Avoid idle gpu

Created by: APJansen

The idea

We observed large gaps between training steps in the tensorboard profile when running on the GPU. After a lot of fiddling were found to be (at least partially) due to a per epoch overhead from tensorflow. This is reduced by redefining one actual training step as a single batch of size 1, not as a whole epoch as it used to be.

Implementation wise, this is done by copying the input up to 100 times, creating a set of 100 identical training inputs. Existing callbacks simply implement on_step_end instead of on_epoch_end and inherit from CallbackStep to take care of the conversion between steps, batches and epochs. One extra issue is that Keras computes metrics cumulatively, they are converted back to per step in CallbackStep.correct_logs. This is the only source of slight numerical differences, which however only appear in the logs and do not propagate at all, training results remain identical.

Performance

Timings for 1000 epochs of the main runcard (NNPDF40_nnlo_as_01180_1000), on Snellius, with 100 replicas on the GPU or 1 replica on the CPU. In brackets the GPU memory used.

branch	commit hash	1 replica	100 replicas
master	0a5fc614	145	91
avoid-idle-gpu	bb366aa6da	145	67

This saves 24 seconds (or 25%) per 1k epochs.

Profile

Note: slightly outdated profiles

This branch: and before this single commit for comparison:

added escience n3fit performance labels

assigned to @enocera

mentioned in merge request !1802 (closed)

mentioned in issue #1977

Created by: APJansen

The issue with the regression test was fixed by simply rebasing. (And then broke another, but I think it's just a fluke that will be fixed by rerunning).

I've updated the timings as well. (Don't worry about the higher time on the CPU, that's just because the CPU nodes I was using before aren't available today, this PR only affects >1 replica.)

Side comment: I added in a slight fix to the logging, where for multiple replicas it would print "Validation chi2s:" and then nothing. Incidentally I don't think the comment # The partial chi2 makes no sense for more than one replica at once is true anymore?

assigned to @enocera

unassigned @enocera

requested review from @enocera

Created by: scarlehoff

I'll try to review this tomorrow and leave it ready for you to have a second look and merge it in case you need to touch n3fit as discussed @RoyStegeman . I need to check a few corner cases but afaics this seems ok.

added run-fit-bot label

Created by: github-actions[bot]

Greetings from your nice fit ! I have good news for you, I just finished my tasks:

• Fit Name: NNBOT-9cf1f0522-2024-06-07
• Fit Report wrt master: https://vp.nnpdf.science/XCW60m6ySVO1cqBCe6Zlig==
• Fit Report wrt latest stable reference: https://vp.nnpdf.science/72Hgjp4vQOaUBgBm6PNnnQ==
• Fit Data: https://data.nnpdf.science/fits/NNBOT-9cf1f0522-2024-06-07.tar.gz

Check the report carefully, and please buy me a , or better, a GPU !

mentioned in issue #2118

removed run-fit-bot label

added redo-regressions label

Created by: scarlehoff

Review: Approved

I think this can be merged. #2188 can be dealt with in a separate PR.

added run-fit-bot label and removed redo-regressions label

Created by: github-actions[bot]

Greetings from your nice fit ! I have good news for you, I just finished my tasks:

• Fit Name: NNBOT-c2dc50df8-2024-07-17
• Fit Report wrt master: https://vp.nnpdf.science/1cLR0izpTQqk193L6Ih_6w==
• Fit Report wrt latest stable reference: https://vp.nnpdf.science/fEvlOQ1QREyxenwccx4zIA==
• Fit Data: https://data.nnpdf.science/fits/NNBOT-c2dc50df8-2024-07-17.tar.gz

Check the report carefully, and please buy me a , or better, a GPU !

Created by: scarlehoff

I'm going to merge this since the tests are passing and it is rebased on top of master (which means it is probably fixing something that has changed since the last merge, probably the TF / np version).

Worst case scenario, it can be reverted. The report in #2127 https://vp.nnpdf.science/WbBCvsjfQV-6ncIQ3GhCVw== was made with a PR which is on top of this one, so we have a reproduction of 4.0 with this changes that seems to do ok.

closed