<html>
  <head>
    <meta content="text/html; charset=ISO-8859-1"
      http-equiv="Content-Type">
  </head>
  <body bgcolor="#FFFFFF" text="#000000">
    On 4/02/2012 5:20 AM, Christoph Klein wrote:
    <blockquote
cite="mid:CAADfth3EYiQmW4-R9yFhF0rju8OkkRfK0jpepN4MU2f_j0s72Q@mail.gmail.com"
      type="cite">
      <div>
        <div><span style="white-space: pre-wrap;">
            <div>
              <div><font face="Verdana">Hi all,</font></div>
              <div><font face="Verdana"><br>
                </font></div>
              <div><font face="Verdana">I am running a water/surfactant
                  system with just under 100000 atoms using MPI on a
                  local cluster and not getting the scaling I was hoping
                  for. The cluster consists of 8 core xeon nodes and I'm
                  running gromacs 4.5 with mpich2-gnu. I've tried
                  running a few benchmarks using 100ps runs and get the
                  following results:</font></div>
              <div><font face="Verdana"><br>
                </font></div>
              <div><b style="font-family:Verdana"><span
style="font-size:13px;font-weight:normal;text-decoration:none;vertical-align:baseline">Threads:
                    8 &nbsp;&nbsp;16 &nbsp;&nbsp;24 &nbsp;&nbsp;32 &nbsp;&nbsp;40 &nbsp;&nbsp;48 &nbsp;&nbsp;56 &nbsp;&nbsp;64</span></b>
              </div>
              <div><b style="font-family:Verdana"><span
                    style="font-weight:normal;text-decoration:none;vertical-align:baseline">hr/n</span><span
style="font-size:13px;font-weight:normal;text-decoration:none;vertical-align:baseline">s:
                    &nbsp;&nbsp;&nbsp;15 18 &nbsp;&nbsp;53 &nbsp;&nbsp;54 &nbsp;&nbsp;76 &nbsp;&nbsp;117 &nbsp;98 &nbsp;&nbsp;50 </span></b></div>
            </div>
          </span></div>
      </div>
    </blockquote>
    <br>
    Are you sure you have the right performance number (and not ns/day
    or something)!<br>
    <br>
    <blockquote
cite="mid:CAADfth3EYiQmW4-R9yFhF0rju8OkkRfK0jpepN4MU2f_j0s72Q@mail.gmail.com"
      type="cite">
      <div>
        <div><span style="white-space: pre-wrap;">
            <div><b style="font-family:Verdana"><span
style="font-size:13px;font-weight:normal;text-decoration:none;vertical-align:baseline"><br>
                </span></b></div>
            <div><b style="font-family:Verdana"><span
style="font-size:13px;font-weight:normal;text-decoration:none;vertical-align:baseline">Each
                  set of 8 threads is being sent to one node and the 8
                  threaded run was performed without MPI. I have tried
                  changing the -npme settings for all permissible values
                  on runs with 16 threads. In every instance the results
                  were worse than if I didn't specify anything.</span></b></div>
            <div><b style="font-family:Verdana"><span
style="font-size:13px;font-weight:normal;text-decoration:none;vertical-align:baseline"><br>
                </span></b></div>
            <div><b style="font-family:Verdana"><span
style="font-size:13px;font-weight:normal;text-decoration:none;vertical-align:baseline">The
                  fact that I am getting negative scaling leads me to
                  believe that something is wrong with my set up. Any
                  tips on what I could try?</span></b></div>
          </span></div>
      </div>
    </blockquote>
    <br>
    The simplest explanation is that your network (or MPI settings for
    it) is not up to the job. Very low latency is required. Gigabit
    ethernet is not good enough.<br>
    <br>
    You could try installing OpenMPI, also.<br>
    <br>
    Mark<br>
  </body>
</html>