Prediction of Distributed Systems State Based on Monitoring Data

Autonomic behavior has emerged as a solution for the issues related to performance improvement and resource-usage optimization in large scale distributed systems. This solution relies on monitoring services to keep track of the states of the managed systems. However, most of the monitoring services are designed to provide general resource information and do not consider specific information for higher-level services, lacking important control capabilities. In this context, a dynamic adaptation layer is required. Based on the collected monitoring information in conjunction with some planning and prediction algorithms, it should be able to reactive and proactive deal with detected or predicted conditions. This paper presents a prediction architecture developed within the MonALISA monitoring framework, providing methods for estimating future values for different parameters on various periods of time. The predictions are used to enhance the self-adaptive behavior of several data intensive applications. Our research was focused on machine learning algorithms correlated with statistical techniques for data mining purposes in order to perform n-step-ahead time series predictions and to evaluate their performances dynamical.