by Roger Evans
The traditional vector supercomputing service at RAL is now complemented by a machine of approximately equal performance based on the most modern microprocessors. The machine is targeted at the scientific community served by the Engineering and Physical Sciences Research Council (EPSRC) and offers them supercomputer performance with the familiarity and ease of use of a desktop workstation.
The new service comprises a Digital 8400 supercomputer with six of the latest EV5 (21164) Alpha processors, 2 GByte memory and 150 GByte of disk space and is known as the EPSRC Super-scalar Computing Service.
The EV5 Alpha processor is substantially faster than previous versions of the Alpha and has a theoretical peak performance of 600 Mflop/s corresponding to two floating point operations each cycle of its 300MHz clock. The processor is supported by two levels of on-chip cache and a 4 MByte tertiary cache. Improved memory pipelining means that up to two memory requests can be outstanding before the processor stalls and the system bus bandwidth of 1800 MByte/s allows for very efficient utilisation of multiple cpus.
The machine is being installed at RAL as the first of the Research Council specific services (the Cray Y-MP8 is available to all six UK Research Councils) and is targeted at work which is too large for local university computer systems but not quite requiring the largest national facilites. The increase in computer performance since the Y-MP was installed means that the six processor Alpha 8400 system with its notional peak performance of 3.6 Gflops can out perform the Y-MP on several important applications.
Having the alternative computer architecture available to a subset of the Cray Y-MP user community is an important advance in being able to match the algorithm to the architecture and make substantial gains in efficiency. One very large grant which had highly recursive algorithms for the solution of stiff equations runs six times faster on the Alpha EV5. Being able to move this work to the 8400 gives the grant holder more effective resource for his work and frees up the Cray Y-MP time for users who are better able to make use of the vector processing and memory bandwidth of the Y-MP.
Batch job management on the 8400 will be through Platform Computing's Load Sharing Facility (LSF) and as the range of computing hardware available to our users increases we look forward to using some of the more advanced features of LSF for distributing a heterogeneous workload to the most appropriate platform for its solution.