Evaluating Supercomputers

Evaluating Supercomputers

After you've bought a few of these new generation, multi-million dollar supercomputers, you begin to notice that some are better than others, and you can't always tell by the color. Here is help in the form of 18 papers by contributors from the US and Europe presenting a wide variety of views on evaluative techniques. Among the topics they address are some of the innovative architectures and how to exploit their full potential, established benchmarks and application-specific ones developed by users, and the modifications being made in the mathematical software libraries to account for the new vector and parallel processing. No index. Distributed by VNR. Annotation copyrighted by Book News, Inc., Portland, OR

Performance Evaluation of Supercomputers

Although supercomputer systems are faster, and have larger memory hierarchies than other computer systems, such characteristics merely imply the existence of great potential power. How much of that power can be harnessed productively is the central theme of performance evaluation. A set of methods for evaluating the performance of applications on supercomputers has not yet been rigorously defined. This volume is a compilation of research approaches and techniques that are a promising means to that end. The contributions are grouped into three sections. Contributors to Performance looks at applications, algorithms, compilers, operating systems, and memory issues. Measurements and Metrics addresses some of the current techniques and methods of performance evaluation. Among the topics are: the performance monitoring capabilities of the CEDAR system, the methodology behind the Livermore loops, the empirical analysis of system performance, and a software simulator developed in connection with the RP3 project. Methods, Models, and Directions looks at ways of establishing a general and theoretical framework for supercomputer performance evaluation.

An Agenda for Improved Evaluation of Supercomputer Performance