Granger Causality-Aware Prediction and Diagnosis of Software Degradation

Software that continuously runs over a long period has been frequently reported encountering "gradual degradation" issues. As time progresses, software tends to exhibit degraded performance, deflated capacity, exhausted physical resource or deteriorated QoS (Quality of Service). Different from transient software anomalies, this issue is a chronic degrading process and usually persists until the software is eventually unavailable. We name it "Software Degradation" or "Degradation" for short. In this paper, we propose a framework GVAR, utilizing Granger Causalities to predict and diagnose software degradation. GVAR is evaluated via an 8-day experiment on a VoD (Video on Demand) platform Helix-Serv. The experimental results show that GVAR can predict the TTF (Time to Failure) of degraded software in an accuracy of 80.1%, remarkably outweighing the widely used ARMA and Sen\´s Slope Estimator approaches. Moreover, GVAR can guide diagnosing the potential root cause of degradation issues.