[Paraview] Paraview and Multicore
Andy Bauer
andy.bauer at kitware.com
Thu May 20 10:55:21 EDT 2010
If I remember correctly, after you partition your points, you try to create
a grid with the Delaunay2D filter. It looks like the Delaunay2D filter
isn't a "true" parallel filter meaning that each process does the
triangulation on all of the points that are assigned to it but doesn't take
into account creating cells between close points that are on different
processes. The best suggestion I have is to try the following:
1) use D3 to partition your points
2) triangulate on each process using Delaunay2D
3) extract the edges on each process using Extract Edges
4) try the Delaunay2D to stitch these edges together (this may not work as
desired)
5) append the datasets/polydatas together using the Append Datasets filter
I have no idea whether or not this will work for you. The main problem is
that VTK/ParaView wasn't designed to be used as a parallel grid generation
tool.
Andy
On Wed, May 19, 2010 at 5:49 AM, Guido Staub <gstaub at udec.cl> wrote:
> So, with D3 things run faster now. Solved as well the mentioned
> problem, it was due to not scaling the z values. However, maybe a
> logical behaviour, the result of the triangulation now consists of
> four tiles. Anybody knows how to for example connect them? I have
> attached a screenshot showing the separation (vertical lines) of the
> tiles.
>
> Guido
>
> Am Tue, 18 May 2010 17:17:06 +0000 schrieb Guido Staub <gstaub at udec.cl>:
>
> > Using Process ID Scalars results in an Data Array ProcessId which in
> > case of np 3 ranges from 0 to 2 (and from 0 to 3 in case of np 4). I
> > have selected colorBy ProcessID to see if the input is well balanced
> > or not, unfortunately all the points have the same color. So I
> > conclude that the data is not well balanced, right?
> >
> > If I apply the D3 filter I almost see the same results. The only
> > difference I have noticed is that there is no new Data Array.
> >
> > Delaunay 2D is now about three times faster after I applied D3 filter.
> > However I do not like the resulting triangulation, seems like some
> > isosurfaces have been calculted.
> >
> > Guido
> >
> > Am Thu, 13
> > May 2010 20:48:11 -0400 schrieb Andy Bauer <andy.bauer at kitware.com>:
> >
> > > Process Id Scalars doesn't do any load balancing, it just shows what
> > > cells are assigned to what process. You can try the D3 filter to do
> > > actual load-balancing (your reader may already be doing this though
> > > in which case you would see the same results from a Process Id
> > > Scalars filter used after the D3 filter). I'm not that familiar
> > > with the internals of Delauney2D for distributed points but it's
> > > quite possible that for the set of points you have that the
> > > partitioning is not appropriate. As Ken said, using the Process Id
> > > Scalars to see if the input is fairly well balanced.
> > >
> > > Andy
> > >
> > > On Thu, May 13, 2010 at 3:33 PM, Guido Staub <gstaub at udec.cl> wrote:
> > >
> > > > Process Id Scalars didn't do the trick, still no performance gain
> > > > with Delaunay 2D.
> > > > My data is stored in a h5 file which gets loaded through a xdmf
> > > > file. After running the Process Id Scalars filter I do not see any
> > > > changes in the distribution of the data. However I guess that they
> > > > are already "evenly distributed" by default.
> > > >
> > > > So maybe exporting the file to an other format will do it or is
> > > > there a problem with the Delaunay implementation in general when
> > > > running in parallel?
> > > >
> > > > Guido
> > > >
> > > > Am Thu, 13 May 2010 15:08:04
> > > > -0600 schrieb "Moreland, Kenneth" <kmorel at sandia.gov>:
> > > >
> > > > > Ah, I see. It sounds like your data is not balanced. Many of
> > > > > the "non parallel" file formats will do something stupid when
> > > > > loading data in parallel. For example, they might load
> > > > > everything on process 0 or load everything everywhere. (And
> > > > > now that I think about it, the Delaunay filter may have trouble
> > > > > in parallel.)
> > > > >
> > > > > Try running the "Process Id Scalars" filter on your data. Do
> > > > > the points look evenly distributed?
> > > > >
> > > > > -Ken
> > > > >
> > > > >
> > > > > On 5/13/10 10:17 AM, "Guido Staub" <gstaub at udec.cl> wrote:
> > > > >
> > > > > Well if I start pvserver by mpirun -np 4 pvserver I have 3 cores
> > > > > running at almost 100%. Now I connect to the server and start a
> > > > > Delaunay 2D calculation on one of my datasets. As a result all
> > > > > of the 4 cores are showing 100%. However I assume that there is
> > > > > only one core doing the job, because on the one hand
> > > > > calculation is really slow. I have done similar processing on
> > > > > an other PC (an outdated one) and there is no significant
> > > > > performance advantage as one would expect. And on the other
> > > > > hand, running pvserver with e.g. -np 2 results in 100% for 2
> > > > > CPUs when starting the Delaunay 2D calc (1 core at 100% when
> > > > > cpu is idle).
