[Paraview] Questions Regarding PV2.6 in Cluster configuration

[Randall Hand]

We just got our new cluster setup and i've been experimenting with Paraview
2.6 (compiled from source with OpenMPI support & Linked Against Mesa).

I've got a few questions for you guys:

1) When using the new Structured Volume Rendering in a Client-Server
(Parallel) configuration, it looks like it doesn't composite the images
correct.  I have verified that the "Disable Ordered Compositing" checkbox is
OFF.

2) Running with OSMesa is working Fantastic (no mangling used, just
OSMesa).  However, when running with a single node (or directly on the
cluster) running while linked against the Hardware Accelerated OpenGL
drivers is (obviously) much faster (1.5Mill Tris/Sec vs 150Mil Tris/Sec).
Is there any way to have Mesa there for client-server
(--use-offscreen-rendering) support, but have Hardware there for single-user
stuff?

3) I'm currently working with loading 2 datasets simultaneously
    a) Dataset A: VTI File, Uniform Rectilinear, 245,993,000 cells;
    b) Dataset B: Legacy VTK file, Polygonal, 20,801,406 Cells (all Vertex
Cells, Point Cloud Basically).
   When I run both the Client & Multiple Servers on the same Cluster
Machine, everything works perfectly.  When I run multiple servers on the
Linux Machine and a Windows Client (that I've compiled with Visual Studio
.NET 2003, no OSMesa, no MPI), it crashes when I load the second dataset.
It doesn't matter which of the 2 I load first, when I try to load the second
one it crashes the Client (the servers report no error and appear to be
functioning normally).  I can fix this by Setting the "Composite Above"
slider to 0.0, which makes it composite all the time.  The problem seems to
be related to the switch between Geometry collection & compositing that
happens when I load the 2nd dataset.

4) I know Paraview has support for automagically distributing non-parallel
data across the nodes.  Albeit sometimes it does a poor job of it, you've
supplied the D3 filter to fix it.  However, I'm loading a Uniform
Rectilinear grid, one of the easiest datatypes in the world to distribute,
yet I still get 1-cell wide seams in my data along each processor partition
when extracting an isosurface from it.  Are ghost cells missing in this
distribution?  Is there any way to generate the ghost-cells while keeping it
Uniform Rectilinear (so that I can continue to use the awesomely fast
FixedPoint Raycaster) ?


Any clues?  Overall, I'm blown away by how well this is working :)
-- 
----------------------------------------
Randall Hand
Visualization Scientist
ERDC MSRC-ITL
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://public.kitware.com/pipermail/paraview/attachments/20070124/05fc4356/attachment.html