Performance Analysis of Shared-Memory Parallel Applications Using Performance Properties

摘要

Tuning parallel code can be a time-consuming and difficult task. We present our approach to automate the performance analysis of OpenMP applications that is based on the notion of performance properties. Properties are formally specified in the APART specification language (ASL) with respect to a specific data model. We describe a data model for summary (profiling) data of OpenMP applications and present performance properties based on this data model. We evaluate the usability of the properties on several example codes using our OpenMP profiler ompP to acquire the profiling data.