Ghost points do not make sense since it is not clear how they should be set. 
This these arrays are not formed via a domain decomposition, there is no source 
process that would set them.

Another point to keep in mind is recovery on a different number of processes. 
If each process can have different data, then it does not make sense to recover 
these arrays on a different number of processes, since either information would 
be lost, or would have to be invented.

-erik

On Mar 9, 2014, at 22:01 , Roland Haas <[email protected]> wrote:

> Signed PGP part
> Hello all,
> 
> while debugging some code using grid arrays with DISTRIB=CONSTANT I
> noticed that Cactus's IOASCII output code outputs data from all
> processes which differs from Carpet's CarpetIOASCII which outputs only
> data from a single process.
> 
> Carpet (and Cactus) seems to assume (see comments in line 89 in
> CarpetIOHDF5/src/Output.cc) that "DISTRIB=CONSTANT arrays
> (including grid scalars) are assumed to be the same on all
> processors". Certainly many grid arrays with DISTRIB=CONSTANT (eg. the
> SphericalSurface radius arrays) behave like this.
> 
> PUGH/IOASCII on the other hand seems to not make this assumption. In
> PUGH's case according to line 721 of SetupPGV.c
> --8<--
>     /* A -ve size means constant load per proc */
>     if (sh[dir] < 0)
>     {
>       if (local_size_includes_ghosts)
>       {
>         this->nsize[dir] = (nprocs[dir]-2) *
>           (-sh[dir] - 2*nghosts[dir]) +
>           2 * (-sh[dir] - nghosts[dir]);
>       }
>       else
>       {
>         this->nsize[dir] = nprocs[dir] * (-sh[dir]);
>       }
>     }
> --8<--
> Cactus would even allow ghost points for DISTRIB=CONSTANT arrays. This
> seems strange to me (and makes trying to make IOASCII behave the way
> CarpetIOASCII does much harder). If the array is not distributed
> across processes, why would there be ghost zones? Would it not make
> more sense to actively forbid GHOSTSIZE>0 for DISTRIB=CONSTANT grid
> arrays? In that case the local_size_includes_ghosts stuff is also not
> needed since nghosts would always be zero when sh[dir] < 0 (which PUGH
> uses to indicate DISTRIB=CONSTANT arrays).
> 
> Yours,
> Roland
> 
> --
> My email is as private as my paper mail. I therefore support encrypting
> and signing email messages. Get my PGP key from http://keys.gnupg.net.
> 
> _______________________________________________
> Users mailing list
> [email protected]
> http://cactuscode.org/mailman/listinfo/users

-- 
Erik Schnetter <[email protected]>
http://www.perimeterinstitute.ca/personal/eschnetter/

My email is as private as my paper mail. I therefore support encrypting
and signing email messages. Get my PGP key from http://pgp.mit.edu/.

Attachment: signature.asc
Description: Message signed with OpenPGP using GPGMail

_______________________________________________
Users mailing list
[email protected]
http://cactuscode.org/mailman/listinfo/users

Reply via email to