APPLIED MACHINE LEARNING IN LOAD BALANCING

Abstract

A common way to maintain the quality of service on systems that are growing rapidly is by increasing server specifications or by adding servers. The utility of servers can be balanced with the presence of a load balancer to manage server loads. In this paper, we propose a machine learning algorithm that utilizes server resources CPU and memory to forecast the future of resources server loads. We identify the timespan of forecasting should be long enough to avoid dispatcher's lack of information server distribution at runtime. Additionally, server profile pulling, forecasting server resources, and dispatching should be asynchronous with the request listener of the load balancer to minimize response delay. For production use, we recommend that the load balancer should have friendly user interface to make it easier to be configured, such as adding resources of servers as parameter criteria. We also recommended from beginning to start to save the log data server resources because the more data to process, the more accurate prediction of server load will be.