2021-11-05 QC meeting notes

Updates from MolSSI

Ben

• BP: how’s the server doing?

• JW: still seeing sporadic failures, or at least have been last couple months

• BP: have a dedicated(?) firewall that routes traffic to the QCArchive server; used for traffic inbound from outside VT

  • recently that firewall stopped forwarding traffic reliably

  • IT has not been helpful

  • currently traffic being routed via the backup server as a workaround

• JW: seeing better success than before, but understand this isn’t a permanent solution

• BP: Long-term options could be renting cloud compute (eg AWS) or building a server in the office.

• DD: Lots of advantages to AWS

• BP: Planning a QCFractal release+upgrade tomorrow

  • DD: Thanks for this. And thanks JHorton for opening qcelemental PR.

  • JH: You’re welcome. BP did a lot of the work on the PR.

  • BP: Conda env solve times have been pretty large too.

  • JH: Switching to mamba shaved off 2 hours from 5 hour total test time.

  • BP: Old managers should still work. Change should be backwards-compatible.

  • DD: It’s not hard for us to update workers.

  • BP: I don’t think it will be necessary, but it can’t hurt. I’ll let you know on Slack. Could pin QCEngine 0.20 and QCElemental 0.23(.1?)

    • steal pins from QCFractal

• BP – Is anyone using Parsl to manage qcf workers? They’re up for renewal and want to report usage stats.

  • (General) – Basically everyone uses “pool”

Compute status

David

• DD – All the worker jobs on Lilac consumed. Just spun up more QM workers on PRP.

• PB – I could use more XTB and ANI workers

  • DD – I can spin those up. JH, are you running these?

  • JH – Not ANI since they keep running out of memory. But I can run XTB.

  • DD – I will spin up some ANI workers on PRP

• DD – I recall there were issues with single point calcs that required high memory allocation and manual observation. PB, are those still being a problem?

  • PB – I think we’re in good shape; only about 6 jobs remaining

  • DD – will fire off more high-memory jobs on Lilac
    

New submissions

• CC – Dipeptide set – I’m waiting for QCSubmit #172 to be merged
  

  • DD – QCSubmit 172 depends on the new QCF release, should be cut around this weekend. Then we can push out new QCSubmit version.

  • CC – Understood. Once that release happens, I’ll ping you if there are further problems with submitting the dataset

• DD – Worked with Willa Wang before I was out last week on qca-dataset-submission #235. Tried to pull down results from dataset. Didn’t see wavefunctions in the results. Not sure what’s going on here.

  • JH – psi4#2242 seemed possibly related, but I haven’t looked closely.

  • DD – I looked at that, seemed to find something similar to what Simon found before (no wavefunction data attached)

    • Dataset is OptimizationDataset: OpenFF RESP Polarizability Optimizations v1.1

  • BP: possible e don’t make it into the QCEngine execution for the gradient

  • DD – I can chase this down – It seems like the store_wavefunction keyword isn’t making it all the way down the stack.

  • BP – We need to be careful that we don’t store the wavefunction for every result record. I did an analysis of what takes up space in the database, and a recent deposition of wavefunction info made the db size shoot up.

  • JW – refresh my memory: we were originally thinking let’s store grids, but that would have been way too large; now we’re thinking orbitals and eigenvalues, which are large but not as bad?

    • BP – scale as order N**2 with number of orbitals

  • BP – Could simplify(?) the process of not recording wavefunction for every step by doing a single point calc only at the end of an optimization.

  • DD +BP – (Some question about where the logic would need to live to record only the wavfunction for the final step)

  • DD – Best thing to do now would be to take the final structures of WWAng’s optimizations and submit single points for all of them

• PB – may submit another dataset soon; will be a single point dataset

• JW – saw that JH had opened a PR for turning an OpenFF topology into SMILES:

  • might introduce a regression fairly soon, with multi-component topology generation from multiple SMILES

  • JH – I do depend on this, so might need to have a discussion on implementation

ML dataset for OpenMM

Pavan

• PB – Had an issue with B93 functional where it didn’t have dispersion. DD and BP put in a fix for this and so I’m just waiting on a release. JH’s upcoming release should solve this.