A kubernetes-based load scheduling method and system

Abstract

The application provides a kubernetes-based load scheduling method and system, which can obtain demand annotations of a game process container to be deployed, GPU computing power, network delay and load indicators of each cluster node in a kubernetes cluster; input the GPU computing power, network delay and load indicators of each cluster node into a node state recognition model to obtain a state recognition result of each cluster node; select candidate cluster nodes from the cluster nodes according to the demand annotations and the state recognition result of each cluster node; input the state recognition result of each candidate cluster node, the demand annotations and game business scenario features into a deep reinforcement learning model to obtain a priority score of each candidate cluster node; and select a target cluster node with the highest priority score from the candidate cluster nodes to deploy the game process container. The method can meet the load scheduling requirements of high-performance computing scenarios such as game services.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to a load scheduling method and system based on Kubernetes. Background Art

[0002] With the rapid development of cloud-native technologies, Kubernetes has become the de facto standard for deploying and managing enterprise-level containerized applications, thanks to its container orchestration capabilities, high scalability, and standardized interfaces. However, existing scheduling strategies often rely on hard-coded rules to filter nodes, such as those based on fixed latency thresholds. These methods fail to dynamically evaluate the real-time status of nodes, causing scheduling decisions to lag behind actual load demands. Consequently, they are unable to meet the load scheduling requirements of high-performance computing scenarios such as gaming services. Summary of the Invention

[0003] The technical problem to be solved by this application is to provide a Kubernetes-based load scheduling method and system that can meet the load scheduling requirements of high-performance computing scenarios such as gaming services. The specific solution is as follows:

[0004] A load scheduling method based on Kubernetes, comprising:

[0005] Obtain the requirements annotations for the game process container to be deployed, the GPU computing power, network latency, and load indicators of each cluster node in the Kubernetes cluster;

[0006] Input the GPU computing power, network latency, and load indicators of each cluster node into a pre-trained node state recognition model to obtain a state recognition result of each cluster node;

[0007] Selecting a candidate cluster node from each of the cluster nodes according to the requirement annotation and the status identification result of each of the cluster nodes;

[0008] Inputting the status recognition result of each candidate cluster node, the requirement annotation, and the current game business scenario characteristics into a deep reinforcement learning model to obtain a priority score for each candidate cluster node;

[0009] A target cluster node with the highest priority score is selected from each of the candidate cluster nodes to deploy the game process container.

[0010] In the above method, optionally, the node state recognition model includes a shared encoder, a first prediction branch network, a second prediction branch network, and a third prediction branch network;

[0011] Inputting the GPU computing power, network latency, and load indicators of each cluster node into a pre-trained node state recognition model to obtain a state recognition result of each cluster node includes:

[0012] For each of the cluster nodes, the GPU computing power, network delay and load indicators of the cluster node are input into the shared encoder of the node state recognition model to obtain the state fusion features of the cluster node; the state fusion features are respectively input into the first prediction branch network, the second prediction branch network and the third prediction branch network to obtain the GPU state prediction information output by the first prediction branch network, the network state prediction information output by the second prediction branch network and the load state prediction information output by the third prediction branch network; the GPU state prediction information, the network state prediction information and the load state prediction information are used as the state recognition results of the cluster node.

[0013] In the above method, optionally, the training process of the node status recognition model includes:

[0014] Obtaining an initial model to be trained and a first training data set, where the first training data set includes a plurality of first training sample data and a sample label for each first training sample data;

[0015] Training the initial model using a first training data set;

[0016] In a case where the initial model satisfies the first training completion condition, the initial model that satisfies the first training completion condition is determined as the trained node state recognition model.

[0017] Optionally, in the above method, selecting a candidate cluster node from each cluster node based on the requirement annotation and the status identification result of each cluster node includes:

[0018] Generate a first constraint condition according to the hardware performance threshold and the network delay threshold in the requirement annotation;

[0019] Determining a node that satisfies the first constraint condition among the cluster nodes as an initial cluster node;

[0020] Screening each of the initial cluster nodes according to a state identification result of each of the initial cluster nodes and a business scenario type indicated by the requirement annotation to obtain an optimized node set;

[0021] The optimized node set is prioritized according to a preset node sorting strategy, and a node subset that meets a preset quantity constraint is intercepted as a candidate cluster node set. The candidate cluster node set includes multiple candidate cluster nodes, and the sorting strategy is dynamically adjusted according to the matching degree between the resource supply capability of each node in the optimized node set and the container resource demand.

