GlueX Software Meeting, March 16, 2021
GlueX Software Meeting
Tuesday, March 16, 2021
3:00 pm EDT
BlueJeans: 968 592 007
- Review of Minutes from the Last Software Meeting (all)
- Minutes from the Last HDGeant4 Meeting (all)
- Report from the SciComp Meeting (Mark)
- The Pull-Request Tests System (Sean)
- Software Testing Discussion (all)
- Review of recent issues and pull requests:
- Review of recent discussion on the GlueX Software Help List (all)
- Action Item Review (all)
Present: Alexander Austregesilo, Thomas Britton, Mark Dalton, Sean Dobbs, Mark Ito (chair), Igal Jaegle, David Lawrence, Simon Taylor, Nilanga Wickramaarachchi, Beni Zihlmann
Review of Minutes from the Last Software Meeting
We went over the minutes from March 2nd. We talked about the new scheme to assign photon energies to tagger hits.
- Sean volunteered to look at patching the halld_recon code used in previous reconstruction launches to reflect the new scheme.
- 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.]
- 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.
- Mark has started meeting regularly with Dmitry Romanov on CCDB 2.0.
Minutes from the Last HDGeant4 Meeting
We reviewed the minutes from the HDGeant4 Meeting on March 9 without significant comment.
Report from the SciComp Meeting
Mark presented slides from Bryan Hess, shown at the SciComp meeting on March 4th. Bryan had covered a lot of topics:
- Network configuration changes
- 100 Gb upgrade for our ESNet connection
- Lustre issues
- New Work Disk server
- LT08 Tape Issue
- Farm node expansion
- OSG jobs running at JLab
- ifarm slowness problem
- Miscellaneous Notes
- New SciComp computing status pages
- Swif2 call for testers
- GitHub contract changes
- Upgrade path from CentOS 7
Please see his slides (linked above) for the details.
The Pull-Request Tests System
Sean described the system we run for all new pull requests for halld_recon and halld_sim. Both build and function tests are done. See his slides for the details.
Mark asked about whether we want to look at Travis CI, 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...
Software Testing Discussion
The discussion never got of the ground due to network problems that put halldweb out of touch. We will try again next time.
GPU-Equipped Farm Nodes
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.
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.
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.
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.
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.
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.
Action Item Review
- Make sure that the automatic tests of HDGeant4 pull requests have been fully implemented. (Mark I., Sean)
- Finish conversion of halld_recon to use JANA2. (Nathan)
- Release CCDB 2.0 (Dmitry, Mark I.)