CN114398148A - Power industry K8S dynamic container arrangement method and storage medium - Google Patents

Abstract

The invention discloses a K8S dynamic container arranging method and a storage medium in the power industry, aiming at the application characteristics of high availability, high safety, high concurrency and intranet in the power industry, by researching the task characteristics and occupied resources of clusters and containers, according to the load of the clusters and the load condition of the similar containers, based on an improved random forest model, the tasks in the same category are uniformly distributed to proper resource nodes in a sub-cluster to be executed, and when the load of the sub-cluster reaches a threshold value, a new sub-cluster is restarted, so that the dynamic, efficient and rapid arrangement of the containers is realized. The invention realizes the maximum utilization of the existing resources for the containers and solves the problem of optimal arrangement of the calculation containers under the condition of limited resources.

Description

A K8S dynamic container arrangement method and storage medium for the power industry

technical field

The invention relates to the field of Internet application in the electric power industry, in particular to a K8S dynamic container arrangement method and storage medium in the electric power industry, in particular to a K8S dynamic container arrangement method and storage medium in the electric power industry using an improved random forest.

Background technique

In general, applications in the power industry have certain particularities, requiring applications with high availability, high security, and high concurrency, and can only be applied on the intranet. Many K8S container allocation strategies that can be applied on the Internet cannot be applied to the power industry. As the de facto standard of containerization, Docker is the basic resource task scheduling unit. Each container is allocated corresponding resources, and the container encapsulates the runtime of the entire software, which can realize the integrated management of DevOps. However, the current container orchestration has problems in the differential allocation of resource nodes' load, ignoring the characteristics of different tasks, which can easily lead to unbalanced K8S container allocation, resulting in low utilization of system resources, slow execution speed, and long execution time.

In view of the allocation of containers in the K8S environment of the power grid, it is necessary to study the task attributes and load balancing from the aspects of container engine management, performance monitoring, anomaly detection, network configuration, I/O performance isolation, and Web deployment.

SUMMARY OF THE INVENTION

In response to the above problems, the inventor provides a K8S dynamic container orchestration method in the power industry. The inventor proposes that according to the utilization rate of each resource node, different services Request, update the dynamic weighted allocation algorithm of resource node weights, and arrange K8S containers. The improved random forest method is used for container arrangement, which realizes the maximum utilization of existing resources for container arrangement, and solves the problem of optimal arrangement of computing containers under the condition of limited resources.

The K8S dynamic container arrangement method in the power industry provided by the present invention includes the following steps:

Step S1: Classify Docker tasks. Divide tasks with execution time between (0, 50000) into 5 different levels, namely Super Long Task (SLT), Long Task (LT), and Normal Task (NT) ), Short Task (ST) and Super Short Task (SST), these five levels are used as the five class labels of the container task set, and the binary labels of the five class labels are shown in the following table:

Binary labels for tasks [10000] [01000] [00100] [00010] [00001] task type SLT LT NT ST SST

Therefore, the class attribute set of the container task set is expressed as Class={SLT,LT,NT,ST,SST}.

Step S2: Using the Class set of Step S1 as a classification condition, classify the task according to the improved random forest classifier training task classification model, and the classification process is divided into two steps.

Step S21: Construct a task set with a class label, create a Docker task container, and put it into a task resource pool to form a task set.

Step S22: Take the task set with class labels as input parameters, and return the optimal random forest classification model. When the container cloud task classification model based on the improved random forest is trained, it represents n base classifiers, and the leaf nodes of each base classifier represent a classification result. For example, [10000] represents the first classification result. That is, SLT, [01000] represents the second classification result, namely LT, and so on, the result is the final category of the task.

In step S3, K8S will use the final result of step S2. Divide the requested task set Task into sub-task sets Docker_Task={SLT, LT, NT, ST, SST} with different execution time lengths. The monitoring platform that comes with K8S obtains the resource usage of each resource node in the cluster, forming a set Resource={R1, R2, R3 R4}. K8S evaluates whether allocating the batch of tasks to each resource node will be overloaded, and if it is not overloaded, it will be allocated according to the plan. If it is overloaded, the resource nodes that may be overloaded will be removed from the resource set, and the batch of tasks will be divided into the remaining resource nodes to re-evaluate whether they are overloaded. Repeat this until the resource set is empty or the task set is empty. If the task set is null.

Beneficial effects of the present invention:

According to the characteristics of high availability, high security, high concurrency, and intranet applications in power industry applications, through the research on the task characteristics and occupied resources of clusters and containers, and the application of an application to improve the power industry K8S dynamic container arrangement method of random forest is in In terms of load balancing and the minimum time span, it has remarkable results, and can be widely used in scenarios where large-scale docker containers are orchestrated to improve resource utilization.

