Parallel Processing and the Sustained Production Performance of the Cray Y-Mp: Benchmarks Using Optimized Microtasked Lattice Su(3) Code

The standardized-maximalist approach to supercom puter benchmarking consists in optimizing a standard production code on the supercomputer, then measur ing a wall-clock-based figure-of-merit that is relevant to users of the code in question. Since 1982, one highly efficient algorithm for simulating SU(3) lattice gauge theory has been used in such benchmarks, tracing the progress of supercomputers from the CDC 7600 to the CRAY X-MP and the NEC SX-2. Here we report on the performance of the CRAY Y-MP/8128 at the Cray Re search Corporate Computing Center under this bench marking procedure. The code was optimized and mi crotasked, taking advantage of the hardware and soft ware features of the Y-MP. The link-update time was measured with the code running on 1, 2, 4, and 8 CPUs. With 8 CPUs, it was 3.1 μsec. This corresponds to a sustained performance of 1.349 GFLOPS com puted on the basis of theoretical operation counts. (Hardware performance monitoring yields an estimate of 1.54 GFLOPS.) It represents an improvement of a factor 3.55 over a maximal CRAY X-MP configuration (four X-MP processors).