Uploaded image for project: 'Lustre'
  1. Lustre
  2. LU-18562

batch allocation in osd_write()

XMLWordPrintable

    • Icon: Improvement Improvement
    • Resolution: Fixed
    • Icon: Minor Minor
    • Lustre 2.17.0
    • None
    • None
    • 3
    • 9223372036854775807

      osd_write() performance can be improved allocating few blocks at once. this can be useful for llog which suffers from frequent serialization due to the allocation's semaphore. preliminary benchmark shows promising results writing 32K llog records into a single plain llog in 1 and 4 threads:

      indexed llog + new write, 16 blocks alloc:
took 51ms, 1594 per rec in 1 threads
took 18ms, 576 per rec in 4 threads

indexed llog + new write, 8 blocks alloc:
took 50ms, 1578 per rec in 1 threads
took 18ms, 583 per rec in 4 threads

indexed llog + new write, 4 blocks alloc:
took 50ms, 1567 per rec in 1 threads
took 19ms, 607 per rec in 4 threads

indexed llog + new write, 2 blocks alloc:
took 51ms, 1622 per rec in 1 threads
took 31ms, 988 per rec in 4 threads

indexed llog + old write:
took 51ms, 1624 per rec in 1 threads
took 34ms, 1090 per rec in 4 threads

non-indexed llog + old write:
took 68ms, 2145 per rec in 1 threads
took 53ms, 1669 per rec in 4 threads

            bzzz Alex Zhuravlev
            bzzz Alex Zhuravlev
            Votes:
            0 Vote for this issue
            Watchers:
            8 Start watching this issue

              Created:
              Updated:
              Resolved: