dataset, along with the resource data of the computing server nodes, as constraints, input them into a multi-objective optimization scheduling algorithm to obtain a subtask server node allocation scheme that satisfies the constraints and minimizes the multi-objective optimization function. The constraints include: The assignment uniqueness constraint is: This indicates that each subtask is assigned to only one computing server node; the resource capacity constraint is: , and , representing the j-th computing server node The total CPU usage, total memory usage, and total bandwidth usage requirements for all subtasks allocated shall not exceed the current remaining available CPU utilization, memory utilization, and bandwidth utilization of the computing server node, respectively. , , These are the j-th computing server nodes. The current CPU utilization, memory utilization, and bandwidth utilization. These represent the remaining available percentage of the corresponding resources; the dependency constraint is: if there is a predecessor-successor dependency relationship between subtasks... Then the normative constraint formula is: ,in, For the precursor mission Assign decision variables to subtask nodes. For the precursor mission The estimated execution time, For successor subtasks The estimated start time indicates the precursor subtask. The overall execution completion time shall not be later than that of the subsequent subtask. Overall startup time; The multi-objective optimization scheduling algorithm is any one of NSGA-II, MOPSO, and MOACO; S04. Combine the set of dependencies between subtasks in the directed acyclic graph of subtasks. That is, the dependency set D of the directed acyclic graph of subtasks. The subtasks in the subtask server node allocation scheme Scheme are sorted according to the topological sorting rules for execution order. The dependency set D is traversed; if a dependency pair exists... Then determine the predecessor subtask. The execution priority of the first subtask is higher than that of the subsequent subtask. Based on the results of traversing all dependencies, generate each subtask. Unique corresponding predecessor sub-task identifier list , The canonical set is expressed as: in For the precursor mission Unique identifier, list When empty, it indicates a subtask. No prerequisite dependencies; S05. Integrate the unique identifiers of each subtask. Subtask server node allocation scheme in Scheme and subtasks The target computing server node to be bound and Precursor Sub-task Identifier List This generates a structured list of subtask processing logic, the canonical set of which is expressed as follows: Each entry in the subtask processing logic list corresponds to the complete execution constraint information of a single subtask, which is expressed as follows: .
[0011] Preferably, the cross-node collaboration unit determines whether a subtask is a dependent subtask or a subtask with dependencies by checking whether the number of elements in the predecessor subtask identifier list of the subtask is empty. The dependent subtask is then sent to the target computing power server node in parallel for direct execution, while the dependent subtask is temporarily stored in the local task queue of the target computing power server node according to its dependency priority. The execution status of its predecessor subtask is monitored in real time, and execution is automatically triggered after all predecessor subtasks are completed.
[0012] Preferably, the processing task calculation result sequence is formed by gradually aggregating the subtask results of each subtask according to the topological sorting order of the directed acyclic graph of the subtasks, following the rules of prioritizing independent subtasks and progressively increasing dependencies. During the aggregation process, the input and output data flow of each subtask is synchronously associated, ultimately forming a structured and traceable processing task calculation result sequence.
[0013] Preferably, the list of abnormal computing power server nodes consists of computing power server node IDs corresponding to abnormal computing power server nodes, and the abnormal computing power server node is a computing power server node whose resource data obtained in 9 consecutive sampling periods all exceed a preset resource threshold. The preset resource thresholds are: CPU utilization > 95%, memory utilization > 98%, bandwidth utilization > 90%, and number of pending subtasks > 10.
[0014] Preferably, the computing server nodes in S01 do not include abnormal computing server nodes in the abnormal computing server node list.
[0015] Preferably, the maintenance list is sent to maintenance personnel via SMS, email, or a pop-up window from the maintenance monitoring platform.
[0016] Preferably, the visual interactive interface is used for processing task submission, manual adjustment of the directed acyclic graph of subtasks, and display of the sequence of processing task calculation results. The function of processing task submission is interconnected with the task decomposition unit. The processing task entered by the user is synchronized to the task decomposition unit after verification. The function of adjusting the directed acyclic graph of subtasks is linked with the task decomposition unit to manually adjust the directed acyclic graph of subtasks. The function of displaying the sequence of processing task calculation results receives the sequence of processing task calculation results from the task result feedback unit and visually presents the results of each subtask and their relationships.
