Run-time monitoring of concurrent programs on the Cedar multiprocessor

A prototype run-time performance monitoring environment for the Cedar multiprocessor has been developed. The authors describe how the Cedar performance monitoring infrastructure is designed to achieve run-time monitoring and distributed communication of parallel program execution data. The underlying tracing procedures used in Cedar to gather execution data are described, and the procedures by which traces are dynamically off-loaded from the Cedar system and sent to remote processes are explained. The application of these tools to capture and visualize matrix data from an application at run-time is discussed. Finally, results regarding the performance of the run-time monitoring system in in terms of execution data bandwidth and program perturbation are presented