[Paraview] [EXTERNAL] ParaView Llvm on Supercomputer

[Scott, W Alan]

Noah,
I am not going to comment on implementing LLVM/Mesa – compile and parallel run issues are a pain and generally over my head.  A few thoughts.


·         If you are not using MPI, you are only using 1 core of your node.  ParaView is not multithreaded.  Even more than having more nodes, you need to get the cores you have available being used.  Then, if you need more speed again, obviously add nodes.

·         If you don’t need the memory, and have the data available, do your rendering on a workstation.  That should have hardware rendering, which will be much faster.

From: ParaView [mailto:paraview-bounces at paraview.org] On Behalf Of Noah Taylor
Sent: Friday, August 07, 2015 1:17 PM
To: ParaView
Subject: [EXTERNAL] [Paraview] ParaView Llvm on Supercomputer

Hi All,

I have worked with one of the IT persons to finally be able to use the capabilities of ParaView to capture a high resolution image of each time step at the end of my model run on a supercomputer. The problem that I am seeing is that Llvm is taking quite some time per rendered image. I currently only have 1 node with 8 cores that I use to my disposal. I run the script using pvbatch and the xmf read to get like 10 images in 10 minutes.

I am asking my IT guy to try and redo the compile so that it enables MPI so I can take advantage of multiple cores. Has any one else tried using Llvm with pvbatch to save out an image per time step? I saw the graph here - http://www.paraview.org/Wiki/ParaView/ParaView_And_Mesa_3D
- that talks about having 8 - 10 cores seems to be efficient.

Would it make sense to span Llvm across mutlple nodes / would it be quicker? Thanks for any input - I am just a little confused on how pvbatch would work between multiple nodes / what command to use.

Thanks,
Noah
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://public.kitware.com/pipermail/paraview/attachments/20150807/b36985aa/attachment-0001.html>