[0017] Due to the adoption of the above technical solution, the technical progress achieved by this invention compared to the prior art is as follows: 1. This invention employs a multi-objective optimization scheduling algorithm, which combines subtask dependencies and node resource data to construct an optimization function. It takes into account task execution time, load balancing, and resource utilization, thus solving the problems of cluster load imbalance and resource waste caused by the single-dimensional scheduling of existing technologies, and significantly improving the efficiency of cluster resource utilization.
[0018] 2. This invention monitors node resource data through continuous multi-sampling cycles, sets clear thresholds to accurately identify abnormal nodes, excludes abnormal nodes during scheduling, and, in conjunction with the distribution of maintenance lists through multiple channels, solves the shortcomings of existing technologies such as delayed identification of abnormal nodes and untimely maintenance response, thus ensuring the stable operation of the cluster.
[0019] 3. This invention decomposes tasks through syntax analysis and semantic parsing, constructs a directed acyclic graph of subtasks with dependencies, and combines topological sorting verification and loop elimination mechanisms to systematically process subtask dependencies. This solves the problems of disordered dependency processing and chaotic execution order in existing technologies, ensuring that tasks are executed in an orderly and efficient manner.
[0020] 4. This invention provides a visual interactive interface, realizing full-process visualization of task submission, manual adjustment of directed acyclic graphs and result display. The interface and core unit data are interconnected, solving the problems of lack of convenient interactive means and low efficiency of manual operation in the prior art, and improving the flexibility and operability of cluster management. Attached Figure Description
[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0022] Figure 1 This is a schematic diagram of the system functional modules of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0024] Examples, such as Figure 1 The aforementioned resource management system for a multi-node computing server cluster includes computing server nodes, a control center, and an operation and maintenance monitoring platform. The control center includes a resource perception unit, a task decomposition unit, a subtask scheduling unit, a cross-node collaboration unit, and a task result feedback unit. The operation and maintenance monitoring platform includes a task submission unit, a cluster resource status monitoring unit, and an operation and maintenance unit. The computing server cluster consists of at least one computing server node. The resource sensing unit acquires resource data from each computing server node; The task decomposition unit decomposes the processing task into several sub-tasks and constructs a directed acyclic graph of the sub-tasks. The subtask scheduling unit allocates each subtask to the optimal computing server node through a multi-objective optimization scheduling algorithm based on the subtask directed acyclic graph and the resource data of each computing server node, thereby obtaining a subtask processing logic list. The cross-node collaborative unit controls the computing server nodes to execute each subtask in an orderly manner according to the subtask processing logic list, and records the subtask results of each subtask. The task result feedback unit aggregates the sub-task results of each sub-task to form a sequence of processing task calculation results; The task submission unit provides a visual interactive interface for handling task submission, manual adjustment of the directed acyclic graph of subtasks, and display of task calculation result sequences. The cluster resource status monitoring unit monitors the resource data of each computing server node, identifies and generates a list of abnormal computing server nodes. The operation and maintenance unit generates an operation and maintenance list based on the list of abnormal computing power server nodes, and distributes the operation and maintenance list to the operation and maintenance personnel for offline operation and maintenance.
[0025] Furthermore, the working principle of the present invention will be illustrated below using a data analysis task processing scenario of a multi-node computing cluster of a certain enterprise as an example: The computing server cluster in this embodiment contains 5 computing server nodes, denoted as follows: , , , and The hardware configuration of each computing server node is standardized based on the total number of CPU cores. Total number of cores and memory GB, bandwidth limit After the computing server cluster starts up (Mbps), each computing server node feeds back resource data in real time through the deployed resource collection agent. The resource data collected from each computing server node at a certain moment are as follows: , , , , ,in Because the resource data for nine consecutive sampling periods exceeded the preset resource threshold, the node was identified as an abnormal computing power server node and was automatically excluded during the scheduling process. Subtask allocation is carried out only based on the remaining 4 normal computing server nodes.
