Bridging the Gap between Application Performance Analysis and System Monitoring

Performance analysis has a long history in the high-performance computing community. On the one hand, the traditional application analysis focuses on scalable yet detailed instrumentation of parallel execution. On the other hand, per-node or cluster-wide monitoring solutions are used in data center operation. However, performance anomalies resulting from the interaction between applications, background processes and the operating system, are difficult to analyze with tools that reveal only part of the issue. In this paper, we present a novel approach that covers all aspects of individual nodes. We extend a monitoring tool to combine call stack sampling, process monitoring, and syscall recording into a symbiotic view of the application execution, background activity, and the operating system.