Software state monitoring model studies based on multivariate HPM

Hardware Performance Monitor counters (HPM) are an emerging analysis technology in the area of software performance analysis. This paper proposes a method of software state monitoring based on HPM from the perspective of software fault diagnosis. Compared with traditional methods, the method does not depend on test case and expected result, and it can detect abnormal behavior in real-time based on software performance data. By the use of Performance API (PAPI), the method can gather CPU performance data. These data are recorded in HPM and can reflect software state at the running time of software. With Hidden Markov Model (HMM), the method can learn prior probability of software state and conditional probability of performance data readings in each interval. Finally, based on the above parameters, the method classifies the follow-up multivariate observations by Naïve Bayesian classifier (NBC) so as to monitor software state in real-time. The experiment shows that based on predefined monitoring event set, our method can effectively identify abnormal behavior which may occur in the running time of software.