GPU computing power resource scheduling method and system

Abstract

The invention relates to the technical field of data analysis, and discloses a GPU computing power resource scheduling method and system, and the method comprises the steps: collecting node hardware parameters and dynamic load indexes of a GPU cluster to construct a multi-dimensional resource feature vector of the GPU cluster, and constructing a resource portrait of the GPU cluster; establishing a node health degree scoring model of the GPU cluster, and generating a health degree score of a cluster node corresponding to the GPU cluster; analyzing a video memory demand of the GPU task request, and calculating an intensive identifier and a communication dependency relationship; determining the SLA weight of the GPU task request, calculating the resource shortage sensitivity of the GPU task request based on the video memory demand, and calculating the target task priority of the GPU task request in combination with the SLA weight; and determining a resource scheduling node group requested by the GPU task in the resource portrait, generating resource scheduling parameters of the resource scheduling node group, and executing scheduling of computing power resources of the GPU cluster based on the resource scheduling parameters. According to the method, the scheduling efficiency of the GPU computing power resources can be improved.

Description

Technical Field

[0001] The present invention relates to a method and system for scheduling GPU computing resources, and belongs to the technical field of data analysis. Background Art

[0002] GPU computing resource scheduling dynamically allocates and manages graphics processing unit (GPU) computing power, memory, bandwidth, and other resources through software or hardware mechanisms to meet the needs of different tasks while optimizing the overall cluster utilization, energy efficiency, and task execution efficiency. Its core goal is to resolve resource competition and heterogeneous demand issues, which is particularly crucial in high-concurrency scenarios such as AI training, scientific computing, and graphics rendering.

[0003] Traditional GPU computing resource scheduling is typically simplistic. The scheduler primarily allocates resources based on the requested memory and number of GPUs, searching for idle nodes in the cluster that meet basic resource requirements. This approach often employs basic strategies such as first-come, first-served or round-robin. This approach lacks in-depth analysis of node health, historical performance, hardware details (such as GPU model and network bandwidth), and task characteristics (such as compute intensity and communication dependencies). It also fails to effectively prioritize tasks and meet SLA requirements, making it difficult to meet the needs of high-priority tasks. Summary of the Invention

[0004] The present invention provides a method and system for scheduling GPU computing resources, the main purpose of which is to improve the scheduling efficiency of GPU computing resources.

[0005] To achieve the above objectives, the present invention provides a method for scheduling GPU computing resources, comprising: Collecting node hardware parameters and dynamic load indicators of the GPU cluster to construct a multidimensional resource feature vector of the GPU cluster, and constructing a resource profile of the GPU cluster based on the multidimensional resource feature vector; Based on the historical monitoring data of the GPU cluster, a node health scoring model of the GPU cluster is established, and a health score of the cluster node corresponding to the GPU cluster is generated; Authenticate the received GPU task request, and if the authentication is successful, analyze the GPU task request's video memory requirements, computationally intensive identifiers, and communication dependencies; Determine an SLA weight of the GPU task request according to the task type of the GPU task request, calculate a resource shortage sensitivity of the GPU task request based on the video memory requirement, and calculate a target task priority of the GPU task request in combination with the SLA weight; Through the computationally intensive identifier, the communication dependency and the health score, the resource scheduling node group requested by the GPU task in the resource portrait is determined, and resource scheduling parameters of the resource scheduling node group are generated. The scheduling of the computing power resources of the GPU cluster is performed based on the resource scheduling parameters.

[0006] Optionally, constructing the multi-dimensional resource feature vector of the GPU cluster includes: Standardizing the node hardware parameters and dynamic load indicators of the GPU cluster to obtain standardized node hardware parameters and standardized dynamic load indicators; Extracting cluster timing characteristics of the GPU cluster based on the standardized node hardware parameters and the standardized dynamic load indicators; Performing topological feature encoding on the cluster time series features to obtain topological encoding features; A multi-dimensional resource feature vector of the GPU cluster is constructed according to the topological coding feature.

[0007] Optionally, constructing a resource profile of the GPU cluster based on the multi-dimensional resource feature vector includes: Analyzing node-level features and cluster features of the GPU cluster based on the multi-dimensional resource feature vector; Define the image tags, computing logic, and application scenarios of the GPU cluster; A resource profile of the GPU cluster is constructed by combining the node-level features, the cluster features, the portrait tags, the computing logic, and the application scenario.

[0008] Optionally, establishing a node health scoring model for the GPU cluster based on historical monitoring data of the GPU cluster includes: Preprocessing the historical monitoring data to obtain processed historical monitoring data; Extracting node health characteristics of the processed historical monitoring data; Identifying historical health failures of the GPU cluster based on the historical monitoring data to generate a health label for the GPU cluster; Based on the node health characteristics, a preset node health scoring model architecture is trained in a temporal-static joint manner to obtain a node health scoring model.

