Memory aware scheduling - #357
Merged
Merged
Conversation
The 2026-09-08 wasm-jit run took the machine out of memory: peak 104.9 GiB, 19176 OOM kills, 5856 models at finalphase 0 where every earlier run had ~894, and 202 killed at their timeout where no earlier run had any. Two causes, neither visible in `maxrss`. A job that runs into the outer timeout is killed by its process group, and OMPython starts omc with `preexec_fn=os.setsid`, so omc is in a session of its own and never gets the signal. testmodel.py, which does find omc through psutil, is dead by then. Each timeout therefore leaks a running omc for the rest of the run. test.py now takes the descendants before it kills the parent and kills them after, and testmodel.py handles SIGTERM: it writes the result file and takes its children with it, inside the ten seconds before the SIGKILL. The 16 jobs also ran with no regard for what they need. Summing the largest measurement each model has, over the newest runs of master, wasm-jit, cpp and newInst-newBackend: | scheduling | peak concurrent | wall | |-----------------------|-----------------|-------| | as before | 121 GiB | 3.18h | | >=4 GiB capped at 4 | 62.7 GiB | 3.18h | The cap is free because the heavy models are 148 of 19800: `runCapped` keeps two queues, and a worker that may not start a heavy job takes a light one instead of waiting for a slot. Only when heavy jobs are all that is left does anyone wait. `configs/heavy-models.json` is that list, meant to be read and edited rather than regenerated: a model that dies early on one branch reports what it reached, not what it wants, and the 33 entries that stopped at exactly 7.01 GiB are sitting on the 8 GB `ulimit -v`. heavy-models.py proposes it and says which entries those are. Children also get oom_score_adj 1000, so that the kernel takes omc and leaves python to report the phase it had reached. Raising a child is allowed unprivileged; lowering ourselves is not. Assisted-by: Claude Opus 5 (1M context)
A run of `--wasmjitrunner` or `--fmisimulator` is one job filling several tables under one date: wasm-jit, wasm-jit-me and wasm-jit-cs. Removing a bad run a table at a time leaves the siblings describing a run that no longer exists, and the rows in omcversion and libversion behind either way. `remove-run.py` takes the branch, finds the newest run of it (or the one `--date`/`--omcversion` names), and deletes from every table of that date at once. It prints what it would delete and does nothing until it is given `--write`, then reads the counts back to check. ./remove-run.py --db postgresql://om@localhost/omdb wasm-jit ./remove-run.py --db postgresql://om@localhost/omdb wasm-jit --write Siblings are found by name and skipped when they have no rows of that date; `--also` names a table the prefix does not reach, which is how the `--solver` runners are stored. job_claim is left alone. Its rows say who tested a library last, the next run overwrites them, and a claim that is not running blocks nothing. Assisted-by: Claude Opus 5 (1M context)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.