Name the forward pass thread in the trainer loop #895
Closed
Conversation
This file contains 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
|
This pull request was exported from Phabricator. Differential Revision: D61924982 |
|
This pull request was exported from Phabricator. Differential Revision: D61924982 |
JKSenthil
added a commit
to JKSenthil/tnt
that referenced
this pull request
Sep 9, 2024
Summary: Pull Request resolved: pytorch#895 Internal # Context With the sched_ext effort we are trying to build custom Linux schedulers that provide a small performance boost to AI training and improve the resource isolation on the trainer hosts. The latter is necessary to avoid cases when noisy neighbor processes, like data loaders, slow down the GPU training. More details in this note: https://fb.workplace.com/notes/1118655556176038 By naming the forward pass thread we can use its name and assign it a higher priority at the linux scheduler level. The backward pass is named inside the Pytorch implementation but the forward pass needs to be named at the application level. We did the same thing in PyPer, APS, MVAI which are the largest trainer frameworks for reco models, consuming 70%+ of fleet level GPU hours for recommender systems. # This Diff Adds core lines ``` if torch.multiprocessing._get_thread_name() != "trainer_main": torch.multiprocessing._set_thread_name("trainer_main") ``` to train/eval/predict scripts. We can check the preexisting name to avoid renaming the same thread. Reviewed By: diego-urgell Differential Revision: D61924982
JKSenthil
force-pushed
the
export-D61924982
branch
from
3b32360
to
7388ecd
Compare
Summary: Pull Request resolved: pytorch#895 Internal # Context With the sched_ext effort we are trying to build custom Linux schedulers that provide a small performance boost to AI training and improve the resource isolation on the trainer hosts. The latter is necessary to avoid cases when noisy neighbor processes, like data loaders, slow down the GPU training. More details in this note: https://fb.workplace.com/notes/1118655556176038 By naming the forward pass thread we can use its name and assign it a higher priority at the linux scheduler level. The backward pass is named inside the Pytorch implementation but the forward pass needs to be named at the application level. We did the same thing in PyPer, APS, MVAI which are the largest trainer frameworks for reco models, consuming 70%+ of fleet level GPU hours for recommender systems. # This Diff Adds core lines ``` if torch.multiprocessing._get_thread_name() != "trainer_main": torch.multiprocessing._set_thread_name("trainer_main") ``` to train/eval/predict scripts. We can check the preexisting name to avoid renaming the same thread. Reviewed By: diego-urgell Differential Revision: D61924982
|
This pull request was exported from Phabricator. Differential Revision: D61924982 |
JKSenthil
force-pushed
the
export-D61924982
branch
from
7388ecd
to
05bd14c
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Summary:
Internal
Context
With the sched_ext effort we are trying to build custom Linux schedulers that provide a small performance boost to AI training and improve the resource isolation on the trainer hosts. The latter is necessary to avoid cases when noisy neighbor processes, like data loaders, slow down the GPU training.
More details in this note: https://fb.workplace.com/notes/1118655556176038
By naming the forward pass thread we can use its name and assign it a higher priority at the linux scheduler level. The backward pass is named inside the Pytorch implementation but the forward pass needs to be named at the application level.
We did the same thing in PyPer, APS, MVAI which are the largest trainer frameworks for reco models, consuming 70%+ of fleet level GPU hours for recommender systems.
This Diff
Adds core lines
to train/eval/predict scripts. We can check the preexisting name to avoid renaming the same thread.
Reviewed By: diego-urgell
Differential Revision: D61924982