[0009] Optionally, performing temporal-static joint training on a preset node health scoring model architecture based on the node health characteristics to obtain a node health scoring model includes: Based on the node health characteristics, the node health scoring model architecture is subjected to temporal-static joint training to obtain a trained node health scoring model to output a node health category probability value of the GPU cluster corresponding to the node health characteristics; Based on the node health category probability value of the GPU cluster, the training node health score of the GPU cluster is calculated using the following formula:

[0010] in, express The health score of the training nodes in the GPU cluster at any given moment. represents the smoothing coefficient, express The node health category probability value of the GPU cluster at this moment, express The health score of the training nodes in the GPU cluster at any given moment; Analyzing the reliability of the training node health score; When the scoring reliability meets a preset scoring reliability threshold, the trained node health scoring model is used as the node health scoring model.

[0011] Optionally, parsing the video memory requirement, computation-intensive identifier, and communication dependency of the GPU task request includes: Extracting the video memory parameters requested by the GPU task; and normalizing the video memory parameters to obtain standardized video memory parameters to determine the video memory requirements of the GPU task request; Extracting the intensive identifier of the GPU task request to calculate the computing density set of the GPU task request; Determining a computationally intensive identifier of the GPU task request based on the computational density set; Extracting a communication field of the GPU task request to identify a communication mode of the GPU task request; The communication dependency of the GPU task request is determined through the communication mode.

[0012] Optionally, the calculating the resource shortage sensitivity of the GPU task request based on the video memory requirement includes: Identify the cluster memory pressure of the GPU cluster corresponding to the GPU task request; Analyzing the normalized task requirements of the GPU task request based on the video memory requirements; According to the cluster memory pressure and the normalized task requirements, the resource shortage sensitivity of the GPU task request is calculated using the following formula:

[0013] in, Indicates that the GPU task request is in Sensitivity to resource scarcity at all times, Indicates The cluster memory pressure at all times, represents the normalized task requirements, represents the waiting sensitivity coefficient, represents the waiting penalty factor, represents the smoothing term.

[0014] Optionally, the calculating the target task priority of the GPU task request based on the combination of the SLA weight includes: Identifying user information and a deadline for the GPU task request; Determining a user priority level of the GPU task request based on the user information; Determining a time criticality of the GPU task request based on the deadline; Define the SLA weight, the user information, the deadline, and the SLA weight coefficient, user information weight coefficient, time shortage weight coefficient, and resource shortage weight coefficient corresponding to the resource shortage sensitivity of the GPU task request to calculate the target task priority of the GPU task request.

[0015] Optionally, the determining the resource scheduling node group requested by the GPU task in the resource portrait by combining the computationally intensive identifier, the communication dependency, and the health score includes: Extracting a first candidate node that meets the GPU task request from the resource profile; Calculating a computing power matching capability of the first candidate node based on the computing intensiveness identifier of the GPU task request; Based on the communication dependency, calculating a network topology of the first candidate node; Screening a second candidate node for the first candidate node based on the computing power matching capability and the network topology; The health score is used to screen the health of the second candidate nodes to obtain the resource scheduling node group.

[0016] In order to solve the above problems, the present invention further provides a GPU computing resource scheduling system, the system comprising: A resource profile building module, configured to collect node hardware parameters and dynamic load indicators of a GPU cluster to build a multi-dimensional resource feature vector of the GPU cluster, and to build a resource profile of the GPU cluster based on the multi-dimensional resource feature vector; A node health scoring module is used to establish a node health scoring model for the GPU cluster based on the historical monitoring data of the GPU cluster, and generate a health score for the cluster node corresponding to the GPU cluster; A task request analysis module is used to authenticate the received GPU task request and, if authentication is successful, analyze the GPU task request's video memory requirements, computationally intensive identifiers, and communication dependencies. a task priority analysis module, configured to determine an SLA weight of the GPU task request based on the task type of the GPU task request, calculate the resource shortage sensitivity of the GPU task request based on the video memory requirement, and calculate a target task priority of the GPU task request in combination with the SLA weight; A computing power resource scheduling module is used to determine the resource scheduling node group requested by the GPU task in the resource portrait through the computing intensive identifier, the communication dependency and the health score, and generate resource scheduling parameters for the resource scheduling node group, and execute the scheduling of the computing power resources of the GPU cluster based on the resource scheduling parameters.

[0017] This method constructs a multi-dimensional resource feature vector by collecting the node hardware parameters and dynamic load indicators of the GPU cluster, and then forms a comprehensive and dynamic GPU cluster resource portrait. This process makes the cluster resource status visual and quantifiable, provides a solid data foundation for subsequent scheduling decisions, and effectively improves the transparency and accuracy of resource management. It uses historical data to establish a node health scoring model and generates health scores in real time, which significantly enhances the reliability of the cluster. The scheduling system can dynamically adjust resource allocation based on the health score and give priority to nodes in good health, thereby reducing the risk of task failure due to node failure and ensuring the stability and continuity of task execution. In terms of task request processing, this method first performs strict authentication to ensure that only legitimate requests can enter the system, thereby enhancing the security of the cluster. Then, through The graphics memory requirements, computationally intensive identifiers, and communication dependencies of the task request are parsed, and the priority of the target task is calculated in combination with the SLA weight and resource scarcity sensitivity determined by the task type. This priority calculation method that comprehensively considers task attributes and resource characteristics makes scheduling more fair and efficient, and can give priority to meeting the resource requirements of high-priority tasks while making rational use of resources. In determining the resource scheduling node group, the computationally intensive identifiers, communication dependencies, and health scores are comprehensively considered to generate resource scheduling parameters and execute scheduling. This method not only considers the computing power and health status of the nodes, but also pays attention to the communication efficiency between nodes, effectively reduces communication delays, and improves the overall performance of task execution. At the same time, the generation and application of resource scheduling parameters make the scheduling process more automated and intelligent, reduce manual intervention, and improve scheduling efficiency. Therefore, the present invention can improve the scheduling efficiency of GPU computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 A flowchart of a method for scheduling GPU computing resources provided by one embodiment of the present invention; Figure 2 A schematic diagram of historical health failure analysis for implementing the GPU computing resource scheduling method provided in one embodiment of the present invention; Figure 3 A schematic diagram of a module for implementing a method for scheduling GPU computing resources provided in one embodiment of the present invention.

