Ivy Zhang

>I see, well I guess we can use 3000 iterations and it would be 1.5-2 hours on a 1080Ti (we have to check our GPU CI has a similar performance...

Oh.. I see the problem. Its an 8x difference per iteration (without considering the extra transformations)...

Interestingly, when comparing the neutral vs charge test iteration time with the LangevinSplittingDynamicsMove and openmm 7.7, charge is in fact 3x slower than neutral. For LangevinDynamicsMove + openmm nightly, I...

>I've made a few comments on https://github.com/choderalab/perses/pull/1065 to suggest some ways we might be able to speed things up a bit. @jchodera : Thanks for the feedback! I just responded...

To address the slowdown problem, we don't actually want to run separate scripts for the test, so Iván and I discussed that the next short term action item is to...

> Can someone point me to the code that sets up the HybridTopologyFactory that exhibits this weird behavior? Here's the neutral test: https://github.com/choderalab/perses/blob/main/perses/tests/test_repex.py#L6-L84 You should be able to reproduce the...

>The GPU is definitely in use during the setup phase. This is likely using one or more ContextCaches Yes, Iván and I discussed this yesterday -- we are wondering whether...

>However, I don't have something to compare it to where I serialize out the state and reload it in a different process to see it get faster by >2x. Do...

Suggestions for names other than `Protocol`: experiment, runner, workflow, pipeline

Thanks for the clarification! I still think it would be better to rename `Protocol` (i.e. to `Procedure` or `InstructionSet`) because if gufe is going to replace the perses API, "protocol"...