[Paraview] ParaView in parallel

[burlen]

Hi Paul,

If you have graphics cards available then you likely don't want to use 
mesa, rather hardware rendering.

You'd be better of reading you data in parallel than using D3. Why add 
an extra stage to your pipeline, and one that involves interprocess 
communication?

Depending on the data the vtk pipeline can tell you how to partition, 
image data for example. Otherwise you can devise your own approach. 
Perhaps giving each process the same number of points such that the 
points are clustered near each other is a good approach. I don't think 
filters need to do anything very special to work in parallel, unless 
interprocess communication is required. There is some documentation in 
the PV guide for parallel readers.

In my experience NFS is suboptimal for parallel readers, especially 
those using MPI IO. But I doubt nfs is entirely to blame for your troubles.
> I also tried saving my data once it was partitioned but then loading 
> it actually took longer (is this due to a lack of parallel filesystem?)
In your serial reader does each process load all of the data? If yes 
then I don't think you can blame nfs. What writer/reader did you use? 
Stay away from XML file format if you care at all about performance.

I have seen readers that divide the available blocks among available 
processes. However, I am a fan of the single file approach as it's 
easier to divide the data up such that one gets balanced loading 
independent of the number of processes used. MPI2 has optimization for 
non-contiguous access on parallel fs.

I often use PV in parallel with Mesa and lustre fs. The system I use 
doesn't have graphics hardware and if I don't run parallel then I run 
out of memory pretty quickly. I have image data and the vtk pipeline 
tells how to partition. It works well. My rule of thumb is to use as few 
processes as possible not to run out of memory. In my experience this 
gives the best performance, but it may have something to do with the 
particular data and filters I use, as I have heard Ken say that PV 
scales well to 100s of processes.


Paul Edwards wrote:
> Hi,
>
> I've been experimenting with ParaView in parallel without much.  Below 
> I have listed my setup, experiences and questions!
>
> *Setup*
> My setup is a 6 node gigabit cluster (without gfx cards), compiled 
> with OSMesa.  The filesystem is only a RAID on the frontend exported 
> with NFS.  Each node has 8 cores and I experimented running with both 
> 1 core per node and also 8 cores per node
>
> *Data*
> The data I am reading is multiple blocks, where each contains an 
> unstructured grid and a multiblock structure of polydata surfaces.  
> The data is about 10 million points in total.
>
> *Reader*
> My reader is not written for parallel so I loaded in the mesh and 
> partitioned it with D3.  The partitioning took a long time - approx 2 
> mins - is this normal?  When the data was partitioned I didn't really 
> notice much difference in speed and the rendering performance was 
> considerably worse than a single node for the same data.  Is this just 
> the result of using OSMesa?  If so, does anyone have any suggestions 
> for the number of nodes to be running on?  I also tried saving my data 
> once it was partitioned but then loading it actually took longer (is 
> this due to a lack of parallel filesystem?)
>
> *Filters
> *I tried to run some of my custom filters but one that calculates 
> scalar variables for pointdata didn't display the variables in the GUI 
> after running the filter.  Do I to do something different for parallel 
> filters?
>
> Finally, is there any documentation for implementing parallel readers 
> (or simple examples) in ParaView?  And, does anyone have any 
> suggestions for how to split up the data?
>
> Thanks,
> Paul
> ------------------------------------------------------------------------
>
> _______________________________________________
> 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
>