Performance Analysis and Prediction for Large-Scale Scientific Applications

By Adolfy Hoisie and Harvey Wasserman (Los Alamos National Laboratory, USA)
Adolfy Hoisie's Homepage
Harvey Wasserman's Homepage
Tutorial's Homepage

Main Fields:
        Large Scale Simulations in all areas of Engineering and Science
        Parallel and Distributed Computing
        Performance Analysis

Abstract:

        We present a methodical but simplified approach to performance analysis of large-scale parallel scientific applications. The core of the tutorial centers on methods for analytical modeling of performance and scalability, showing how to model applications, and why it is crucial to do so, in order to reveal scalability bottlenecks in current systems and in order to predict performance on future systems. Three case studies will be presented. One involves an accurate prediction of how processor speed and network performance will affect overall runtimes in future, hypothetical 100-TeraOp peak systems.
        A second shows how modeling can be used for point-design studies of SMP architectures, and a third shows how models can be used to accurately predict improvement in performance as a result of algorithmic changes. The unifying thread in all three studies is the use of modeling to reveal true understanding of architecture/application matching, in what we call "experimenting ahead:" screening algorithms and machines before installation.
        We will not emphasize any particular machine in the tutorial, nor performance rankings; rather, we will generally address performance of RISC processors and of widely utilized parallel systems such as the SGI Origin 2000, IBM SP2/3, Compaq HPC systems, clusters, and Cray T3E.
        The target audience is a mixture of computational scientists, computer scientists and code developers interested in performance analysis of "real-life" applications. The tutorial will also be useful to those trying to define needs for future-generation, high-end computing systems, from either the buyers or the designers point of view.
        This tutorial was presented in a half-day version at SC98 and at HPDC8 (1999), and in a full-day version at SC99 and SC2000. Each of the authors has presented portions of the tutorial separately (HJW at NPACI, 1998 & 1999, and AH at International Conference on Computational Physics, 1997, and at the Compaq User Group in November, 2001).
        Attendance has been high each year and evaluations have been overwhelmingly positive.

Keywords:
        Performance, scalability, modeling, performance prediction