Page Comparison

...

• David Dotson

• Pavan Behara

• Joshua Horton

• Trevor Gokey

Goals

• New advancements

  • Compute expansion rework - considering how to improve it

• New submissions

  • PEPCONF compute expansion submission failureTorsiondrive benchmark update

  • Genentech set2 - molcules that have more than three rotors fragment

• Upcoming infrastructure improvements

  • STANDARDSv3 submission machinery in QCSubmit

  • STANDARDSv3 submission machinery in qca-dataset-submission

  • Multiple PR templates

    • New submission

    • Compute expansion

    • Infrastructure modification

• Upcoming science support

• Larger advances

  • Automated FF coverage gap identification, torsion prioritization, submission generation

  • Benchmarking (dashboard, etc.)

Discussion topics

...

Item	Presenter	Notes
geomeTRIC memory limits	Trevor	geomeTRIC has no memory constraint machinery TG: decided to let a molecule run indefinitely, found that geomeTRIC had held on to more than 1000 B-matrices made Lee-Ping aware; suggested these don’t need to be held on to [proposal] profile what geomeTRIC is doing, start with B-matrices limit for storing TG: will create an issue on geomeTRIC detailing the observations; DD will pursue solution
High-mem, low-mem split	Trevor	May make sense to split out high-mem, low-mem jobs DD: we could do this with compute tags; on my list is to make this easy to declare from GitHub
Torsiondrive priority resets	Trevor	Need to make it possible to change priority, compute tag for services TG: if we’re going to have many compute tags, need a nice way to query what exists, how many tasks on each
Torsiondrive submission broken in QCSubmit with threading!=1	Josh	In an awkward place for QCSubmit; supporting benchmarking work, but also need to switch to using openff-toolkit DD: should be good to leave the 0.2.x as it is for benchmarking, not anticipating changes needed since most partners not using a server approach use 0.3.x as the openff-toolkit basis, breaking change JH: also planning to migrate packaging to conda-forge for 0.3.x would like to get STANDARDSv3 machinery in there as well if possible On the threading for torsiondrives, not clear what the problem is at the moment TG: when you have a shared self, a dictionary is being modified by the function called; this is shared state that can be unpredictable since it amounts to a race condition TG: we’re using threads because it’s low-memory, but could sidestep this easily with switching to a process pool DD: probably makes sense to switch to processes since these submissions do a good deal of client-side processing TG: will need to be able to set processes used on submission, since memory usage can be high DD: [committed] will pursue switch to process pool for submission
Compute expansion rework	Josh	Would like feedback on proposal; definitely want to improve the picture there, though we have time
Optimization and Torsiondrive starter notebooks	Trevor	TG: been working on some candidate notebooks; finding issues in QCSubmit and raising them, Josh largely addressing waiting for fixes before PRing notebooks
PR templates	David	Fairly easy task, will be adding multiple PR templates to improve submitter experience
New submissions	David	DD: reviewed torsiondrive benchmarks, just need an input file

Action items

Decisions