2020-10-20 QCFractal Meeting notes

Updates from MolSSI

• BP – Expecting some brief downtime, not worth announcing/halting managers. Will just be a few seconds.

• BP – May restart servers periodically to attach profilers. Will probably be on the order of hours. Managers may not be able to pull new tasks.

  • JW – Could you post a @here message on the openff slack (#qcfractal-compute) if this will take more than an hour? We shouldn’t be shutting down managers but it’ll be good to have a heads up.

• BP – Searching for some tricky bug that’s showing up in tests, something with torsiondrives.

  • JW – Would be best to contact LPW with questions.

  •  

Queue/Manager status

• DD –

• DD – Just moved some datasets with incompletes into error cycling – Recent updates or manager shapes are making these pass again.

• DD – 4 datasets that are being processed on lots of compute. Working on preparing submissions.

  • (General) – Protein torsiondrives are still far from complete

User questions

• TG – Update on stale jobs?

  • BP – Nothing immediate. Looking at pushing at update at the end of the month.

  • TG – We send everything in msgpack, should be binary/compressed. Big job that was sent was 25MB. Should have just been one task.

  • BP – I saw the payload as ~200MB. Seems to have had multiple tasks in it.

  • TG – For background, if one job goes stale, the entire manager seems to shut down.

• DD – TG tried out some ANI2x torsiondrives (which failed to converge on QCF) locally. He could increase maxiter and get some to complete. Raising maxiter is hard, and would require making a new entry.

  • JW – What is the target for ANI2x completion? Will we adjust maxiter until we reach 100%?

  • TG – B3LYP finishes 100% of these jobs. So there’s nothing wrong with the molecules.

  • DD – I’m running maxiter=1000 on my computer, and it hasn’t finished.

  • DD – Maybe we could relax tolerance?

  • TG – I’ve been seeing substantial changes in geometry/energy at every step.

  • JW – Do we resubmit the entire dataset with maxiter=1000, or do we treat this as an experiment and only resubmit a few of the “trickiest” cases with maxiter=1000

  • DD – I’ll talk to Dominic Rufa and give him the code to run these locally, and ask what he wants to do.

  • DD will contact Dominic and John in the #qcfractal channel to discuss next steps.

• MT – Basis set exchange is now on conda forge Huge thanks to BP.

•