A Performance Monitoring Tool for Predicting Degradation in Distributed Systems

Continuous performance monitoring is critical for detecting software aging and enabling performance tuning. In this paper we design and develop a performance monitoring system called PerfMon. It makes use of the /proc virtual file system's kernel-level mechanisms and abstractions in Linux-based operating system, which provides the building blocks for implementation of efficient, scalable and multi-level performance monitoring. Using PerfMon, we show that (1) monitoring functionality can be customized according to clients' requirements, (2) by filtering of monitoring information, the trade-offs can be attained between the quality of the information monitored and the associated overheads, and (3) by performing monitoring at application-level, we can predict software aging by taking into account the multiple resources used by applications. Finally, we evaluate PerfMon by experiments.