Library cache latches gone in Oracle 11g

Tanel Poder

2008/08/02

In Oracle 11g even more library cache operations have been changed to use KGX mutexes instead of latches.

In Oracle 10.2.0.2+ the library cache pin latch usage was replaced with mutexes whenever _kks_use_mutex_pin was true, also few other things like V$SQLSTATS arrays and parent cursor examination were protected by mutexes. However the traversing of library cache hash chains (the right child cursor lookup using kksfbc()) was still protected by library cache latches which could become a problem with frequent soft parsing combined with too little cursor cache and long library cache hash chains (remember, the library cache latches were always taken exclusively even for plain hash chain scanning).

In 11g all library cache related latches except “library cache load lock” are gone and corresponding operations are protected by mutexes instead. The “library cache” latches have been replaced by “Library Cache” mutexes for example.

Here are couple queries which illustrate the change.

**Executed on 10.2.0.3:

**

SQL> select name from v$latch where lower(name) like '%library%';

NAME
--------------------------------------------------
library cache pin allocation
library cache lock allocation
library cache hash chains
library cache lock
library cache
library cache pin
library cache load lock

7 rows selected.

SQL> select name from v$event_name where name like '%library%';

NAME
----------------------------------------------------------------
latch: library cache
latch: library cache lock
latch: library cache pin
library cache pin
library cache lock
library cache load lock
library cache revalidation
library cache shutdown

8 rows selected.

Same queries executed on 11.1.0.6 and the bold lines above are gone:

SQL> select name from v$latch where lower(name) like '%library%';

NAME
----------------------------------------------------------------
library cache load lock

SQL> select name from v$event_name where name like '%library%';

NAME
----------------------------------------------------------------
library cache pin
library cache lock
library cache load lock
library cache: mutex X
library cache: mutex S
OSD IPC library
library cache revalidation
library cache shutdown

8 rows selected.

Looks like the developers have thrown out library cache latching mechanism in 11g as the mutexes introduced in 10.2 have proven to work fine (and they have managed to implement mutexes for protecting (almost) the full set of library cache operations).

So there is no way to revert back to old behaviour using _kks_use_mutex_pin=false (and you wouldn’t need to do this anyway, btw). The parameter is still there though and out of interest I checked what happens if I set it to false and bounced the instance.

As, expected, I started getting error messages like following right after startup:

ORA-03113: end-of-file on communication channel

ORA-00600: internal error code, arguments: [kglGetSessionUOL], [7], [], [], [], [], [], []

This is a good example of dangers with undocumented parameters – they may work ok in one version (and platform) but could cause serious trouble anywhere else.

Anyway, back to mutexes. In 10g you see there are 3 types of mutexes used:

SQL> select * from v$mutex_sleep;

MUTEX_TYPE                       LOCATION                                     SLEEPS  WAIT_TIME
-------------------------------- ---------------------------------------- ---------- ----------
Cursor Stat                      kksFindCursorStat [KKSSTALOC3]                  339        775
Cursor Parent                    kkspsc0 [KKSPRTLOC26]                          1507     123969
Cursor Parent                    kksLoadChild [KKSPRTLOC5]                       170        372
Cursor Parent                    kksLoadChild [KKSPRTLOC4]                       385       3799
Cursor Parent                    kksfbc [KKSPRTLOC2]                            1649      22484
Cursor Parent                    kksfbc [KKSPRTLOC1]                             128       1599
Cursor Pin                       kksLockDelete [KKSCHLPIN6]                     3505     928387
Cursor Pin                       kkslce [KKSCHLPIN2]                           15343  160394917
Cursor Pin                       kksfbc [KKSCHLFSP2]                            3219    9065433

9 rows selected.

In 11g there are couple additional mutexes, one (and most important) of them is Library Cache mutex:

SQL> select distinct mutex_type from v$mutex_sleep_history;

MUTEX_TYPE
--------------------------------
Library Cache
Cursor Pin
Cursor Parent
Cursor Stat

(I had to sample v$mutex_sleep_history instead of v$mutex_sleep on 11g as the latter was empty on 11g… it may be that the mutex get operations have been tuned further to not maintain the counters just to save even few more CPU cycles every get)

So, starting from 11g, each library cache bucket is protected by a separate mutex (yes all 131072 of them!).

Previously we had all those buckets were hashed to and protected by max 67 library cache latches which inevitably were prone to unnecessary false contention in case of many cursors being executed concurrently. Now this issue should be resolved once and for all. Of course the library cache mutexes still don’t solve all problems in the world (especially the ones related to excessive hard parsing!), for example there’s still a chance of hash collision of two entirely different cursors. Also, if there are many child cursors under a parent and the application cursor management is poor (e.g. cursors are closed after every execution and no session cursor caching is done) then you could still have contention on the mutex due continuous library cache hash chain traversing.

Comments powered by Talkyard.