search

BSDExabio / PSP

protein structure prediction
1 stars 0 forks source link

Dask-ify the post-AF minimization of structures #14

Closed rbdavid closed 2 years ago

rbdavid commented 2 years ago

Need to clean up minimization code and implement a process to run the code across many nodes of multiple GPUs.

rbdavid commented 2 years ago

Dask pipeline has been created and tested on a small number of nodes on Summit. Now running a test run using 922 nodes (leadership scale). Will check for errors/issues and then likely push to main.

rbdavid commented 2 years ago

922 Summit nodes is honestly overkill for the speed at which the OpenMM minimizations take. We're talking 1s to 10s of seconds per structure for the whole process, which consists of adding missing atoms (rate limiting step), creating the OpenMM simulation object, and running the minimization. Have successfully run the pipeline for number of nodes from 1 to 16 with no issues. Caveat: I'm still getting a huge amount of INFO dumps that seem to be issues with the workers, yet the pipeline runs fine. It may just be how I clean up the dask workers/scheduler that is causing the errors/alarming messages.