Sparse Regression Model to Predict a Server Load for Dynamic Adjustments of Server Resources
Takaya Miyazawa, Hiroaki Harai, Yusuke Yokota, Yasushi Naruse · 2019
Along with rapid prevalence of Internet of things (IoT) and mobile network services, network function virtualization (NFV) based infrastructure requires dynamic computational resource adjustments in response to time-varying environments (network traffic, resource usage, failure status, etc). To provide agile resource control and adaptiveness, it is effective to predict a virtual server load by means of machine learning technologies for proactive control. In this paper, we propose a regression analysis model utilizing a sparse modeling to predict the average server load in a future specific time period on the basis of the server load in a past specific time period. We target at the number of access to a Web server as a virtual server load. The model can reduce the prediction error compared to a traditional least square method. Besides, we propose two methods to further reduce prediction errors: (A) an algorithm for determining the time period targeted for explanatory variables, and (B) two-stage regression analyses. As a result of MATLAB calculations, we show that application of the sparse modeling can reduce the prediction error by more than 40%, and besides, the above two proposed methods are effective at further reducing the prediction error with a few minutes’ learning time. Our model also contributes to make humans’ posterior analysis easier by making a short list of explanatory variables of data than conventional methods.