Squeezing the most performance out of your computer

Vladimir Getov (Westminster), Tony Hey (Southampton) and Roger Hockney (Independent Consultant) were amongst the first to programme parallel machines, way back in the eigthies. They shared their vaste knowledge of the topic with an audiance of (future) experts in parallel processing at EuroPar'98.

This tutorial gave an excellent introduction to performance evaluation methodology as well as an overview of performance analysis and optimisation techniques. Various issues related with parameterisation of performance measurements, characterisation of applications and architectures, and interpretation of results were discussed. The methods of analytical performance modelling and estimation were introduced in detail, and case studies were used to demonstrate application usefulness.

These issues were illustrated with latest benchmark results including PARKBENCH and NPB codes for the state-of-the-art parallel computers - Origin-2000, SP2, T3E, Tera, and others. The Tera results are only a few weeks old (See also our coverage of the Terabenchmarks in PrimeurLive! from Supercomputer'98 in Mannheim.

In one of the case studies, the characteristics of the Cray T3D and the SP2 were compared using segments from the NAS parallel benchmarks and employing the White-box technique.

The white-box technique essentially times computation code sections and specific chosen classes of other machine activities such as data movement, communication latency and other overheads.

The timings obtain data which are applicable across computers and therefore allows us to assess the balance of each machine and compare it with other systems.

A balanced system performs well across the whole spectrum of applications even when some of its components are not as fast as those of its competitors. This is why this method gives a good indication of what performance can be achieved.

In this case study the results show that the SP2 is a well balanced system and is getting good performance results with most applications despite its slow computation chip. The Cray T3D is less balanced with a communication bottleneck inherent in the earlier version of the Alpha chip. It compensates for this with a superior computation performance, however. The difference shows in the data movement internal to the processors.

The parallel programming experts
Vladimir Getov is a Senior Lecturer in the School of Computer Science at the University of Westminster and Head of the Computer Systems Performance Engineering Group there. Dr Getov has considerable experience in the performance analysis and evaluation of parallel and distributed systems. He has actively participated in the work of the PARKBENCH Committee since its creation in 1992 and is now involved in the activities of the Java Grande Forum.

Tony Hey is Professor of Computation and Head of the Department of Electronics and Computer Science at the University of Southampton. He is also Chairman of the Parallel Applications Centre (PAC), which works closely with industry in a technology transfer role. Professor Hey has been a leading researcher in parallel computing for many years. He is on the editorial board of a number of scientific journals and has been on the organising committee of many international conferences. He is also the co-founder, with Jack Dongarra, and the current chairman of the PARKBENCH Committee initiated at Supercomputing'92 in Minneapolis.

Roger Hockney is Emeritus Professor of Computer Science at Reading University, and visiting Professor at Westminster and Southampton Universities. Professor Hockney has a track record of pioneering work in performance evaluation. He is now working as an independent consultant in parallel computing, concentrating on the benchmarking of parallel computers and the development of the PARKBENCH Interactive Curve-Fitting Tool on the Web. He was the first chairman of the PARKBENCH committee, and one of the co-founders of Euroben.