[petsc-users] KSP_SpeedUP

Pedro Torres torres.pedrozpk at gmail.com
Wed Apr 14 17:07:10 CDT 2010

Thanks, for a quickly reply. I have four nodes (GigaEthernet), each node
with two Quad Core E5410 at 2.33GHz, Mem 16Gb - DDR2 667Mhz., and definitely
I'm not have the enough memory bandwidth for a reasonable speedup.

This may be a dummy question but in the second bullet says "its own memory
bandwith of roughly 2 or more gigabytes", this means gigabytes/seconds, or
refers to amount of memory per core?. Thanks a lot.


2010/4/14 Barry Smith <bsmith at mcs.anl.gov>

>  Second bullet at
> http://www.mcs.anl.gov/petsc/petsc-as/documentation/faq.html#computers
>   Barry
> On Apr 14, 2010, at 4:10 PM, Jed Brown wrote:
>  On Wed, 14 Apr 2010 18:02:11 -0300, Pedro Torres <
>> torres.pedrozpk at gmail.com> wrote:
>>> Hello,
>>> Sorry if this questions its not appropiate for the petsc-list, but I
>>> really
>>> want to known what happen when I'm getting differente KSP time results.
>>> For
>>> example, allocating two process in the same node I get 6.24 sec, and when
>>> allocating two process in two nodes (1 process per node) I get 4.7sec.
>>> Is
>>> there a memory contention problem in my node?? The problem get worst when
>>> increase the number of process.
>> Sparse matrix kernels are primarily limited by memory bandwidth which
>> does not increase much with multicore hardware (vendors rarely mention
>> this).  When you use multiple cores per socket, they have to share the
>> available bandwidth, so you get lower performance.  It's *usually* still
>> faster to use the available cores, but the per-core performance is
>> definitely lower than when using only one core per socket.
>> Jed

Pedro Torres
Rua Fonseca Teles 121, São Cristóvão
Rio de Janeiro - Brasil
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20100414/46b52aff/attachment.htm>

More information about the petsc-users mailing list