[Paraview] minimal linux cluster for paraview, volume rendering

[Andy Cedilnik]

Hi Dan,

You should think of graphic card as two things:

1. Almost general purpose computer wich is optimized for doing graphics
stuff, such as OpenGL, ...
2. Something that takes frame buffer and sends it to the monitor. 

You can take frame buffer and send it back to the main CPU though. That
is what ParaView does.

On Fri, 2003-08-22 at 10:05, Dr. Daniel James White PhD wrote:
> So is it possible to have a good openGL enable graphics card on each 
> node to do some number crunching, and send the results back to the 
> master node.
> I thought  everything on the graphics card went in only 1 direction: 
> out to the monitor?

Depending on where the bottleneck is. If it is in the displaying, then
you will lose. If it is in the number crunching, then you win. The
second thing is especially the case when the number crunching part can
be parallelized. ParaView does a great job in that. If you do not have
space problem, then I would get cheap nVidia cards and put them in. They
will speed-up the rendering, so no Mesa, more CPU for the other stuff.

> I was thinking 4 Athalon equipped motherboards with 1Gb ram and normal 
> ethernet cards, that each boot off the master node, so no hard disks or 
> video cards.
> I guess without the videocard hardware acceleration on each node, the 
> nodes wouldn't really speed up the openGL display on the master node 
> (with a nice videocard), since software rendering is so much slower 
> than hardware?

ParaView at this point does not support volume rendering. That said,
volume rendering is primarly cpu intensive and not graphics intensive.
(Except of-course if you use VolumePro or 3D texture mapper). In most
cases all you need to do is display a 2D image, which can be almost as
fast with no graphics card (if you do not do any zooming and panning).

If you do isosurface, then you need to consider two things:

1. Isosurface tends to be processor intensive and slow operation (if you
do not use patented algorithms).

2. It can produce lot's of polygons, which could be slow on software
rendering.

			Andy

> What i am looking for is fast interactive display of large datasets, 
> using isosurfaces or even better true volume rendering (which I know 
> isn't in paraview yet, but i can do it with some other vtk tools, ie 
> Mayavi with the experimental volume module. Volume rendering with 
> composite ray cast function is slow at high quality settings, that is 
> what I want to speed up. It would be nice if the hardware stereo wasn't 
> swapped LR in the vtk raycastmapper... but that is another story...)
> Are you telling me that parrallelisation of the vtkVolumeRayCastMapper 
> or whatever it is called would not really help, since it is the 
> graphics card that does most of the work anyway...?