Monitoring Parallel Programs for Performance Tuning in Cluster Environments

This chapter surveys the issues of collecting monitoring data for performance debugging of parallel programs executed in cluster environments. The main categories of monitoring tools are either clock driven or event driven, the latter including timing, counting or tracing. The paper focuses on software tracing, deemed the most general and portable event driven monitoring technique. The lack of global clock and the tracing intrusion are identified as the two major problems hindering the quality of the traced information collected by software tracers. A global time implementation by software and an approach for compensating the software tracing intrusion are then described.