[0019] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0020] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0021] The embodiment of the present application provides a method for scheduling GPU computing resources. The execution subject of the GPU computing resource scheduling method includes but is not limited to at least one of the electronic devices such as a server, a terminal, etc. that can be configured to execute the method provided by the embodiment of the present application. In other words, the GPU computing resource scheduling method can be executed by software or hardware installed on a terminal device or a server device. The server includes but is not limited to: a single server, a server cluster, a cloud server, or a cloud server cluster, etc.

[0022] Example 1: Reference Figure 1 FIG. 1 is a flow chart of a method for scheduling GPU computing resources according to an embodiment of the present invention. In this embodiment, the method for scheduling GPU computing resources includes: S1. Collect node hardware parameters and dynamic load indicators of a GPU cluster to construct a multidimensional resource feature vector of the GPU cluster, and construct a resource profile of the GPU cluster based on the multidimensional resource feature vector.

[0023] It should be explained that the GPU cluster refers to a computer system composed of multiple computing nodes and used for parallel computing or deep learning tasks. The node hardware parameters include video memory capacity, number of computing units, NVLink bandwidth, PCIe topology position and other parameters. The dynamic load indicators include video memory occupancy, core utilization, temperature, power consumption and other parameters.

[0024] The present invention constructs a multi-dimensional resource feature vector of a GPU cluster to serve as an improvement basis for subsequent portrait construction.

[0025] In detail, the multi-dimensional resource feature vector of the GPU cluster is constructed, including: Standardizing the node hardware parameters and dynamic load indicators of the GPU cluster to obtain standardized node hardware parameters and standardized dynamic load indicators; Extracting cluster timing characteristics of the GPU cluster based on the standardized node hardware parameters and the standardized dynamic load indicators; Performing topological feature encoding on the cluster time series features to obtain topological encoding features; A multi-dimensional resource feature vector of the GPU cluster is constructed according to the topological coding feature.

[0026] Among them, the standardized node hardware parameters refer to the data obtained after standardization of the originally collected node hardware parameters (such as GPU model, quantity, video memory size, number of CPU cores, memory capacity, network bandwidth, etc.); the standardized dynamic load indicators refer to the data obtained after standardization of the originally collected node dynamic load indicators (such as GPU utilization, video memory occupancy, CPU usage, memory usage, disk I / O, network traffic, etc.); the cluster time series features refer to the information extracted from the standardized dynamic load indicators that can reflect the trend and pattern of cluster or node changes over time; the topology encoding features refer to the process of encoding the network topology structure information of the GPU cluster (such as the physical connection relationship between nodes, network delay, bandwidth limitation, etc.) into numerical features; the multidimensional resource feature vector refers to the final single data structure that integrates the node hardware capabilities, real-time load status, dynamic change trends and network topology location information.

[0027] Optionally, the standardizing of the node hardware parameters and the dynamic load index of the GPU cluster to obtain the standardized node hardware parameters and the standardized dynamic load index can be achieved by normalization.

[0028] The present invention constructs a resource profile of the GPU cluster based on the multi-dimensional resource feature vector, which can effectively provide support for subsequent resource scheduling.

[0029] In detail, constructing the resource profile of the GPU cluster based on the multi-dimensional resource feature vector includes: Analyzing node-level features and cluster features of the GPU cluster based on the multi-dimensional resource feature vector; Define the image tags, computing logic, and application scenarios of the GPU cluster; A resource profile of the GPU cluster is constructed by combining the node-level features, the cluster features, the portrait tags, the computing logic, and the application scenario.

[0030] Among them, the node-level feature refers to an indicator extracted from the multidimensional resource feature vector, which can reflect the specific attributes or status of a single node in the cluster, such as a single GPU computing power, the number of CPU cores, the total memory of a single node, the size of a single GPU video memory, the current video memory occupancy rate and other indicators; the cluster feature refers to an indicator extracted from the multidimensional resource feature vector, which can reflect the overall attributes or status of the entire GPU cluster, such as the overall resource capacity, the overall load level, the resource distribution characteristics and other indicators; the portrait label refers to a classification with a specific meaning assigned to a node or cluster based on the node-level features and cluster features obtained through analysis, such as "high-load computing node", "memory bottleneck node", "idle node", "network bottleneck node", "critical mission node" and other classifications; the calculation logic refers to the specific algorithm or rule used to generate portrait labels and calculate node-level / cluster features, such as "if the GPU utilization of the node is > 80% and lasts for 10 minutes, then the label is 'high-load computing node'", "cluster average GPU utilization = The application scenario processes specific tasks, such as the preferred node for multi-card communication tasks, automatic frequency reduction or migration tasks, and other task scenarios. The resource profile refers to the final comprehensive resource management view that integrates node-level features, cluster features, profile tags, computing logic, and application scenarios.

