[Paraview-developers] [EXTERNAL] Re: Threading

[Scott, W Alan]

Thanks Berk – this sounds closer to my notes than what I wrote below.  I just thought my notes were wrong.  And as always, your team is on top of it!  I shall patiently await.

Alan

From: Berk Geveci [mailto:berk.geveci at kitware.com]
Sent: Friday, May 16, 2014 2:48 PM
To: Scott, W Alan
Cc: paraview-developers at paraview.org
Subject: [EXTERNAL] Re: [Paraview-developers] Threading

Hi Alan,

I don't think that this is exactly accurate. What they have done is to use threaded parallelism within each MPI rank in the cases where there are multiple blocks/parts to parallelize over. Meaning that they parallelize each filter's execution over blocks using posix threads. For example, a CTH dataset would benefit from a similar approach in ParaView (which we can also support). On the other hand, the way we load Exodus, we don't generate enough blocks to parallelize over so this approach would not result in any performance improvement.

Also, I very seriously doubt that they experience any significant performance improvement (if any) over using multi-threaded parallelism over MPI. Meaning if they used as many MPI ranks as threads on each node, I would bet that the performance would be comparable. This will not continue to be the case as the number of cores increase but it still holds for the kind of workload we have. Using threads reduces memory overhead however.

We have been exploring parallelism in similar ways both through VTK and VTK-m and I expect to see ParaView benefit from this work starting this summer. So don't worry :-)

-berk

On Fri, May 16, 2014 at 3:57 PM, Scott, W Alan <wascott at sandia.gov<mailto:wascott at sandia.gov>> wrote:
Utkarsh/All,
I head that VisIt started communicating between different cores/ processes on a single node using interthread communication (rather than MPI).  I would assume this is using openMP or some such.  Anyway,  my notes also say that they picked up a 2x speedup.  Would it be worthwhile moving ParaView to not use MPI to communicate within a node?  As nodes pick up more and more cores, this may become more important.

Alan


--------------------------------------------------------
W. Alan Scott
ParaView Support Manager

SAIC
Sandia National Laboratories, MS 0822
Org 9326 - Building 880 A1-C
(505) 284-0932<tel:%28505%29%20284-0932>   FAX (505) 284-5619<tel:%28505%29%20284-5619>
---------------------------------------------------------




_______________________________________________
Paraview-developers mailing list
Paraview-developers at paraview.org<mailto:Paraview-developers at paraview.org>
http://public.kitware.com/mailman/listinfo/paraview-developers

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/paraview-developers/attachments/20140516/9c892d8e/attachment-0001.html>