HOUSSEN Franck <[email protected]> writes:

> MatMult            13216 1.0 5.4614e+01 1.0 5.52e+10 1.0 0.0e+00 0.0e+00 
> 0.0e+00 21 56  0  0  0  21 56  0  0  0  1010

Much of your time is spent here, which is bandwidth limited.  It needs 6
bytes per flop (plus a little, if the vector is perfectly reused) , so
this number is about 6 GB/s.

> MatMult            13550 1.0 3.8204e+01 1.0 2.85e+10 1.0 2.7e+04 3.5e+04 
> 0.0e+00 23 55 79 93  0  23 55 79 93  0  1494

Here with two processes, you have about 9 GB/s.

Was your STREAM test (getting 11 GB/s) using multiple threads/processes?
Can you send STREAM results for one and for two threads?  50% of STREAM
is not very good (though it's actually the best you can do on some funny
architectures), 70-85% is what we expect.

If you're getting a low fraction of peak in MatMult, try reordering your
matrix to have lower bandwidth.  You can use MatGetOrdering with RCM for
this.

Attachment: pgpCY4yWeX8Yy.pgp
Description: PGP signature

Reply via email to