Implementation and Performance of Scalable Scienti c

We report progress to date on our project to implement high-impact scientiic subroutines on Fujitsu's parallel-vector VPP500. Areas covered in the project are generally between the level of basic building blocks and complete applications, including such things as random number generators, fast Fourier transforms, various linear equation solvers, and eigenvalue solvers. Highlights so far include a suite of fast Fourier transform methods with extensive functionality and performance of approximately one third of peak; a parallel random number generator guaranteed to not repeat sequences on diierent processors, yet reproducible over separate runs, that produces randoms in 2.2 machine cycles; and a Gaussian elimination code that has achieved over a GGop per processor for 32 processors of the VPP500, and 124.5 GGops total on the Fujitsu-built Numerical Wind Tunnel, a machine very similar architecturally to the VPP500.