Using computational intelligence to identify performance bottlenecks in a computer system

Abstract

System administrators have to analyze a number of system parameters to identify performance bottlenecks in a system. The major contribution of this paper is a utility - EvoPerf - which has the ability to autonomously monitor different system-wide parameters, requiring no user intervention, to accurately identify performance based anomalies (or bottlenecks). EvoPerf uses Windows Perfmon utility to collect a number of performance counters from the kernel of Windows OS. Subsequently, we show that artificial intelligence based techniques - using performance counters - can be used successfully to design an accurate and efficient performance monitoring utility. We evaluate feasibility of six classifiers - UCS, GAssist-ADI, GAssist-Int, NN-MLP, NN-RBF and J48 - and conclude that all classifiers provide more than 99% classification accuracy with less than 1% false positives. However, the processing overhead of J48 and neural networks based classifiers is significantly smaller compared with evolutionary classifiers. © 2010 Springer-Verlag.