<html>
  <head>

    <meta http-equiv="content-type" content="text/html; charset=UTF-8">
  </head>
  <body>
    <p>Folks,</p>
    <p>Please find the minutes <a moz-do-not-send="true"
href="https://halldweb.jlab.org/wiki/index.php/GlueX_Software_Meeting,_March_16,_2021#Minutes">here</a>
      and below.</p>
    <p>  -- Mark</p>
    <p>     ______________________________________</p>
    <p>
    </p>
    <div id="globalWrapper">
      <div id="column-content">
        <div id="content" class="mw-body" role="main">
          <h2 id="firstHeading" class="firstHeading" lang="en"><span
              dir="auto">GlueX Software Meeting, March 16, 2021, Minutes</span><span
              class="mw-headline" id="Minutes"></span></h2>
          <div id="bodyContent" class="mw-body-content">
            <div id="mw-content-text" dir="ltr" class="mw-content-ltr"
              lang="en">
              <p>Present: Alexander Austregesilo, Thomas Britton, Mark
                Dalton, Sean Dobbs, Mark Ito (chair), Igal Jaegle, David
                Lawrence, Simon Taylor, Nilanga Wickramaarachchi, Beni
                Zihlmann
              </p>
              <p>There is a <a rel="nofollow" class="external text"
                  href="https://bluejeans.com/s/y80vA_yXz9L/">recording
                  of this meeting</a>. Log into the <a rel="nofollow"
                  class="external text"
                  href="https://jlab.bluejeans.com">BlueJeans site</a>
                first to gain access (use your JLab credentials).
              </p>
              <h3><span class="mw-headline"
                  id="Review_of_Minutes_from_the_Last_Software_Meeting">Review
                  of Minutes from the Last Software Meeting</span></h3>
              <p>We went over the <a
href="https://halldweb.jlab.org/wiki/index.php/GlueX_Software_Meeting,_March_2,_2021#Minutes"
                  title="GlueX Software Meeting, March 2, 2021">minutes
                  from March 2nd</a>. We talked about the new scheme to
                assign photon energies to tagger hits.
              </p>
              <ul>
                <li> Sean volunteered to look at patching the
                  halld_recon code used in previous reconstruction
                  launches to reflect the new scheme.</li>
                <li> Sean suggested that we do some testing of the
                  changes in Richard's pull requests before merging. We
                  agreed to contact Drew Smith and ask if the changes
                  solve his issue. Also Sean agreed to look at doing
                  some tests. [Added in press: Alex did some testing,
                  looking at the new REST format. He then went ahead and
                  merged in those changes in halld_recon.]</li>
                <li> Mark mentioned the idea of having the system
                  stagger the start time of farm jobs at the SciComp
                  meeting last week. This appeared to be a new concept
                  to the gathered; it will have to be raised again if it
                  is a feature we really want.</li>
                <li> Mark has started meeting regularly with Dmitry
                  Romanov on CCDB 2.0.</li>
              </ul>
              <h3><span class="mw-headline"
                  id="Minutes_from_the_Last_HDGeant4_Meeting">Minutes
                  from the Last HDGeant4 Meeting</span></h3>
              <p>We reviewed the <a
href="https://halldweb.jlab.org/wiki/index.php/HDGeant4_Meeting,_March_9,_2021#Minutes"
                  title="HDGeant4 Meeting, March 9, 2021">minutes from
                  the HDGeant4 Meeting on March 9</a> without
                significant comment.
              </p>
              <h3><span class="mw-headline"
                  id="Report_from_the_SciComp_Meeting">Report from the
                  SciComp Meeting</span></h3>
              <p>Mark presented <a rel="nofollow" class="external text"
href="https://halldweb.jlab.org/talks/2021/scicomp-ops-2021-03-edit2.pdf">slides
                  from Bryan Hess</a>, shown at the SciComp meeting on
                March 4th. Bryan had covered a lot of topics:
              </p>
              <ul>
                <li> Network configuration changes</li>
                <li> 100 Gb upgrade for our ESNet connection</li>
                <li> Lustre issues</li>
                <li> New Work Disk server</li>
                <li> LT08 Tape Issue</li>
                <li> Farm node expansion</li>
                <li> OSG jobs running at JLab</li>
                <li> ifarm slowness problem</li>
                <li> Miscellaneous Notes
                  <ul>
                    <li> New SciComp computing status pages</li>
                    <li> Swif2 call for testers</li>
                    <li> GitHub contract changes</li>
                    <li> Upgrade path from CentOS 7</li>
                  </ul>
                </li>
              </ul>
              <p>Please see his slides (linked above) for the details.
              </p>
              <h3><span class="mw-headline"
                  id="The_Pull-Request_Tests_System">The Pull-Request
                  Tests System</span></h3>
              <p>Sean <a rel="nofollow" class="external text"
href="https://halldweb.jlab.org/wiki/images/9/9c/Sdobbs_offline_20210316.pdf">described
                  the system we run for all new pull requests</a> for
                halld_recon and halld_sim. Both build and function tests
                are done. See his slides for the details.
              </p>
              <p>Mark asked about whether we want to look at <a
                  rel="nofollow" class="external text"
                  href="https://travis-ci.org/">Travis CI</a>, which is
                integrated with GitHub. The consensus was that using our
                local resources to do the test fits well with the
                structure of our software and if it ain't broke...
              </p>
              <h3><span class="mw-headline"
                  id="Software_Testing_Discussion">Software Testing
                  Discussion</span></h3>
              <p>The discussion never got of the ground due to network
                problems that put halldweb out of touch. We will try
                again next time.
              </p>
              <h3><span class="mw-headline" id="GPU-Equipped_Farm_Nodes">GPU-Equipped
                  Farm Nodes</span></h3>
              <p>David told us that in the next round of purchases, in
                addition to spending $120 k on traditional CPUs, another
                $120 k will go to nodes with GPUs on-board. He asked if
                GlueX had plans to exploit this resource in the near
                future. He mentioned two areas of which he was already
                aware: machine learning based differentiation between
                photon showers and hadronic split-offs in the FCAL and
                amplitude analysis with AmpTools.
              </p>
              <p>Alex reported that he has already been using AmpTools
                GPU capabilities on some of the farm nodes to do his
                spin density matrix analysis and that is working well
                for him. He suspects that he is the only one an JLab
                doing this. He mentioned that it is not obvious how to
                use GPUs for REST production.
              </p>
              <p>David reported on a project he and others are working
                on to do charged particle tracking with a machine
                learning system. That is work in progress.
              </p>
              <p>Simon thought that to reconfigure the current tracking
                algorithm to use GPUs was challenging and is not
                something that can be done in the near term.
              </p>
              <p>Mark mentioned that he had heard of efforts at other
                labs to do tracking with GPUs. If machinery could be
                imported for use with our software, that might cut down
                on development time for us. He also remarked that track
                reconstruction is really the area where we would benefit
                the most from a reduction in processing time. </p>
              <p>Sean wondered if kinematic fitting might be something
                that could be configured to run on a GPU. That sounded
                like an interesting direction to explore.
              </p>
              <h3><span class="mw-headline" id="Action_Item_Review">Action
                  Item Review</span></h3>
              <ol>
                <li> Make sure that the automatic tests of HDGeant4 pull
                  requests have been fully implemented. (Mark I., Sean)</li>
                <li> Finish conversion of halld_recon to use JANA2.
                  (Nathan)</li>
                <li> Release CCDB 2.0 (Dmitry, Mark I.)</li>
              </ol>
            </div>
            <br>
          </div>
        </div>
      </div>
    </div>
  </body>
</html>