Page Comparison

Updates from MolSSI

• BP: No major updates, working on documentation. Server is going good.

  • DD – Last week I worked with Lorenzo to update the instructions for accessing the old industry datasets. This is using the test server, so that’s a vote of confidence!

  • On the QCSubmit side, we want to use some test fixtures in QCFractal. Is that allowed/recommended?

    • BP – Yes. The function signatures may change a little bit but the functionality will remain.

Infrastructure advances

• DD – Now have MPI cluster working at full capacity. It took me a while to figure out how to set up jobs well - network file system isn’t very fast and there isn’t a ton of scratch space. So now I copy conda env to local disk, which takes about 15 min but really speeds up execution. I’m also reducing the frequency at which I dispatch these jobs so that we don’t cause trouble in the shared NFS.

Throughput status

1. New OpenFF sets from Jessica:

   • OpenFF multiplicity correction optimization set v1.0 - 400 opts, complete.

   • OpenFF multiplicity correction torsion drive data v1.0 - 92/99 TDs done, remaining persistent errors. PB will check with JM about next steps for this dataset.

2. New OpenFF sets from Chapin:

   • OpenFF Protein Capped 1-mers 3-mers Optimization Dataset v1.0 - 753/759 opts complete.

   • OpenFF Protein Capped 3-mer Backbones v1.0 - 0/54 TDs complete. 19493 opts done.

3. SPICE sets: around 25K calcs last week

   • SPICE PubChem Set 4 Single Points Dataset v1.2: From 34 remaining to 26, most are persistent errors.

   • SPICE PubChem Set 5 Single Points Dataset v1.2: 80892 from 55874, around 42K more remaining.
     

Science support needs

• PB – Can we use intel MKL?

  • PB – For QC workers?

    • DD – I don’t see a legal problem with this. Will XTB work correctly with intel MKL?

      • PB – I think so. See

        Github link macro
        link https://github.com/grimme-lab/xtb/issues/401

      • JW – It’s noteworthy that here even running with intel MKL still scales really poorly - Adding 8 or 16 cores only improves runtime by a factor of 2.

      • DD – Is anyone familiar with the syntax in that discussion like conda create --name xtb-mkl xtb "libblas=*=*mkl"?

      • (General) - No

    • DD – Do we have big XTB datasets that will need XTB workers?

      • PB – Not now

      • DD – Let’s punt on this until then.

  • JW – For Bespokefit?

• DD – PB, could you ask TG to spin up workers tageting the openff-tscc tag?

  • PB – Yes, I think he’s doing that now. 26 workers with 8 cores and 30GB each. I’m running 20 workers now from my user account.

  • DD – Thanks for the update. Could I ask you to update the currently running info in qcfractal-compute?