<html>
  <head>
    <meta content="text/html; charset=ISO-8859-1"
      http-equiv="Content-Type">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <div class="moz-cite-prefix">Hi Andy,<br>
      <br>
      data that small should be fairly fast, and nersc's global scratch
      shouldn't blink when 24 procs access file in read only mode. maybe
      PV is reading all the data on a single process(or worse all of
      them) then doing a redistribution behind the scenes?? That would
      certainly explain your results. either way good luck.<br>
      <br>
      Burlen<br>
      <br>
      On 02/07/2013 09:51 AM, Andy Bauer wrote:<br>
    </div>
    <blockquote
cite="mid:CAMaOp+Hmmq0Qrv991uTguifkCCh3d4+zVPxhPr2__F=-GO4MUw@mail.gmail.com"
      type="cite">Hi Burlen,<br>
      <br>
      I got the data from a different user and that's where he put the
      data. I thought about copying it to $SCRATCH. I just thought
      though that it was really funky that trying to read in data that
      was under 4 MB for a single time step should be pretty fast for
      when I only have 24 processes asking for data. I was thinking that
      using the scratch space would just be covering up some deeper
      problem too in that I want to scale up to much more than 24
      processes. After all, any run that can't scale beyond 24 processes
      shouldn't be running on Hopper anyways!<br>
      <br>
      Andy<br>
      <br>
      <div class="gmail_quote">On Thu, Feb 7, 2013 at 11:57 AM, Burlen
        Loring <span dir="ltr">&lt;<a moz-do-not-send="true"
            href="mailto:bloring@lbl.gov" target="_blank">bloring@lbl.gov</a>&gt;</span>
        wrote:<br>
        <blockquote class="gmail_quote" style="margin:0 0 0
          .8ex;border-left:1px #ccc solid;padding-left:1ex">
          <div text="#000000" bgcolor="#FFFFFF">
            <div>Hi Andy,<br>
              <br>
              do you have a strong reason for using the global scratch
              fs? if not you may have better luck using hopper's
              dedicated lustre scratch. Spec quote &gt; 2x bandwidth[*].
              In reality I'm sure it depends on the number of user's
              hammering it at the time in question. may help to use
              lustre scratch while you're working on parallelization of
              the netcdf readers.<br>
              <br>
              Burlen<br>
              <br>
              * <a moz-do-not-send="true"
href="http://www.nersc.gov/users/computational-systems/hopper/file-storage-and-i-o/"
                target="_blank">http://www.nersc.gov/users/computational-systems/hopper/file-storage-and-i-o/</a>
              <div>
                <div class="h5"><br>
                  <br>
                  On 02/06/2013 03:35 PM, Andy Bauer wrote:<br>
                </div>
              </div>
            </div>
            <blockquote type="cite">
              <div>
                <div class="h5">Hi Ken,<br>
                  <br>
                  I think it's more than just a file contention issue.
                  On hopper@nersc I did set DVS_MAXNODES to 14 and that
                  helped out a lot. Without that set before I was able
                  to run with 480 processes accessing the same data file
                  (the <span>17*768*1152 with 324 time steps data set)
                    but with the "bad" one that was </span><span>768*1152

                    with 9855 time steps</span> I had problems with just
                  24 processes.<br>
                  <br>
                  I have some things which I want to try out but I think
                  you're right that using a parallel netcdf library
                  should help a lot, if it doesn't cause conflicts.<br>
                  <br>
                  Thanks,<br>
                  Andy<br>
                  <br>
                  <div class="gmail_quote"> On Wed, Feb 6, 2013 at 5:20
                    PM, Moreland, Kenneth <span dir="ltr">&lt;<a
                        moz-do-not-send="true"
                        href="mailto:kmorel@sandia.gov" target="_blank">kmorel@sandia.gov</a>&gt;</span>
                    wrote:<br>
                    <blockquote class="gmail_quote" style="margin:0 0 0
                      .8ex;border-left:1px #ccc solid;padding-left:1ex">
                      <div
style="font-size:14px;font-family:Calibri,sans-serif;word-wrap:break-word">
                        <div>
                          <div>
                            <div>This does not surprise me. &nbsp;The current
                              version of the netCDF reader only uses the
                              basic interface for accessing files, which
                              is basically a serial interface. &nbsp;You are
                              probably getting a lot of file request
                              contention.</div>
                            <div><br>
                            </div>
                            <div>At the time I wrote the netCDF reader,
                              parallel versions were just coming online.
                              &nbsp;I think it would be relatively
                              straightforward to update the reader to
                              use collective parallel calls from a
                              parallel netCDF library. &nbsp;Unfortunately, I
                              have lost track on the status of the
                              parallel netCDF library and file formats.
                              &nbsp;Last I looked, there were actually two
                              parallel netCDF libraries and formats.
                              &nbsp;One version directly added collective
                              parallel calls to the library. &nbsp;The other
                              changed the format to use hdf5 under the
                              covers and use the parallel calls therein.
                              &nbsp;These two libraries use different formats
                              for the files and I don't think are
                              compatible with each other. &nbsp;Also, it
                              might be the case for one or both
                              libraries that you cannot read the data in
                              parallel if it was not written in parallel
                              or written in an older version of netCDF.</div>
                            <div><br>
                            </div>
                            <div>-Ken</div>
                            <div>
                              <div><br>
                              </div>
                            </div>
                          </div>
                        </div>
                        <span>
                          <div style="border-right:medium
                            none;padding-right:0in;padding-left:0in;padding-top:3pt;text-align:left;font-size:11pt;border-bottom:medium
                            none;font-family:Calibri;border-top:#b5c4df
                            1pt
                            solid;padding-bottom:0in;border-left:medium
                            none"> <span style="font-weight:bold">From:
                            </span>Andy Bauer &lt;<a
                              moz-do-not-send="true"
                              href="mailto:andy.bauer@kitware.com"
                              target="_blank">andy.bauer@kitware.com</a>&gt;<br>
                            <span style="font-weight:bold">Date: </span>Wednesday,