[0031] Optionally, the node-level features and cluster features of the GPU cluster analyzed based on the multidimensional resource feature vector can be determined by clustering the multidimensional resource feature vector to determine the node-level features, and the cluster features can be determined by aggregating and statistically calculating the global resource indicators of the cluster.

[0032] S2. Based on the historical monitoring data of the GPU cluster, a node health scoring model of the GPU cluster is established, and a health score of the cluster node corresponding to the GPU cluster is generated.

[0033] It should be explained that the historical monitoring data refers to the various operating status, performance indicators, event logs and other data collected and saved for the above-mentioned GPU cluster and its nodes over the past period of time. The node health scoring model refers to a model that can output a quantitative score based on current or recent node status data to represent the current health status of the node.

[0034] In detail, the node health scoring model of the GPU cluster is established based on the historical monitoring data of the GPU cluster, including: Preprocessing the historical monitoring data to obtain processed historical monitoring data; Extracting node health characteristics of the processed historical monitoring data; Identifying historical health failures of the GPU cluster based on the historical monitoring data to generate a health label for the GPU cluster; Based on the node health characteristics, a preset node health scoring model architecture is trained in a temporal-static joint manner to obtain a node health scoring model.

[0035] Among them, the processing historical monitoring data refers to the data set obtained by cleaning, converting and integrating the original collected GPU cluster historical operation data (including historical records of node hardware parameters, historical records of dynamic load indicators, system logs, error reports, etc.), and the node health characteristics refer to the representative feature set extracted from the processed historical data that can reflect the health status of a single GPU node, such as GPU model, video memory capacity, number of CPU cores, historical average / peak GPU utilization, video memory occupancy, node temperature, CPU load, network I / O, disk I / O, task failure rate, error log frequency, specific performance indicators (such as the historical average inference delay of P100 nodes), volatility of resource utilization (standard deviation), temperature change rate over time, etc. The historical health characteristics are the following: A failure refers to an identifiable node or cluster-level health problem that has occurred in the historical operation records of the GPU cluster. The health label refers to a label assigned to the node status at each time point or each time window in the historical data. The node health scoring model architecture refers to the initial structure used to construct the node health scoring model. The time-series-static joint training refers to the process of training the same model architecture by simultaneously using the node's dynamic time-series features (such as historical utilization and temperature changes) and static features (such as hardware configuration) as input during model training. The node health scoring model refers to the final model with a specific "node health scoring model architecture" obtained through the above-mentioned "time-series-static joint training" process using processed historical data, node health features, and health labels.

[0036] Optionally, the historical health failures of the GPU cluster can be identified based on the historical monitoring data. Figure 2 A schematic diagram of historical health fault analysis for implementing the scheduling method for GPU computing resources provided in one embodiment of the present invention: First, the historical monitoring data of the GPU cluster is loaded, and then data preprocessing, including cleaning, normalization and other operations, is performed to ensure data quality. Next, feature engineering is performed to extract historical fault feature vectors to provide a basis for subsequent analysis. After that, an anomaly detection model is constructed, and based on the historical fault feature vectors, a machine learning algorithm is used to identify abnormal patterns. After the anomaly is detected, a root cause analysis of the fault is performed to locate the source of the problem. Finally, a health report is generated to determine the historical health faults.

[0037] Furthermore, based on the node health characteristics, a preset node health scoring model architecture is subjected to temporal-static joint training to obtain a node health scoring model, including: Based on the node health characteristics, the node health scoring model architecture is subjected to temporal-static joint training to obtain a trained node health scoring model to output a node health category probability value of the GPU cluster corresponding to the node health characteristics; Based on the node health category probability value of the GPU cluster, the training node health score of the GPU cluster is calculated using the following formula:

[0038] in, express The health score of the training nodes in the GPU cluster at any given moment. represents the smoothing coefficient, express The node health category probability value of the GPU cluster at this moment, express The health score of the training nodes in the GPU cluster at any given moment; Analyzing the reliability of the training node health score; When the scoring reliability meets a preset scoring reliability threshold, the trained node health scoring model is used as the node health scoring model.

[0039] The trained node health scoring model refers to a node health scoring model in the verification or evaluation phase, obtained through training using the "time-series-static joint training" method based on extracted "node health features" and corresponding "health labels." The node health category probability value refers to a vector or array output by the trained node health scoring model after prediction based on the input "node health features" (including time-series and static features). The trained node health score is a numerical score representing the health level of the corresponding node in the GPU cluster at a specific time t, calculated using a given recursive formula combined with the "node health category probability value." The smoothing coefficient is a value between 0 and 1 (usually close to 1, such as 0.9 or 0.95). It controls the degree to which a new prediction affects the current score. The larger it is, the greater the impact of the new prediction value, and the faster the score responds to state changes; The smaller it is, the greater the weight of the historical score, the smoother the score, and the stronger the anti-interference ability. The score reliability refers to the credibility of the "training node health score" output by the trained node health scoring model. The score reliability threshold refers to a pre-set standard for judging whether the "score reliability" reaches an acceptable level.

