<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
</head>
<body bgcolor="#ffffff" text="#000000">
Hello all,<br>
<br>
I've been making some changes to the caching/interpolation used in
vtkStreamTracer and mostly vtkInterpolatedVelocityField - which I hope
will give a performance increase - but I have tweaked so many
things over a relatively long period, that I'm not entirely confident
of benchmarks - and to be honest I don't want to spend the time setting
up a series of (rigorous) tests to confim anything.<br>
Other groups have also compared Ensight Streamline generation
performance to ParaView and found between 2x and 8x performance
difference between the two packages. <br>
<br>
Is there anyone out there that can help me out, by doing the following <br>
1) Use a <b>Release </b>build of <b>CVS head</b> paraview or raw VTK
C++ (from very recently - or now)<br>
2) To generate streamlines on the biggest data they generally work
with. <br>
The more cells and the more multiblock datasets the better as my
caching is partly aimed at multiblock improvements. One addition is the
ability to use a locator for initial seeding. This adds a few seconds
to the initialization - but improves performcnce when injecting many
seeds.<br>
3) Provide a couple of timing benchmarks which are<br>
a) A large number of relatively small streamlines (where Injection of
the initial
seed points has a big impact on the calculation)<br>
b) A large-or-small number of long streamlines on big datasets where
the streamlines pass through many
blocks of multiblock data - the inner loop velocity interpolation
generally dominates when the integration produces long streamlines.<br>
4) Compare the perfomance with a similar/identical streamline
generation in Ensight (optional request, if you can that'd be great)<br>
<br>
5) Do a cvs update after I commit some changes and rebuild paraview
with exactly the same settings for release/optimization etc<br>
6) Re run identical tests which will show if the changes have made any
difference.<br>
7) The tester should be reasonably familar with all the parameters that
can be tweaked in the vtkStreamline module so that they can tune the
test so that it is realistic.<br>
8) The tester(s) should be able to set up a test without help - I'd
hope that if someone can dedicate an hour or so to testing before CVS
commits and then a short time after updating from cvs (assuming test is
now setup and needs no adjustment) that they will require no help from
me.<br>
<br>
the fixes I have made are mostly for the benefit of the ParticleTracer
module, but in the process I have made some tweaks which should help
the Streamline module too. I'm keen to find out if they really have
made
any difference (5% or 50% - I have no idea?) . If ParaView can compete
with Ensight, that'd also be
useful to know. It may be a week or more before I check in my stuff as
I have to synchonize my code with the multiblock/composite changes of a
few weeks back, but if a handful of people can do tests this week and
then again in a week or so, that'd be great.<br>
<br>
Thanks in advance<br>
<br>
JB<br>
<pre class="moz-signature" cols="78">--
John Biddiscombe, email:biddisco @ cscs.ch
<a class="moz-txt-link-freetext"
href="http://www.cscs.ch/about/BJohn.php">http://www.cscs.ch/about/BJohn.php</a>
CSCS, Swiss National Supercomputing Centre | Tel: +41 (91) 610.82.07
Via Cantonale, 6928 Manno, Switzerland | Fax: +41 (91) 610.82.82</pre>
</body>
</html>