[vtkusers] Existing data parallel example in contrib director

Stanley Dunn smd at occlusal.rutgers.edu
Tue Aug 8 09:01:33 EDT 2000


Jim, et al -

On Tue, 1 Aug 2000, Jim Ahrens wrote:
> The example program ParallelIso.cxx found in the contrib directory is an
> example of a data parallel program. Multiple processes each create a

I found this sample program to be quite useful.  Before Jim's email, I had
gotten a version of it running on our linux cluster with MPI. For my own
edification, I also reworked it into a versions with multiple methods
instead of the single method, as written.

I've been working on parallelizing other common routines and I'm stuck
right now on streamlines.  As a first cut, I have a single process as in
the ParallelIso.cxx example.  Each process computes some subset of
the streamlines, having read the whole dataset. There is no call to
app->ParallelStreamingOn() to partition the dataset. Each process has it's
own streamline source, set by the controller via an RMI.

> Running this example in parallel with multiple processes should speedup
> the reading, isosurfacing and coloring filters of the program. (One
> caveat, too many processes can cause performance problems, for example,
> too many readers may overwhelm the file system). If you have question on
> this, please post back to the list and let's continue this thread.

The problem with the streamlines example is that when the computation is
finished and the streamlines are rendered, The render window interactor
behavior is different, depending on the number of processes across the
network.  I didn't expect this at all; rather, I would have expected that
once the compute nodes were finished and the data collected via the
AppendPolyDataFilter, that the rendering and interaction would be
independent of the number of processes.

I can see how too many processes can cause performance problems,
especially with readers, but I'm experiencing a performance problem that I
didn't expect.  I think it has something to do with the vtk execution
model and the implication of that across the cluster, but I'm not sure.
Anyone have any ideas?

Stan Dunn
Rutgers Univ/Princeton Plasma Physics Lab





More information about the vtkusers mailing list