[0040] S3. Authenticate the received GPU task request. When the authentication is successful, analyze the video memory requirement, computation-intensive identifier, and communication dependency of the GPU task request.

[0041] It should be explained that the authentication of received GPU task requests means that when the GPU cluster management system receives a request (usually a computing task that requires GPU resources to run, such as deep learning training, large-scale parallel computing, etc.), it performs a series of verification steps to ensure the legitimacy of the request, the reliability of the source, and whether the requester has the right to use GPU resources. The authentication includes identity verification, permission verification, and request integrity / source verification.

[0042] The present invention analyzes the video memory requirements, computationally intensive identifiers, and communication dependencies of the GPU task request to accurately match task requirements with cluster resources and optimize scheduling decisions.

[0043] Specifically, parsing the video memory requirements, computationally intensive identifiers, and communication dependencies of the GPU task request includes: Extracting the video memory parameters requested by the GPU task; and normalizing the video memory parameters to obtain standardized video memory parameters to determine the video memory requirements of the GPU task request; Extracting the intensive identifier of the GPU task request to calculate the computing density set of the GPU task request; Determining a computationally intensive identifier of the GPU task request based on the computational density set; Extracting a communication field of the GPU task request to identify a communication mode of the GPU task request; The communication dependency of the GPU task request is determined through the communication mode.

[0044] The "memory parameter" refers to the amount of GPU memory required for the task, as explicitly specified in the GPU task request. The "normalized memory parameter" refers to the value obtained by converting the original memory parameter (memory parameter) into a uniform format or unit. For example, the string "16GB" is converted to the value 16384 (MB). The "memory requirement" refers to the total amount of GPU memory required for the task, which has been standardized (normalized memory parameter) and possibly verified or adjusted. The "intensiveness flag" refers to a tag that may be explicitly included in the GPU task request. The "computational density set" is an intermediate result of comprehensively evaluating the computational intensity of the task by analyzing other information in the task request (such as the task script, used framework libraries, input data size, and the ratio of the requested number of CPU / GPU cores to memory). The "computational intensity flag" refers to a clear classification or label of the task's computational characteristics, ultimately derived from the analysis of the computational density set. For example, it may be "high" (high computational intensive), "medium" (medium), or "low". The communication field refers to the information related to communication between nodes or between processes within a node during task runtime contained in the GPU task request. The communication mode refers to the communication method that the task will adopt during runtime, as identified by parsing the communication field. The communication dependency refers to the communication connection or data exchange relationship that must exist between the various parts (or nodes) of the task during runtime, which is derived based on the identified communication mode.

[0045] Optionally, the computing density set of the GPU task request may be calculated by inferring computing density by analyzing a task script or a framework (such as PyTorch / TensorFlow).

[0046] S4. Determine the SLA weight of the GPU task request according to the task type of the GPU task request, calculate the resource shortage sensitivity of the GPU task request based on the video memory requirement, and calculate the target task priority of the GPU task request in combination with the SLA weight.

[0047] The present invention determines the SLA weight of a GPU task request based on the task type of the GPU task request, thereby prioritizing tasks with higher quality of service requirements during the scheduling process, thereby better managing cluster resources and meeting user expectations. The SLA weight is used to quantify or represent the relative importance of the SLA requirement of a GPU task request in scheduling decisions. This weight is predefined based on the task type. The scheduling system maintains a mapping table or rule that associates each task type with a default SLA weight value.

[0048] The present invention calculates the resource shortage sensitivity of the GPU task request based on the video memory demand, which can provide a basis for subsequent priority optimization.

[0049] Specifically, the calculating the resource shortage sensitivity of the GPU task request based on the video memory requirement includes: Identify the cluster memory pressure of the GPU cluster corresponding to the GPU task request; Analyzing the normalized task requirements of the GPU task request based on the video memory requirements; According to the cluster memory pressure and the normalized task requirements, the resource shortage sensitivity of the GPU task request is calculated using the following formula:

[0050] in, Indicates that the GPU task request is in Sensitivity to resource scarcity at all times, Indicates The cluster memory pressure at all times, represents the normalized task requirements, represents the waiting sensitivity coefficient, represents the waiting penalty factor, represents the smoothing term.

[0051] Among them, the cluster memory pressure refers to the global tension of the memory resources in the current GPU cluster, the normalized task demand refers to the ratio of the memory demand of the task request to the maximum memory capacity of a single card, the waiting sensitivity coefficient refers to the amplification intensity of the resource sensitivity of the control waiting time, which can be 0.5, the waiting penalty factor refers to the ratio of the waiting time of the task to the preset timeout threshold, which quantifies the urgency of waiting, the smoothing term refers to a very small positive constant, which prevents the formula from outputting a zero value in extreme cases, and the resource shortage sensitivity refers to the responsiveness of the quantitative task to the tension of the memory resources.

