Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-18891

increase default "lfs setdirstripe --max-inherit-rr"

Details

    • Improvement
    • Resolution: Unresolved
    • Major
    • None
    • None
    • 3
    • 9223372036854775807

    Description

      Recent testing has shown that increasing the default round-robin depth for directories created near the root directory from the current default of 3 to some higher depth (5/6) or larger can improve performance for larger filesystems.

      However, care must be taken to avoid adding unnecessary overhead for systems with only a small number of MDTs, because creating a remote directory can take up to 10x longer to finish.

      It may be that the ideal situation is to tune the max-inherit-rr value set on the root directory based on the number of MDTs formatted in the filesystem. If the number of MDTs is 10+ then the extra overhead for creating remote directories is compensated by the improved performance from getting more MDTs involved. Otherwise, something like "3 + MDT_COUNT" might be reasonable.

      Attachments

        Activity

          [LU-18891] increase default "lfs setdirstripe --max-inherit-rr"
          adilger Andreas Dilger added a comment - - edited

          The NCSA LUG presentation reports using a default max-inherit-rr=10 with their 48-MDT Delta filesystem to provide good MDT space balance (45/48 MDTs are within 6% usage of each other).

          The HRT LUG presentation reports using max-inherit-rr=17 for each of the project directories on their 40-MDT filesystem and maintain MDT balance within 3% usage.

          Conversely, we have seen imbalanced MDT usage on smaller clusters using only the max-inherit-rr=3 setting.

          adilger Andreas Dilger added a comment - - edited The NCSA LUG presentation reports using a default max-inherit-rr=10 with their 48-MDT Delta filesystem to provide good MDT space balance (45/48 MDTs are within 6% usage of each other). The HRT LUG presentation reports using max-inherit-rr=17 for each of the project directories on their 40-MDT filesystem and maintain MDT balance within 3% usage. Conversely, we have seen imbalanced MDT usage on smaller clusters using only the max-inherit-rr=3 setting.

          People

            core-lustre-triage Core Lustre Triage
            adilger Andreas Dilger
            Votes:
            0 Vote for this issue
            Watchers:
            5 Start watching this issue

            Dates

              Created:
              Updated: