[vtk-developers] MPI_Comm_split

[Moreland, Kenneth]

vtkMPIController::PartitionController is a specialization of the more generic vtkMultiProcessController::PartitionController.  You should not modify the vtkMPIController version without also making equivalent changes in vtkMultiProcessController.  The correct implementation will do the following:


 1.  Create a constant like vtkMultiProcessController::NO_PARTITION.
 2.  Modify vtkMultiProcessController::PartitionController to recognize the NO_PARTITION color and return NULL instead of create a new controller for that group.
 3.  Modify vtkMultiProcessController::PartitionController to convert NO_PARTITION to MPI_UNDEFINED and then convert any MPI_COMM_NULL to return NULL instead of a new vtkMPIController.  (It is a bad idea to create controllers that do not behave properly.)
 4.  Add code to VTK/Parallel/Testing/Cxx/ExerciseMultiProcessController.cxx to use this new feature with PartitionController.  This code will be used to test both the vtkMultiProcessController and vtkMPIController versions of that method.

-Ken


On 6/11/10 4:18 PM, "pat marion" <pat.marion at kitware.com> wrote:

VTK wraps MPI_Comm_split with vtkMPIController::PartitionController(), which calls vtkMPICommunicator::SplitInitialize().  Currently there is no support for passing MPI_UNDEFINED as a color.  If you pass MPI_UNDEFINED as a color you should get back a null communicator (MPI_COMM_NULL).

There is handling of null communicators in vtkMPIController::CreateSubController() which wraps MPI_Comm_create.  So I can follow the same pattern to make PartitionController support it too, no problem.

But I do a lot of prototyping in python, and in python I can't easily use MPI_UNDEFINED.  So I'm thinking about adding a static method called vtkMPICommunicator::GetMPIUndefined() that returns MPI_UNDEFINED  (it's just an int).  Objections?  My other thought was to have vtkMPIController::PartitionController automatically convert negative color values to MPI_UNDEFINED, but that behavior could cover up programming errors, so I don't like that.

Pat



   ****      Kenneth Moreland
    ***      Sandia National Laboratories
***********
*** *** ***  email: kmorel at sandia.gov
**  ***  **  phone: (505) 844-8919
    ***      web:   http://www.cs.unm.edu/~kmorel

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/vtk-developers/attachments/20100614/3e3fa8bd/attachment.html>