Alternate implementation of MaxConcurrentIO parameter #41
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 previous implementation split the MPI ranks into groups and prevented any ranks in the next group from proceeding until all ranks in the current group have finished. This will waste a lot of time if one (or a few) ranks are very slow. This new implementation tries to make sure that we always have MaxConcurrentIO ranks reading. When any one rank finishes the next is allowed to start immediately.
|
I also found that on COLIBRE L400M7 DMO the code suddenly started taking a long time to write the output at a particular snapshot. This pull request might help a bit if it's due to certain OSTs being slow. I think it might also be worth looking into increasing the metadata block size to avoid small writes (H5Pset_meta_block_size) and paged file space management (H5Pset_file_space_strategy) to align data blocks with lustre stripes. |
# for free
to join this conversation on GitHub.
Already have an account?
# 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Outdated suggestions cannot be applied.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.
On the FLAMINGO 10k run I've been finding that if not all ranks are allowed to read at the same time then the code is very slow. I think this might be because if the system is busy and a few ranks suffer long delays then the others are forced to wait. The current implementation divides the MPI ranks into groups and only one group at a time may read. None of the ranks in the next group can start until ALL ranks in the previous group finish.
This pull request modifies the code so that as soon as any one rank finishes reading another is immediately allowed to start. This is implemented by having the first rank which finishes reading become responsible for signalling the others to start.