[Paraview] Paraview Benchmark and IceT timings

[Moreland, Kenneth]

There are some internal measurements taken, but I'm not sure there is any
way to get to them without hacking on some code.


You might be able to get a rough estimate of where the time is taken by
playing with the input a bit. If you put a small number of large polygons
on each process and force compositing on, that can give you an idea of the
time to do (mostly) just compositing. Conversely you could load in a
smaller number of polygons and turn compositing off.

Not sure if that helps...

-Ken

On 8/22/13 2:40 AM, "Tijs de Kler" <tijs.dekler at surfsara.nl> wrote:

>Hey all,
>
>I am trying to do a benchmark, in order to be able to compare our GPU
>cluster and our Supercomputer. The GPU cluster has one Nvidia GTX680 per
>node 
>and 10G ethernet, while the supercomputer has no GPU's and Infiniband
>
>In short, i open a file with 31 million triangles, run D3() to divide it
>over the processes, and render it to a 1600x1200 renderview, using a
>python 
>script (attached) and pvpython (paraview 3.98.1).
>
>However, i am running into some unexpected deviations in my frame render
>times on the GPU-cluster, when using multiple (8+) nodes. Most frames
>render 
>in a constant time, but every so many frames i get one 2 to 3 times as
>slow.  The output of the benchmark module doesnt shine much light on the
>subject: All processes report the same, much higher frame time for these
>frames, so it is not clear if one process is slowing the whole down.
>
>Our initial hunch was that it might have to do with the compositing step:
>Since the infiniband is much faster in terms of latency and throughput,
>this 
>step might be more efficient on the Supercomputer.
>
>My question is, is there a way to separate the benchmark times of the
>rendering and the compositing part of each frame?
>Or is there any other advice on pinpointing the performance culprit?
>
>regards,
>Tijs de Kler
>SURFsara Visualization
>
>
>