[0022] Optionally, the above method further includes, after selecting the target cluster node with the highest priority score from among the candidate cluster nodes to deploy the game process container, the following steps are further included:

[0023] Monitor runtime indicator data of the game process container on the target node and generate a monitoring data set including actual resource utilization, network communication quality, and load change rate;

[0024] By comparing the monitoring data set with the predicted data of the state identification result, a node state prediction deviation index is calculated, and the node label update frequency of the target cluster node is adjusted according to the node state prediction deviation index;

[0025] When the load growth rate in the monitoring data set continuously exceeds the corresponding prediction value and reaches a preset threshold, the target cluster node is marked as a prediction failure state.

[0026] A Kubernetes-based load scheduling system, including:

[0027] The acquisition unit is used to obtain the requirement annotations of the game process container to be deployed, the GPU computing power, network latency, and load indicators of each cluster node in the Kubernetes cluster;

[0028] A first execution unit is configured to input the GPU computing power, network latency, and load indicators of each cluster node into a pre-trained node state recognition model to obtain a state recognition result of each cluster node;

[0029] a first selection unit, configured to select a candidate cluster node from each of the cluster nodes according to the requirement annotation and a status identification result of each of the cluster nodes;

[0030] A second execution unit is configured to input the state recognition result of each candidate cluster node, the requirement annotation, and the current game business scenario characteristics into a deep reinforcement learning model to obtain a priority score for each candidate cluster node;

[0031] The third execution unit is configured to select a target cluster node with the highest priority score from among the candidate cluster nodes to deploy the game process container.

[0032] In the above system, optionally, the node state recognition model includes a shared encoder, a first prediction branch network, a second prediction branch network, and a third prediction branch network;

[0033] The first execution unit includes:

[0034] The first execution sub-unit is used to input the GPU computing power, network delay and load indicators of the cluster node into the shared encoder of the node state recognition model for each of the cluster nodes to obtain the state fusion features of the cluster node; input the state fusion features into the first prediction branch network, the second prediction branch network and the third prediction branch network respectively to obtain the GPU state prediction information output by the first prediction branch network, the network state prediction information output by the second prediction branch network and the load state prediction information output by the third prediction branch network; and use the GPU state prediction information, the network state prediction information and the load state prediction information as the state recognition result of the cluster node.

[0035] In the above system, optionally, the first execution unit includes:

[0036] A first acquisition subunit is configured to acquire an initial model to be trained and a first training data set, where the first training data set includes a plurality of first training sample data and a sample label for each first training sample data;

[0037] A training subunit, configured to train the initial model using a first training data set;

[0038] The second execution subunit is configured to determine the initial model that meets the first training completion condition as a trained node state recognition model when the initial model meets the first training completion condition.

[0039] In the above system, optionally, the first selection unit includes:

[0040] A generating subunit, configured to generate a first constraint condition according to a hardware performance threshold and a network delay threshold in the requirement annotation;

[0041] a determination subunit, configured to determine, among the cluster nodes, a node that satisfies the first constraint condition as an initial cluster node;

[0042] A third execution subunit is configured to screen each of the initial cluster nodes according to a state identification result of each of the initial cluster nodes and a business scenario type indicated by the requirement annotation to obtain an optimized node set;

[0043] A sorting subunit is used to prioritize the optimized node set according to a preset node sorting strategy, and intercept a node subset that meets the preset quantity constraint as a candidate cluster node set. The candidate cluster node set includes multiple candidate cluster nodes, and the sorting strategy is dynamically adjusted according to the matching degree between the resource supply capacity of each node in the optimized node set and the container resource demand.

[0044] The above system may optionally further include:

[0045] A monitoring unit, configured to monitor runtime indicator data of the game process container on the target node and generate a monitoring data set including actual resource utilization, network communication quality, and load change rate;

[0046] a calculation unit, configured to calculate a node state prediction deviation index by comparing the monitoring data set with the prediction data of the state identification result, and adjust the node label update frequency of the target cluster node according to the node state prediction deviation index;

[0047] The fourth execution unit is configured to mark the target cluster node as being in a prediction failure state when the load growth rate in the monitoring data set continuously exceeds the corresponding prediction value and reaches a preset threshold.

