[visit-users] VisIt memory usage scale with number cores?

Miller, Mark C. miller86 at llnl.gov
Mon Jul 20 13:01:47 EDT 2015


Hmm. How many 'blocks' do you have in the VTK files?

In order for VisIt to 'parallelize' well, the input database must already have been broken into chunks (typically, we call these chunks 'blocks'). If you have a VTK file with only a single block, then adding more processors does nothing useful. The other processors will idle, doing nothing, because there are no independent mesh chunks to give them to work on.

Next, the VTK reader in VisIt can wind up caching 2 or 3 copies of the input VTK file *before* the data gets to VisIt to even do anything with it. This happens for the *first*block* on *every* processor. If you have only one block, then that VTK data is getting cached nprocs * 2 times even though only one processor will ever operate on it (due to there just being one chunk). This behavior has to do with an impedance mismatch between VisIt's approach to caching input data and VTK readers.

If you have a multi-block VTK file, then everything I've said here does not apply (at least not as significantly).

Mark


From: Danny Sale <dsale at uw.edu<mailto:dsale at uw.edu>>
Reply-To: VisIt software users community <visit-users at elist.ornl.gov<mailto:visit-users at elist.ornl.gov>>
Date: Saturday, July 18, 2015 11:32 AM
To: "visit-users at elist.ornl.gov<mailto:visit-users at elist.ornl.gov>" <visit-users at elist.ornl.gov<mailto:visit-users at elist.ornl.gov>>
Subject: [visit-users] VisIt memory usage scale with number cores?

hello VisIt enthusiasts,

I compiled VisIt 2.9.2 so I could run in parallel mode.  When I start VisIt in parallel, like ./visit -np $nCPU, I notice that RAM usage scales with number of processors.  Why does this occur?

I am trying to visualize a list of VTK files I converted from many OpenFOAM time steps.  Here are a few measurements I take:

$nCPU    ~RAM usage (GiB)
-------------------------
 1   11
 2   16
 4   24
16   90

I was thinking that running in parallel mode would speed up the rendering time?  I'm running on a shared-memory workstation (2 CPUs at 8 cores each).  Can you advise how to speed up rendering time without excess RAM usage?  thanks!

best,
Danny Sale
phd student, mechanical engineering
University of Washington, Seattle
Northwest National Marine Renewable Energy Center


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://elist.ornl.gov/pipermail/visit-users/attachments/20150720/398d07bb/attachment.html>


More information about the visit-users mailing list