Are ScaLAPACK anc BLAS optimized for OMP?

[spiga.filippo]

Hi,
today I want to test cluster's performances using HPCC benchmark suite. With HPCC I can use ScaLAPACK and BLAS, linked to ACML libraries.

My question is: are ScaLAPACK and BLAS optimized for OpenMP?

Thanks a lot

[Danesh_D]

Spiga,

OpemMP is an API for shared memory programming while ScaLAPACK is designed to work on distributed systems. You can use OpenMP whenever you want in your code along with ScaLAPACK functions. If you take a look at ScaLAPACK functions you will see that they are based on message passing protocols which use BLACS routines for their communications while OpenMP is based on multi-thread paradigm. In fact if you assume that your code uses n nodes where each node has two processors, Scalapack functions can be called to handle your calculations over the whole grid and OpenMP can be used to handle parallel implementation of your code on each node by running parts of your code e.g. loops within several threads in parallel.

Regards,

Danesh

[PHinker]

The HPCC benchmark already handles the details of cluster distribution (normally handled by ScaLAPACK/BLACS). You can link the HPCC benchmark to a vendor supplied, optimized version of LAPACK/BLAS (e.g. ACML, MKL, GotoBLAS, Sunperf) to get higher performance for a particular processor. All the vendor supplied implementations can run parallel on-node using environment variables. OMP_NUM_THREADS is typically the environment variable used to contol the number of threads spun up on a particular node.