[Trilinos-Users] Partitioning Tpetra matrices

[Hoemmen, Mark]

On Dec 6, 2013, at 12:00 PM, <trilinos-users-request at software.sandia.gov>
 wrote:
> Message: 1
> Date: Fri, 6 Dec 2013 18:27:33 +0100
> From: " Pavel Jir?nek " <pavel.jiranek at gmail.com>
> Subject: [Trilinos-Users] Partitioning Tpetra matrices
> To: "trilinos-users at software.sandia.gov"
> 	<trilinos-users at software.sandia.gov>
> Message-ID:
> 	<CAL_nKQ7psSgtkcQ-d7BYLtEc2=ViPmh=VvDfAKypk9pxqbsgrg at mail.gmail.com>
> Content-Type: text/plain; charset="iso-8859-1"
> 
> Hello,
> 
> I was wondering what are the options for partitioning Tpetra matrices (no
> geometry, pure graph partitioning) in Trilinos (in particular with long
> integer global indices).
> 
> I've noticed that there might be some support for Tpetra through Isorropia
> (although only for normal integers, I guess because Zoltan is hard-coded
> with int types), however, Trilinos does not compile after using
> Isorropia_ENABLE_TPETRA=ON in my config script.
> 
> I've also tried Zoltan2. I understand that, at this moment, the only option
> for graph partitioning is through the interface to PT-Scotch (I suppose
> working as well for long integers as soon as PT-Scotch is built with the
> proper option for integer length). It seems to work well for improving
> partitionings of already distributed matrices/graphs, however, the program
> crashes (it seems that somewhere in the construction of a GraphModel
> instance) on processes which do not contain any rows. For instance, this
> happens when reading a (rather small) matrix on process 0 and computing a
> partitioning from scratch. Interestingly, this problem can be fixed by
> first naively redistributing the matrix to a uniform map and then applying
> Zoltan2 to the redistributed matrix.
> 
> Thanks for any feedback.

That could be a Zoltan2 bug.  I'll forward it on to the appropriate developers.

Thanks!
mfh