Oracle 内存回滚 (In Memory Undo)
[English]
作者：
fuyuncat
来源：
www.HelloDBA.com
分享到
IMU是10g引入的一项新技术，并且是Oracle的专利技术。但是，在10g中似乎没有完全激活，以下的测试在10.2.0.3中无法通过，在11g中可以进行。
在传统的事务更新过程中，如果一条数据记录被更新，就会从buffer cache中读取/分配一块UNDO数据块，并且立即会写入一条UNDO条目。如果同一个事务中有多条记录被更新，则undo buffer数据块中就会写入多条undo条目。引入IMU后，会从shared pool中分配出一个新的内存池——IMU pool。当一条数据记录被更新，仍然会从buffer cache中读取/分配一块undo数据块，但是，这块undo块并不会立即被更新，而是会在IMU pool中产生一个IMU node，IMU节点通过IMU map与数据记录更新对应。如果事务中有多条记录被修改，则IMU pool中就生产多个IMU nodes，而buffer中的undo block不会发生任何变化。当发生IMU commit或IMU flush时，才会通过IMU map将这些IMU node记录的undo信息写入undo buffer block中。并且，所有这些redo信息会和commit vector一起作为一个Redo条目写入Redo log中。整个过程中UNDO所产生的redo信息则大大减少。
隐含参数_in_memory_undo用于控制IMU特性的开关，可以在会话/系统级立即生效，默认为true。另外一个隐含参数_IMU_pools则控制IMU pool的数量，默认为3。此外，目前IMU的使用还存在一些限制，如undo管理方式（undo_management）必须为auto，在RAC中无效，
SQL代码
- --建立测试表、数据
- HELLODBA.COM>create table ttt (a number, b varchar2(20));
- Table created.
- HELLODBA.COM>begin
- 2 for i in 1..2000 loop
- 3 insert into ttt values (i, ''||i);
- 4 end loop;
- 5 commit;
- 6 end;
- 7 /
- PL/SQL procedure successfully completed.
- HELLODBA.COM>select a
- 2 from (select a, dbms_rowid.rowid_block_number(ROWID) block_id, lag(dbms_rowid.rowid_block_number(ROWID)) over (order by rowid) as pre_block_id from ttt)
- 3 where block_id != pre_block_id;
- A
- ----------
- 1124
- 1643
- 1
IMU Commit
让我们看下IMU commit与传统事务commit时产生的redo size的变化。首先看传统模式下，
SQL代码
- HELLODBA.COM>conn demo/demo@ora11
- Connected.
- HELLODBA.COM>alter session set "_in_memory_undo"=false;
- Session altered.
- HELLODBA.COM>update ttt set b='X' where a=1124;
- 1 row updated.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# and b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 4
- redo size 1600
- IMU commits 0
- HELLODBA.COM>update ttt set b='Y' where a=1643;
- 1 row updated.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# and b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 5
- redo size 1960
- IMU commits 0
- HELLODBA.COM>update ttt set b='Z' where a=1;
- 1 row updated.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# and b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 6
- redo size 2320
- IMU commits 0
- HELLODBA.COM>commit;
- Commit complete.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# and b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 7
- redo size 2416
- IMU commits 0
可以看到，每一条数据被update都产生一条redo 条目。
然后，我们激活IMU，再重复上述事务过程，
SQL代码
- HELLODBA.COM>conn demo/demo@ora11
- Connected.
- HELLODBA.COM>alter session set "_in_memory_undo"=true;
- Session altered.
- HELLODBA.COM>update ttt set b='X' where a=1124;
- 1 row updated.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# an
- d b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 3
- redo size 1084
- IMU commits 0
- HELLODBA.COM>update ttt set b='Y' where a=1643;
- 1 row updated.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# an
- d b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 3
- redo size 1084
- IMU commits 0
- HELLODBA.COM>update ttt set b='Z' where a=1;
- 1 row updated.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# an
- d b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 3
- redo size 1084
- IMU commits 0
- HELLODBA.COM>commit;
- Commit complete.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# an
- d b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 4
- redo size 2176
- IMU commits 1
可见redo数量并没有随着数据的更新而增加，而是在IMU commit时增加。而当1条DML语句更新多条记录时，也可以使用到IMU：
SQL代码
- HELLODBA.COM>conn demo/demo@ora11
- Connected.
- HELLODBA.COM>alter session set "_in_memory_undo"=true;
- Session altered.
- HELLODBA.COM>update ttt set b='X' where a in (1643, 1124, 1);
- 3 rows updated.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# an
- d b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 3
- redo size 1084
- IMU commits 0
- HELLODBA.COM>commit;
- Commit complete.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# an
- d b.name in ('redo entries', 'redo size', 'IMU commits');
- NAME VALUE
- ---------------------------------------------------------------- ----------
- redo entries 4
- redo size 2344
- IMU commits 1
从上面的例子中你也许注意到了，尽管UPDATE过程中redo size没有变化，但是，在IMU commit时，redo size的变化却很大，比传统模式下的commit产生的redo大许多。这是因为在IMU commit中，不仅仅包含了commit vector，还包含了commit之前数据变化，并且这些redo数据的写入是一次批量写入。我们可以将这个redo条目dump出来观察其内容：
SQL代码
- HELLODBA.COM>conn demo/demo@ora11
- Connected.
- HELLODBA.COM>set serveroutput on
- HELLODBA.COM>var v_bt number;
- HELLODBA.COM>var v_et number;
- HELLODBA.COM>alter session set "_in_memory_undo"=false;
- Session altered.
- HELLODBA.COM>update tt set x=1 where rownum <= 1;
- 1 row updated.
