2021-02-26 QCA Submission Meeting notes

Errors in production

Trevor

• TG: JC reporting zero successes on Lilac

  • think this may be a case where we are seeing it manifested as QCEngine Unknown Error.

  • memory shouldn’t be an issue for these WBO cases

  • somewhere in geomeTRIC there may be an infinite loop

  • might have some awkward geometries in this set of torsiondrives

  • 70% success would be acceptable

  • DD: is problem isolated to-lilac?

  • TG: maybe; tracebacks look good so far today

  • TG: would like to be able to troubleshoot the queue in greater detail

    • query compute tags, e.g. get counts for each, etc.

Use SCF_MEM_SAFETY_FACTOR
for all datasets?

Trevor

• Should we set this for SCF_MEM_SAFETY_FACTOR as a keyword?

  • small performance hit

• DD: perhaps we add env variable input to the Psi4Harness?

  • it’s an execution detail, avoids hardcoding in dataset

• TG: I like the idea of making it standard in the dataset submission; less to keep track of if you’re running a manager

  • just have to make sure we don’t give workers <4GiB of memory

• TG: DD, please refactor Psi4Harness output handling to not drop output on the floor, keeping only the msgpack output; for segfaults, really need stdout and stderr

• DD [committed] will pursue PR on QCEngine

• BP: QCFractal Manager can read env variables

  • so, route is there at the top level

  • needs to make it down the stack through QCEngine, psi4

• DD: perhaps introduce a mechanism for variances

New submission

Pavan

• PB: question on how to get past a recursion error in new submission:

• JH: haven’t seen this error before; will try and reproduce locally

default should always mean what our current best practice is

• DD: should we switch our “default” spec to default-dlc? Currently default with latest QCSubmit does DLC, but this differs from older cases

• [decision] keep default as what we are using today; version bumps of dataset indicate updates

• TG: would prefer re-using same directory for dataset version bumps; add a new dataset-v${VERSION}.json

Can we change README.md

Pavan

• README is outdated; would prefer either LIFECYCLE or INDEX to replace it

  • DD: [committed] can do; good impetus to get the automation for INDEX generation up finally

psi4's internal optimizer is being exposed in QCEngine.compute_procedure

Josh

• Planning to test this out; would be of interest to benchmark against geomeTRIC, see if we can find success against hard datasets