Workload Characterization Using the TAU Performance System

摘要

Workload characterization is an important technique that helps us understand the performance of parallel applications and the demands they place on the system. It can be used to describe performance effects due to application parameters, compiler options, and platform configurations. In this paper, workload characterization features in the TAU parallel performance system are demonstrated for elucidating the performance of the MPI library based on the sizes of messages. Such characterization partitions the time spent in the MPI routines used by an application based on the type of MPI operation and the message size involved. It requires a two-level mapping of performance data, a unique feature implemented in TAU. Results from the NPB LU benchmark are presented. We also discuss the use of mapping for memory consumption characterization.