2020-11-20 QCA Submission Meeting notes

PCM-based implicit solvent

Simon

• PCM appears to be working on the COH submisson

• JH: Also first dataset storing wavefunctions/eigenvalues, so another first

• SB: storage and retrieval working just fine!

• DD: would be worth showing this off at next show-and-tell; I’ll find out from Jeff

Submissions

• SB: COH is about 50% complete

  • don’t have error cycling in place for basic DataSets yet; will get today

• PB: genentech optimization; working on first submission. Only 20% of dataset would be submitted in this first run

  • is this acceptable?

  • Yes, we’ll proceed with the smaller, 127 molecule (20%) subset for the first submission

  • DD: feel free to reach out to me when desired; we’ll re-roll the PR off of master (DD messed up long-lived branches with squash merges)

• PB: protomers/tautomers

  • JH: fewer tasks than there are conformers; due to QCF index not being case sensitive, and some of the SMILES clash when reduced to lowercase

    • Do we have another solution? Do we drop the use of SMILES for the index?

    • TG: For torsiondrives, this is still useful.

    • JH: Still want to be able to group molecules that are just peer conformers

  • JH: change how we index molecules, just do molecule-0, conformer-0, basically avoid SMILES for OptimizationDatasets, Basic DataSets; keep SMILES as index on TorsionDrives

  • TG: May still run into issues on this with TorsionDrives, but like this because we tag the driven torsion

  • DD+BP: Could also go with removing the lowercase-casting on indices; would be almost a trivial change, and non-destructive for database access (we’ll pursue this)

    • Issue raised:

• DD: PEPCONF

  • We’re getting some user pressure; why is it proceeding slowly?

  • Decide on a rebalancing of priorities for datasets:

    • reduce priority to low for some optimization sets

  • TG: Many of these molecules will take a lot of memory > 50GiB

  • DD: Perhaps time to scale up all our nodes to a minimum amount of memory for QM jobs

    • Do we know if there are ways to reduce the memory usage of Optimizations?

    • BP: Psi4 can write to disk if needed when memory gets constrained

    • DD: I will reduce the memory offered to the manager to below the constraints given to each worker; may trigger writing to local storage

      • also increase the total memory of each replica to 64GiB

      • Could also scale the CPUs to 32, perhaps even 64

  • We’ll increase the priority of PEPCONF to high

  • TG: will reduce number of workers deployed, see if this reduces pre-emption frequency

• Phenyl Dataset - will start to starve others

  • DD: I’ll touch base with Jessica, find out timeline needs for Phenyl set

Strategies for user timelines, expectations

David

• JH: I think we can be faster in merging datasets now, especially with STANDARDS coming into place

• DD: we’re already defaulting to ‘high’ priority for fitting datasets, more discretionary for others

• JH: Some of the datasets were from PI pressure to get things running; could be re-tagged to ‘low’ priority

• DD: compute tags are an avenue for controlling flows, but dangerous if we park tasks in a compute tag for which we have no managers

Dataset index

Josh

• Probably good to merge; can’t find the script used to generate

• DD: we can merge and manually curate for now, add automation later

Error Cycling

David

• TG: Restarts of SCF convergence, optimization convergence appear to clear often enough, probably don’t want to exclude these

  • High memory for psi4 can be dealt with through better configuration of workers (setting memory available to less than memory allocated on the node)

• DD: We’ll close for now; can chew on more ways to utilize compute tags for routing, how we want to filter error cycling

Enforced C1 symmetry

Josh

• C1 symmetry is coming in Psi4, old datasets where we didn’t do this will still work

  • if method requires a specific symmetry, psi4 will set it itself