                            February 6, 2013 10:38 AM<br>
                            <span style="font-weight:bold">To: </span>"<a
                              moz-do-not-send="true"
                              href="mailto:paraview@paraview.org"
                              target="_blank">paraview@paraview.org</a>"
                            &lt;<a moz-do-not-send="true"
                              href="mailto:paraview@paraview.org"
                              target="_blank">paraview@paraview.org</a>&gt;,

                            Kenneth Moreland &lt;<a
                              moz-do-not-send="true"
                              href="mailto:kmorel@sandia.gov"
                              target="_blank">kmorel@sandia.gov</a>&gt;<br>
                            <span style="font-weight:bold">Subject: </span>[EXTERNAL]

                            vtkNetCDFCFReader parallel performance<br>
                          </div>
                          <div>
                            <div>
                              <div><br>
                              </div>
                              <blockquote style="BORDER-LEFT:#b5c4df 5
                                solid;PADDING:0 0 0 5;MARGIN:0 0 0 5">
                                <div>
                                  <div>Hi Ken,<br>
                                    <br>
                                    I'm having some performance issues
                                    with a fairly large NetCDF file
                                    using the vtkNetCDFCFReader. The
                                    dimensions of it are 768 lat, 1152
                                    lon and 9855 time steps (no
                                    elevation dimension). It has one
                                    float variable with these dimensions
                                    -- pr(time, lat, lon). This results
                                    in a file around 33 GB. I'm running
                                    on hopper and for small amounts of
                                    processes (at most 24 which is the
                                    number of cores per node) and the
                                    run time seems to increase
                                    dramatically as I add more
                                    processes. The tests I did read in
                                    the first 2 time steps and did
                                    nothing else. The results are below
                                    but weren't done too rigorously:<br>
                                    <br>
                                    numprocs -- time<br>
                                    1&nbsp; -- 1:22<br>
                                    2 -- 1:52<br>
                                    4 -- 7:52<br>
                                    8 -- 5:34<br>
                                    16 -- 10:46<br>
                                    22 -- 10:37<br>
                                    24 -- didn't complete on hopper's
                                    "regular" node with 32 GB of memory
                                    but I was able to run it in a
                                    reasonable amount of time on
                                    hopper's big memory nodes with 64 GB
                                    of memory.<br>
                                    <br>
                                    I have the data in a reasonable
                                    place on hopper. I'm still playing
                                    around with settings (things get a
                                    bit better if I set DVS_MAXNODES --
                                    <a moz-do-not-send="true"
href="http://www.nersc.gov/users/computational-systems/hopper/performance-and-optimization/hopperdvs/"
                                      target="_blank">
http://www.nersc.gov/users/computational-systems/hopper/performance-and-optimization/hopperdvs/</a>)
                                    but this seems a bit weird as I'm
                                    not having any problems like this on
                                    a data set that has spatial
                                    dimensions of 17*768*1152 with 324
                                    time steps.<br>
                                    <br>
                                    Any quick thoughts on this? I'm
                                    still investigating but was hoping
                                    you could point out if I'm doing
                                    anything stupid.<br>
                                    <br>
                                    Thanks,<br>
                                    Andy<br>
                                    <br>
                                    <br>
                                  </div>
                                </div>
                              </blockquote>
                            </div>
                          </div>
                        </span> </div>
                    </blockquote>
                  </div>
                  <br>
                  <br>
                  <fieldset></fieldset>
                  <br>
                </div>
              </div>
              <pre>_______________________________________________
Powered by <a moz-do-not-send="true" href="http://www.kitware.com" target="_blank">www.kitware.com</a>

Visit other Kitware open-source projects at <a moz-do-not-send="true" href="http://www.kitware.com/opensource/opensource.html" target="_blank">http://www.kitware.com/opensource/opensource.html</a>

Please keep messages on-topic and check the ParaView Wiki at: <a moz-do-not-send="true" href="http://paraview.org/Wiki/ParaView" target="_blank">http://paraview.org/Wiki/ParaView</a>

Follow this link to subscribe/unsubscribe:
<a moz-do-not-send="true" href="http://www.paraview.org/mailman/listinfo/paraview" target="_blank">http://www.paraview.org/mailman/listinfo/paraview</a>
</pre>
            </blockquote>
            <br>
          </div>
        </blockquote>
      </div>
      <br>
    </blockquote>
    <br>
  </body>
</html>