Can I run LAPACK on shared memory computers

[wjiang]

Dear all,

As stated on the homepage of lapack,

The original goal of the LAPACK project was to make the widely used EISPACK and LINPACK libraries run efficiently on shared-memory vector and parallel processors

.

But I cannot figure out how to do it. For example, when I use Subroution DGESV to solve linear algebra equations I do not know how to make it run on multi processors on shared-memory computers. It seems there is no choice for the number of processors.

Could anyone give me some help? Thanks,

Wen

[buttari]

Dear all,

As stated on the homepage of lapack,

The original goal of the LAPACK project was to make the widely used EISPACK and LINPACK libraries run efficiently on shared-memory vector and parallel processors

.

But I cannot figure out how to do it. For example, when I use Subroution DGESV to solve linear algebra equations I do not know how to make it run on multi processors on shared-memory computers. It seems there is no choice for the number of processors.

Could anyone give me some help? Thanks,

Wen

Hello Wen,
the bulk of work in LAPACK is done inside BLAS routines. Thus if you link LAPACK to a multithreaded version of BLAS (for example, ATLAS, MKL, ACML or GotoBLAS) you obtain a rather efficient parallelization. These parallel BLAS usually let you control the number of processes through an environment variable. However some of these parallel BLAS libraries include parallel implementations of the LAPACK libraries; in these implementations, parallelism is extracted at an higher level than the BLAS routines which results in a much higher performance. ATLAS, MKL and ACML have all (or nearly all)of the LAPACK routines, GotoBLAS only has LU and Cholesky factorization (at least the last time I used it); IBM machines have ESSL.
Best regards

Alfredo

[wjiang]

Thanks a lot, Alfredo.