Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Initial grid stripe #514

Merged
merged 9 commits into from
Jan 28, 2021
Merged

Conversation

markusbattarbee
Copy link
Contributor

This is to fix #509
Quite possible though that bulk files already are written with the system default striping setting, but now it can be manually set.
(not tested yet)

@markusbattarbee markusbattarbee changed the base branch from master to dev November 26, 2020 14:15
@markusbattarbee
Copy link
Contributor Author

testing this now on Vorna

@markusbattarbee
Copy link
Contributor Author

Well, on Vorna, setting either stripe setting (bulk or restart) did not affect the actual striping of the file. I tried setting them to 8 and 12 respectively, with the directory stripe value set to 16 (maximum) or -1, and either way around, the files report 16 with lfs getstripe -c. Hmm. Lustre-dependent settings preventing an override?

@ursg
Copy link
Contributor

ursg commented Jan 27, 2021

My suspicion is that the MPI implementation on Vorna is simply not aware of lustre, and thus ignores this MPI hint. In these cases, it probably makes sense to set the striping factor on the working directory with the lfs tool before submitting, or adjust it in the job script.

@markusbattarbee
Copy link
Contributor Author

Anyhow can confirm testpackage zero diffs.

@markusbattarbee
Copy link
Contributor Author

Tested this on Mahti, there also the settings didn't propagate into an actual file striping. Will need to test on Puhti and Hawk as well.

@ykempf
Copy link
Contributor

ykempf commented Jan 28, 2021

Can you at the same time check whether restart striping hints propagate to the files or not?

@markusbattarbee
Copy link
Contributor Author

Yeah, testing both initial, restart and bulk, all of those had the directory striping value set to them. Also checked with directory striping values of -1 (default) and a manually set value (10).

@markusbattarbee
Copy link
Contributor Author

Well, on Hawk this PR works as intended.

@ursg
Copy link
Contributor

ursg commented Jan 28, 2021

I'm wondering: can we somehow query from the MPI implementation, whether it supports and implements this? So that we could give a friendly warning message in those situations?

@ursg
Copy link
Contributor

ursg commented Jan 28, 2021

This slide set is listing some code (on page 101-109), to query all predifined mpi info values for a newly opened file. We could do that, and check if a striping parameter even exists there, but I'm not sure this is sufficient to test for it.

This Nasa document on the other hand warns that "The hints are only advisory and may not be honored. For example, SGI MPT 2.12r26 honors these hints, but MPT 2.14r19 does not. Intel MPI 5.0x honors these hints when the I_MPI_EXTRA_FILESYSTEM and I_MPI_EXTRA_FILESYSTEM_LIST variables are set appropriately", so this might be completely arbitrary.

@markusbattarbee
Copy link
Contributor Author

I think verifying if striping is set to the file is a further code improvement which can be potentially tackled later, separate from this PR.

@ursg
Copy link
Contributor

ursg commented Jan 28, 2021

Fair enough. Merge early, merge often. 👍

@markusbattarbee markusbattarbee merged commit e2f143b into fmihpc:dev Jan 28, 2021
@markusbattarbee markusbattarbee deleted the initial_grid_stripe branch January 28, 2021 12:34
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

initial-grid writing should be striped too
3 participants