A Software Aging Prediction Based On Ant Colony Neural Networks

Software aging can produce system crashes, undesirable hangs or a lot of storage space debris and other phenomena in long-term uninterrupted operation process, which lead to software performance degradation. Also, it is called software aging. Software aging phenomenon exists not only in normally use Web server or universal server, including high reliability, availability of the application system. Software aging caused extensive damage for security key fields including human life. Timely forecasting software can take corresponding recovery strategy software. According to the degree of the exhausted temporal of system resources, we can avoid the loss caused by software aging. In this paper, we proposed the algorithm ACONN based on BP neural network (BPNN) in order to forecast software aging trend, which makes it easy to determine when to perform software rejuvenation.Firstly, we discuss the research significance and research status about software aging, built the software aging experimental platform using Apache web server. Secondly, we introduces linear forecast and nonlinear prediction methods, including Sen slope, ARMA model, BP neural network etc. In this paper, various prediction algorithms make a comprehensive comparison. In predicting aging trends, applying prediction methods forecast system performance data collected from experiment platform firstly, experiments show that the accuracy of BP neural network fitting power and accuracy of prediction. However, it has some limit due to the long delay in calculation of initial weight. Therefore, ant colony neural network (ACONN) is proposed and established to predict the software aging trend. Finally, we compare the results of BP and ACONN, which show that ACONN has a better predicting power than BP.Experimental results show that because of ant colony algorithm global optimization characteristics, ACONN can quickly find initial choice of BP neural network weights, which reduced BPNN training frequency and training time. Comparing results show that ACONN is better than BPNN in prediction accuracy and fitting precision.