Resource load balancing control method and cluster scheduler

Patent metadata

Field	Value
Publication number	US-10924535-B2
Application number	US-201916397674-A
Country	US
Kind code	B2
Filing date	Apr 29, 2019
Priority date	Oct 31, 2016
Publication date	Feb 16, 2021
Grant date	Feb 16, 2021

Abstract

Embodiments of the present invention disclose a resource load balancing control method and a cluster scheduler. The method includes: predicting performance data that is of an application deployed on each cluster node and that is in a preset time period; calculating a first standard deviation of the cluster system according to the predicted performance data of each cluster node; when the first standard deviation of the cluster system is greater than a preset threshold, determining an application migration solution according to a resource load balancing rule; and sending a application migration solution to the cluster application manager, to trigger the cluster application manager to perform resource load balancing control on the cluster system according to a application migration solution.

First claim

What is claimed is:

1. A resource load balancing control method, comprising:
predicting performance data that is of an application deployed on each cluster node and that is in a preset time period, wherein the performance data represents resource usage that is of an application deployed on each cluster node and that is in the preset time period;
calculating a first standard deviation of a cluster system according to the predicted performance data of each cluster node, wherein the first standard deviation represents a resource load balance degree of the cluster system in the preset time period; and
in response to determining that the first standard deviation of the cluster system is greater than a preset threshold, determining an application migration solution according to a resource load balancing rule to balance a current resource load of the cluster system or a resource load of the cluster system in the preset time period after a application migration solution is executed, wherein determining the application migration solution comprises:
sorting all cluster nodes according to a resource load of each cluster node;
classifying the sorted cluster nodes into either a heavy-load cluster node or a light-load cluster node;
traversing an application deployed on the heavy-load cluster node and an application deployed on the light-load cluster node, so as to exchange an application deployed on the heavy-load cluster node with an application deployed on the light-load cluster node;
determining a first application deployed on a first cluster node in the heavy-load cluster node and a second application deployed on a second cluster node in the light-load cluster node by maximizing a difference between a third standard deviation of the cluster system before an application exchange and a fourth standard deviation of the cluster system after the application exchange, wherein the third standard deviation is greater than the fourth standard deviation; and
generating an application migration solution instructing to exchange the first application deployed on the first cluster node with the second application deployed on the second cluster node.

2. The method according to claim 1 , wherein the method further comprises:
constructing training data for each cluster node by using historical performance data of an application deployed on the cluster node;
performing training by using the training data to obtain a data model; and
predicting the performance data that is of an application deployed on each cluster node and that is in a preset time period comprises:
for each cluster node, determining multiple applications deployed on the cluster node;
for each application, constructing historical performance data of a application that uses a current period as a reference, and using the historical performance data as input data;
inputting the input data into the data model to predict performance data that is of a application and that is in the preset time period; and
adding predicted performance data of all applications on the cluster node up to obtain the performance data that is of the cluster node and that is in the preset time period.

3. The method according to claim 2 , wherein performing training by using the training data to obtain the data model comprises:
performing training by using the training data according to a support vector machine (SVM) algorithm to obtain the data model; or
performing training by using the training data according to a linear regression algorithm to obtain the data model.

4. The method according to claim 1 , wherein the preset time period is in a unit of period; and when the preset time period is multiple periods, calculating the first standard deviation representing the resource load balance degree of the cluster system in the preset time period according to the predicted performance data of each cluster node comprises:
for each cluster node, calculating a resource load of the cluster node in each period according to the predicted performance data of the cluster node;
obtaining an average resource load value by averaging the calculated resource loads of the cluster node; and
determining the average resource load value as the resource load of the cluster node in the multiple periods.

5. The method according to claim 1 , wherein determining the application migration solution according to the resource load balancing rule comprises:
determining, according to the resource load balancing rule, the application migration solution by using current performance data of the application deployed on each cluster node, so that the current resource load of the cluster system is balanced after the application migration solution is executed; or
determining, according to the resource load balancing rule, the application migration solution by using performance data that is of the application deployed on each cluster node and that is in the preset time period, so that the resource load of the cluster system in the preset time period is balanced after the application migration solution is executed.

6. The method according to claim 5 , wherein the determining the application migration solution comprises:
selecting a cluster node with heaviest resource load from among multiple cluster nodes as a first cluster node;
traversing applications on the first cluster node to select a first application, wherein a difference between resource load of the first cluster node after a migration of the first application is simulated and average resource load of the cluster system before the migration of the first application is simulated is minimum;
traversing cluster nodes other than the first cluster node in the multiple cluster nodes to select a second cluster node, wherein after a migration of the first application to the second cluster node is simulated, a second standard deviation of the cluster system is minimum and the second standard deviation of the cluster system is less than the preset threshold; and
generating an application migration solution instructing to migrate the first application deployed on the first cluster node to the second cluster node.

7. A cluster scheduler, comprising at least one processor, at least one memory, and a communications interface, wherein:
the at least one memory is configured to store computer executable instructions; and
the at least one processor is connected to the at least one memory by using the communications interface, wherein the computer executable instructions instruct the at least one processor to perform one or more operations comprising:
predicting performance data that is of an application deployed on each cluster node and that is in a preset time period, wherein the performance data represents resource usage that is of an application deployed on each cluster node and that is in the preset time period;
calculating a first standard deviation of the cluster system according to the predicted performance data of each cluster node, wherein the first standard deviation represents a resource load balance degree of the cluster system in the preset time period; and
in response to determining that the first standard deviation of the cluster system is greater than a preset threshold, determining an application migration solution according to a resource load balancing rule to balance a current resource load of the cluster system or a resource load of the cluster system in the preset time period after a application migration solution is executed, wherein determining the application migration solution comprises:
sorting all cluster nodes according to a resource load of each cluster node;
classifying the sorted cluster nodes into either a heavy-load cluster node or a light-load cluster node;
traversing an application deployed on the h

Assignees

• Huawei Tech Co Ltd

Inventors

• Du Shubao
• Jin Xibo

Classifications

• G06F9/5088 Primary

  involving task migration · CPC title

• H04L67/1008 Primary

  based on parameters of servers, e.g. available memory or workload (monitoring of computer activity G06F11/30) · CPC title

• H04L67/1029

  using data related to the state of servers by a load balancer · CPC title

• G06F2209/5022

  Workload threshold · CPC title