umfpack acml 4-4-0 performance problems

Discussion created by cn_an on May 17, 2011
umfpack acml performance problems


currently i am evaluating the sparse linear solver

umfpack from T.Davis's Suite Sparse. Opposed to my

expectation i get a big performance drop, when i

statically linked against libacml_mp.a (gfortran 4.4. 64Bit), 

instead of using LAPACK 'standard' BLAS implementation.

The used sparse matrices vary in size and are taken from

the UF Sparse Matrix Collection ( Someone a idea, what i might do wrong?

Used System:

Hardware: AMD Phenom(tm) X6 1045T

OS: Fedora Core 13 64 Bit

Development System:

Fedora Core 14 64 Bit running in VM VirtualBox with

gcc 4.5.1 20100924