do you consider 'distributed' to make use of MPI? If so, you can check out sB_BLAS (www.cs.utexas.edu/users/rvdg/sw/sB_BLAS) or look into what the folks at Argonne National Labs (US) have done. They have significant number of massively parallel machines, and have extended some of the basic linear algebra libraries. www.alcf.anl.gov

Thanks for your answer.

By "distributed" I mean "not having physical shared memory", which usually results in using MPI, indeed.

For most of the time, doing BLAS in parallel over MPI is a loss. This is mostly due to the size of the matrices and the amount of data exchange that needs to happen. In this particular case, it would be helpful to know which BLAS functions you're particularly in need of and what are the dimensions of the matrices. And even better question is what is the higher order algorithm you're trying to solve? Are you doing systems of linear questions? Some other factorization? You'd be much better suited by using libraries that implement that routine such that they can schedule low-level linear algebra operations appropriately and take advantage of as much locality as possible.

The NETLIB PBLAS was invented a while back, and there are now much better ways to do it in parallel - task graph driven approaches as used in libflame, MAGMA, PLASMA and Intel's Cluster Math library. These approaches reduce or laminate the latency by using a dataflow style approach, Rather than a static schedule.

Phil Mucci

Minimal Metrics