WEB SOFTWARE TRAFFIC CHARACTERISTICS AND FAILURE PREDICTION MODEL SELECTION

摘要

In this paper, we analyze web traffic characters and the relationship with software failures. Results indicate hourly web access traffic is the lowest from 3:00 to 4:00 am, while the traffic load gradually reaches peak between 14:00 and 16:00, before declining. In daily base, web traffic fluctuates in the 25 observed days. The hourly access hits appear in similar patterns to the software failures. The web site reliability is 0.9878. The mean time between failures is 82.03 hits. Five popular software reliability models are calibrated with real data. The validations show that Goel-Okumoto and Gompertz models have good predictions to describe web software failures. The coefficient variation analyses for the two selected models show that the models are stable when the values are less than 1.0. Further investigations indicate that both models have some deviations in prediction accuracy starting from the 20th day. Using similar approach to change-point solutions, we recalibrate the models with different parameter values after 20th day. The results appear that two sets of parameter values greatly improve model prediction accuracy.