[petsc-users] Not getting scalability.

[Matthew Knepley]

On Mon, Mar 7, 2011 at 8:43 AM, Gaurish Telang <gaurish108 at gmail.com> wrote:

> Oh thank you, this was helpful.  I am interested in iterative solvers, so
> what is the minimum matrix size you
> think that strong scalability will show up for such methods?
>

Such blanket predictions are not worth much for strong scaling since they
depend on the
architecture, interconnect, etc. What is most important is to understand the
timing output
in -log_summary and see what is not scaling correctly. Dave pointed out that
linear iterations
must also scale correctly.

   Matt


>
> On Mon, Mar 7, 2011 at 9:38 AM, Matthew Knepley <knepley at gmail.com> wrote:
>
>> On Mon, Mar 7, 2011 at 8:20 AM, Gaurish Telang <gaurish108 at gmail.com>wrote:
>>
>>> Hi,
>>>
>>> I have been testing PETSc's scalability on clusters for matrices of sizes
>>> 2000, 10,000, uptill 60,000.
>>>
>>
>> 1) These matrices are incredibly small. We usually recommend 10,000
>> unknowns/process for weak scaling. You
>>     might get some benefit from a shared memory implementation on a
>> multicore.
>>
>>
>>> All I did was try to solve Ax=b for these matrices. I found that the
>>> solution time dips if I use upto 16 or 32 processors. However for a larger
>>> number of processors however the solution time seems to go up rather than
>>> down. IS there anyway I can make my code strongly scalable ?
>>>
>>
>> 2) These are small enough that direct factorization should be the fastest
>> alternative. I would try UMFPack, SuperLU, and MUMPS.
>>
>>     Matt
>>
>>
>>> I am measuring the total time (sec) and KSP_SOLVE time in the
>>> -log_summary output. Both times show the same behaviour described above.
>>>
>>> Gaurish
>>>
>> --
>> 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
>>
>
>


-- 
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/20110307/54c8204f/attachment-0001.htm>