[0026] The resource perception unit of the control center receives resource data from each computing server node via TCP / IP protocol and synchronizes it to the relevant units; the task decomposition unit first uses the LL(1) parser and the BERT-based task semantic parsing model to parse the user-submitted "multi-dimensional data modeling processing task" and generate an execution logic script in JSON format, and then decomposes it into 4 sub-tasks according to the principle of functional independence. Data preprocessing, Feature engineering Model training, (Data cleaning), forming a set of subtasks The analysis yielded the dependency set. After constructing the initial directed graph, it was verified by topological sorting. Since the length of the topological sequence was equal to 4, it was determined to be a directed acyclic graph of subtasks. The subtask scheduling unit extracted the identifiers of each subtask. Required computing resources ( : 2 cores, 2GB RAM, 100Mbps, 30 seconds; : 4 cores, 8GB RAM, 200Mbps, 60 seconds; : 8 cores, 32GB, 300Mbps, 120 seconds; The scheduling input dataset is composed of a set of parameters (2 cores, 4GB RAM, 150Mbps bandwidth, 45 seconds) and dependencies, combined with resource data from normal computing power server nodes. A multi-objective optimization function (weight coefficients) is then defined. The algorithm uses three types of constraints—uniqueness of allocation, resource capacity, and dependencies—to solve for the subtask server node allocation scheme. Then, it generates a predecessor subtask identifier list based on dependency sorting, and finally integrates these to form a subtask processing logic list. After the cross-node collaboration unit reads the subtask processing logic list, it will process the dependent subtasks. and Parallel distribution to and Execution will depend on subtasks. , Save to The local task queue is triggered sequentially after the predecessor subtask is completed, and the results of each subtask are recorded in real time. The task result feedback unit is sorted according to the topological order of the directed acyclic graph of subtasks. The results are aggregated following the rule of "prioritizing non-dependent subtasks and progressively increasing dependencies", and the input and output data flows are associated to form a structured and traceable sequence of processing task calculation results.
[0027] The operation and maintenance monitoring platform communicates in real time with the control center, providing support for multi-dimensional data modeling and processing tasks and the operation and maintenance of the computing server cluster. The task submission unit provides a visual interactive interface that allows users to input and submit task parameters for multi-dimensional data modeling and processing tasks. After format verification, the data is synchronized to the task decomposition unit of the control center. It can also load the initial sub-task directed acyclic graph for manual adjustment and receive and visualize the processing task calculation result sequence from the task result feedback unit for easy user review and verification. The cluster resource status monitoring unit continuously monitors the resource data of each computing server node, accurately identifies abnormal computing servers, generates a list of abnormal computing server nodes, and synchronizes it to the operation and maintenance unit and the sub-task scheduling unit. The operation and maintenance unit generates an operation and maintenance checklist based on the abnormal node list and distributes it to operation and maintenance personnel through both SMS and email channels to ensure that abnormal computing server nodes receive timely offline operation and maintenance processing, ensuring the stable operation of the computing server cluster.
[0028] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A resource management system for a multi-node computing server cluster, comprising computing server nodes, a control center, and an operation and maintenance monitoring platform, characterized in that, The control center includes a resource perception unit, a task decomposition unit, a subtask scheduling unit, a cross-node collaboration unit, and a task result feedback unit. The control center is communicatively connected to an operation and maintenance monitoring platform and at least one computing server node. The operation and maintenance monitoring platform includes a task submission unit, a cluster resource status monitoring unit, and an operation and maintenance unit. The resource sensing unit acquires resource data from each computing server node; The task decomposition unit decomposes the processing task into several sub-tasks and constructs a directed acyclic graph of the sub-tasks. The subtask scheduling unit allocates each subtask to the optimal computing server node through a multi-objective optimization scheduling algorithm based on the subtask directed acyclic graph and the resource data of each computing server node, thereby obtaining a subtask processing logic list. The cross-node collaborative unit controls the computing server nodes to execute each subtask in an orderly manner according to the subtask processing logic list, and records the subtask results of each subtask. The task result feedback unit aggregates the sub-task results of each sub-task to form a sequence of processing task calculation results; The task submission unit provides a visual interactive interface for handling task submission, manual adjustment of the directed acyclic graph of subtasks, and display of task calculation result sequences. The cluster resource status monitoring unit monitors the resource data of each computing server node, identifies and generates a list of abnormal computing server nodes. The operation and maintenance unit generates an operation and maintenance list based on the list of abnormal computing power server nodes, and distributes the operation and maintenance list to the operation and maintenance personnel for offline operation and maintenance.
