Skip to Content.
Sympa Menu

charm - [charm] net-smp vs multicore builds performance

charm AT lists.cs.illinois.edu

Subject: Charm++ parallel programming system

List archive

[charm] net-smp vs multicore builds performance


Chronological Thread 
  • From: Eduard Llamosí <llamosi.e AT husky.neu.edu>
  • To: charm AT cs.illinois.edu
  • Subject: [charm] net-smp vs multicore builds performance
  • Date: Wed, 20 Mar 2013 13:16:55 -0400
  • List-archive: <http://lists.cs.uiuc.edu/pipermail/charm/>
  • List-id: CHARM parallel programming system <charm.cs.uiuc.edu>

Greetings,

I've been collecting performance data from the multicore, net and net-smp builds of a synthetic charm++ test with an all-to-all communication pattern where actual computation time is negligible. The 3 different builds were run on a single quad core processor, with 3 worker chares, leaving 1 free core for the communication thread on the multicore and net-smp builds.

As I understand it from what I've read about Charm++ so far, the performance of multicore and net-smp builds on a single processor (no network usage at all) should be about the same, but it appears that the net-smp build consistently under performs relative to the multicore build, unless sent messages grow large enough in size, in which case the net-smp build seems to slightly outperform the multicore build.

There must be some overhead with the net-smp build that I'm not aware of, but I was unable to find detailed enough information of what happens behind the scenes on each case for me to understand this trend.

The following graph might better illustrate what I'm trying to describe: https://dl.dropbox.com/u/1002586/charmpp.png

If this is explained in detail somewhere, I would really appreciate being pointed to such document. If it's not, some quick comment of what might be occurring will be more than welcome.

Thank you for your assistance.



  • [charm] net-smp vs multicore builds performance, Eduard Llamosí, 03/20/2013

Archive powered by MHonArc 2.6.16.

Top of Page