07341 Executive Summary – Code Instrumentation and Modeling for Parallel Performance Analysis

Abstract

Given the exponential increase in the complexity of modern parallel systems, parallel
applications often fail to exploit the full power of the underlying hardware. At scale, it is not
uncommon for applications to run at parallel efficiencies in the low single digits. Moreover,
their optimization is extremely difficult due to the inherent complexity of the systems and the applications themselves. Therefore, a variety of projects have been developing tools and
techniques for the measurement, analysis, and visualization of parallel program performance
in order to help guide users in the optimization process.


This meeting was the third in a series of seminars related to the topic "Performance Analysis
of Parallel and Distributed Programs", with previous meetings being the Dagstuhl Seminar
02341 on "Performance Analysis and Distributed Computing" held in August 2002 and
Seminar 05501 on "Automatic Performance Analysis" in December 2005. While these
seminars concentrated on the "analysis" part of performance analysis, at the most recent
seminar the focus was on the building blocks of program instrumentation and modeling that
are prerequisites for the analysis phase. As a result, the presentations of the participants
concentrated on several fundamental issues related to instrumentation for generating high-quality performance data, methodologies for performance modeling leading to accurate
predictions for the performance, and on the ways in which these techniques are combined for the performance analysis of applications and systems.