[Paraview] Paraview on a cluster ; Streamtrace question

[Tim Bowker]

Hi all,

I am running paraview on a cluster with some success, however, it
seems incredibly slow and crashes often. The cluster has no graphics
hardware so I compiled and installed OSmesa, I then downloaded the
paraview CVS source and changed all open GL and mesa settings as per
the server wiki page. To circumvent firewall issues I ssh to the
cluster and start pvserver with the --use-offscreen-rendering flag. I
found I still got the bad X server connection message but this did not
lead to an immediate crash and although it was slow I could manipulate
everything up to a point (see below). Then, I downloaded and compiled
3.2.3 and I no longer observe the bad X server connection error but it
still seems very slow and still crashes frequently. I have
successfully performed the test under 'Paraview does not scale' on the
server wiki page. I tried a 6M then 1M cell unstructured grid (ensight
file with one zone) containing a velocity vector and 4 other variables
running on various combinations up to 112 processors (14 nodes x 8
processors per node & 8GB RAM per node) but it still seems slow. When
I look at the CPU load logs all processors are maxed out more that 95%
of the time. Typically my work flow is something as follows:
Import File
D3
Extract Surface
Perform 2 clips on extracted surface
Add streamlines (this frequently causes a crash however coarse I make
the settings)
Make 1 of the clips opaque (almost always crashes)
Often it also hangs after viewer manipulation or even just upon
selection of a different filter.

My questions:
Perhaps my expectations of the MESA library are too optimistic and
paraview really requires graphics hardware to run smoothly?
Has anyone got experience of which MPI library performs best? We have
openmpi gnu & intel, mpich2 & SCore and willing to install more :-)
Are there any benchmarks/tools that would give an indication of
whether the cluster is under performing / where the bottleneck is?
It's a new cluster and we don't have any software that we use
regularly with enough licences to test over all the cores.
On smaller grids on my desktop with a fairly recent graphics card the
serial paraview client zips along and does almost everything (see
below) I could wish for so I hope there are some tips out there to
improve the parallel performance.

Second:
When I have a streamtrace with attached glyph can I control the glyph
seeding so that spheres are plotted at fixed time intervals along the
streamtrace i.e. in low velocity regions spheres are bunched up and
conversely spread out in high velocity regions.

Thanks for reading this far and thanks in advance for any
comments/tips/solutions,
Tim