Uploaded image for project: 'MariaDB Server'
  1. MariaDB Server
  2. MDEV-15058

Remove multiple InnoDB buffer pool instances

    XMLWordPrintable

    Details

      Description

      This came up during the MDEV-15016 review.

      I started to wonder whether multiple InnoDB buffer pools actually help with any workloads. Yes, it probably was a good idea to split the buffer pool mutex when Inaam Rana introduced multiple buffer pools in MySQL 5.5.5, but since then, there have been multiple fixes to reduce contention on the buffer pool mutex, such as Inaam's follow-up fix in MySQL 5.6.2 to use rw-locks instead of mutexes for the buf_pool->page_hash.

      In MySQL 8.0.0, Shaohua Wang implemented one more thing that MariaDB should copy: MDEV-15053 Split buf_pool_t::mutex.

      I think that we should seriously consider removing all code to support multiple buffer pools or page cleaners.
      Should multiple buffer pools be needed in the future (for example, on NUMA machines), it should be designed better from the ground up. Currently the partitioning is arbitrary; buffer pool membership is basically determined by a hash of the page number.

      The description of WL#6642: InnoDB: multiple page_cleaner threads seems to imply that it may have been a mistake to partition the buffer pool.

      Note: partitioning or splitting mutexes often seems to be a good idea. But partitioning data structures or threads might not be.

      Axel Schwenke, please test different workloads with innodb_buffer_pool_instances=1 and innodb_page_cleaners=1, and compare the performance to configurations that use multiple buffer pools (and page cleaners). If using a single buffer pool instance never seems to causes any regression, I think that we should simplify the code.

        Attachments

        1. 1bp.txt
          81 kB
          Vladislav Vaintroub
        2. 4bp.txt
          86 kB
          Vladislav Vaintroub
        3. MDEV-15058.ods
          82 kB
          Axel Schwenke
        4. MDEV-15058.pdf
          56 kB
          Axel Schwenke
        5. MDEV-15058-10.4.10.ods
          88 kB
          Axel Schwenke
        6. MDEV-15058-10.4vs10.5.ods
          140 kB
          Axel Schwenke
        7. MDEV-15058-10.5.ods
          88 kB
          Axel Schwenke
        8. MDEV-15058-10.5-34dafb7e3a8.ods
          49 kB
          Axel Schwenke
        9. MDEV-15058-10.5-dev.ods
          68 kB
          Axel Schwenke
        10. MDEV-15058-B.ods
          77 kB
          Axel Schwenke
        11. MDEV-15058-B.pdf
          51 kB
          Axel Schwenke
        12. MDEV-15058-RAM-ARM.ods
          77 kB
          Axel Schwenke
        13. MDEV-15058-RAM-Intel.ods
          82 kB
          Axel Schwenke
        14. MDEV-15058-singleBP.ods
          51 kB
          Axel Schwenke
        15. MDEV-15058-SSD-ARM.ods
          79 kB
          Axel Schwenke
        16. MDEV-15058-SSD-Intel.ods
          80 kB
          Axel Schwenke
        17. MDEV-15058-thiru.ods
          73 kB
          Axel Schwenke
        18. MDEV-15058-thiru.pdf
          53 kB
          Axel Schwenke
        19. MDEV-15058-tpcc.ods
          45 kB
          Axel Schwenke
        20. ramdisk-ro1.svg
          428 kB
          Axel Schwenke
        21. ramdisk-ro4.svg
          395 kB
          Axel Schwenke
        22. ramdisk-rw1.svg
          782 kB
          Axel Schwenke
        23. ramdisk-rw4.svg
          581 kB
          Axel Schwenke

          Issue Links

            Activity

              People

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

                Dates

                Created:
                Updated:
                Resolved: