<html>
  <head>
    <meta http-equiv="Content-Type" content="text/html;
      charset=windows-1252">
  </head>
  <body>
    <p>p.s</p>
    <p>sorry, I forgotted to say that we built the environment (both
      Darshan and benchmark) using IntelMPI release mpiifort script:<br>
    </p>
    <p> mpiifort -v<br>
      mpiifort for the Intel(R) MPI Library 2018 Update 4 for Linux*<br>
      Copyright(C) 2003-2018, Intel Corporation.  All rights reserved.</p>
    <p><br>
    </p>
    <p>regards</p>
    <p><br>
    </p>
    <p>Piero<br>
    </p>
    <div class="moz-cite-prefix">Il 12/02/2020 16:54, Piero LANUCARA ha
      scritto:<br>
    </div>
    <blockquote type="cite"
      cite="mid:e1b69b39-89a8-3826-7ae5-80414f31476b@cineca.it">
      <meta http-equiv="Content-Type" content="text/html;
        charset=windows-1252">
      <p>Hi Phil</p>
      <p><br>
      </p>
      <p>in attach</p>
      <p>dn stands for different names....</p>
      <p><br>
      </p>
      <p>cheers</p>
      <p>Piero<br>
      </p>
      <div class="moz-cite-prefix">Il 12/02/2020 16:13, Carns, Philip H.
        ha scritto:<br>
      </div>
      <blockquote type="cite"
cite="mid:BN8PR09MB3620FFD6FEEA11FFD44E2196F71B0@BN8PR09MB3620.namprd09.prod.outlook.com">
        <meta http-equiv="Content-Type" content="text/html;
          charset=windows-1252">
        <style type="text/css" style="display:none;"> P {margin-top:0;margin-bottom:0;} </style>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> Ah, great, thank you for the
          confirmation.</div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> In that case it looks like Darshan is
          instrumenting properly at run time, but I think Kevin is on
          the right track that Darshan's heuristics for calculating
          performance in post processing are getting confused for some
          reason.<br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> Probably GPFS is client-side caching
          aggressively in the single client case, but that wouldn't
          explain why the benchmark output reports a much different
          number than Darshan, though.  They should both perceive
          roughly the same performance; neither the benchmark itself nor
          Darshan know if caching is happening or not.</div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> It's hard to see where the performance
          heuristic went wrong from looking at the log, in large part
          because the app repeatedly opens a file with the same name
          (there is a clue to this in the OPEN counters; the same file
          name is opened 20 times):</div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <span># WARNING: POSIX_OPENS counter
            includes both POSIX_FILENOS and POSIX_DUPS counts<br>
          </span><span>POSIX 0 6563482044800691889 POSIX_OPENS 20
/gpfs/scratch/userinternal/planucar/benchio-master/shared-file/source/benchio_files/serial.dat
            /gpfs/scratch gpfs</span></div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> Every time the file is opened
          (regardless of whether it was unlinked in between or not),
          Darshan keeps adding counters to the same record, which are
          associated with that serial.dat file name.  So things like
          close() timestamps become nonsensical because Darshan records
          when the first close() starts and when the last one finishes:</div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <span>[carns@carns-x1-7g Downloads]$
            darshan-parser benchio_1202.darshan |grep CLOSE<br>
          </span>
          <div>POSIX 0 6563482044800691889 POSIX_F_CLOSE_START_TIMESTAMP
4.536853/gpfs/scratch/userinternal/planucar/benchio-master/shared-file/source/benchio_files/serial.dat
            /gpfs/scratch gpfs<br>
          </div>
          <div>POSIX 0 6563482044800691889 POSIX_F_CLOSE_END_TIMESTAMP
            43.041987
/gpfs/scratch/userinternal/planucar/benchio-master/shared-file/source/benchio_files/serial.dat
            /gpfs/scratch gpfs<br>
          </div>
          <span></span></div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> (This doesn't mean there was one close()
          that took ~40 seconds; in this case there were many close()
          calls and ~40 seconds elapsed between the start of the first
          one and completion of the last one).</div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> If it is possible for you to modify the
          benchmark (as an experiment) so that it chooses a new file
