No clouds, bright sunshine for upgraded weather forecasting Fujitsu VPP5000Mannheim 09 jun 2000 Since 1992, Engineer Matthias Nethe is working as a systems analyst at the computer division of the European Centre for Medium-range Weather Forecasts (ECMWF). Ing. Nethe was invited at Supercomputer 2000 to highlight the supercomputing resources at ECMWF in general and of the recently acquired Fujitsu vector parallel processor VPP5000 in particular. This machine is actively being used in the centre's Integrated Forecasting System (IFS). Ing. Nethe equally presented some valuable information on ECMWF's plans for upgrading its high-performance computing facilities in 2000, as well as on the performance figures for runs of the IFS T213L19 benchmark on a variety of computing platforms, including the VPP5000.
The ECMWF is an international organisation established by a Convention which was signed in October 1973. Currently, the institution is supported by 18 European Member States. Its prime objectives are the operational production of analyses and medium-range forecasts of weather and ocean waves, and scientific and technical research directed to the improvement of these forecasts. Before 1996, the Centre applied several Cray parallel vector processor (PVP)shared-memory computers for these purposes. In that year, these systems were replaced by a Fujitsu vector parallel processor (VPP) distributed-memory computer, the VPP700. In 1999, ECMWF installed a VPP5000 computer, which is Fujitsu's follow-on system to the VPP700.
The Fujitsu VPP5000 is primarily used to run the Integrated Forecasting System. IFS is a numeral weather prediction package which was jointly developed by METEO France and ECMWF. The programme includes the forecast model, involving spectral, semi-Lagrangian, and semi-implicit algorithms; 4D variational analysis using tangent-linear and adjoint versions of the forecast model; and the Ensemble Prediction System (EPS). The IFS model is utilised for Medium-range forecasts spanning 10 days, for 4D variational analysis, for the EPS application which is used by 51 members, and for seasonal forecasts, which are being coupled to the HOPE ocean model.
In the early nineties, the IFS T213L19 benchmark was ported to distributed memory machines, as Ing. Nethe explained. Initially, IFS was developed on Cray Parallel Vector Processors (PVPs), such as XMP, YMP, and C90, using Cray MACROTASKING. IFS is highly vectorised and very well parallelised. The package makes use of a message-passing programming paradigm. At the start, this was PARMACS but now the ECMWF workers use MPI with Open/MP design for SMPs. IFS has been run on Fujitsu VPPs, Cray T3Es, IBM SPs, SGI Origins, NEC SXs, and many other systems. The programme even has been demonstrated on a PC. It runs reasonably efficiently on a wide variety of PEs, ranging from a single VPP5000 PE up to over 1400 PEs on a Cray T3E.
Ing. Nethe also made the effort to compare the VPP architecture against the scalar systems. In VPPs, the vectorisation is very well understood. They achieve a high PE peak performance, sustaining a rate of about 30% to 50% of peak. Only a low degree of parallelism, involving some 16 to 24 PEs, is required against over 256 PEs in scalar systems. In contrats with their scalar colleagues, VPPs show a higher parallel efficiency with a small number of PEs. As a result, several time-critical jobs are able to run in a single PE.
On the other hand, scalar systems have a better scalar performance than VPPs. Here, however, many time-critical jobs will have to be parallelised.
In scalar machines, the programming for cache is rather difficult. These systems also have a lower PE peak performance, sustaining a rate of about 10% to 15% of peak. As for the upgrading of the VPP5000 system, the IMPEs will climb from 4 to 8, the SIOPEs from 0 to 4, the SPEs from 33 to 87, the HIPPI networks from 4 to 10 with no longer the assistance of FDDIs, whereas Gigabit Ethernet facilities will be upgraded from 6 to 16. This power enhancement will help ECMWF to cope with growing supercomputing demands.
Leslie Versweyveld |
 |