On the scalability of tracing mechanisms1

Abstract. Performance analysis tools are an important component of the parallel program development and tuning cycle. To obtain the raw performance data, an instrumented application is run with probes that take measures of specific events or performance indicators. Tracing parallel programs can easily lead to huge trace files of hundreds of Megabytes. Several problems arise in this context: The storage requirement of the high number of traces from executions under slightly changed conditions; visualization packages have difficulties in showing large traces efficiently leading to slow response time; large trace files often contain huge amounts of redundant information. In this paper we propose and evaluate a dynamic scalable tracing mechanism for OpenMP based parallel applications. Our results show: With scaled tracing the size of the trace files becomes significantly reduced. The scaled traces contain only the non-iterative data. The scaled trace reveals important performance information faster to the performance analyst and identifies the application structure.