fs/ufs: change default locking protocol - v4.1.x #12760
Merged
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
The fs/ufs component by default disabled all file locking before read/write operations (except for NFS file systems). This was based on the assumption, that the file system itself performs the required locking operation and hence we don't have to add to it. This assumption is incorrect when using data sieving. In data sieving, the code 'ignore' small gaps when we write to a file, and perform instead a read-modify-write sequence ourselves for performance reasons. The problem is however that even within a collective operation not all aggregators might want to use data sieving. Hence, enabling locking just for the data-sieving routines is insufficient, all processes have to perform the locking. Therefore, our two options are: a) either disable write data-sieving by default, or b) enable range-locking by default. After some testing, I think enabling range-locking be default is the safer and better approach. It doesn't seem to show any significant performance impact on my test systems. Note, that on Lustre file systems, we can keep the default to no-locking as far as I can see, since the collective algorithm used by Lustre is unlikely to produce this pattern. I did add in however an mca parameter that allows us to control the locking algorithm used by the Lustre component as well, in case we need to change that for a particular use-case or platform. Fixes Issue open-mpi#12718 Signed-off-by: Edgar Gabriel <[email protected]> (cherry picked from commit c697f28)
edgargabriel
changed the title
lrbison
approved these changes
Aug 14, 2024
jsquyres
reviewed
Sep 24, 2024
| (void) mca_base_component_var_register(&mca_fs_lustre_component.fsm_version, | ||
| "lock_algorithm", "Locking algorithm used by the fs ufs component. " | ||
| " 0: auto (default), 1: skip locking, 2: always lock entire file, " | ||
| "3: lock only specific ranges", |
There was a problem hiding this comment.
If this MCA parameter only accepts specific values, is there any reason not to use the enum MCA type? (such that the MCA base will take care of ensuring that only allowable values are set)
There was a problem hiding this comment.
@jsquyres you are technically correct that we could enum MCA types, this is more a feature of what I am used to doing vs. what is possible. That being said, the same pr has already been merged to 5.0.x as well, do we really want to go back and change that as well? We could still fix it in main branch for future releases.
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.
The fs/ufs component by default disabled all file locking before read/write operations (except for NFS file systems). This was based on the assumption, that the OS itself performs the required locking operation and hence we don't have to add to it.
This assumption is incorrect when using data sieving. In data sieving, the code 'ignore' small gaps when we write to a file, and perform instead a read-modify-write sequence ourselves for performance reasons. The problem is however that even within a collective operation not all aggregators might want to use data sieving. Hence, enabling locking just for the data-sieving routines is insufficient, all processes have to perform the locking. Therefore, our two options are: a) either disable write data-sieving by default, or b) enable range-locking by default.
After some testing, I think enabling range-locking be default is the safer and better approach. It doesn't seem to show any significant performance impact on my test systems.
Note, that on Lustre file systems, we can keep the default to no-locking as far as I can see, since the collective algorithm used by Lustre is unlikely to produce this pattern. I did add in however an mca parameter that allows us to control the locking algorithm used by the Lustre component as well, in case we need to change that for a particular use-case or platform.
Fixes Issue #12718
Signed-off-by: Edgar Gabriel [email protected]
(cherry picked from commit c697f28)