Resource Consumption Analysis of Virtualised Server Consolidation System

Abstract

Resource consumption analysis is necessary because of continuous performance degradation of any long-running computing systems. Performance degradation is due to operating system's resource shrinkage. The most common causes of performance degradation include memory resource leakages, unreleased le descriptors, and numerical approximation errors. It is observed from literature that memory exhaustion has contributed majorly to the system failure. Resource consumption analysis is essential in a virtualized server consolidation system because Virtual Machines (VMs) use resources on demand. Another reason for selecting virtualized server consolidation system is due to the increased popularity of cloud computing. The key motivation behind this work is to help the system administrators to avoid accidental outage due to resource crunch. The key challenges in analyzing resource consumption data in server virtualized system are the volatility of the data and structural changes in the data. First, this thesis focussed on establishing performance degradation/aging e ect in virtualized server consolidation system. Then, we studied the e ectiveness of ARIMA models for forecasting the resource consumption data of virtualized server consolidation system; we found the presence of heteroscedasticity in the residuals of ARIMA model. The presence of heteroscedasticity in the residuals motivated us to try heteroscedastic models like ARCH and GARCH for resource forecasting. Another hybrid model namely ARIMA-ANN is also tried for resource forecasting. By combining di erent models, it is possible to capture various aspects of the underlying patterns. But we have experienced a slackness of t in all these models namely ARIMA, ARIMA-ARCH, ARIMA-GARCH, and ARIMA-ANN for the considered data. This slackness of t is due to the presence of structural changes in the resource consumption data. Further, Regime-Switching Models like MS-GARCH and SETAR are also used to analyze the data and found that these models have reasonably tted the considered data very well. Since there is no clear strategy for nding the order of GARCH and ARCH models, hence we tried di erent models and thus selected one model with least AIC, BIC, and log likelihood values for resource forecasting. An interested statistician could further investigate other mechanisms for nding the order of ARCH and GARCH models. As an extension, we would like to try these models and study the reasons for software aging in mobile platforms like Android systems in the near future.