[Dune-devel] dune-istl umfpack support

Dominic Kempf dominic.r.kempf at gmail.com
Thu Oct 31 13:45:50 CET 2013


Hey Christoph,
Hey devel-list,

while doing a writeup of my work for Markus, I reran some of the
performance tests and realized that infact SuperLU and UMFPack
performance results are NOT within the same order of magnitude.
UMFPack shows significantly better performance, see this table (I hope
its somehwat readable)

dofs                   1e4           4e4         1.6e5          6.4e5
T_{superlu}         0.076        0.395       2.707         19.839
T_{umfpack}       0.064        0.245       1.367         9.176
time savings      15.79%      37.98%    49.51%      53.75%
$M_{superlu}      30.78 MB  123.23 MB 493.2 MB
$M_{umfpack}    25.72 MB  88.66 MB  418.9 MB
mem savings     16.44%     28.06%     15.07%
$P_{superlu}      13.4 MB    55.64 MB  253.18 MB  1.16 GB
$P_{umfpack}    11.8 MB    41.33 MB  177.14 MB  771.16 MB
peakmem sav.   11.95%     25.72%     30.04%       33.53%

T is the wallclock time here, M the total allocation size as given by
valgrind and P the peak memory usage as given by /usr/bin/time -v.
As you can see, there is something to be gained here!

Best,
Dominic

On Thu, Oct 17, 2013 at 10:33 AM, Dominic Kempf
<dominic.r.kempf at gmail.com> wrote:
> Hey Christoph,
> thanks for testing the branch. My system gave me memory with zeroes,
> yours didnt... thats exactly where I need you.
>
> I did a bit of measurements to compare superlu and umfpack, though not
> on hard problems. In those tests umfpack was a bit faster and needed
> less memory, within the same order of magnitude though. The original
> reasons to implement umfpack support were conceptional difficulties
> with superlu in multithreaded environment. Markus can elaborate more
> on that (and whether umfpack actually solved the issue).
>
>
> Best,
> Dominic
>
> 2013/10/17 Christoph Grüninger <christoph.grueninger at iws.uni-stuttgart.de>:
>> Hi Dominic,
>> thanks for the quick patch. Using commit ...da3510863 it does no longer
>> segfault.
>>
>> Bye
>> Christoph
>>
>> --
>> Une science n'était vraiment développée que quand elle
>> pouvait utiliser les mathématiques.    (Paul Lafargue)
>> *********************************************
>> CMWR 2014: 10th - 13th June 2014 in Stuttgart
>>          Please visit www.cmwr14.de
>> *********************************************
>>
>> _______________________________________________
>> Dune-devel mailing list
>> Dune-devel at dune-project.org
>> http://lists.dune-project.org/mailman/listinfo/dune-devel




More information about the Dune-devel mailing list