From: Kohei Yoshida <[email protected]> Subject: Re: [sc-dev] Microsoft testing Excel for supercomputers Date: Thu, 19 Nov 2009 00:07:57 -0500
> So, I know little about parallel computing or clustered computing, and I > do know that this is your area of expertise. With that in mind, I'd > like to ask you this question. How can an application make use of > clustered computers to perform parallel computations? Is there any > library that takes care of the complexity of sending/receiving > computation requests to multiple machines? > > It's sort of my pet peeves to make Calc more appealing to the scientific > community, but unfortunately I don't have enough expertise to follow > through with that. So, I would much like someone to educate me in this > area. My idea is just a not a special and very simple. If we do vector-vector, vector-matrix, matrix-matrix multiplication etc, *just call BLAS and LAPACK* Some of (not all) BLAS and LAPACK routines are highly optimized, using threads, we can call them as a blackbox. GotoBLAS reaches almost theoretical peak. If you need parallel version, then call PBLAS, BLACS and ScaLAPACK. We may also need MPI for that. summary; what we need are very simple: --- isolate vector-vector, vector-matrix, matrix-matrix manipulations; then call BLAS. isolate LU/Cholesky factorization, Eigenvalue problems, Singular value decomposition part; then call LAPACK. --- If we need parallele computing just call ScaLAPACK. --- Pros 1. BLAS, LAPACK, ScaLAPACK, and MPI are freely available, just define API. 2. BLAS, LAPACK... are very reliable and tested with many users. 3. ATLAS is a free software, and provides optimized BLAS and a part of LAPACK routines for many architectures, and also utilizes multi-cores via pthreads. Cons 1. BLAS, LAPACK, ScaLAPACK, and MPI are freely available but they are not tuned. 2. Usually machine vendors provide optimized BLAS and LAPACK codes, as proprietary libraries, and they are very fast. 3. performance tuning is a big problem. 4. We may need FORTRAN compiler as well :-) --- Best, -- Nakata Maho http://accc.riken.jp/maho/ , http://ja.openoffice.org/ Nakata Maho's PGP public keys: http://accc.riken.jp/maho/maho.pgp.txt --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
