Aberrant software-aging server detection and analysis using sliding window over LOF

As software grows rapidly in size and complexity, software aging has been observed more frequently in long-running computer system failures. Ever since the notion of software aging was introduced, lots of researches have devoted in studying software aging empirically by learning aging trend under controlled experiments. However, we notice that how software aging evolves within online system is seldom covered in the literature. To fill the blank and improve usability of software-aging works, we in this paper propose an approach to detect aberrant software aging in online system, i.e., servers grow old more quickly or slowly than the rest servers. We believe studying aberrant aging can contribute to better understanding of software aging in real world. The mechanism utilizes random sliding windows over LOF (Local Outlier Factor) algorithm to analyze the CPU and workload indicators, which displays good accuracy and recall rate in a real world application. Moreover, we analyze how the software trend to age aberrantly and verify if aberrant aging also obeys the rules in controlled experiment. We conclude three aging patterns from these aberrant aging servers, and find out CPU-oriented indicators shows higher precision and recall rate than memory-oriented in analyzing software aging.