[petsc-users] Recommended unknowns per processor

[Matthew Knepley]

On Fri, Aug 3, 2012 at 4:52 AM, TAY wee-beng <zonexo at gmail.com> wrote:

> On 3/8/2012 11:10 AM, thomas wrote:
>
>> Hi,
>> look at the FAQ
>>
>> http://www.mcs.anl.gov/petsc/**documentation/faq.html#**slowerparallel<http://www.mcs.anl.gov/petsc/documentation/faq.html#slowerparallel>
>>
>> "There must be enough work for each process to overweigh the
>> communication time. We recommend an absolute minimum of about 10,000
>> unknowns per process, better is 20,000 or more."
>>
>
> Hi Thomas,
>
> Thanks for the info.
>
> Btw, is there an upper limit? When the unknowns are 100,000 or more per
> process, is it better (as in faster) to use more processes instead (e.g.
> 20,000 per process)?


That depends on the machine. The lower limit is the minimum number of
unknowns to cover the latency of network communication.

   Matt


>
>> Regards
>> Thomas Huxhorn
>>
>> On 08/03/2012 09:03 AM, TAY wee-beng wrote:
>>
>>> Hi,
>>>
>>> I remembered that in one of the mails, someone mentions the recommended
>>> unknowns on each processor when doing parallel solving.
>>>
>>> However, I can't find that mail. Can someone enlighten me? Also, is this
>>> number problem dependent?
>>>
>>> Btw, I'm using PETSc to solve linear equations using BCGS.
>>>
>>> Thanks!
>>>
>>>
>>
>


-- 
What most experimenters take for granted before they begin their
experiments is infinitely more interesting than any results to which their
experiments lead.
-- Norbert Wiener
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20120803/56d0c828/attachment.html>