          name on each iteration, then I think it would probably
          disentangle the counters enough for us to tell what went
          wrong.<br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> <br>
        </div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> thanks,</div>
        <div style="font-family: Calibri, Arial, Helvetica, sans-serif;
          font-size: 12pt; color: rgb(0, 0, 0); background-color:
          rgb(255, 255, 255);"> -Phil<br>
        </div>
        <hr style="display:inline-block;width:98%" tabindex="-1">
        <div id="divRplyFwdMsg" dir="ltr"><font style="font-size:11pt"
            face="Calibri, sans-serif" color="#000000"><b>From:</b>
            Piero LANUCARA <a class="moz-txt-link-rfc2396E"
              href="mailto:p.lanucara@cineca.it" moz-do-not-send="true"><p.lanucara@cineca.it></a><br>
            <b>Sent:</b> Wednesday, February 12, 2020 9:47 AM<br>
            <b>To:</b> Carns, Philip H. <a
              class="moz-txt-link-rfc2396E"
              href="mailto:carns@mcs.anl.gov" moz-do-not-send="true"><carns@mcs.anl.gov></a>;
            Snyder, Shane <a class="moz-txt-link-rfc2396E"
              href="mailto:ssnyder@mcs.anl.gov" moz-do-not-send="true"><ssnyder@mcs.anl.gov></a>;
            Harms, Kevin <a class="moz-txt-link-rfc2396E"
              href="mailto:harms@alcf.anl.gov" moz-do-not-send="true"><harms@alcf.anl.gov></a><br>
            <b>Cc:</b> <a class="moz-txt-link-abbreviated"
              href="mailto:darshan-users@lists.mcs.anl.gov"
              moz-do-not-send="true">darshan-users@lists.mcs.anl.gov</a>
            <a class="moz-txt-link-rfc2396E"
              href="mailto:darshan-users@lists.mcs.anl.gov"
              moz-do-not-send="true"><darshan-users@lists.mcs.anl.gov></a><br>
            <b>Subject:</b> Re: [Darshan-users] Darshan & EPCC
            benchio different behaviour</font>
          <div> </div>
        </div>
        <div>
          <p>Hi Phil.....POSIX</p>
          <p>this is a well known benchmark ....you can easily verify
            it!</p>
          <p><br>
          </p>
          <p>by the way it's something like that</p>
          <p><br>
          </p>
          <p>! Serial write is unconditionally compiled<br>
            subroutine serialwrite(filename, iodata, n1, n2, n3,
            cartcomm)<br>
            <br>
              character*(*) :: filename<br>
            <br>
              integer :: n1, n2, n3<br>
              double precision, dimension(0:n1+1,0:n2+1,0:n3+1) ::
            iodata<br>
            <br>
              integer :: cartcomm, ierr, rank, size<br>
              integer, parameter :: iounit = 10<br>
            <br>
              integer :: i<br>
            <br>
              call MPI_Comm_size(cartcomm, size, ierr)<br>
              call MPI_Comm_rank(cartcomm, rank, ierr)<br>
            <br>
            !  Write same amount of data as the parallel write but do it
            all from rank 0<br>
            !  This is just to get a baseline figure for serial IO
            performance - note<br>
            !  that the contents of the file will be differnent from the
            parallel calls<br>
            <br>
              if (rank == 0) then<br>
            <br>
                 open(file=filename, unit=iounit, access='stream')<br>
            <br>
                 do i = 1, size<br>
                    write(unit=iounit) iodata(1:n1, 1:n2, 1:n3)<br>
                 end do<br>
            <br>
                 close(iounit)<br>
            <br>
              end if<br>
            <br>
            end subroutine serialwrite<br>
            <br>
          </p>
          <p><br>
          </p>
          <p>Piero<br>
          </p>
          <div class="x_moz-cite-prefix">Il 12/02/2020 14:13, Carns,
            Philip H. ha scritto:<br>
          </div>
          <blockquote type="cite">
            <style type="text/css" style="display:none">
<!--
p
        {margin-top:0;
        margin-bottom:0}
-->
</style>
            <div style="font-family:Calibri,Arial,Helvetica,sans-serif;
              font-size:12pt; color:rgb(0,0,0);
              background-color:rgb(255,255,255)"> Hi Piero,</div>
            <div style="font-family:Calibri,Arial,Helvetica,sans-serif;
              font-size:12pt; color:rgb(0,0,0);
              background-color:rgb(255,255,255)"> <br>
            </div>
            <div style="font-family:Calibri,Arial,Helvetica,sans-serif;
              font-size:12pt; color:rgb(0,0,0);
