Batch TIBDExchangeMove - Githubissues

badisa commented 5 months ago

Appears to about 12-15x (depends on the GPU, 12x on A4000, 15x on RTX 2080) faster with a batch size of 250 on the Buckyball test case, though haven't explored what the right batch size is here.
Looked at HSP90 (Woodhead and Kung) and BRD4 which have some open questions, but doesn't look unreasonable.
There are some things that could be cleaned up later (some extra buffers lying around, the addition of the new translations buffer, etc), and will leave some of these changes to deboggling.

I ran 6 runs of the buckyball simulation (batch size of 250), three starting from 0 waters and three starting from 6 waters

Ran a batch size of 1 for 20k iterations with seed 2024 and got the same result as the batched version of 250 (based on logs).

The results seems similar to https://github.com/proteneer/timemachine/pull/1281#issuecomment-2023789699 though wouldn't say this is converged at 10k MC -> 10k MD for 20k iterations.

Using the buckyball system to benchmark the ideal value for the batch size The error bars are the std deviation of 30 rounds of running the TIBDExchange mover

badisa commented 5 months ago

Re-ran the scan of batch sizes on an A10 rather than an A4000 and seeing slightly different timings. Less noisy since this isn't being run on my work machine.

batch_size_to_ns

badisa commented 5 months ago

@maxentile noticed inconsistency in the naming of a variable, renamed in https://github.com/proteneer/timemachine/pull/1283/commits/fc4c5f5d9f86bc4865ff147c9ad0836ff6ed867b

maxentile commented 5 months ago

Sorry, was just saying I hadn't looked at it closely enough to provide substantive comments, only things that jumped out were superficial / naming-related such as this example.

A few other instances where the same value is referred to as both a weight and a log_weight:

https://github.com/proteneer/timemachine/pull/1283/files#diff-9a201b070e92eb5a50a13e2f43084da3041d25f075a4dc94565b980c478c2036R33-R47

https://github.com/proteneer/timemachine/pull/1283/files#diff-d869f683a02521d1d94fd7f6f0f5ae649114ea7971c2deda24964f75632c5235R135-R140

https://github.com/proteneer/timemachine/pull/1283/files#diff-7ebce44dc6d240c4cfe9ae169eac3f3e9a4b6b18c50fc34f932588407fba88f6R520-R531

https://github.com/proteneer/timemachine/pull/1283/files#diff-7ebce44dc6d240c4cfe9ae169eac3f3e9a4b6b18c50fc34f932588407fba88f6R496

(didn't spot any errors, but mixing log vs. not-log weights could get confusing in the future)

badisa commented 5 months ago

Sorry, was just saying I hadn't looked at it closely enough to provide substantive comments, only things that jumped out were superficial / naming-related such as this example.

A few other instances where the same value is referred to as both a weight and a log_weight:

https://github.com/proteneer/timemachine/pull/1283/files#diff-9a201b070e92eb5a50a13e2f43084da3041d25f075a4dc94565b980c478c2036R33-R47

https://github.com/proteneer/timemachine/pull/1283/files#diff-d869f683a02521d1d94fd7f6f0f5ae649114ea7971c2deda24964f75632c5235R135-R140

https://github.com/proteneer/timemachine/pull/1283/files#diff-7ebce44dc6d240c4cfe9ae169eac3f3e9a4b6b18c50fc34f932588407fba88f6R520-R531

https://github.com/proteneer/timemachine/pull/1283/files#diff-7ebce44dc6d240c4cfe9ae169eac3f3e9a4b6b18c50fc34f932588407fba88f6R496

(didn't spot any errors, but mixing log vs. not-log weights could get confusing in the future)

Fixed up in https://github.com/proteneer/timemachine/pull/1283/commits/638b242fc8c6ce8606f1e5f97844de5e48c48b4f

proteneer / timemachine

Batch TIBDExchangeMove #1283