Measurement-based Prediction On Software Aging

Software aging can result system crashes, undesirable hangs or a lot of storage spacedebris and other phenomena in long-term uninterrupted operation process, which lead tosoftware performance degradation. Also, it is called software aging. Software agingphenomenon exists not only in current Web server or other servers, but also in the applicationsystems that require high reliability and availability. Software aging may cause hugeeconomic losses and even personnel casualties in commercial area or military area. Therefore,by predicting software aging trend in time, and taking appropriate recovery strategy, we canavoid the loss caused by software aging.In this paper, the ARMA model is used to predict software aging. Firstly, we use ARMAto predict the trends of the used swap space parameter. And establish multi-dimensional AR(mAR) model through the analysis of the relationship between system parameters (freephysical memory, used swap space, and response time), and predict the used swap space withthe mAR model. Secondly, response time is an important factor affecting system performancefrom a user perspective, and the system resource (load average, free physical memory, CPUutilization and free swap space) affect the system performance too. Therefore, a SPE(software performance evaluation) model is proposed based on BP neural network to evaluatethe trends for software performance parameters. And on the basis of the SPE model, we usethe robust locally weighted algorithm to identify the inflexion point of software aging. Finally,with weighted Markov chain and fuzzy comprehensive evaluation theory, we predict the trendof resource parameters in short-term which can reflect software aging.Experimental results show that, ARMA model and multi-dimensional AR model canpredict software aging trend effectively. And we can evaluate the trends of softwareperformance parameters reflecting software aging effectively with the SPE model. We canpredict data accurately in short-term with weighted Markov chain and fuzzy comprehensiveevaluation theory.