              background-color:rgb(255,255,255)"> In the serial case, is
              the rank that's doing I/O still using MPI-IO, or is it
              making calls directly to POSIX in that case?</div>
            <div style="font-family:Calibri,Arial,Helvetica,sans-serif;
              font-size:12pt; color:rgb(0,0,0);
              background-color:rgb(255,255,255)"> <br>
            </div>
            <div style="font-family:Calibri,Arial,Helvetica,sans-serif;
              font-size:12pt; color:rgb(0,0,0);
              background-color:rgb(255,255,255)"> The Darshan log for
              the serial case doesn't show any MPI-IO activity, but I'm
              not sure if that's accurate, or if it's an indication that
              we missed some instrumentation.</div>
            <div style="font-family:Calibri,Arial,Helvetica,sans-serif;
              font-size:12pt; color:rgb(0,0,0);
              background-color:rgb(255,255,255)"> <br>
            </div>
            <div style="font-family:Calibri,Arial,Helvetica,sans-serif;
              font-size:12pt; color:rgb(0,0,0);
              background-color:rgb(255,255,255)"> thanks,</div>
            <div style="font-family:Calibri,Arial,Helvetica,sans-serif;
              font-size:12pt; color:rgb(0,0,0);
              background-color:rgb(255,255,255)"> -Phil<br>
            </div>
            <hr tabindex="-1" style="display:inline-block; width:98%">
            <div id="x_divRplyFwdMsg" dir="ltr"><font
                style="font-size:11pt" face="Calibri, sans-serif"
                color="#000000"><b>From:</b> Darshan-users <a
                  class="x_moz-txt-link-rfc2396E"
                  href="mailto:darshan-users-bounces@lists.mcs.anl.gov"
                  moz-do-not-send="true">
                  <darshan-users-bounces@lists.mcs.anl.gov></a> on
                behalf of Piero LANUCARA <a
                  class="x_moz-txt-link-rfc2396E"
                  href="mailto:p.lanucara@cineca.it"
                  moz-do-not-send="true"> <p.lanucara@cineca.it></a><br>
                <b>Sent:</b> Wednesday, February 12, 2020 5:29 AM<br>
                <b>To:</b> Snyder, Shane <a
                  class="x_moz-txt-link-rfc2396E"
                  href="mailto:ssnyder@mcs.anl.gov"
                  moz-do-not-send="true"> <ssnyder@mcs.anl.gov></a>;
                Harms, Kevin <a class="x_moz-txt-link-rfc2396E"
                  href="mailto:harms@alcf.anl.gov"
                  moz-do-not-send="true"> <harms@alcf.anl.gov></a><br>
                <b>Cc:</b> <a class="x_moz-txt-link-abbreviated"
                  href="mailto:darshan-users@lists.mcs.anl.gov"
                  moz-do-not-send="true">
                  darshan-users@lists.mcs.anl.gov</a> <a
                  class="x_moz-txt-link-rfc2396E"
                  href="mailto:darshan-users@lists.mcs.anl.gov"
                  moz-do-not-send="true">
                  <darshan-users@lists.mcs.anl.gov></a><br>
                <b>Subject:</b> Re: [Darshan-users] Darshan & EPCC
                benchio different behaviour</font>
              <div> </div>
            </div>
            <div class="x_BodyFragment"><font size="2"><span
                  style="font-size:11pt">
                  <div class="x_PlainText">Hi Shane, Kevin<br>
                    <br>
                    thanks for the update.<br>
                    <br>
                    I attached a new upated files (log and pdf) to this
                    email.<br>
                    <br>
                    Also, the log from BENCHIO is attached.<br>
                    <br>
                    thanks again<br>
                    <br>
                    regards<br>
                    <br>
                    Piero<br>
                    <br>
                    <br>
                    Il 11/02/2020 20:15, Shane Snyder ha scritto:<br>
                    > Definitely looks like something strange is
                    happening when Darshan is <br>
                    > estimating the time spent in I/O operations (as
                    seen in the very first <br>
                    > figure, observed write time barely even
                    registers) in the serial case, <br>
                    > which it is ultimately used to provide the
                    performance estimate.<br>
                    ><br>
                    > If you could provide them, the raw Darshan logs
                    would be really <br>
                    > helpful. That should make it clear whether it's
