[CIG-MC] Benchmark of convection codes including CPU timing and scaling

[Matthew Knepley]

On Thu, Aug 12, 2010 at 2:01 PM, John Hernlund <hernlund at gmail.com> wrote:

> Hi Shijie,
>
> I think this is a good thing to do, since,
>
> 1) there is a lot of discussion about the speed of various codes, but
> no comprehensive quantitative comparisons have been performed to my
> knowledge,
>
> 2) solver efficiency is an important benchmark of progress in mantle
> convection-related computing and code development and is one of our
> most important computational issues in general, and,
>
> 3) if we can sort out which solution approaches work best (e.g.,
> treatment of pressure, various multigrid schemes, etc.) then we can
> perhaps understand which direction future development might want to go.
>
> It is not always simple to compare speed for different codes running
> on different architectures and different CPUs (unless you know of a
> good way of measuring the number of floating point operations, FPOs,
> which is the more robust measure), and running these cases will take a
> little time, but perhaps for the above 3 reasons everyone with a code
> can run some of these comparisons.
>

4) PAPI is a nice and easy way to measure the floating point operations. It
    works particularly well on some architectures (BG, Linux), and you only
    need to run it once to get your count.

5) All the time measurements are necessary for a benchmark, but just as
    crucial is a complete specification of the stopping criteria for the
iterative
    solver. Without this, it is impossible to compare the efficiency of
different
    solvers and preconditioners.

  Thanks,

     Matt


> Paul Tackley (cc) or somebody in his group might be able to run these
> comparisons relatively quickly using his code, and this could be a
> good place to begin.
>
> Cheers!
> John
>
> On Aug 12, 2010, at 10:19 AM, Shijie Zhong wrote:
>
> >
> > I apologize in advance if this topic is not of interest to you.
> >
> > At the GLADE meeting, there were some good discussions on the need
> > for more
> > benchmark, particularly regarding timing and speed. For CitcomS, our
> > 2008 G^3
> > paper provides quite extensive benchmark calculations for Stokes'
> > flow,
> > isochemical and thermochemical convection with large viscosity
> > contrast (up to
> > 1e7). In Table 1 of this paper, we also include some information on
> > the speed
> > and CPU time for a typical calculation, and scaling up to 3000 cores.
> >
> > Needless to say, more benchmarks and comparisons with other codes
> > will be
> > helpful. If any of you are interested in doing this, I would be more
> > than happy to
> > participate or take the lead. I can provide all those solutions
> > published in the
> > paper including analytical solutions for Stokes' flow.
> >
> >
> >
> > Shijie Zhong
> > Department of Physics
> > University of Colorado at Boulder
> > Boulder, CO 80309
> > Tel: 303-735-5095; Fax: 303-492-7935
> > Web: http://anquetil.colorado.edu/szhong
>
> _______________________________________________
> CIG-MC mailing list
> CIG-MC at geodynamics.org
> http://geodynamics.org/cgi-bin/mailman/listinfo/cig-mc
>



-- 
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://geodynamics.org/pipermail/cig-mc/attachments/20100812/de49ab82/attachment.htm