On Wed, 22 Jul 2009, Roy Stogner wrote:
>> Conclusion: New PetscVector implementation speeds up ghosted vectors
>> considerably, while leaving non-ghosted vectors approximately unchanged.
>> If the new PetscVector implementation is enabled, the ghosted vectors on
>> equal number of cores are approximately equally fast as the non-ghosted
>> vectors. If memory is the limitation, ghosted vectors will allow more cores
>> per processor and therewith speed up the computation considerably.
One of the runtimes I reported last time was actually wrong; here is
the corrected table. The change is in line 4. The conclusions stated
above remain essentially true, but the new PetscVector implementation
indeed also speeds up the non-ghosted vector application.
PetVec ghost #cores #nodes hh:mm:ss
new ghost 18 3 07:02:54
new ghost 9 3 08:19:55
new no 9 3 08:35:28
old no 9 3 08:57:34
old ghost 9 3 09:36:12
old ghost 18 3 07:30:14
>> Are there any more tests you would like me to perform?
> Presuming you verified correctness too on those runs,
Yes, I did. They coincide up to numerical differences. Actually, the
largest difference is between 9 cores and 18 cores (of course, because
the solver behaves different then).
> there's nothing else I can think of. Unless another developer
> objects, I'd say we should switch to your new implementation and
> turn on ghosted vectors by default.
> Would you send me a new patch against the SVN head? I'd hate to dig
> through my email for the last patch and then accidentally get the
> wrong version.
Sure, here you are.
Dr. Tim Kroeger
tim.kroeger@... Phone +49-421-218-7710
tim.kroeger@... Fax +49-421-218-4236
Fraunhofer MEVIS, Institute for Medical Image Computing
Universitaetsallee 29, 28359 Bremen, Germany