- HELLODBA.COM>update tt set x=2 where rownum <= 1;
- 1 row updated.
- HELLODBA.COM>update tt set x=3 where rownum <= 1;
- 1 row updated.
- HELLODBA.COM>begin
- 2 select current_scn into :v_bt from v$database;
- 3 dbms_output.put_line(''||:v_bt);
- 4 end;
- 5 /
- 6328064
- PL/SQL procedure successfully completed.
- HELLODBA.COM>commit;
- Commit complete.
- HELLODBA.COM>begin
- 2 select current_scn into :v_et from v$database;
- 3 dbms_output.put_line(''||:v_et);
- 4 end;
- 5 /
- 6328067
- PL/SQL procedure successfully completed.
- HELLODBA.COM>declare
- 2 v_log varchar2(2000);
- 3 v_sql varchar2(4000);
- 4 begin
- 5 select a.member into v_log from v$logfile a, v$log b where a.group#=b.group# and b.status='CUR
- RENT' and rownum <= 1;
- 6 execute immediate 'alter system switch logfile';
- 7 v_sql := 'alter system dump logfile '''||v_log||''' SCN MIN '||:v_bt||' SCN MAX '||:v_et;
- 8 execute immediate v_sql;
- 9 end;
- 10 /
- PL/SQL procedure successfully completed.
可以看到，在trace日志中，这一redo条目包含了多个change：
SQL代码
- REDO RECORD - Thread:1 RBA: 0x0000c8.00000f39.0010 LEN: 0x046c VLD: 0x0d
- SCN: 0x0000.00608ed4 SUBSCN: 1 11/16/2009 14:59:10
- CHANGE #1 TYP:2 CLS: 1 AFN:4 DBA:0x010016cf OBJ:74952 SCN:0x0000.00602dc7 SEQ: 4 OP:11.19
- KTB Redo
- ...
- CHANGE #2 TYP:0 CLS:17 AFN:3 DBA:0x00c00009 OBJ:4294967295 SCN:0x0000.00608e9b SEQ: 2 OP:5.2
- ...
- CHANGE #8 TYP:0 CLS:18 AFN:3 DBA:0x00c006f7 OBJ:4294967295 SCN:0x0000.00608ed4 SEQ: 2 OP:5.1
- ...
IMU Flush
IMU pool也是按照LRU算法管理的。当IMU pool没有足够空闲内存可分配时，会将buffer链上LRU段的buffer块flush出来。其他一些事件也会导致IMU flush的发生，如switch logfile、rollback。但是，尽管IMU pool是从shared pool中分配的，手动flush shared pool并不会导致IMU flush。当IMU flush发生时，也会将undo、redo数据批量写入。
SQL代码
- HELLODBA.COM>conn demo/demo@ora11
- Connected.
- HELLODBA.COM>alter session set "_in_memory_undo"=true;
- Session altered.
- HELLODBA.COM>update tt set x=1;
- 1 row updated.
- HELLODBA.COM>update tt set x=2;
- 1 row updated.
- HELLODBA.COM>update tt set x=3;
- 1 row updated.
- HELLODBA.COM>select b.name, a.value from v$sysstat a, v$statname b where a.statistic#=b.statistic# and b.name like '%IMU%';
- NAME VALUE
- ---------------------------------------------------------------- ----------
- IMU commits 320
- IMU Flushes 159
- IMU contention 19
- ...
- 13 rows selected.
- HELLODBA.COM>alter system switch logfile;
- System altered.
- HELLODBA.COM>select b.name, a.value from v$sysstat a, v$statname b where a.statistic#=b.statistic# and b.name like '%IMU%';
- NAME VALUE
- ---------------------------------------------------------------- ----------
- IMU commits 320
- IMU Flushes 160
- IMU contention 20
- ...
- 13 rows selected.
Tips: 通过dump出事务的undo block，可以比较IMU commit/flush前后undo block的变化——commit/flush之前没有写入数据。
IMU CR
在传统事务中，需要进行一致性读时，会从相应的UNDO数据块中读入undo数据进行undo操作。而在IMU中，在发生IMU commit或IMU flush之前，这些undo数据并未写入UNDO数据块中，此时一致性读就从IMU pool中读取相应的IMU node中的undo信息。
SQL代码
- --Session 1:
- HELLODBA.COM>conn demo/demo@ora11
- Connected.
- HELLODBA.COM>alter session set "_in_memory_undo"=true;
- Session altered.
- HELLODBA.COM>update tt set x=1;
- 1 row updated.
- HELLODBA.COM>update tt set x=2;
- 1 row updated.
- HELLODBA.COM>update tt set x=3;
- 1 row updated.
- --Session 2:
- HELLODBA.COM>conn demo/demo@ora11
- Connected.
- HELLODBA.COM>alter system flush buffer_cache;
- System altered.
- HELLODBA.COM>alter session set tracefile_identifier=IMU_CR;
- Session altered.
- HELLODBA.COM>alter session set events '10046 trace name context forever, level 8';
- Session altered.
- HELLODBA.COM>select * from tt;
- X
- ----------
- 3
- HELLODBA.COM>alter session set events '10046 trace name context off';
- Session altered.
- HELLODBA.COM>select b.name, a.value from v$mystat a, v$statname b where a.statistic#=b.statistic# and b.name like '%IMU%';
- NAME VALUE
- ---------------------------------------------------------------- ----------
- ...
- IMU CR rollbacks 3
- ...
- 13 rows selected.
从trace文件可以看到，我们将buffer cache的内容都flush了，但是并未从undo文件中读取undo 信息（没有相应的IO等待）。
--- Fuyuncat Mark ---