[Paraview] "Number of cores" vs. "Data size"

[David E DeMarle]

On Sat, Sep 15, 2012 at 10:03 AM, Arya Mazaheri <aryanet at gmail.com> wrote:
> Hi there,
> 1) I have installed and ran paraview in parallel client/server mode. My
> server has 48 CPU cores and 2 high-end GPUs. But I'm confused that how many
> cores should I assign to my dataset considering it's size. Is there any
> equation available which can answer my question?

Amdahls law vs Gustafson's law.

ParaView's parallelism is for data scalability - large data is made
visualizable because of abundant RAM not abundant CPU. Since most vis
algorithms are not computationally intensive if you can fit the data
(including the derived datasets produced by algorithms along the
pipeline)  in RAM and process it with one core, throwing more cores at
the problem is very likely to slow things down. If the data doesn't
fit in RAM, you can throw another workstation (or thousands of them)
with its own disk and RAM at the problem and make it possible to
process it.

> 2) I really don't feel any upgrade in performance in parallel mode. Should I
> install my GPU driver? I don't have OpenGL installed. Does it affect the
> performance dramatically?

Yes - if rendering is a bottleneck then the GPU will render faster
than mesa will on the CPU.

>
> For example I created a sphere source with 90MB data size. Then I tried to
> render it as 3d Glyphs. My Macbook Pro with single mode rendered it faster
> than rendering the same state with 24 cores in parallel mode.
>

90MB will almost certainly not benefit from more than one cpu. There
are some rules of thumb in the ParaView book section on best practices
for how big of a cluster you need for how big of a dataset, and
guidelines for how to avoid data size explosion.

> I appreciate any response in advance.
>
> Regards,
>
> _______________________________________________
> Powered by www.kitware.com
>
> Visit other Kitware open-source projects at
> http://www.kitware.com/opensource/opensource.html
>
> Please keep messages on-topic and check the ParaView Wiki at:
> http://paraview.org/Wiki/ParaView
>
> Follow this link to subscribe/unsubscribe:
> http://www.paraview.org/mailman/listinfo/paraview
>