Pre-compiled reference BLAS is outperforming OpenBLAS?!

[barnburnr]

Thought we might get a cheap performance hit by replacing our reference BLAS library, the MinGW compiled one offered on your site, with a MinGW compiled version of OpenBLAS. We are compiling with MinGW 4.7.1 and turned on threads for OpenBLAS. The computation ends up being almost an order of magnitude slower using OpenBLAS in comparison to the reference BLAS library! For kicks, we downloaded and compiled reference BLAS and obtained almost the the same computation time as with the pre-compiled one. Will give ATLAS BLAS a shot next but am wondering if anyone has any idea why OpenBLAS would be so miserably slow? This is all in Win 7 and linking occurs with VS2010.

[barnburnr]

An update: the OpenBLAS folks suggested turning *off* threading support and that seems to bring run-time down to near reference BLAS performance level. One might speculate that the overhead of doing threading in Windows is so high that it is not worth it for this problem, at least.

[admin]

Actually, I asked the PLASMA team, and here is what Jakub said:

Has to be a classic case of a performance "glitch/mishap".
For instance, a drastic oversubscription of threads, which may be the case if a multithreaded library calls a multithreaded BLAS.
Can be a case of a drastic affinity problem, e.g. all threads pinned to one core.
It has to be a problem of that sort.
I would focus on the number of threads first, and then affinity.
Just an educated guess.
As a side note, GNU + Windows + OpenBLAS sounds like an explosive combination.
I would attempt that if I was trying to break things