> > > > >
> > > > > Guido
> > > > >
> > > > >
> > > > > Am Thu, 13 May 2010
> > > > > 12:39:43 -0600 schrieb "Moreland, Kenneth" <kmorel at sandia.gov>:
> > > > >
> > > > > > I am afraid I simply don't understand the question. You said
> > > > > > in (1) that you have three cores running at 100%. Then in (2)
> > > > > > you said that you only have one core running. Is it happening
> > > > > > when you start the client, connect the client to the server,
> > > > > > launch the server from the client, or something else? Is
> > > > > > something running or is the client sitting idle waiting for
> > > > > > the user?
> > > > > >
> > > > > > -Ken
> > > > > >
> > > > > >
> > > > > > On 5/13/10 8:15 AM, "Guido Staub" <gstaub at udec.cl> wrote:
> > > > > >
> > > > > > Thanks Ken, but I have already read this thread, therefore I
> > > > > > started the client process anyway without taking care of cpu
> > > > > > usage for now.
> > > > > >
> > > > > > However my second question still keeps me busy. Isn't it
> > > > > > possible to use all 4 cores?
> > > > > >
> > > > > > Guido
> > > > > >
> > > > > > Am Thu, 13 May 2010
> > > > > > 10:29:13 -0600 schrieb "Moreland, Kenneth"
> > > > > > <kmorel at sandia.gov>:
> > > > > >
> > > > > > > The question about why the pvserver processes are always at
> > > > > > > 100% CPU comes up frequently on the mailing list (such as
> > > > > > >
> http://www.paraview.org/pipermail/paraview/2008-December/010338.html
> > > > ).
> > > > > > > I've added some information to the Wiki about it to provide
> > > > > > > an explanation:
> > > > > > >
> > > >
> http://www.paraview.org/Wiki/Setting_up_a_ParaView_Server#Server_processes_always_have_100.25_CPU_usage
> > > > > > >
> > > > > > > -Ken
> > > > > > >
> > > > > > >
> > > > > > > On 5/13/10 5:06 AM, "Guido Staub" <gstaub at udec.cl> wrote:
> > > > > > >
> > > > > > > Hi all,
> > > > > > >
> > > > > > > I have succesfully compiled Paraview with MPI support on my
> > > > > > > Workstation (Quad Core). I have read that paraview runs
> > > > > > > serial, pvserver parallel, so I started the server by mpirun
> > > > > > > -np 4 pvserver and connected through X. Everything seems to
> > > > > > > work fine.
> > > > > > >
> > > > > > > But there are two strange behaviours I have noticed:
> > > > > > >
> > > > > > > 1. CPU usage on workstation is almost 100% on three of the
> > > > > > > four cores although no client is connected (when I type
> > > > > > > mpirun -np 3 pvserver there are 2 out of 4 running at 100%;
> > > > > > > with -np 2 only 1). I have noticed this using MPICH2 and
> > > > > > > OpenMPI.
> > > > > > >
> > > > > > > 2. When I now start a client process the server uses only
> > > > > > > one core (-np 4/3/2/1). Why?
> > > > > > >
> > > > > > > Does MPI not work on multicore systems as on multiprocessor
> > > > > > > systems or is this a Paraview issue?
> > > > > > >
> > > > > > > Thanks,
> > > > > > > Guido
> > > > > > > _______________________________________________
> > > > > > > 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
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > >
> > > > > > > **** Kenneth Moreland
> > > > > > > *** Sandia National Laboratories
> > > > > > > ***********
> > > > > > > *** *** *** email: kmorel at sandia.gov
> > > > > > > ** *** ** phone: (505) 844-8919
> > > > > > > *** web:
> > > > > > > http://www.cs.unm.edu/~kmorel<http://www.cs.unm.edu/%7Ekmorel>
> <http://www.cs.unm.edu/%7Ekmorel>
> > > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > >
> > > > > > **** Kenneth Moreland
> > > > > > *** Sandia National Laboratories
> > > > > > ***********
> > > > > > *** *** *** email: kmorel at sandia.gov
> > > > > > ** *** ** phone: (505) 844-8919
> > > > > > *** web:
> > > > > > http://www.cs.unm.edu/~kmorel <http://www.cs.unm.edu/%7Ekmorel><
> http://www.cs.unm.edu/%7Ekmorel>
> > > > > >
> > > > >
> > > > >
> > > > >
> > > > >
> > > > > **** Kenneth Moreland
> > > > > *** Sandia National Laboratories
> > > > > ***********
> > > > > *** *** *** email: kmorel at sandia.gov
> > > > > ** *** ** phone: (505) 844-8919
> > > > > *** web:
> > > > > http://www.cs.unm.edu/~kmorel <http://www.cs.unm.edu/%7Ekmorel><
> http://www.cs.unm.edu/%7Ekmorel>
> > > > >
> > > > _______________________________________________
> > > > 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
> > > >
> > _______________________________________________
> > 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
> >
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.paraview.org/pipermail/paraview/attachments/20100520/400aef1f/attachment.htm>
More information about the ParaView
mailing list