]> www.infradead.org Git - users/jedix/linux-maple.git/commit
dm btree: improve btree residency
authorJoe Thornber <ejt@redhat.com>
Thu, 8 Apr 2021 12:47:08 +0000 (13:47 +0100)
committerMike Snitzer <snitzer@redhat.com>
Tue, 25 May 2021 20:23:02 +0000 (16:23 -0400)
commitdee29bf9192b7e7fec69191951286fe6a5cea89a
treea9695574b82324269634fb830c0361114bf0d5f7
parent7e768532b2396bcb7fbf6f82384b85c0f1d2f197
dm btree: improve btree residency

This commit improves the residency of btrees built in the metadata for
dm-thin and dm-cache.

When inserting a new entry into a full btree node the current code
splits the node into two.  This can result in very many half full nodes,
particularly if the insertions are occurring in an ascending order (as
happens in dm-thin with large writes).

With this commit, when we insert into a full node we first try and move
some entries to a neighbouring node that has space, failing that it
tries to split two neighbouring nodes into three.

Results are given below.  'Residency' is how full nodes are on average
as a percentage.  Average instruction counts for the operations
are given to show the extra processing has little overhead.

                         +--------------------------+--------------------------+
                         |         Before           |         After            |
+------------+-----------+-----------+--------------+-----------+--------------+
|    Test    |   Phase   | Residency | Instructions | Residency | Instructions |
+------------+-----------+-----------+--------------+-----------+--------------+
| Ascending  | insert    |        50 |         1876 |        96 |         1930 |
|            | overwrite |        50 |         1789 |        96 |         1746 |
|            | lookup    |        50 |          778 |        96 |          778 |
| Descending | insert    |        50 |         3024 |        96 |         3181 |
|            | overwrite |        50 |         1789 |        96 |         1746 |
|            | lookup    |        50 |          778 |        96 |          778 |
| Random     | insert    |        68 |         3800 |        84 |         3736 |
|            | overwrite |        68 |         4254 |        84 |         3911 |
|            | lookup    |        68 |          779 |        84 |          779 |
| Runs       | insert    |        63 |         2546 |        82 |         2815 |
|            | overwrite |        63 |         2013 |        82 |         1986 |
|            | lookup    |        63 |          778 |        82 |          779 |
+------------+-----------+-----------+--------------+-----------+--------------+

   Ascending - keys are inserted in ascending order.
   Descending - keys are inserted in descending order.
   Random - keys are inserted in random order.
   Runs - keys are split into ascending runs of ~20 length.  Then
          the runs are shuffled.

Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Colin Ian King <colin.king@canonical.com> # contains_key() fix
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
drivers/md/persistent-data/dm-btree.c
drivers/md/persistent-data/dm-transaction-manager.c
drivers/md/persistent-data/dm-transaction-manager.h