I'm actually trying to get best performances of Paraview, for the moment on one machine (core i7 920, 2x GTX295, 6Go DDR3, Ubuntu Jaunty 64bits), and maybe more computers in future.
I'm writing here to request some help.
To begin, take care of this : VirtualGL uses his own opengl libs, Paraview alone uses normals Opengl libs.
I got 2 same versions of paraview 3.4.0 > The precompiled version, and a version I compile myself with Qt 4.3.0 and MPI support.
With help of Roland by emails (thanks
) my quadSLI works with 185.18.04 (beta) 64bits version of NVIDIA's drivers. (I can see on each GPU in nvidia X server settings the mention (SLI) ).
So, if I launch paraview precompiled without the OpenGL SLI heads in nvidia driver, I got my rendering in 2,1seconds.
If I launch paraview precompiled WITH the OpenGL SLI heads in nvidia driver, I can see in the rendering window of paraview "Quad SLI", with lines showing the type of quadSLI (vertical lines on left and right sides for AFR, and 3 horizontal lines for SFR), I timed for a rendering 2,1 seconds too (so where's is the SLI's performance ?)
Then now with my compiled paraview, I got the same problem (same time for rendering with and without SLI) BUT with 1 more seconds to render (3,1 seconds), I don't understand why.
Ok, now I try exactly same test WITH virtuagl to run paraview (vglrun ./paraview), with the 2 versions of paraview, I don't see any type of logo "QuadSLI" (normal, because it's not the opengl lib uses by nvidia's drivers), but I got exactly sames perfs, 2,1sec for paraview precompiled, and 3,1sec for myself's compile.
if you've read so far, already bravo
I'm searching if I didn't do anything wrong, or if it's normal to don't get better performances with SLI.
And maybe if there's a soft for linux who show informations about current use of GPU, or any other informations about graphics cards in real time.
Thanks for help, and good day