2021-09-10 QCA Submission meeting notes

Updates from MolSSI

Ben

• Fixed issues with inconsistent state

  • 10s of thousands of jobs in task queue associated with complete records

  • if you resubmit a record that already exists, even if COMPLETE, a task will get created

    • know where this hole is; can fix and produce new release, deploy to public QCArchive

• BP: removed all duplicate tasks

  • DD: will turn managers back on to compute industry datasets, put them on their own compute tag and observe if backward-forward behavior persists

User issues, new submissions

• PB: #220 - need new qcsubmit

  • JH: just have one PR on qcsubmit blocking release; working on this

  • PB: have one compute spec on this submission DF-CCSD(T)/CBS that will take up to 150 GiB of memory for 16 heavy atoms

  • PB: typically also use 48 cores

• JH: working on #223, blocked by validation issues as well

  • going to add some ANI specs, will need some ANI workers

    • will still use ANI2x we had issues with, but relaxing convergence criteria

    • DD: until we have a release from Adrian, unfortunately can’t use the latest fixes to this in prod

• JH: ML stuff, adding HDF5 support for QCSubmit

  • instead of a ton of SDFs, can use one file

  • JW: what are the contents

  • JH: conformers and mapped SMILES

  • JW: is this file going to contain the same content as the other files, or is there something fundamentally different here?

    • one thing that makes SDF safer is that readers and writers are not something we’re defining

    • JH: there’s a lot of repeated info in the SDF

      • also want to pave the way for multimolecule support, dimers, etc.

    • JW: good point

    • JH: understand concerns on future variability; would like to get a spec down as much as possible

  • JH: any feedback anyone has on this issue ( ) appreciated

• DD: concerned about collection size; will run into same issue as before

  • SB+JH: not clear if it’s a single collection with a million conformers, or spread across several collections, or multiple million conformer collections

  • BP: the metadata object for a collection gets very big as more and more objects are involved (molecules, specs), so this becomes an issue in the way collections are currently implemented

    • is getting fixed in the next branch

  • SB: can see this taking another month for John and Peter to resolve; what is the timeline for next branch deployment?

    • BP: end of the year earliest? Can’t make a guarantee there, though

  • JW: DD, would you be willing to jump onto next OpenMM call to lay out constraints?

    • DD: yes

• JH: is a test submission still in play?

  • SB: yeah, can push for this, also as a way to prove the core idea of the dataset works before we push through a massive set

• SB: Chapin’s dataset; what’s the status?

  • DD: worked with him to set up manager on UCSD resources; can switch on and off at will; waiting for word on new submission status

  • SB: think there may still be some ambiguity on what data, how it will be different from the Cerutti sets; will coordinate with Chapin and see where we’re at

Science support needs

Infrastructure needs

• BP: with this fix for the submission deduplication, can also include the fix for the slow queries encountered recently

  • this is adding indices to a single table; remove combined index, add a bunch of single indices to columns

  • is a DB migration in practice

  • will require more memory on the server; shouldn’t be an issue

• PB: question for Josh

  • using one spec in new submission where method is a “ + “ joined method, no basis (https://psicode.org/psi4manual/master/cbs.html )

  • JW: short form for this method looks like it might present a discoverability issue; perhaps use a long form instead?

  • PB: should I leave it like this, or use long form?

    • JH: I think long form is supported through keywords; qcsubmit won’t like None for basis in psi4