2. The resource management system for a multi-node computing server cluster according to claim 1, characterized in that, The resource data includes CPU utilization, memory usage, and bandwidth utilization.
3. The resource management system for a multi-node computing server cluster according to claim 1, characterized in that, The method for constructing the directed acyclic graph of the subtasks is as follows: S1. The user-submitted processing task is parsed using a syntax analyzer and a task semantic parsing model to obtain the execution logic script for the processing task; S2. Based on the execution logic script, the processing task is decomposed into several subtasks, forming a set of subtasks. It also analyzes the dependencies between subtasks to form a dependency set. The set of subtasks Each subtask Includes subtask identifier and the required computing resources; S3. Using subtasks in the subtask set T as nodes and dependency pairs in the dependency set D as directed edges, construct an initial directed graph. Where the node set V=T and the edge set E=D; S4. For the initial directed graph Perform topological sort verification, wherein the topological sort verification is as follows: if the initial directed graph If a cycle exists, identify the sub-task set of the cycle, update the dependency set D by splitting the dependencies between subtasks within the sub-task set, and then return to S3 to rebuild the initial directed graph. If the initial directed graph If there are no cycles, then the initial directed graph will be... It has been determined to be a directed acyclic graph for subtasks.
4. A resource management system for a multi-node computing server cluster according to claim 2 or 3, characterized in that, The process of obtaining the subtask processing logic list is as follows: S01. Extract the subtask identifiers, required computing resources, and dependency sets of all subtasks in the directed acyclic graph of subtasks, and combine them with the resource data of all computing server nodes to form the scheduling input dataset; S02. Set optimization objectives including minimizing the total execution time of subtasks, balancing the resource load of computing server nodes, and maximizing resource utilization, and construct a multi-objective optimization function based on the optimization objectives; S03. The set of required computing resources and dependencies of subtasks in the scheduling input dataset, as well as the resource data of computing server nodes, are used as constraints and input into a multi-objective optimization scheduling algorithm to solve for the subtask server node allocation scheme that satisfies the constraints and minimizes the multi-objective optimization function. The multi-objective optimization scheduling algorithm is any one of NSGA-II, MOPSO, and MOACO. S04. Based on the dependency set of each subtask in the directed acyclic graph, sort the execution order of the subtasks to obtain the order of execution of each subtask. List of identifiers; S05. Integrate the subtask identifiers, target computing power server nodes, and... The identifier list generates a structured list of subtask processing logic.
5. The resource management system for a multi-node computing server cluster according to claim 4, characterized in that, The cross-node collaborative unit identifies sub-tasks. If the identifier list is empty, determine whether the subtask is a dependent subtask or a dependent subtask. Dependent subtasks are distributed to the target computing power server node in parallel for execution, while dependent subtasks are temporarily stored in the local task queue of the target computing power server node and are triggered for execution after all the predecessor subtasks of the dependent subtasks are completed.
6. The resource management system for a multi-node computing server cluster according to claim 5, characterized in that, The sequence of processing task calculation results is obtained by aggregating the subtask results of each subtask according to the topological sorting order of the directed acyclic graph of the subtasks.
7. The resource management system for a multi-node computing server cluster according to claim 6, characterized in that, The list of abnormal computing power server nodes consists of the computing power server node ID corresponding to the abnormal computing power server node. The abnormal computing power server node is the computing power server node whose resource data obtained in 9 consecutive sampling periods exceeds the preset resource threshold. The preset resource thresholds are: CPU utilization > 95%, memory utilization > 98%, bandwidth utilization > 90%, and number of pending subtasks > 10.
8. The resource management system for a multi-node computing server cluster according to claim 7, characterized in that, The computing server nodes in S01 do not include the abnormal computing server nodes in the abnormal computing server node list.
9. The resource management system for a multi-node computing server cluster according to claim 7, characterized in that, The maintenance checklist is sent to maintenance personnel via SMS, email, or a pop-up window from the maintenance monitoring platform.