[petsc-users] speedup on KSPSetUp() using qmd reordering

[Pedro Torres]

2010/5/7 Matthew Knepley <knepley at gmail.com>

> On Fri, May 7, 2010 at 2:06 PM, Pedro Torres <torres.pedrozpk at gmail.com>wrote:
>
>> Hello,
>>
>> I have a sparse linear system, block dense, one block for each process,
>> and solve with cg and block jacobi (ICC(0)) as preconditioners, with
>> different reordering. During my tests I found quasi linear speed on
>> kspsetup() and KSPSetUpOnBlocks() using 1wd,nd,rcm reorderings, but using
>> qmd I get superlinear speedup. My CPU is a Xeon 5410.
>>
>> Is that possible or something is going wrong?, and if I want to explain
>> that, what functions on this stage should I monitor??. I really
>> appreciate any advice.
>>
>
> It is possible (but unlikely) that the ordering on a larger number of
> processes creates a better
> preconditioner, particularly because ICC(0) is so unpredictable. You can
> try and separate
> arithmetic efficiency from algorithmic efficiency by looking at the number
> of iterates between these
>
runs.
>
Thanks for a swift response. I found that given an ordering, the number of
iterations is almost  uniform,with just a little increase, but 1wd have the
minimun numbers of iterations. May be I should look
in the ratio of cache miss of MatCholeskyFactorNumeric to explain this
behavior,
rigth?.
Thanks a lot!

>    Matt
>
>
>> Thanks in advance.
>>
>> --
>> Pedro Torres
>> GESAR/UERJ
>> Rua Fonseca Teles 121, São Cristóvão
>> Rio de Janeiro - Brasil
>>
>
>
>
> --
> 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/20100507/e555e36c/attachment.htm>