Trigger in-memory splits sooner

XMLWordPrintableJSON

    • Type: Improvement
    • Resolution: Done
    • Priority: Major - P3
    • WT2.7.0
    • Affects Version/s: None
    • Component/s: None
    • None
    • None
    • None

      ... so that multi-threaded append workloads don't get forced to evict pages when they fall behind.

      In the past, we had a single size limit (memory_page_max, default 5MB), and when pages grew to that size, threads stalled and tried to evict / split them. We added in-memory splits as a lightweight operation that creates a new page so that appends can move on, leaving the large page behind for eviction threads to write and split.

      That scheme worked well for single-threaded appends, or workloads where threads stayed closely in lock-step. We also added some delays to when pages were considered for splits to try to give threads a chance to move to the new page before the old one was locked for eviction.

      However, if we triggered in-memory splits sooner (e.g., at 80% of memory_page_max, before the last page gets completely full, then other threads should notice the new end-of-tree before the previous page has to be locked down.

              Assignee:
              Unassigned
              Reporter:
              Michael Cahill (Inactive)
              Votes:
              0 Vote for this issue
              Watchers:
              4 Start watching this issue

                Created:
                Updated:
                Resolved: