Uploaded image for project: 'MariaDB Server'
  1. MariaDB Server
  2. MDEV-26827

Make page flushing even faster

    XMLWordPrintable

    Details

      Description

      MDEV-25113 removed the acquisition of buf_pool.flush_list_mutex from buf_page_write_complete(). But, acquiring buf_pool.mutex should not really be necessary except in eviction flushing (request.is_LRU()) or when the block->lock is not available (no uncompressed copy of a ROW_FORMAT=COMPRESSED exists in the buffer pool).

      In the common case of checkpoint flushing, we can actually rely on the block->lock to protect us. The only other reason to hold buf_pool.mutex was updating the count of outstanding page writes (buf_pool.n_flush_list()). That counter is actually redundant, because we can refer to write_slots->pending_io_count().

      Finally, we would remove pthread_cond_broadcast(&buf_pool.done_flush_list) from buf_page_write_complete(). Only the buf_flush_page_cleaner would call that. To wait for page writes to complete, os_aio_wait_until_no_pending_writes() may be invoked.

        Attachments

          Issue Links

            Activity

              People

              Assignee:
              marko Marko Mäkelä
              Reporter:
              marko Marko Mäkelä
              Votes:
              0 Vote for this issue
              Watchers:
              6 Start watching this issue

                Dates

                Created:
                Updated:

                  Git Integration