There are 73 supercomputers in Japan listed in the TOP500. This represents a 14.6% share of the 500 entries - a decrease from 82 systems one year ago. The accumulated Rmax performance of these 73 systems reaches 1.234 TFlops/s which represents 25.6% of the accumulated Rmax performance of the TOP500. These figures show that Japanese supercomputer installations have in average a significantly higher performance than sites in other countries. This trend continues already for several years. In figure xxx the distribution of the number of systems and the accumulated Rmax performance are listed for different vendors.
The market leader in Japan is still Fujitsu with 34.25% of the number of systems and even 57.38% of the accumulated Rmax performance of the TOP500 sites in Japan. This outstanding relation is caused by the success of Fujitsu's VPP500 series. The 15 VPP500 systems in Japan (including the Numerical Wind Tunnel system) account already for 667.762 GFlops/s Rmax performance.
When we now look at the distribution into MPP, PVP and SMP systems, we face the difficulty that to count the VPP500 as an MPP system may lead to a misinterpretation of the Japanese market. The SGI system (installed at SGI Japan) is the only SMP system within the Japanese TOP500 list. But there are many SMP systems installed in Japan which just have too few processors to make it into the TOP500 list. The traditional PVP systems from CRI, Fujitsu, Hitachi and NEC account for 37 systems and 416.036 GFlops/s. The MPP systems including the VPP500 account for 35 systems and 813.814 GFlops/s. Does that mean that MPP systems have achieved a breakthrough in Japan? If we only look at the MPP systems which are based on standard RISC processors we count only 20 systems and 146.025 GFlops/s, nevertheless this is a significant progress compared with 1994.
The usage of the VPP500 systems is still very similar to the usage of the PVP systems. The computing centres are using the processors partly in throughput mode for serial vector jobs. This enables them to use these resources immediately very efficiently. At the same time, users are incrementally parallelizing their applications. Some users do it in the message-passing style with PARMACS or PVM. The majority of parallelization work, however, is done in Fujitsu's VPP-FORTRAN style. This compiler directive based method is similar to the HPF style and allows users to stay very close to their original vector codes. Consequently, the usage of the VPP500 systems can be claimed to be closer to PVP than to MPP although the VPP500 is clearly a scalable system with distributed memory - a main characteristic of MPP. It will be interesting to watch this development in 1996 when more scalable vector systems, also from NEC, will enter the list.