[petr.f.]

Hi all,

I'd like to ask you for help with the following problem: I'm post-processing data from numerical simulation of 3D two-phase flow (water-air, polyhedral mesh with approx. 670 000 cells). I need to use volume mapping in paraview (so I can see the exact motion of water during time).

However both rendering and making any changes in view (redraw after rotating, adding text into the picture field...) are TERRIBLY slow. Rendering of one picture (time-state) takes 6 minutes. I've brand new, 12 cores (Intel Xeon E5-2667, 2.9 GH) machine with 64 GB of RAM and NVidia Quadro 4000 (2 GB) graphic card with Scientific Linux 6.2.

As I understand it, there is no advantage in running paraview in parallel on the local machine. But are there some recommended (best?) settings for speed up paraview in cases like this (I've read this post: http://www.cfd-online.com/Forums/ope...-parallel.html, but it didn't help much)? So far, I'm only testing various configurations and settings of the solver. At the end I can expect very fine mesh with cca 20E6 cells on a complex geometry. With the current speed the rendering would take more time than the computation itself...

[MartinB]

Hi Petr,

do you have the proprietary Nvidia drivers installed? If your system is using the OpenSource Nouveau drivers you don't have a chance to speed it up.

Martin

[petr.f.]

Hi Martin,

I have the newest (but apparently not certified) NVidia drivers installed, OpenGL is working (at least on glxgears). Shoud I try to compile paraview with MESA support?

What I'm not sure about is whether the problems are consequence of bad OpenGL instalation or long read times from harddisk. However, as the mesh has "only" cca 6.7e5 cells, my guess would be the first case.

[MartinB]

Hi Petr,

the attached image is rendered on a NVIDIA Geforce GTX 460 card (several years old by now) and it takes less then 2 seconds for a mesh with 2600000 cells.

I have my doubts, that with MESA support it is faster than with a graphics card, but I have never tried it...

Can you setup another Linux (OpenSUSE for example) and the NVIDIA drivers?

Martin

[petr.f.]

Hi Martin,

thanks for Your reply. I will try Suse on my laptop (don't have root rights for the desktop computer).

However, when You look on the two pictures below, it takes more than 2 minutes when I switch between one to the other (move between time steps) plus 20 seconds of rendering each one...

The water is not pure water but mixture of water and air (9:1), but I don't think that's the main problem.

[sail]

http://paraview.org/Wiki/ParaView:Server_Configuration

take a look at the paraview wiki, you might try to run the pvserver in parallel and see if it increase the performances.

also it could be useful to take a look at what the computer is doing while the rendering occours:

if the cpu is used in userspace you probably are using mesa dirver or the cpu itself to calculte the renderings and it is slow.

another possibility is that the whole system is waiting for the data to be read from the disk, if you have a very slow unit. I say that because you talk of 2 minutes to switch from two timesteps and just 20 secs of rendering time.

[petr.f.]

I've tried to run pvserver in parallel and the performance went down even more :-|
Velocity of harddisk should be OK. Anyway, I'll try another machine, where I have root access - to see, if there's any difference.

[petr.f.]

Unfortunately I didn't solve the problem yet (as we were working with smaller test models, so it wasn't imminent). But now, I've got 3D polyhedral mesh with ~2 600 000 cells, and it is almost "unrenderable". I've decided to use the binary version of ParaView (4.0.1, 64-bit, Scientific Linux 6.2) and tried to figure out if it is using cpus or gpu (NVidia quadro 4000, 2048 MB).

If I let it render the volume (either with 1 or eg. 8 cores) it runs on 100% on each cpu (according to top) and the nvidia-smi -a gives 0%-10% usage. I'd guess that in this case the cpus are used for rendering. Is there any other way how to check it? Thanks.

[Tobi]

Dear Petr,

did you solve your problem?
I have the same error. Always 100% CPU and non GPU further more my mesh size is over 2million cells - horrible

If I read sth. like that:

in 2 seconds

its like a dream

[Tobi]

While rendering, my GPU utilization is avg. 3-5 percent ... I think paraview is rendering on my CPU ? ... how can I change this?

[SteveGoat]

Hi,
did anyone managed to get fast volume rendering leveraging GPUs?
I am running Paraview on a PC machine and it is crazy slow

thanks
Ste