On Thu, Jun 13, 2013 at 9:00 AM, Georg Walther <[email protected]>wrote:

> Hi,
>
>
>
> I'm worried about the performance that FiPy shows for a simple
> three-variable reaction-diffusion system in one space dimension.
>

Hi Georg,

There are certainly issues with FiPy's performance both in terms of speed
and memory. We actually have a summer student working on profiling both
speed and memory and creating better diagnostic tools. I think going back
to the problem you highlight below would be a good test case. We will look
into in it and hopefully have some data in the next few weeks and reply
properly to your request.


>
>
> Going through posts on the mailing list I noticed a discussion from 2010
> entitled "speeding up an RD system":
>
> http://comments.gmane.org/gmane.comp.python.fipy/1913
>
> The messages back then culminated in improved code being posted which I
> copied onto Gist for convenience:
>
> https://gist.github.com/waltherg/86da981f3d9b8f9191d8
>
> and timings posted that compared the original code with this improved code:
>
>
> NEW (s)                         OLD (s)
> --pysparse
> 66                              175
>
> --trilinos
> 119                             691
>
> mpirun -np 4
> 48                              254
>
> mpirun -np 12
> 20                              106
> (on slow network)
>
>
> When I run the NEW code (see Gist link), my timings are in the ballpark
> of 120 seconds.
>
> I use the PySparse solver (haven't gotten around to installing Trillions
> yet) and these are the software versions I have on my systems:
>
>  >>> fipy.__version__
> '3.0'
>  >>> numpy.__version__
> '1.7.1'
>  >>> pysparse.__version__
> '1.2-dev224'
>
>
> I'm interested to learn what timings other users of FiPy have and I
> would greatly appreciate hints on speeding up FiPy computations in general.
>

At least one of the issues is that Trilinos and PySparse are not solving to
the same tolerance (if I remember correctly), they use different noms by
default. Another issues is that there is some Python-C interface overhead
that scales with system size in Trilinos. We'll try and get to the bottom
of some of these issues with the summer profiling project and then write up
something useful on this shortly.

Thanks.

-- 
Daniel Wheeler
_______________________________________________
fipy mailing list
[email protected]
http://www.ctcms.nist.gov/fipy
  [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ]

Reply via email to