[0052] Optionally, the waiting penalty factor may be set according to SLA requirements (eg, 1 hour = 3600 seconds).

[0053] The present invention combines the SLA weight to calculate the target task priority of the GPU task request, which can further determine the importance of the task and improve the rationality of resource scheduling.

[0054] Specifically, calculating the target task priority of the GPU task request in combination with the SLA weight includes: Identifying user information and a deadline for the GPU task request; Determining a user priority level of the GPU task request based on the user information; Determining a time criticality of the GPU task request based on the deadline; Define the SLA weight, the user information, the deadline, and the SLA weight coefficient, user information weight coefficient, time shortage weight coefficient, and resource shortage weight coefficient corresponding to the resource shortage sensitivity of the GPU task request to calculate the target task priority of the GPU task request.

[0055] The user information refers to an identifier associated with the user who submitted the GPU task request; the deadline refers to the time point at which the task specified in the GPU task request must be completed; the user priority is a priority score assigned to the user who submitted the task based on the user information; the time urgency is a quantitative indicator calculated based on the task deadline, indicating the urgency of the task's completion time; the SLA weight coefficient is a weight factor assigned to the task's SLA weight (Service Level Agreement Weight) when calculating the final target task priority; the user information weight coefficient is a weight factor assigned to the user priority when calculating the final target task priority; the time urgency weight coefficient is a weight factor assigned to the time urgency indicator when calculating the final target task priority; the resource scarcity weight coefficient is a weight factor assigned to the resource scarcity sensitivity when calculating the final target task priority; and the target task priority is a final task priority score calculated using the corresponding weight coefficients by comprehensively considering multiple factors, such as the SLA weight, user priority, time urgency, and resource scarcity sensitivity.

[0056] Optionally, the user priority level of the GPU task request is determined based on the user information mainly by combining the overall situation of the user (such as payment level, historical behavior, importance of the organization to which the user belongs).

[0057] S5. Determine the resource scheduling node group requested by the GPU task in the resource portrait through the computationally intensive identifier, the communication dependency, and the health score, and generate resource scheduling parameters for the resource scheduling node group, and execute the scheduling of the computing power resources of the GPU cluster based on the resource scheduling parameters.

[0058] The present invention determines through the computing-intensive identification, the communication dependency and the health score that the resource scheduling node group requested by the GPU task in the resource portrait can dynamically generate a resource scheduling node group that takes into account computing power, communication efficiency and node health, thereby improving the scheduling efficiency and task success rate of the GPU cluster.

[0059] In detail, determining the resource scheduling node group requested by the GPU task in the resource portrait through the computationally intensive identifier, the communication dependency, and the health score includes: Extracting a first candidate node that meets the GPU task request from the resource profile; Calculating a computing power matching capability of the first candidate node based on the computing intensiveness identifier of the GPU task request; Based on the communication dependency, calculating a network topology of the first candidate node; Screening a second candidate node for the first candidate node based on the computing power matching capability and the network topology; The health score is used to screen the health of the second candidate nodes to obtain the resource scheduling node group.

[0060] The first candidate nodes are a group of nodes initially selected from the resource profile of the entire GPU cluster. These nodes meet the minimum resource requirements of the GPU task request in terms of hardware specifications (such as GPU model, video memory size, and number of CPU cores). The computing power matching capability is a quantitative indicator that measures the degree of match between the computing power of the first candidate nodes and the computationally intensive requirements of the GPU task request. The network topology refers to the physical or logical connections between the first candidate nodes and between them and the cluster network infrastructure (such as switches and high-speed interconnects). The second candidate nodes are a subset of nodes further selected from the first candidate node set by comprehensively considering computing power matching capability and network topology. The resource scheduling node group is the final set of GPU nodes determined to actually execute the GPU task request.

[0061] Optionally, the computing power matching capability of the first candidate node can be calculated based on the computing intensiveness flag of the GPU task request by: If the task is marked as "highly computing intensive", the node with high GPU computing power (such as TFLOPS) is preferentially selected.

[0062] Optionally, the network topology of the first candidate node calculated based on the communication dependency relationship may be analyzed by physical location and logical grouping.

[0063] Finally, the present invention generates resource scheduling parameters for the resource scheduling node group, and scheduling the computing resources of the GPU cluster based on the resource scheduling parameters can achieve efficient scheduling of the computing resources of the GPU cluster. The resource scheduling parameters include task priority, number of CPUs, video memory, and inter-node bandwidth.

