[Met_help] More MET questions

[Daniel Schaffer]

Thanks John

I do need to remind you of the FYI I included in my previous email that due to the short time fuse I'm on, I had to move ahead without MET. So I probably won't try your fix at this time.  However, as I also said, it may be appropriate down the road to come back to MET and compare my results with those from MET.  This would be contingent on a solution to the performance problem associated with a large number of points in the region polygons.  

Also, I think a subtext to GSL issue is the apparent fact that there was no way I could configure MET to avoid computing stats at all.  In my case, as I mentioned, I just needed the counts.  So, something else to think about for your long term planning purposes.

Dan

-------------------------------------------------------------------------------

Dan Schaffer, Research Associate            e-mail: Daniel.S.Schaffer at noaa.gov
NOAA/OAR/Earth System Research Lab         phone:  (303) 497-7252
Aviation Branch                             fax:    (303) 497-6301
R/GSD5   325 Broadway                       
Boulder, CO 80305

www-ad.fsl.noaa.gov/ac/schaffer.html

On Wed, 29 Oct 2008, John Halley Gotway wrote:

> Dan,
> 
> I looked more closely at the bug you're seeing when running Grid-Stat over the 1830x918 domain.
> 
> I narrowed it down to a GSL library routine that's choking on a the large number of grid points.  For whatever reason, the GSL routine "gsl_cdf_chisq_Pinv", which MET uses in computing normal
> confidence intervals, is choking when the degrees of freedom passed to it are greater than 1,263,131.  For your data, we're passing it a value of 1,679,940.
> 
> I tried downloading and linking to the latest version of GSL, version 1.11, released 3/30/2008.  But the bug is still present in that version.  So I've submitted a bug report to GSL.
> 
> As n -> infinity, the value returned by this routine -> infinity as well.  As a crude workaround, I've added a check to the wrapper for this GSL routine to check for n being too large.  If it is, the
> routine now returns a value of n*0.998.  With n so large anyway, this should have a negligible affect on the confidence intervals computed.
> 
> I tried this using your sample data files, and it does now work.
> 
> The attached file "gsl_cdf.cc" which includes the workaround replaces the file "METv1.1/lib/vx_gsl_prob/gsl_cdf.cc".  After replacing that file, please rebuild MET being sure to do a "make clean" first.
> 
> Thanks,
> John
> 
> Daniel Schaffer wrote:
> > Hey John
> > 
> > FYI.  Given the difficulties I ran into as we have been discussing and a short time fuse I face, I have gone ahead and written my own code that computes counts for various neighborhood widths.  So at this moment it is not super-critical to resolve these issues.  However, it would be interesting to compare the counts I generate with MET's answers.
> > 
> > I've attached the files that you requested.  Note that I only attached a sample forecast file since the behavior is the same if the forecast and obs files are identical.
> > 
> > As for your response below; I had already seen that.  I'm not clear how it's relevant to the problem I'm having with the GSL gamma function.
> > 
> >>> Regarding GSL, please see the entry about GSL on the known issue page of the MET website:
> >>> http://www.dtcenter.org/met/users/support/known_issues/METv1.1/index.php
> >>> Look in the section on Grid-Stat.
> >>>
> > 
> > Thanks for your continued support.
> > Dan
> > 
> > 
> > ------------------------------------------------------------------------
> > 
> > _______________________________________________
> > Met_help mailing list
> > Met_help at mailman.ucar.edu
> > http://mailman.ucar.edu/mailman/listinfo/met_help
>