Description of drawings

FIG. 1 is a schematic diagram of the container cloud task classification model of the improved random forest of the present invention.

Figure 2 is a flow chart of the method of the present invention.

Detailed ways

Referring to FIG. 1 , the K8S dynamic container arrangement method in the power industry of the application of the present invention improves the random forest, including the following steps:

Step S1: Classify Docker tasks. Divide tasks whose execution duration (ms) is between (0, 50000) into 5 different levels, namely Super Long Task (SLT), Long Task (LT), and Normal Task (Normal Task). Task, NT), short task (Short Task, ST) and super short task (Super Short Task, SST), these five levels are used as the five class labels of the container task set, and the binary labels of the five class labels are shown in the following table :

Binary labels for tasks [10000] [01000] [00100] [00010] [00001] task type SLT LT NT ST SST

Therefore, the class attribute set of the container task set is expressed as Class={SLT,LT,NT,ST,SST}.

Step S2: Classify the task according to the improved random forest classifier training task classification model, and the classification process is divided into two steps.

Step S21: Construct a task set with class labels, create a Docker task container, and put it into the task resource pool to form an unordered task set order_Docker_Task={LT,NT,SLT,SST,ST}.

Step S22: Take the task set with the class label as the input parameter, and return the ordered task set of the optimal random forest classification. When the container cloud task classification model based on the improved random forest is trained, it represents n base classifiers, and the leaf nodes of each base classifier represent a classification result. For example, [10000] represents the first classification result. That is, SLT, [01000] represents the second classification result, that is, LT, and so on. The specific correspondence is shown in Figure 1. The result is the final category of the task, and the ordered set order_Docker_Task={SLT,LT, NT, ST, SST}.

In step S3, K8S will use the above final result. Divide the requested task set Task into sub-task sets with different execution time lengths order_Docker_Task={SLT, LT, NT, ST, SST}. The monitoring platform that comes with K8S obtains the resource usage of each resource node in the cluster, forming a set Resource={R1, R2, R3 R4}. K8S evaluates whether allocating the batch of tasks to each resource node will be overloaded, and if it is not overloaded, it will be allocated according to the plan. If it is overloaded, the resource nodes that may be overloaded will be removed from the resource set, and the batch of tasks will be divided into the remaining resource nodes to re-evaluate whether they are overloaded. Repeat this until the resource set is empty or the task set is empty. If the task set is null.

Claims (3)

1. A K8S dynamic container arrangement method in the power industry, applying random forest algorithm, is characterized in that, comprises the following steps: Step S1, classify according to the execution time of the Docker container task, and divide the tasks whose execution time is between (0, 50000) into 5 different levels, namely Super Long Task (SLT) and Long Task (Long Task). Task, LT), normal task (Normal Task, NT), short task (Short Task, ST) and super short task (Super Short Task, SST), these five levels are used as the five class labels of the container task set, five The binary labels of each class label are shown in the following table: Binary labels for tasks [10000] [01000] [00100] [00010] [00001] task type SLT LT NT ST SST Therefore, the class attribute set of the container task set is expressed as Class={SLT,LT,NT,ST,SST}; Step S2, using the Class set of Step S1 as a classification condition, classify the task according to the improved random forest classifier training task classification model, and the classification process is divided into two steps: Step S21, constructing a task set with a class label, creating a Docker task container, and putting it into a task resource pool to form a task set; Step S22, take the task set with the class label as the input parameter, and return the optimal random forest classification model; Step S3: Divide the requested task set Task into sub-task sets with different execution time lengths Docker_Task={SLT, LT, NT, ST, SST}; the monitoring platform that comes with K8S obtains the resource usage of each resource node under the cluster , forming a set Resource={R1, R2, R3 R4}; K8S evaluates whether the batch of tasks will be overloaded when allocated to each resource node. If it is not overloaded, it will be allocated according to the plan; Remove it from the resource set, re-evaluate whether the batch of tasks is evenly divided into the remaining resource nodes, and so on, until the resource set is empty or the task set is empty, if the task set is empty. 2. The method according to claim 1, wherein step S22 further comprises: When the container cloud task classification model based on the improved random forest is trained, it represents n base classifiers, and the leaf nodes of each base classifier represent a classification result. For example, [10000] represents the first classification result. That is, SLT, [01000] represents the second classification result, namely LT, and so on, the result is the final category of the task. 3. A computer-readable storage medium on which a computer program is stored, characterized in that, when the program is executed by a processor, the steps of the K8S dynamic container arrangement method in the power industry according to claim 1 or 2 are implemented.

Images