[0064] This method constructs a multi-dimensional resource feature vector by collecting the node hardware parameters and dynamic load indicators of the GPU cluster, and then forms a comprehensive and dynamic GPU cluster resource portrait. This process makes the cluster resource status visual and quantifiable, provides a solid data foundation for subsequent scheduling decisions, and effectively improves the transparency and accuracy of resource management. It uses historical data to establish a node health scoring model and generates health scores in real time, which significantly enhances the reliability of the cluster. The scheduling system can dynamically adjust resource allocation based on the health score and give priority to nodes in good health, thereby reducing the risk of task failure due to node failure and ensuring the stability and continuity of task execution. In terms of task request processing, this method first performs strict authentication to ensure that only legitimate requests can enter the system, thereby enhancing the security of the cluster. Then, through The graphics memory requirements, computationally intensive identifiers, and communication dependencies of the task request are parsed, and the priority of the target task is calculated in combination with the SLA weight and resource scarcity sensitivity determined by the task type. This priority calculation method that comprehensively considers task attributes and resource characteristics makes scheduling more fair and efficient, and can give priority to meeting the resource requirements of high-priority tasks while making rational use of resources. In determining the resource scheduling node group, the computationally intensive identifiers, communication dependencies, and health scores are comprehensively considered to generate resource scheduling parameters and execute scheduling. This method not only considers the computing power and health status of the nodes, but also pays attention to the communication efficiency between nodes, effectively reduces communication delays, and improves the overall performance of task execution. At the same time, the generation and application of resource scheduling parameters make the scheduling process more automated and intelligent, reduce manual intervention, and improve scheduling efficiency. Therefore, the present invention can improve the scheduling efficiency of GPU computing resources.

[0065] Example 2: like Figure 3 FIG. 1 is a functional module diagram of a scheduling system for GPU computing resources according to the present invention.

[0066] The GPU computing resource scheduling system 300 described in the present invention can be installed in an electronic device. Depending on the functionality implemented, the GPU computing resource scheduling system may include a resource profile building module 301, a node health scoring module 302, a task request analysis module 303, a task priority analysis module 304, and a computing resource scheduling module 305. The modules described in the present invention, also referred to as units, refer to a series of computer program segments that can be executed by an electronic device processor and perform a fixed function, and are stored in the electronic device's memory.

[0067] In the embodiment of the present invention, the functions of each module / unit are as follows: The resource profile building module 301 is used to collect node hardware parameters and dynamic load indicators of the GPU cluster to build a multi-dimensional resource feature vector of the GPU cluster, and build a resource profile of the GPU cluster based on the multi-dimensional resource feature vector; The node health scoring module 302 is used to establish a node health scoring model for the GPU cluster based on the historical monitoring data of the GPU cluster, and generate a health score for the cluster node corresponding to the GPU cluster; The task request analysis module 303 is used to authenticate the received GPU task request, and when the authentication is successful, analyze the video memory requirement, computation-intensive identifier, and communication dependency of the GPU task request; The task priority analysis module 304 is configured to determine an SLA weight of the GPU task request based on the task type of the GPU task request, calculate the resource shortage sensitivity of the GPU task request based on the video memory requirement, and calculate a target task priority of the GPU task request in combination with the SLA weight; The computing power resource scheduling module 305 is used to determine the resource scheduling node group requested by the GPU task in the resource portrait through the computing intensive identifier, the communication dependency and the health score, and generate resource scheduling parameters for the resource scheduling node group, and execute the scheduling of the computing power resources of the GPU cluster based on the resource scheduling parameters.

[0068] In detail, the modules in the GPU computing resource scheduling system 300 in the embodiment of the present invention adopt the same Figure 1 The scheduling method of GPU computing power resources described in the above is the same technical means and can produce the same technical effects, so it will not be repeated here.

[0069] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for scheduling GPU computing resources, characterized in that: The method comprises: Collecting node hardware parameters and dynamic load indicators of the GPU cluster to construct a multidimensional resource feature vector of the GPU cluster, and constructing a resource profile of the GPU cluster based on the multidimensional resource feature vector; Based on the historical monitoring data of the GPU cluster, a node health scoring model of the GPU cluster is established, and a health score of the cluster node corresponding to the GPU cluster is generated; Authenticate the received GPU task request, and if the authentication is successful, analyze the GPU task request's video memory requirements, computationally intensive identifiers, and communication dependencies; Determine an SLA weight of the GPU task request according to the task type of the GPU task request, calculate a resource shortage sensitivity of the GPU task request based on the video memory requirement, and calculate a target task priority of the GPU task request in combination with the SLA weight; Through the computationally intensive identifier, the communication dependency and the health score, the resource scheduling node group requested by the GPU task in the resource portrait is determined, and resource scheduling parameters of the resource scheduling node group are generated. The scheduling of the computing power resources of the GPU cluster is performed based on the resource scheduling parameters.

2. The method for scheduling GPU computing resources according to claim 1, wherein: The constructing of the multi-dimensional resource feature vector of the GPU cluster includes: Standardizing the node hardware parameters and dynamic load indicators of the GPU cluster to obtain standardized node hardware parameters and standardized dynamic load indicators; Extracting cluster timing characteristics of the GPU cluster based on the standardized node hardware parameters and the standardized dynamic load indicators; Performing topological feature encoding on the cluster time series features to obtain topological encoding features; A multi-dimensional resource feature vector of the GPU cluster is constructed according to the topological coding feature.

3. The method for scheduling GPU computing resources according to claim 2, wherein: The constructing a resource profile of the GPU cluster based on the multi-dimensional resource feature vector includes: Analyzing node-level features and cluster features of the GPU cluster based on the multi-dimensional resource feature vector; Define the image tags, computing logic, and application scenarios of the GPU cluster; A resource profile of the GPU cluster is constructed by combining the node-level features, the cluster features, the portrait tags, the computing logic, and the application scenario.