[0048] Based on the above-mentioned implementation of this application, a load scheduling method and system based on Kubernetes is provided, wherein the requirement annotation of the game process container to be deployed, the GPU computing power, network delay and load index of each cluster node in the Kubernetes cluster are obtained; the GPU computing power, network delay and load index of each cluster node are input into a pre-trained node state recognition model to obtain the state recognition result of each cluster node; based on the requirement annotation and the state recognition result of each cluster node, candidate cluster nodes are selected from each cluster node; the state recognition result of each candidate cluster node, the requirement annotation and the current game business scenario characteristics are input into a deep reinforcement learning model to obtain the priority score of each candidate cluster node; the target cluster node with the highest priority score is selected from each candidate cluster node to deploy the game process container. The method provided by the embodiment of this application can meet the load scheduling requirements of high-performance computing scenarios such as game services. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without any creative work.

[0050] Figure 1 A method flowchart of a kubernetes-based load scheduling method provided in the present application;

[0051] Figure 2 A flowchart of a training process of a node state recognition model provided in the present application;

[0052] Figure 3 A flowchart of a process of selecting a candidate cluster node from each cluster node provided in the present application;

[0053] Figure 4 A structural schematic diagram of a kubernetes-based load scheduling system provided in the present application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0055] In the present application, the term “comprises”, “comprising” or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or equipment including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or equipment. Without more limitation, the element defined by the statement “comprises a” does not exclude the presence of another identical element in the process, method, article or equipment including the element.

[0056] The embodiments of the present application provide a kubernetes-based load scheduling method, applied to an electronic device, and a method flowchart of the method is as shown in Figure 1 The specific steps include:

[0057] S101: Obtain the requirement annotation of a game process container to be deployed, GPU computing power, network delay and load index of each cluster node in a kubernetes cluster.

[0058] S102: Input the GPU computing power, network delay and load index of each cluster node into a pre-trained node state recognition model to obtain a state recognition result of each cluster node.

[0059] S103: Selecting a candidate cluster node from each cluster node according to the requirement annotation and the status identification result of each cluster node.

[0060] S104: Inputting the status recognition result of each candidate cluster node, the requirement annotation, and the current game business scenario characteristics into a deep reinforcement learning model to obtain a priority score for each candidate cluster node.

[0061] S105: Selecting a target cluster node with the highest priority score from among the candidate cluster nodes to deploy a game process container.

[0062] The method provided in the embodiments of the present application can meet the load scheduling requirements of high-performance computing scenarios such as gaming services.

[0063] In an embodiment provided by the present application, based on the above solution, optionally, the node state recognition model includes a shared encoder, a first prediction branch network, a second prediction branch network, and a third prediction branch network;

[0064] Inputting the GPU computing power, network latency, and load indicators of each cluster node into a pre-trained node state recognition model to obtain a state recognition result of each cluster node includes:

[0065] For each of the cluster nodes, the GPU computing power, network delay and load indicators of the cluster node are input into the shared encoder of the node state recognition model to obtain the state fusion features of the cluster node; the state fusion features are respectively input into the first prediction branch network, the second prediction branch network and the third prediction branch network to obtain the GPU state prediction information output by the first prediction branch network, the network state prediction information output by the second prediction branch network and the load state prediction information output by the third prediction branch network; the GPU state prediction information, the network state prediction information and the load state prediction information are used as the state recognition results of the cluster node.

[0066] In an embodiment provided in the present application, based on the above solution, optionally, the training process of the node state recognition model is as follows: Figure 2 Shown, including:

[0067] S201: Obtain an initial model to be trained and a first training data set, where the first training data set includes a plurality of first training sample data and a sample label for each first training sample data;

[0068] S202: Training the initial model using a first training data set;

[0069] S203: When the initial model satisfies the first training completion condition, the initial model that satisfies the first training completion condition is determined as a trained node state recognition model.

