Closed cwwalter closed 7 years ago
I think the workflow engine may not need modifying, but the simulation and DM processing workflows will do. I guess this issue is mainly about the DM Level 2 pipeline, right? Or are you including the simulation pipeline as well? It makes sense to keep the two separate, we found.
Strictly speaking, how we end up running at NERSC could drive changes to the workflow engine. At present it could not handle wrapping jobs in the MPI queue. If the solution is to bulk up the shared queue, we're good to go already.
Two issues mostly as far as I understood @richardxdubois and @tony-johnson should comment more or assign themselves to this issue:
1) DM pipeline. This is related to #7.
2) Really making everything run at NERSC is an issue related to queues etc.
Oops crossed messages. Yes, exactly as Richard says.
With regard to
1) DM pipeline. This is related to #7.
I think we will want to modify any tasks in the Twinkles Level 2 pipeline having to do with the coadds to loop in parallel over "patches" and modify the launching of the forced photometry tasks to loop over sensors as well as over visits. In working on #7, I wrote some tools to help with these.
The workflow engines were modified in order to run DC1 and this is a continuing topic of work and exploration for the CI group.
The current CI workflow needs to be tweaked to work at NERSC. We need to understand how to make both the workflow and DM work on full focal planes and more generally how to run in the time restricted batch environment there.