A statistical approach for software resource leak detection and prediction

Summary form only given. Resource leaks are a common type of software fault. Accruing with time, resource leaks can lead to performance degradation and/or service failures. However, there are few effective general methods and tools to detect and especially predict resource leaks. We propose a lightweight statistical approach to tackling this problem. Without complex resource management and modification to the original application code, the proposed approach simply monitors the target's resource usage periodically, and exploits some statistical analysis methods to extract the useful information behind the usage data. The decomposition method from the field of time series analysis is adopted to identify the different components (trend, seasonal, and random) of resource usage. The Mann-Kendall test method is then applied to the decomposed trend component to identify whether a significant consistent upward trend exists (and thus a leak). Furthermore, we establish a prediction procedure based on the decomposition. The basic idea is to estimate the three different components separately (using such statistical methods as curve fitting and confidence limit), and then add them together to predict the total usage. Several experimental studies that take memory as an example resource demonstrate that our proposed approach is effective to detect leaks and predict relevant leak index of interest (e.g., time to exhaustion, time to crossing some dangerous threshold), and has a very low runtime overhead.