[0070] In an embodiment provided by the present application, based on the above solution, optionally, the process of selecting candidate cluster nodes from each cluster node according to the requirement annotation and the status identification result of each cluster node is as follows: Figure 3 Shown, including:

[0071] S301: Generate a first constraint condition according to the hardware performance threshold and the network delay threshold in the requirement annotation.

[0072] S302: Determine a node that satisfies the first constraint condition among the cluster nodes as an initial cluster node.

[0073] S303: Filter each of the initial cluster nodes according to the status identification result of each of the initial cluster nodes and the business scenario type indicated by the requirement annotation to obtain an optimized node set.

[0074] S304: Prioritize the optimized node set according to a preset node sorting strategy, and intercept a node subset that meets the preset quantity constraint as a candidate cluster node set. The candidate cluster node set includes multiple candidate cluster nodes, and the sorting strategy is dynamically adjusted according to the matching degree between the resource supply capability of each node in the optimized node set and the container resource demand.

[0075] In an embodiment provided by the present application, based on the above solution, optionally, after selecting the target cluster node with the highest priority score from each candidate cluster node to deploy the game process container, the method further includes:

[0076] Monitor runtime indicator data of the game process container on the target node and generate a monitoring data set including actual resource utilization, network communication quality, and load change rate;

[0077] By comparing the monitoring data set with the predicted data of the state identification result, a node state prediction deviation index is calculated, and the node label update frequency of the target cluster node is adjusted according to the node state prediction deviation index;

[0078] When the load growth rate in the monitoring data set continuously exceeds the corresponding prediction value and reaches a preset threshold, the target cluster node is marked as a prediction failure state.

[0079] See also Figure 4, is a structural diagram of a Kubernetes-based load scheduling method system provided in an embodiment of the present application, the system comprising:

[0080] The acquisition unit 401 is used to obtain the requirement annotation of the game process container to be deployed, the GPU computing power, network latency and load indicators of each cluster node in the Kubernetes cluster;

[0081] The first execution unit 402 is configured to input the GPU computing power, network latency, and load indicators of each cluster node into a pre-trained node state recognition model to obtain a state recognition result of each cluster node;

[0082] A first selection unit 403 is configured to select a candidate cluster node from each of the cluster nodes according to the requirement annotation and the status identification result of each of the cluster nodes;

[0083] The second execution unit 404 is configured to input the state recognition result of each candidate cluster node, the requirement annotation, and the current game business scenario characteristics into a deep reinforcement learning model to obtain a priority score for each candidate cluster node;

[0084] The third execution unit 405 is configured to select a target cluster node with the highest priority score from among the candidate cluster nodes and deploy the game process container.

[0085] In an embodiment provided by the present application, based on the above solution, optionally, the node state recognition model includes a shared encoder, a first prediction branch network, a second prediction branch network, and a third prediction branch network;

[0086] The first execution unit includes:

[0087] The first execution sub-unit is used to input the GPU computing power, network delay and load indicators of the cluster node into the shared encoder of the node state recognition model for each of the cluster nodes to obtain the state fusion features of the cluster node; input the state fusion features into the first prediction branch network, the second prediction branch network and the third prediction branch network respectively to obtain the GPU state prediction information output by the first prediction branch network, the network state prediction information output by the second prediction branch network and the load state prediction information output by the third prediction branch network; and use the GPU state prediction information, the network state prediction information and the load state prediction information as the state recognition result of the cluster node.

[0088] In an embodiment provided in the present application, based on the above solution, optionally, the first execution unit includes:

[0089] A first acquisition subunit is configured to acquire an initial model to be trained and a first training data set, where the first training data set includes a plurality of first training sample data and a sample label for each first training sample data;

[0090] A training subunit, configured to train the initial model using a first training data set;

[0091] The second execution subunit is configured to determine the initial model that meets the first training completion condition as a trained node state recognition model when the initial model meets the first training completion condition.

