Analyzing PICL trace data with MEDEA

Abstract

A detailed performance analysis of the behavior of a computer system under its real workload can be achieved by means of event-driven monitors, i.e., tools that capture the events generated by a program and store them into trace files. Execution traces and performance statistics can be collected for parallel applications on a variety of multiprocessor platforms by using the Portable Instrumented Communication Library (PICL). Starting from these measurement data, the construction of accurate workload models requires the application of different types of statistical and numerical techniques interacting together to fully characterize the behavior of the applications submitted to a system. The static and dynamic performance characteristics of performance data can be analyzed easily and effectively with the facilities provided within the MEasurements Description Evaluation and Analysis tool (MEDEA). This paper outlines a case study that uses PICL and MEDEA to characterize the performance of a parallel benchmark code executed on different hardware platforms and using different parallel algorithms and communication protocols.

This research was supported by the Applied Mathematical Sciences Research Program, Office of Energy Research, US Department of Energy, under contract DE-AC05-84OR21400 with Martin Marietta Energy System Inc., by the Italian Research Council (C.N.R.) under Grant 92.01571.PF69, and by the Italian MURST under the 40% and 60% Projects.