[petsc-users] Number of unknowns/proc for scaling tests

[Matthew Knepley]

On Mon, Feb 27, 2017 at 10:29 AM, Casalegno Francesco <
francesco.casalegno at epfl.ch> wrote:

> Hello,
>
> I am pretty new to PETSc and I would like to get some advice on choosing
> the right ratio of unknowns/proc.
>
> I have implemented a finite element solver for 3D problems with
> potentially up to 5B vertices, using CG as iterative solver. To run weak
> scaling tests I have to keep constant the ratio unknowns/proc, but if that
> if this ratio is too small I will not see any scaling as the communication
> overhead would dominate. Similarly, for strong scaling tests I expect to
> observe nearly-linear scaling only up to a certain number of procs, after
> which communications will dominate.
>
> My question is: what is a reasonable minimum number of unknowns per proc
> to get good scaling efficiency? I know that for <50k unknowns a direct
> solver would be preferable, but I do not know if this number is also a
> reasonable guess for the ratio unknown/proc.
>
 We have normally recommended > 10K, but of course this depends on the
speed of the processor/memory vs. the network, and the
arithmetic intensity of your algorithm. Consider using the "performance
spectrum" plot. Here is an example:

  https://hpgmg.org/2016/06/21/isc16-list/

You run a variety of problem sizes, and you can see where strong scaling
starts to tail off on the left edge.

  Thanks,

     Matt

> Thanks in advance for your help!
>
>
> Francesco
>



-- 
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/20170227/bc9811fc/attachment.html>