[0092] In an embodiment provided in the present application, based on the above solution, optionally, the first selection unit includes:

[0093] A generating subunit, configured to generate a first constraint condition according to a hardware performance threshold and a network delay threshold in the requirement annotation;

[0094] a determination subunit, configured to determine, among the cluster nodes, a node that satisfies the first constraint condition as an initial cluster node;

[0095] A third execution subunit is configured to screen each of the initial cluster nodes according to a state identification result of each of the initial cluster nodes and a business scenario type indicated by the requirement annotation to obtain an optimized node set;

[0096] A sorting subunit is used to prioritize the optimized node set according to a preset node sorting strategy, and intercept a node subset that meets the preset quantity constraint as a candidate cluster node set. The candidate cluster node set includes multiple candidate cluster nodes, and the sorting strategy is dynamically adjusted according to the matching degree between the resource supply capacity of each node in the optimized node set and the container resource demand.

[0097] In an embodiment provided in the present application, based on the above solution, optionally, the system further includes:

[0098] A monitoring unit, configured to monitor runtime indicator data of the game process container on the target node and generate a monitoring data set including actual resource utilization, network communication quality, and load change rate;

[0099] a calculation unit, configured to calculate a node state prediction deviation index by comparing the monitoring data set with the prediction data of the state identification result, and adjust the node label update frequency of the target cluster node according to the node state prediction deviation index;

[0100] The fourth execution unit is configured to mark the target cluster node as being in a prediction failure state when the load growth rate in the monitoring data set continuously exceeds the corresponding prediction value and reaches a preset threshold.

[0101] It should be noted that the various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the various embodiments can be referenced to each other.

[0102] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0103] For the convenience of description, the above system is described as being divided into various units according to their functions. Of course, when implementing this application, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0104] Through the description of the above embodiments, it can be seen that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present application or certain parts of the embodiments.

[0105] The above is a detailed introduction to a Kubernetes-based load scheduling method provided by this application. This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method of this application and its core ideas. At the same time, for general technical personnel in this field, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A load scheduling method based on kubernetes, characterized in that: include: Obtain the requirements annotations for the game process container to be deployed, the GPU computing power, network latency, and load indicators of each cluster node in the Kubernetes cluster; Input the GPU computing power, network latency, and load indicators of each cluster node into a pre-trained node state recognition model to obtain a state recognition result of each cluster node; Selecting a candidate cluster node from each of the cluster nodes according to the requirement annotation and the status identification result of each of the cluster nodes; Inputting the status recognition result of each candidate cluster node, the requirement annotation, and the current game business scenario characteristics into a deep reinforcement learning model to obtain a priority score for each candidate cluster node; Selecting the target cluster node with the highest priority score from among the candidate cluster nodes to deploy the game process container; The selecting a candidate cluster node from each of the cluster nodes according to the requirement annotation and the status identification result of each of the cluster nodes includes: Generate a first constraint condition according to the hardware performance threshold and the network delay threshold in the requirement annotation; Determining a node that satisfies the first constraint condition among the cluster nodes as an initial cluster node; Screening each of the initial cluster nodes according to a state identification result of each of the initial cluster nodes and a business scenario type indicated by the requirement annotation to obtain an optimized node set; The optimized node set is prioritized according to a preset node sorting strategy, and a node subset that meets a preset quantity constraint is intercepted as a candidate cluster node set. The candidate cluster node set includes multiple candidate cluster nodes, and the sorting strategy is dynamically adjusted according to the matching degree between the resource supply capability of each node in the optimized node set and the container resource demand.

2. The method according to claim 1, characterized in that The node state recognition model includes a shared encoder, a first prediction branch network, a second prediction branch network and a third prediction branch network; Inputting the GPU computing power, network latency, and load indicators of each cluster node into a pre-trained node state recognition model to obtain a state recognition result of each cluster node includes: For each of the cluster nodes, the GPU computing power, network delay and load indicators of the cluster node are input into the shared encoder of the node state recognition model to obtain the state fusion features of the cluster node; the state fusion features are respectively input into the first prediction branch network, the second prediction branch network and the third prediction branch network to obtain the GPU state prediction information output by the first prediction branch network, the network state prediction information output by the second prediction branch network and the load state prediction information output by the third prediction branch network; the GPU state prediction information, the network state prediction information and the load state prediction information are used as the state recognition results of the cluster node.

