10
$\begingroup$

What's the best way for parallelization in VASP for periodic systems? In the documentation, it mostly explains it for the local computer and molecule calculation. Apparently, these explanations don't help for periodic system calculations using supercomputer systems.

Let's say I want to do an MP2/GW/RPA calculation and I have access to 1000 cores (10 nodes each 100 cores). What would be the best way for me to make a good parallelization and have a fast result?

$\endgroup$
9
  • $\begingroup$ when it comes to parallelisation, the interconnection between the nodes/cores could be critical, can you please specify whether your 10 nodes are connected by specialised connection (e.g. Infiniband) or just LANs? $\endgroup$ – bzbzbz Sep 20 '20 at 11:23
  • $\begingroup$ Must be just LANs but not 100% sure. $\endgroup$ – Alfred Sep 20 '20 at 11:55
  • 1
    $\begingroup$ @bzbzbz it is very unlikely that the nodes are connected by ethernet. Most likely they are connected by Infiniband or Omnipath, and I think it's perfectly find to write your answer based on that assumption. $\endgroup$ – Nike Dattani Oct 14 '20 at 11:25
  • $\begingroup$ @NikeDattani You would hope that is true, but at my current university I have discovered that there are only 10 GBps links over standard ethernet between nodes. As you can imagine, I tend to only run 1 node jobs but in massive parallel. You may also want to see this question, it may be a duplicate. If it is not, then you may want to clarify why it isn't. mattermodeling.stackexchange.com/q/2039/697 $\endgroup$ – Tristan Maxson Oct 14 '20 at 14:01
  • $\begingroup$ @TristanMaxson can you edit your previous comment to contain the next one? The two together fit in one comment and I'd like to reduce the # of comments so as to avoid making unnecessary chat rooms (I think we both agree we don't need a chat room just for this!). Ok I see you're trying to address the first one to me and the second one to Alfred. I'd still combine the comments and just start the 2nd one with something like "Alfred:" $\endgroup$ – Nike Dattani Oct 14 '20 at 14:03