Server Load Prediction Based on Dynamic Neural Networks

Abstract

Predicting server load is involved in distributed system applications such as load balancing and load sharing. Applying machine learning based methods for load prediction in distributed system applications can improve the availability and performance of these applications. Many machine learning methods have been applied for load prediction. However, some researches show that applying Neural Networks (NN) technique is more efficient in predicting the load in future time. This paper is to investigate and compare different dynamic NN models in server load prediction such as Time-Delay Neural Network (TDNN) and Nonlinear Autoregressive Network with eXogenous inputs (NARX). Data used to forecast is acquired from Webmail server of Palestine Polytechnic University (PPU). Results have shown that NARX model provide better performance in comparison to TDNN model in server load prediction.