                    an instrumentation <br>
                    > issue (i.e., under accounting for time spent in
                    I/O operations at <br>
                    > runtime) or if its an issue with the heuristics
                    in the PDF summary <br>
                    > tool you are using, as Kevin points out. If
                    it's the latter, having an <br>
                    > example log to test modifications to our
                    heuristics would be very <br>
                    > helpful to us.<br>
                    ><br>
                    > Thanks,<br>
                    > --Shane<br>
                    ><br>
                    > On 2/11/20 8:36 AM, Harms, Kevin wrote:<br>
                    >> Piero,<br>
                    >><br>
                    >>    the performance estimate is based on
                    heuristics, it's possible the <br>
                    >> 'serial' model is breaking some assumptions
                    about how the I/O is <br>
                    >> done. Is every rank opening the file, but
                    only rank 0 is doing actual <br>
                    >> I/O?<br>
                    >><br>
                    >>    If possible, you could provide the log
                    and we could check to see <br>
                    >> what the counters look like.<br>
                    >><br>
                    >> kevin<br>
                    >><br>
                    >> ________________________________________<br>
                    >> From: Piero LANUCARA <a
                      class="x_moz-txt-link-rfc2396E"
                      href="mailto:p.lanucara@cineca.it"
                      moz-do-not-send="true">
                      <p.lanucara@cineca.it></a><br>
                    >> Sent: Tuesday, February 11, 2020 2:28 AM<br>
                    >> To: Harms, Kevin<br>
                    >> Cc: <a class="x_moz-txt-link-abbreviated"
                      href="mailto:darshan-users@lists.mcs.anl.gov"
                      moz-do-not-send="true">
                      darshan-users@lists.mcs.anl.gov</a><br>
                    >> Subject: Re: [Darshan-users] Darshan &
                    EPCC benchio different behaviour<br>
                    >><br>
                    >> Hi Kevin<br>
                    >><br>
                    >> first of all thanks for the
                    investigation..I did some futher test and it<br>
                    >> seems like the issue may appear using
                    Fortran (MPI, mainly IntelMPI) <br>
                    >> codes.<br>
                    >><br>
                    >> Is this information useful?<br>
                    >><br>
                    >> regards<br>
                    >> Piero<br>
                    >> Il 07/02/2020 16:07, Harms, Kevin ha
                    scritto:<br>
                    >>> Piero,<br>
                    >>><br>
                    >>>     just to confirm, the serial case is
                    still running in parallel, <br>
                    >>> 36 processes, but the I/O is only from
                    rank 0?<br>
                    >>><br>
                    >>> kevin<br>
                    >>><br>
                    >>>
                    ________________________________________<br>
                    >>> From: Darshan-users <a
                      class="x_moz-txt-link-rfc2396E"
                      href="mailto:darshan-users-bounces@lists.mcs.anl.gov"
                      moz-do-not-send="true">
                      <darshan-users-bounces@lists.mcs.anl.gov></a>
                    on <br>
                    >>> behalf of Piero LANUCARA <a
                      class="x_moz-txt-link-rfc2396E"
                      href="mailto:p.lanucara@cineca.it"
                      moz-do-not-send="true">
                      <p.lanucara@cineca.it></a><br>
                    >>> Sent: Wednesday, February 5, 2020 4:56
                    AM<br>
                    >>> To: <a
                      class="x_moz-txt-link-abbreviated"
                      href="mailto:darshan-users@lists.mcs.anl.gov"
                      moz-do-not-send="true">
                      darshan-users@lists.mcs.anl.gov</a><br>
                    >>> Subject: Re: [Darshan-users] Darshan
                    & EPCC benchio different behaviour<br>
                    >>><br>
                    >>> p.s<br>
                    >>><br>
                    >>> to be more "verbose" I add to the
