Contents of
October Issue


Daresbury Laboratory evaluates workstations

London, 17-10 -1996 The annual Daresbury Machine Evaluation Workshop was held last September at Daresbury Laboratory, near Warrington, Cheshire in the UK. This workshop is the leading national technical event for distributed high performance scientific computing. It is now an important part of the Distributed Computing Support Programme (DisCo) of the EPSRC and is operated by the Central Laboratory at Daresbury. This year's important topic was the evaluation of workstations.

The principle objective is to encourage close contact between the research communities from the mathematics, physics, chemistry and materials grant lines of the UK research council (EPSRC), and the major vendors of workstations, software and peripherals.

The workshop comprises three main activities: a two day programme of short talks; a two day technical exhibition; and an opportunity for delegates to access loaned systems and benchmark their own codes.

A total of 67 delegates were at the workshop and 23 companies including major workstation manufacturers, Hewlett Packard, IBM, DEC, Sun and Silicon Graphics.

The workshop lecture session provided a useful forum for delegates to receive an update from vendors. The technical exhibition ran in parallel and included demonstrations of a wide range of technology relevant to distributed computing. Some systems were installed at Daresbury in advance of the workshop and DisCo staff set up a mechanism for delegates to transfer their own codes and run benchmarks. This was appreciated as standard performance measures such as SPEC, Mflop/s, Whetstones and timings on standard codes, whilst providing very useful information, can never be as indicative as the code to be used in production.

A closed delegate's session allowed participants to share experiences and results of benchmarking trials. Martyn Guest, Daresbury presented his benchmark results which are included, in full, in the workshop proceedings and therefore can be reported here.

In summary, these benchmarks, taken mainly from computational chemistry, include the most important intensive operations found in scientific codes, i.e. vector and scalar matrix operations, with and without optimised BLAS where available.

Raw matrix operation benchmarks are complemented 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 paper also includes results on full production runs of the GAMESS-UK ab initio electronic structure code.

Across the board, the new HP K460 with the 160 MHz PA-8000 processor did not live up to expectations. Only on some of the matrix operation benchmarks did the machine show the promise of performance expected from the quoted SPEC figures. Indeed, some delegates had already seen these disparities under their own benchmarks. Newer versions of the compiler and optimised libraries could improve this position; both HP and ABC Workstations are investigating further.

The SGI R10000 at 195 MHz showed good performance on all benchmarks, often closely followed by the 333 MHz DEC 600/5/333 EV5-based system. The Ultra-2 systems were lacking a fully functional BLAS library but demonstrated performance broadly in line with SPEC figures. IBM's Power2 based systems are now mostly surpassed by all other major processors and also by the Pentium Pro in several benchmarks.

PC hardware for HPCN use

For assessing PC hardware, vendors Viglen and NeTpower (via ABC Workstations) were invited to demonstrate their products.

The Viglen machine, a top end Pentium Pro system, was installed with Windows NT and the Linux Operating System, a free version of Unix. The NeTpower system ran on Windows NT, but also included a high performance graphics subsystem and the Intrigue system which provides multi-user login.

The Pentium Pro systems showed considerable potential, although the programming environment in both cases tested presented problems under benchmark. The NT system appeared not to have a full test and debug suite and no timing routine. Multi-user access was not available by default under NT.

In the Linux case, neither the f2c Fortran system nor the GNU software g77 were mature enough for the entire benchmark suite. In neither case were optimised BLAS or other numerical libraries available. Nevertheless, in spite of what are presumably short term deficiencies, the promise of very cheap high performance hardware was noted. This could pose a significant threat to the RISC based workstations.

The theme of Pentium based computing was highlighted by Martin Houston, who gave an informed talk on Linux, entitled "Free Software on Cheap Hardware". He explained the high potential of this system which enables researchers to exploit cheap hardware.

The above is an edited version of a report produce by staff from Daresbury Laboratory. For further details on this and how to get the proceedings of the workshop contact: Royd Whittington at Daresbury Laboratory

Chris Lazou

Top of Article

© The HOISe-NM Consortium 1996