4. The method for scheduling GPU computing resources according to claim 3, wherein: The step of establishing a node health scoring model for the GPU cluster based on historical monitoring data of the GPU cluster includes: Preprocessing the historical monitoring data to obtain processed historical monitoring data; Extracting node health characteristics of the processed historical monitoring data; Identifying historical health failures of the GPU cluster based on the historical monitoring data to generate a health label for the GPU cluster; Based on the node health characteristics, a preset node health scoring model architecture is trained in a temporal-static joint manner to obtain a node health scoring model.

5. The method for scheduling GPU computing resources according to claim 4, wherein: The node health scoring model is obtained by performing temporal-static joint training on a preset node health scoring model architecture based on the node health characteristics, including: Based on the node health characteristics, the node health scoring model architecture is subjected to temporal-static joint training to obtain a trained node health scoring model to output a node health category probability value of the GPU cluster corresponding to the node health characteristics; Based on the node health category probability value of the GPU cluster, the training node health score of the GPU cluster is calculated using the following formula: ; in, express The health score of the training nodes in the GPU cluster at any given moment. represents the smoothing coefficient, express The node health category probability value of the GPU cluster at this moment, express The health score of the training nodes in the GPU cluster at any given moment; Analyzing the reliability of the training node health score; When the scoring reliability meets a preset scoring reliability threshold, the trained node health scoring model is used as the node health scoring model.

6. The method for scheduling GPU computing resources according to claim 5, wherein: The parsing of the video memory requirement, computationally intensive identifier, and communication dependency of the GPU task request includes: Extracting the video memory parameters requested by the GPU task; and normalizing the video memory parameters to obtain standardized video memory parameters to determine the video memory requirements of the GPU task request; Extracting the intensive identifier of the GPU task request to calculate the computing density set of the GPU task request; Determining a computationally intensive identifier of the GPU task request based on the computational density set; Extracting a communication field of the GPU task request to identify a communication mode of the GPU task request; The communication dependency of the GPU task request is determined through the communication mode.

7. The method for scheduling GPU computing resources according to claim 6, wherein: The calculating the resource shortage sensitivity of the GPU task request based on the video memory requirement includes: Identify the cluster memory pressure of the GPU cluster corresponding to the GPU task request; Analyzing the normalized task requirements of the GPU task request based on the video memory requirements; According to the cluster memory pressure and the normalized task requirements, the resource shortage sensitivity of the GPU task request is calculated using the following formula: ; in, Indicates that the GPU task request is in Sensitivity to resource scarcity at all times, Indicates The cluster memory pressure at all times, represents the normalized task requirements, represents the waiting sensitivity coefficient, represents the waiting penalty factor, represents the smoothing term.

8. The method for scheduling GPU computing resources according to claim 7, wherein: The calculating the target task priority of the GPU task request based on the combination of the SLA weight includes: Identifying user information and a deadline for the GPU task request; Determining a user priority level of the GPU task request based on the user information; Determining a time criticality of the GPU task request based on the deadline; Define the SLA weight, the user information, the deadline, and the SLA weight coefficient, user information weight coefficient, time shortage weight coefficient, and resource shortage weight coefficient corresponding to the resource shortage sensitivity of the GPU task request to calculate the target task priority of the GPU task request.

9. The method for scheduling GPU computing resources according to claim 8, wherein: The determining, by combining the computationally intensive identifier, the communication dependency, and the health score, a resource scheduling node group requested by the GPU task in the resource portrait includes: Extracting a first candidate node that meets the GPU task request from the resource profile; Calculating a computing power matching capability of the first candidate node based on the computing intensiveness identifier of the GPU task request; Based on the communication dependency, calculating a network topology of the first candidate node; Screening a second candidate node for the first candidate node based on the computing power matching capability and the network topology; The health score is used to screen the health of the second candidate nodes to obtain the resource scheduling node group.

10. A GPU computing resource scheduling system, characterized in that: The system comprises: A resource profile building module, configured to collect node hardware parameters and dynamic load indicators of a GPU cluster to build a multi-dimensional resource feature vector of the GPU cluster, and to build a resource profile of the GPU cluster based on the multi-dimensional resource feature vector; A node health scoring module is used to establish a node health scoring model for the GPU cluster based on the historical monitoring data of the GPU cluster, and generate a health score for the cluster node corresponding to the GPU cluster; A task request analysis module is used to authenticate the received GPU task request and, if authentication is successful, analyze the GPU task request's video memory requirements, computationally intensive identifiers, and communication dependencies. a task priority analysis module, configured to determine an SLA weight of the GPU task request based on the task type of the GPU task request, calculate the resource shortage sensitivity of the GPU task request based on the video memory requirement, and calculate a target task priority of the GPU task request in combination with the SLA weight; A computing power resource scheduling module is used to determine the resource scheduling node group requested by the GPU task in the resource portrait through the computing intensive identifier, the communication dependency and the health score, and generate resource scheduling parameters for the resource scheduling node group, and execute the scheduling of the computing power resources of the GPU cluster based on the resource scheduling parameters.

Images