3. The method according to claim 2, characterized in that The training process of the node status recognition model includes: Obtain an initial model to be trained and a first training data set, where the first training data set includes a plurality of first training sample data and a sample label for each first training sample data; the first training sample data includes historical GPU computing power, historical network latency, and historical load indicators; Training the initial model using a first training data set; In a case where the initial model satisfies the first training completion condition, the initial model that satisfies the first training completion condition is determined as the trained node state recognition model.

4. The method according to claim 1, wherein After selecting the target cluster node with the highest priority score from among the candidate cluster nodes to deploy the game process container, the method further includes: Monitor runtime indicator data of the game process container on the target cluster node to generate a monitoring data set including actual resource utilization, network communication quality, and load change rate; By comparing the monitoring data set with the predicted data of the state identification result, a node state prediction deviation index is calculated, and the node label update frequency of the target cluster node is adjusted according to the node state prediction deviation index; When the load growth rate in the monitoring data set continuously exceeds the corresponding prediction value and reaches a preset threshold, the target cluster node is marked as a prediction failure state.

5. A load scheduling system based on kubernetes, characterized in that: include: The acquisition unit is used to obtain the requirement annotations of the game process container to be deployed, the GPU computing power, network latency, and load indicators of each cluster node in the Kubernetes cluster; A first execution unit is configured to input the GPU computing power, network latency, and load indicators of each cluster node into a pre-trained node state recognition model to obtain a state recognition result of each cluster node; a first selection unit, configured to select a candidate cluster node from each of the cluster nodes according to the requirement annotation and a status identification result of each of the cluster nodes; A second execution unit is configured to input the state recognition result of each candidate cluster node, the requirement annotation, and the current game business scenario characteristics into a deep reinforcement learning model to obtain a priority score for each candidate cluster node; A third execution unit is configured to select a target cluster node with the highest priority score from among the candidate cluster nodes and deploy the game process container thereto; The first selection unit includes: A generating subunit, configured to generate a first constraint condition according to a hardware performance threshold and a network delay threshold in the requirement annotation; a determination subunit, configured to determine, among the cluster nodes, a node that satisfies the first constraint condition as an initial cluster node; A third execution subunit is configured to screen each of the initial cluster nodes according to a state identification result of each of the initial cluster nodes and a business scenario type indicated by the requirement annotation to obtain an optimized node set; A sorting subunit is used to prioritize the optimized node set according to a preset node sorting strategy, and intercept a node subset that meets the preset quantity constraint as a candidate cluster node set. The candidate cluster node set includes multiple candidate cluster nodes, and the sorting strategy is dynamically adjusted according to the matching degree between the resource supply capacity of each node in the optimized node set and the container resource demand.

6. The system according to claim 5, characterized in that The node state recognition model includes a shared encoder, a first prediction branch network, a second prediction branch network and a third prediction branch network; The first execution unit includes: A first execution subunit is configured to input, for each of the cluster nodes, the GPU computing power, network latency, and load indicators of the cluster node into a shared encoder of the node state recognition model to obtain a state fusion feature of the cluster node; The state fusion features are respectively input into the first prediction branch network, the second prediction branch network and the third prediction branch network to obtain GPU state prediction information output by the first prediction branch network, network state prediction information output by the second prediction branch network and load state prediction information output by the third prediction branch network; the GPU state prediction information, the network state prediction information and the load state prediction information are used as the state recognition results of the cluster node.

7. The system according to claim 6, characterized in that The first execution unit includes: A first acquisition subunit is configured to acquire an initial model to be trained and a first training data set, where the first training data set includes a plurality of first training sample data and a sample label for each first training sample data; A training subunit, configured to train the initial model using a first training data set; The second execution subunit is configured to determine the initial model that meets the first training completion condition as a trained node state recognition model when the initial model meets the first training completion condition.

8. The system according to claim 5, wherein: Also includes: A monitoring unit, configured to monitor runtime indicator data of the game process container on the target cluster node and generate a monitoring data set including actual resource utilization, network communication quality, and load change rate; a calculation unit, configured to calculate a node state prediction deviation index by comparing the monitoring data set with the prediction data of the state identification result, and adjust the node label update frequency of the target cluster node according to the node state prediction deviation index; The fourth execution unit is configured to mark the target cluster node as being in a prediction failure state when the load growth rate in the monitoring data set continuously exceeds the corresponding prediction value and reaches a preset threshold.

Images