[CIG-LONG] testing gale
Walter Landry
walter at geodynamics.org
Thu May 3 17:11:27 PDT 2012
Hi Alex,
I am Cc'ing the list.
Alex T <ataka530 at hotmail.com> wrote:
>
> Hey Walter,
>
> I'm compiling/installing gale on a number of different platforms and
> architectures, and running tests on them to determine whether the
> results are consistent. Is there a particular value that would be
> useful in determining the accuracy of the computation? I'm not sure
> if running yielding.json is sufficient or not for determining the
> effectiveness of gale on a particular platform, or whether it will
> provide enough computation to detect problems platform to platform.
> I've written a general script to ask the testing nodes to provide
> feedback and error reports when numbers deviate beyond a pre-set
> error bound, so If there's an error bound and value you would like
> to use for the tests that would be helpful.
Running yielding.json, non_newtonian.json, and extension3D.json should
exercise enough of the code to tell you whether there is anything
wrong. To compare results, I would look at the residual for the last
solution. For example, running yielding.json with increased verbosity
should give you a line at the end like
Non linear solver - Residual 9.88682846e-03; Tolerance 1.0000e-02 - Converged - 4.635880e-01 (secs)
The timing will be different, but the value of the residual should
match pretty well (at least 4 figures).
The residuals may be different in parallel, so make sure you only
compare parallel runs to each other.
Also, if you are compiling from source, I would recommend using the
version in the repository. It has a number of important fixes.
Cheers,
Walter Landry
More information about the CIG-LONG
mailing list