[Paraview] large differences in rendering times

[Jean M. Favre]

Hi all

I have been measuring the Still Render time of a subset of an EnSight
case file with 46 datasets. The subset is made of 12 parts, for a total
of 85K cells and 43K points, i.e. pretty small stuff.

I tested 3 options:

1) rendering the "Extract Parts" object made of 12 parts.
2) Saving it to a *.pvd file (collecting 12 *.vtu) files, reloading and
rendering
3) Use AppendDatasets, save to disk as a single *.vtu file, reloading
and rendering.

All renderings are done on a dual-cpu Linux desktop, with Parallel
Projection, no strips, immediate mode ON, one scalar variable at the nodes.

With PV 1.8.4 compiled on my linux box (gcc 3.3.5), I get the same
rendering time () of 0.04 seconds. No problem thus far. That is my
reference time.

I switched to PV 2.0.2 and PV nightly and rendering times are up to 30
times longer.

Option 1) takes 1.3 seconds
Option 2) takes 0.10 seconds
Option 3) takes 0.05 seconds (pretty close to the resulsts of PV1.8.4)

Can anyone bring some light to this?

Option 1) being what I use most of the time (to avoid creating all these
additional *vtu) files, I have seen my application render data at about
4/5 seconds per frame (I have many EnSight files, with CAD, external and
internal CFD, thus with hundreds of datasets).

Using LOD does not help and every time the window is obscured (by a menu
pop-up for example), a still render takes place and takes too long...

I have compared the output of "ldd" on all my executables and they all
use exactly the same libs (OGL included).

So, while single dataset visualization seems equal in PV1.8.4 and PV2.*,
multi-dataset handling sems to have changed a lot between the two
releases. Or am I missing something?

Jean
-- 
Dr. Jean M. Favre,                            email:favre @ cscs.ch
http://www.cscs.ch/about/Favre.php
CSCS, Swiss National Supercomputing Centre  | Tel:  +41 (91) 610.82.40
Via Cantonale, 6928 Manno, Switzerland      | Fax:  +41 (91) 610.82.82