Loading ...
Sorry, an error occurred while loading the content.

36Re: [genetic_programming] Beowulf Clusters

Expand Messages
  • Martin Pelikan
    Oct 1 12:30 PM
    • 0 Attachment
      Lee Spector wrote:
      > Mattias,
      > All of the CPUs in our cluster are the same speed so I haven't had to deal
      > with this, but your idea of scaling population sizes with clock speed seems
      > like a good approach to me. Restarting a node and then importing
      > individuals from late in other runs seems like a bad idea (as you seem to
      > think as well).
      > -Lee

      We did some work on parallel machines, and the truth is that the different
      populations can start being apart even by 10 generations or more, even if the
      processors are of the same speed. Things like caching, multitasking, etc may
      affect the speed significantly. I do not see this as a big problem though.

      But let's say we want synchronize the processing of the sub-populations so that
      they are processed at the same speed, and we have very different processors. One
      of the things to consider then might be to use some of the programming models
      that allow for *load balancing*. One of these is Charm++ which allows you to
      create more sub-populations than you have processors (say 4 per processor),
      monitors the load on different processors, idle time, and other performance, and
      reschedules the sub-population processes as required. The best processor could
      therefore process 8 sub-populations while the slowest could end up with 1 only.
      The communication between the processes can be done by sending messages and the
      system would handle the rest. More on charm++ can be found at

      I know there exist some MPI versions that also allow load balancing to take
      place, I don't really remember any at the moment.

      Anyway, I think the best thing to do would be use one of these languages, and
      take the advantage of their being able to solve the discussed problem without
      additional heuristics.

      all the best,

      Martin Pelikan
      Illinois Genetic Algorithms Laboratory
      University of Illinois at Urbana Champaign
      117 Transportation Building
      104 S. Mathews Avenue, Urbana, IL 61801
      tel: (217) 333-2346, fax: (217) 244-5705
    • Show all 19 messages in this topic