Vector Processing RenaissanceLondon 15 Dec 99 From the DisCo benchmark results, the new EV7-based machines from Compaq/Digital appear to have the edge over leading CPUs from the other workstation vendors. The promise of very cheap high performance PC-based hardware was again dangled at the workshop delegates, with a 32 Node Beowulf system being offered as competitive to the current Cray T3E at the University of Edinburgh. This year however, it was admitted that production quality Beowulf systems are difficult to put together. The annual Machine Evaluation Workshop at Daresbury aims to encourage close contact between the research communities supported by the EPSRC grant lines (Mathematics, Physics, Chemistry, Materials and Engineering), and the major vendors of PCs, workstations, mid-range systems, software and peripherals. The Workshop comprises three main activities, (i) a two day programme of short talks, (ii) a two day technical exhibition and (iii) an opportunity for delegates to access loaned systems and benchmark their own codes via the internet and on site. The lecture programme provides a forum for delegates to receive an update from all the major vendors. The technical exhibition runs in parallel and includes demonstrations of a wide range of technology of importance to scientists using distributed computing. The benchmarking exercise is pivotal, because although many standard performance measures are available, such as SPECfp95, MFLOPS, Whetstones and timings on standard codes, and whilst they provide very useful information, these can never be as precise as results from user codes used in production. Focusing on supercomputing, Mike O'Neill of Fujitsu described the VP5000 claiming a fivefold increase in performance from the VP700 and Alan Baldock of Hitachi described the SR8000 Super Technical Server. Christian Lantwin of NEC described how the new parallel vector processor technologies incorporated in systems, such as the successful NEC SX-5, have stimulated a vector processing renaissance in Europe. He went on to say that HPC users are better served by vector parallel systems because they deliver high efficiency and are easy to use. The recent SX-5 achievement of 126 Gflop/s from 16 CPUs attest to this. He then went on to quote recent ECMWF results which show that by using OpenMP without MPI weather model applications scale very well. In addition he informed the audience that the NEC Japan Earth Simulator project aims to deliver a 40 Tflop/s system using parallel vector technologies. This illustrates that the US scalable ASCI model for Tflop/s systems is not the only option available. In another presentation, Robert Henry described the multi-threaded architecture of the Tera MTA machine which is struggling to preserve native vector supercomputing in the USA. A closed delegate's session provided an opportunity for speakers and workshop delegates to present experiences and the results of the benchmarking trials. Dr. Martyn Guest (Daresbury) presented his benchmark results which will be included, in full, in the forthcoming proceedings. Note that these benchmarks, are taken mostly from computational chemistry and include the most important intensive operations found in this kind of scientific code, i.e. vector and scalar matrix operations both with and without optimised BLAS where available. Raw matrix operation benchmarks are complimented by a set of codes taken from the numerically demanding kernels of larger applications including self-consistent field calculations using Gaussian basis functions, molecular dynamics, quantum Monte-Carlo, and an iterative Jacobi linear equation solver. The presentation also included results on full production runs of both the GAMESS-UK {ab initio} electronic structure code and DL_POLY molecular dynamics simulation code. Across the board, the new EV7-based machines from Compaq/Digital, such as, the Compaq ES40/667 outperformed the leading CPUs from the other workstation vendors, but HP, SGI, SUN, IBM are not far behind. The position of these vendors varies for different benchmarks, and users need to ensure the machine they choose performs best for their application domain. Several vendor presentations were given, Ben Ralston talked about the latest IBM RS/6000 Models with names like nighthawk used in the ASCI-White system and winterhawk, promised for next year, which will upgrade the ASCI-White system to 10 Tflop/s peak. Andrew Grant from Silicon Graphics on their workstation and server range; Phillip Bull of Sun Microsystems explaining their thinking and technology advances in the HPC arena, based on their experience with the ASCI -pathforward project; Tim Bush of Hewlett Packard on the PA-8800 processor which promises 2Gflop/s in an HP workstation on your desk by year 2002; and Ken Atkinson of Compaq, talking about Compaq's future products from "The view from mount improbable". Yes you guess it, their products are named after mountains, Sierra, and Matahorn. Other talks focused on software and networking, from Jonathon Hill of Sychron on "Virtual Private Servers", and Joel Barnes of Visual Numerics, who described their IMSL Library software. During the two days, delegates attended vendor demonstrations at the Exhibition. The latter proved a very successful component of the Workshop as several newly announced systems were on show. The following hardware and software was demonstrated and most made available for benchmarking: Information including benchmark results are accessible from the main DisCo page at URL: www.cse.clrc.ac.uk/Activity/DisCo ), or, Martyn Guest, Robert Allan, Barry Searle and Kevin Maguire at Daresbury Laboratory.
[News on Advanced IT] [Calendar] [Analysis] [IT in Medicine] |
|