2022-05-24 QC meeting notes

Updates from MolSSI

• BP: Data migration last week, working on speeding up datasets access this week, documentation, tests, etc., will reach out to users for testing.

• BP – Making a script to compare pre- and post-migration database as a sanity check. This should be done today or tomorrow.

  • DD – BP, want to do a working session this week? Or is this too early?

  • BP – I should be able to do one on Friday. I don’t think you do single point opts very often, right?

    • DD – We do a good number of single-points.

    • BP – Single points do have some issues on the staging server. Stuff related to composition planner and D3 setting.

    • DD – For QCSubmit we have the “submission side” and the “results side” - which one was this?

    • BP – “Results side”.

    • DD – Let’s work together on this on Friday. What was up with the composition planner?

    • BP – It was a really bad design, this won’t be supported in the next release. It did things like split up complex evaluations into two evaluations, which would work in some use cases but wouldn’t work in others. But this can be handled in psi4 now. The advantage of handling it on the QCF level was that we could deduplicate PARTS of an evaluation if the same mol was submitted with multiple partially-overlapping methods, but this didn’t end up being worth it performance-wise and adds a lot of complexity.

  • PB – So the new version will or wont split the calculations?

    • BP – The new version WON’T, but you can always submit them yourself as two separate evaluations and handle the recombination yourself.

Infrastructure needs/advances

• Psi4 v1.6 released last week - need to be tested

  • COSX-Chain of spheres method for exchange calculation may speed up SCF runs (completely in-core, SCF_TYPE=COSX).

  • “ADIIS/EDIIS for RHF/UHF. Now the default.”

  • Newer libint2 is the new integrals library (under-the-hood detail but may improve some runs).

  • libecpint is supported now if anyone is gonna use ECP (effective core potential, instead of all-electron basis) basis sets in their work, may speed up some gradient calculations where analytical derivatives were not present in previous versions

Throughput status

SPICE sets: Around 40K jobs last week

• SPICE PubChem Set 2 Single Points Dataset v1.2: 19 left (16 errored with scf convergence failures, 3 incomplete)

• SPICE PubChem Set 3 Single Points Dataset v1.2: 31 left (29 consistent errors, 2 incomplete)

• SPICE PubChem Set 4 Single Points Dataset v1.2: 114606 from 70787, 7K more remaining.

• SPICE PubChem Set 5 Single Points Dataset v1.2: 123135 calcs waiting in queue.

• PB – BP, could you run script to resubmit all stale jobs?

  • BP – Yes, I’ll run that.

• DD – Are we getting sufficient throughput PB?

  • PB – No, the openff user on UCI HPC isn’t enough to get this throughput. So I’m submitting managers from my user account.

  • JW – I think it’s OK if the SPICE datasets go slow, I wouldn’t set a minimum throughpiut there and I’m happy tot alk to Eastman about this if needed. But if Chapin’s next set goes slow then we can scramble to find more compute.

User questions/issues

• QCEngine issues

  • JW: QCEngine uses CMILES IIRC, so connectivity information in the molecule isn’t considered.

  • DD: In QCEngine the user’s molecule is going through Openmm harness, and passing through RDKitHarness since no CMILES is present.

  • JW: I had a chat with Abi-Mansour before on a similar issue. I will follow up on that and I will dig into the rdkit harness again and check whether formal charge is handled properly.

  •  

Science support needs