NARX-Based Locally Distributed Web-Servers Load Prediction

No Thumbnail Available

Date

2020-02

Journal Title

Journal ISSN

Volume Title

Publisher

Addis Ababa University

Abstract

With the continuous development of network technology, the need to have a page view of users and the load on servers has grown exponentially, resulting in temporary loss of services. Distributed computing systems are becoming a widely used paradigm to provide high performance and uninterrupted services to users. In such system, it is crucial to use effective resource management techniques to handle a large number of requests and provide dependable services with high quality constantly. Indeed, both service interruptions and resource waste can be reduced with the implementation of an eff ective prediction system. One promising approach is realizing artificial neural network algorithm to predict resource usage series of server. Several research has been conducted using different combination of load descriptor to predict resource usage series of server. Yet, queue time, is believed to be a good load descriptor of a server, because it gives a good estimate of job response time. It strives to produce a global improvement in system performance. However, this load descriptor, still not be seen in the study of resource usage prediction of server. Thus, in this research, we have investigated nonlinear autoregressive network with exogenous inputs (NARX) neural network multi-step ahead predictability of web server load. Besides it, three different training algorithms: Lavenberg-Marquardt (LM), Bayesian Regularization (BR) and Scaled Conjugate Gradient (SCG) forecasting accuracy were evaluated. We have collected resource usage series of a week data from locally distributed web servers of ethio telecom business support system. Two Cases (Case-1 and Case-2) of experiments were conducted to evaluate the performance of the algorithms; using as input in the first Case CPU and memory, and in the second Case CPU, memory and queue time. MATLAB was employed to verify the prediction accuracy of the algorithms. The results of the simulation show that for 12-step ahead web server load prediction, LM learning algorithm, Case-2 approach has registered the best prediction accuracy with MAPE of 4.459%, followed by the BR learning algorithm with MAPE of 4.649%. SCG was the lowest performer with MAPE of 5.610%. Thus, accuracy in prediction is necessary since the more efficient resources can be managed in data centers. Therefore, having such a model, enhances the process of working towards reliable services.

Description

Keywords

NARX neural network, training algorithm, web server load, multi-step ahead prediction

Citation