[Carpet] Bug in restriction

[Erik Schnetter]

On Apr 25, 2006, at 13:19:39, Christian David Ott wrote:

> Thomas Radke and I found a problem with Erik's patch of
> Feb 26th.:

[schnipp]

> which does not fix the following problem we observe:
>
> when running CarpetExtra/WaveToyMol (you also need CarpetExtra/ 
> IDScalarWaveMoL; parameter file attached) on a 50x50x50 basegrid  
> and 1 additional refinement level (2 levels total), 3 ghost zones  
> and 6 buffer zones, the results of runs using 1 to 8 processors  
> match to round-off error. However, runs using more than 8  
> processors yield results (pointwise and in the reduction results)  
> that are different in significant, non-round-off error digits.

Yes, there is a problem with this Feb. 26 patch "CarpetLib: Do not  
restrict to points that are used for boundary prolongation".  It has  
a logical error in the way in which it determines which points are  
the source for boundary prolongations -- it takes only points from  
the current processor into account.  That means that the restriction  
region depends on the number of processors.  Bad.

This patch has also caused some controversy in the past.  Ian Hinder  
reported that it helps with (but does not solve) certain problems  
that he was seeing.  Wolfgang Kastaun did not like that there  are  
now regions covered by both a fine and a coarse grid, and which are  
neither prolongated nor restricted.  I have in the mean time  
convinced myself that the prolongation operation before this patch  
was actually well defined.

I will ponder this problem a bit more, and I currently favour the  
idea of introducing a parameter (so that people can experiment), with  
the default setting disabling this patch.

-erik

-- 
Erik Schnetter <schnetter at cct.lsu.edu>

My email is as private as my paper mail.  I therefore support encrypting
and signing email messages.  Get my PGP key from www.keyserver.net.



-------------- next part --------------
A non-text attachment was scrubbed...
Name: PGP.sig
Type: application/pgp-signature
Size: 186 bytes
Desc: This is a digitally signed message part
Url : /archives/developers/attachments/20060503/021462cc/attachment.pgp