Clusters of C90 systems that are connected via Hippi channels seem to be an interesting architecture for the parallel IFS. For clusters showing more than two systems, a network of Hippi channels is required. In the ideal case, each pair of systems should be connected by a separate channel. Nevertheless, an appropriate partitioning, mapping, and some other actions to minimise the data transfer via Hippi are required. The parallel program should be carefully vectorised. The best results are obtained by a partitioning that prefers to subdivide the latitudes. An efficiency in the range of 80 to 90% can be expected for clusters showing up to 64 processors and for large test cases.
Just the effort required for a good parallelisation shows that essentially slower interconnection systems like Ethernet do not promise efficient runs of the IFS on clustered supercomputers. Even if the capacity of the Ethernet is fully available for the cluster computing, we should expect an efficiency less than 30%, i.e. to run the IFS on ethernetted clusters is useless.
Acknowledgment. The RAPS version 2.0 of the IFS we used for our investigation was developed at the ECMWF. Essential parts of the work reported here and the development of the basic parallel algorithms have been carried out in close co-operation with the ECMWF (Prof. Geerd Hoffmann). The authors would like to thank Saulo de Barros, David Dent, and Lars Isaksen for their advice. The performance expected for C90 clusters has been investigated in co-operation with Cray Research. We would like to thank Michael O'Neill and Markus Kienemund who provided us with measurements on the IFS and on communication for the C90.