                    discussion:<br>
                    >>><br>
                    >>> Darshan output for the "serial" run
                    (serial.pdf)<br>
                    >>><br>
                    >>> Darshan output for the MPI-IO run
                    (mpiio.pdf)<br>
                    >>><br>
                    >>> benchio output for "serial" run
                    (serial.out)<br>
                    >>><br>
                    >>> benchio output for "MPI-IO" run
                    (mpi-io.out)<br>
                    >>><br>
                    >>> thanks<br>
                    >>><br>
                    >>> Piero<br>
                    >>><br>
                    >>> Il 04/02/2020 19:44, Piero LANUCARA ha
                    scritto:<br>
                    >>>> Dear all<br>
                    >>>><br>
                    >>>> I'm using Darshan to measure EPCC
                    benchio benchmark<br>
                    >>>> (<a
                      href="https://github.com/EPCCed/benchio"
                      moz-do-not-send="true">https://github.com/EPCCed/benchio</a>)
                    behaviour on a given x86 Tier1<br>
                    >>>> machine.<br>
                    >>>><br>
                    >>>> running two benchio tests (MPI-IO
                    and serial) a different behaviour<br>
                    >>>> appear<br>
                    >>>><br>
                    >>>> while Darhsan pdf log file is able
                    to recover the estimated time and<br>
                    >>>> bandwidth in the MPI-IO case, the
                    "serial" run is completely<br>
                    >>>> underestimated by Darshan (the time
                    and bandwidth are less/greater<br>
                    >>>> than benchio output).<br>
                    >>>><br>
                    >>>> Suggestions are welcomed<br>
                    >>>><br>
                    >>>> thanks<br>
                    >>>><br>
                    >>>> Piero<br>
                    >>>><br>
                    >>>>
                    _______________________________________________<br>
                    >>>> Darshan-users mailing list<br>
                    >>>> <a
                      class="x_moz-txt-link-abbreviated"
                      href="mailto:Darshan-users@lists.mcs.anl.gov"
                      moz-do-not-send="true">
                      Darshan-users@lists.mcs.anl.gov</a><br>
                    >>>> <a
                      href="https://lists.mcs.anl.gov/mailman/listinfo/darshan-users"
                      moz-do-not-send="true">https://lists.mcs.anl.gov/mailman/listinfo/darshan-users</a><br>
                    >>
                    _______________________________________________<br>
                    >> Darshan-users mailing list<br>
                    >> <a class="x_moz-txt-link-abbreviated"
                      href="mailto:Darshan-users@lists.mcs.anl.gov"
                      moz-do-not-send="true">
                      Darshan-users@lists.mcs.anl.gov</a><br>
                    >> <a
                      href="https://lists.mcs.anl.gov/mailman/listinfo/darshan-users"
                      moz-do-not-send="true">https://lists.mcs.anl.gov/mailman/listinfo/darshan-users</a><br>
                    ><br>
                    > _______________________________________________<br>
                    > Darshan-users mailing list<br>
                    > <a class="x_moz-txt-link-abbreviated"
                      href="mailto:Darshan-users@lists.mcs.anl.gov"
                      moz-do-not-send="true">
                      Darshan-users@lists.mcs.anl.gov</a><br>
                    > <a
                      href="https://lists.mcs.anl.gov/mailman/listinfo/darshan-users"
                      moz-do-not-send="true">https://lists.mcs.anl.gov/mailman/listinfo/darshan-users</a><br>
                  </div>
                </span></font></div>
          </blockquote>
        </div>
      </blockquote>
      <br>
      <fieldset class="mimeAttachmentHeader"></fieldset>
      <pre class="moz-quote-pre" wrap="">_______________________________________________
Darshan-users mailing list
<a class="moz-txt-link-abbreviated" href="mailto:Darshan-users@lists.mcs.anl.gov">Darshan-users@lists.mcs.anl.gov</a>
<a class="moz-txt-link-freetext" href="https://lists.mcs.anl.gov/mailman/listinfo/darshan-users">https://lists.mcs.anl.gov/mailman/listinfo/darshan-users</a>
</pre>
    </blockquote>
  </body>
</html>