On the effectiveness of Mann-Kendall test for detection of software aging

Software aging (i.e. progressive performance degradation of long-running software systems) is difficult to detect due to the long latency until it manifests during program execution. Fast and accurate detection of aging is important for eliminating the underlying defects already during software development and testing. Also in a deployment scenario, aging detection is needed to plan mitigation methods like software rejuvenation. The goal of this paper is to evaluate whether the Mann-Kendall test is an effective approach for detecting software aging from traces of computer system metrics. This technique tests for existence of monotonic trends in time series, and studies of software aging often consider existence of trends in certain metrics as indication of software aging. Through an experimental study we show that the Mann-Kendall test is highly vulnerable to creating false positives in context of aging detection. By increasing the amount of data considered in the test, the false positive rate can be reduced; however, time to detect aging increases considerably. Our findings indicate that aging detection using the Mann-Kendall test alone is in general unreliable, or may require long measurement times.