Load balancing method, device and system based on dynamic probability model

Abstract

The invention provides a load balancing method based on a dynamic probability model. The load balancing method comprises the following steps: periodically acquiring state information of all service nodes; scoring the service nodes according to the state information of the service nodes, and obtaining score values corresponding to the service nodes; calculating a corresponding hit probability according to the score value of each service node; dividing the intervals (0,1] into hit intervals corresponding to the service nodes according to the hit probabilities; according to the service request, generating a uniformly distributed random number as a projection value in an interval (0,1] range; and obtaining the service node corresponding to the hit interval where the projection value is locatedas the load node of the service. According to the load balancing method provided by the invention, the instantaneous state of the server does not need to be completely depended on, the information ofthe suboptimal nodes can be effectively utilized, the peak value problem of node access under high concurrency can be effectively optimized, hard decision is avoided, and the load of each service node becomes balanced and smooth.

Description

technical field [0001] The present invention relates to the technical field of distributed services, in particular to a load balancing method, device and system based on a dynamic probability model. Background technique [0002] Distributed service deployment has been widely used at present, such as the current popular microservice architecture. In actual production, a service usually has multiple servers that are actually running. When a user actually calls a service, this task should be specified Which service node entity is distributed to belongs to the category of load balancing. Commonly used load balancing methods include polling, minimum number of connections, fastest response time, priority weight, source information hash, etc. They rely on the status information of each server node or the source information of the applicant to determine the attribution node of the service. Among the above strategies, the polling strategy has a poor balance effect, and other strateg...

AI Technical Summary

Problems solved by technology

Among the above strategies, the polling strategy has a poor balance effect, and other strategies have the problem of instantaneous peak value; the instantaneous peak value is mainly caused by the transient state dependence and hard judgment, and the instantaneous state dependence is due to the load information of the service node (such as the number of connections, Response time) cannot be perceived instantaneously, but must rely on the periodic message interaction of service nodes to obtain it. When the server has a high degree of concurrency and a huge amount of instantaneous access, because the state information cannot be updated instantaneously, the load strategy that relies on these state information will As a result, a large number of users may be allocated to the same node in an instant, resulting in an instantaneous peak value; because the hard decision only selects the optimal node under a certain strategy each time, and ignores the suboptimal node, resulting in the loss of effective information

[0003] Due to the shortcomings of instantaneous state dependence and hard judgment, the existing load balancing method has the problem of instantaneous peak value in practical applications, and it is particularly prominent in some scenarios with extremely high instantaneous business, such as "second kill" scenarios and ticket grabbing scenarios. The momentary peak access of the node may directly cause the node to be paralyzed, which will seriously affect the quality of service

Images