[petsc-users] CPU-scaling

[Matthew Knepley]

On Fri, Oct 18, 2013 at 4:49 PM, Torquil Macdonald Sørensen <
torquil at gmail.com> wrote:

> Hi!
>
> You probably hit the nail on the head with the debugging suggestion. After
> reinstalling petsc using --with-debugging=0, I'm getting more encouraging
> results. For the same problem as before, but with a larger matrix (now
> 10000x10000), I'm now getting the following scaling:
>
> # n_core    time (sec)
> 1                1540
> 2                870
> 3                660
>

Here you are very likely running into the memory bandwidth bottleneck.
Unfortunately, there is not
a whole lot you can do about this. This is not something vendor's like to
acknowledge.

     Matt


> 4                571
>
> which is much more promising than before. So thanks!
>
> - Torquil
>
> On 18/10/13 20:49, Jose E. Roman wrote:
>
>> No, you don't need MUMPS for ex1.c with default options.
>> Then I don't know what is happening. Did you build PETSc
>> --with-debugging=0 ? You should send performance data obtained with
>> -log_summary, otherwise we cannot guess what happens.
>>
>> Jose
>>
>>
>> El 18/10/2013, a las 17:21, Torquil Macdonald Sørensen escribió:
>>
>>  It is a standard eigenvalue problem, and I'm only using default
>>> algorithm options, so I don't think I'm using any nontrivial spectral
>>> transformation.
>>>
>>> The program is just ex1.c from the documentation, available here:
>>>
>>> http://www.grycap.upv.es/**slepc/documentation/current/**
>>> src/eps/examples/tutorials/**ex1.c<http://www.grycap.upv.es/slepc/documentation/current/src/eps/examples/tutorials/ex1.c>
>>>
>>> FAQ #6 mentions MUMPS, which I don't have the moment, so I'll have to
>>> look into that.
>>>
>>> - Torquil
>>>
>>>
>>>
>>> On 18 October 2013 16:46, Jose E. Roman <jroman at dsic.upv.es> wrote:
>>>
>>> El 18/10/2013, a las 16:42, Torquil Macdonald Sørensen escribió:
>>>
>>>  Hi!
>>>>
>>>> I'm just starting to use SLEPc on a quad-core workstation computer
>>>> (Intel Core i5-2500 CPU @ 3.30GHz).
>>>>
>>>> But I'm not getting the speedup that I was expecting, when increasing
>>>> the number of cores. E.g. when running ex1.c from the documentation, on 1,
>>>> 2, 3 and 4 cores. When asking for the 500 eigenvalues of smallest
>>>> magnitude, for a 5000x5000 matrix, I'm getting the following approximate
>>>> timings:
>>>>
>>>> # n_core  time_usage
>>>> 1              480s
>>>> 2              400s
>>>> 3              500s
>>>> 4              580s
>>>>
>>>> Should I not expect a speedup for n_core > 2?
>>>>
>>>> The example is using the krylovschur algorithm. For all four runs, the
>>>> program states "Number of iterations of the method: 33".
>>>>
>>>> I'm using PETSc 3.4.3, SLEPc 3.4.3, OpenMPI 1.6.5, and everything is
>>>> compiled using GCC 4.8.1 or 4.8.2.
>>>>
>>>> Best regards
>>>> Torquil Sørensen
>>>>
>>>>  Is it a generalized eigenproblem and/or are you doing
>>> shift-and-invert? In that case, have a look at FAQ #6
>>> http://www.grycap.upv.es/**slepc/documentation/faq.htm<http://www.grycap.upv.es/slepc/documentation/faq.htm>
>>>
>>> Jose
>>>
>>>
>>>
>


-- 
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/20131018/9994a4aa/attachment.html>