[20190214]11g Query Result Cache RC Latches.txt

[20190214]11g Query Result Cache RC Latches.txt

--//昨天我重复链接http://www.pythian.com/blog/oracle-11g-query-result-cache-rc-latches/的测试，
--//按照我的理解如果sql语句密集执行，使用Result Cache反而更加糟糕，这是我以前没有注意到的。
--//联想我们生产系统也存在类似的问题，我们有1个判断连接的语句select count（*) from test_connect;
--//在业务高峰它执行可以达到1600次/秒。另外一个简单的select sysdate from dual; 也达到800次/秒。
--//而实际上业务高峰sql语句执行率3000次/秒。这样的2条语句就占了2400次/秒。我以前一直以为将表设置
--//为result cache,可能提高执行效率，还是通过例子测试看看。

1.环境：
SCOTT@book> @ ver1
PORT_STRING                    VERSION        BANNER
------------------------------ -------------- --------------------------------------------------------------------------------
x86_64/Linux 2.4.xx            11.2.0.4.0     Oracle Database 11g Enterprise Edition Release 11.2.0.4.0 - 64bit Production

SCOTT@book> show parameter job
NAME                TYPE    VALUE
------------------- ------- ------
job_queue_processes integer 200

SCOTT@book> select * from v$latchname where name like 'Result Cache%';
LATCH# NAME                          HASH
------ ----------------------- ----------
   436 Result Cache: RC Latch  1054203712
   437 Result Cache: SO Latch   986859868
   438 Result Cache: MB Latch   995186388
--//我看到Result Cache名字与作者的不同，命名为Result Cache: RC Latch。

SCOTT@book> select name,gets from v$latch where lower(name) like '%result cache%';
NAME                                 GETS
------------------------------ ----------
Result Cache: RC Latch                  0
Result Cache: SO Latch                  0
Result Cache: MB Latch                  0

SCOTT@book> select count(*) from v$latch_children where lower(name) like '%result cache%';
  COUNT(*)
----------
         0

--//可以注意一个细节,Result Cache没有children latch。也仅仅1个Result Cache: RC Latch 父latch。从这里也可以看出如果
--//做了result cache的表，多个用户并发执行，可能反而不能获得好的性能,可能出现大量的Result Cache: RC Latch争用的情况.

2.建立测试例子：

create table t as select rownum id from dual ;
create unique index pk_t on t(id);
--//分析略。

SCOTT@book> create table job_times ( sid   number, time_ela number);
Table created.

--//按照源链接的例子修改如下：
create or replace procedure do_work(
 p_iterations in number
) is
 l_rowid  rowid;
 v_t number;
begin
 insert into job_times
  values (sys_context('userenv', 'sid'), dbms_utility.get_time)
  returning rowid into l_rowid;

for i in 1 .. p_iterations
 loop
     select count(*) into v_t from t;
 end loop;

update job_times set
   time_ela=dbms_utility.get_time-time_ela
  where rowid=l_rowid;
 commit;
end;
/

3.测试:
--//首先测试不做result cache的情况：
--//alter table t result_cache (mode default);

declare
 l_job number;
begin
 for i in 1 .. 50
 loop
  dbms_job.submit(
   job => l_job,
   what => 'do_work(1000000);'
  );
 end loop;
end;
/

SCOTT@book> commit ;
Commit complete.

--//注意一定要写提交，不然dbms_job.submit要等很久才执行。

SCOTT@book> select count(*),avg(TIME_ELA),sum(TIME_ELA) from job_times ;
  COUNT(*) AVG(TIME_ELA) SUM(TIME_ELA)
---------- ------------- -------------
        50        9235.1        461755

4.测试:

--///测试做result cache的情况，为了测试的准确，我重启数据库。
SCOTT@book> delete from job_times;
50 rows deleted.

SCOTT@book> commit ;
Commit complete.

SCOTT@book> alter table t result_cache (mode force);
Table altered.

--//重启数据库.

SCOTT@book> select name, gets, misses, sleeps, wait_time from v$latch where name like 'Result Cache%';
NAME                                 GETS     MISSES     SLEEPS  WAIT_TIME
------------------------------ ---------- ---------- ---------- ----------
Result Cache: RC Latch                  0          0          0          0
Result Cache: SO Latch                  0          0          0          0
Result Cache: MB Latch                  0          0          0          0

declare
 l_job number;
begin
 for i in 1 .. 50
 loop
  dbms_job.submit(
   job => l_job,
   what => 'do_work(100000);'
  );
 end loop;
end;
/

SCOTT@book> commit ;
Commit complete.

SCOTT@book> select count(*),avg(TIME_ELA),sum(TIME_ELA) from job_times ;
  COUNT(*) AVG(TIME_ELA) SUM(TIME_ELA)
---------- ------------- -------------
        50       7135.96        356798

SCOTT@book> select name, gets, misses, sleeps, wait_time from v$latch where name like 'Result Cache%';
NAME                                 GETS     MISSES     SLEEPS  WAIT_TIME
------------------------------ ---------- ---------- ---------- ----------
Result Cache: RC Latch           54232541    3499238          0          0
Result Cache: SO Latch                202          0          0          0
Result Cache: MB Latch                  0          0          0          0

--//很明显,即使存在Result Cache: RC Latch的争用,但是WAIT_TIME=0,不过我发现这样测试的一个缺点,就是50个job并不是同时运行.
--//$ ps -ef | grep ora_[j]|wc ,看看数量是不断增加的过程.
--//而且采用Result Cache后效果还是增强的.

5.换一个方式测试:
SCOTT@book> delete from job_times;
53 rows deleted.

SCOTT@book> commit ;
Commit complete.

--//设置result_cache=default
SCOTT@book> alter table t result_cache (mode default);
Table altered.

$ seq 50 | xargs -I{} echo 'sqlplus -s -l scott/book <<< "execute do_work(1000000)" & '| bash

--//等全部完成...

SCOTT@book> select count(*),avg(TIME_ELA),sum(TIME_ELA) from job_times ;
  COUNT(*) AVG(TIME_ELA) SUM(TIME_ELA)
---------- ------------- -------------
        50      10588.26        529413

SCOTT@book> delete from job_times;
50 rows deleted.

SCOTT@book> commit ;
Commit complete.

--//设置result_cache=force
SCOTT@book> alter table t result_cache (mode force);
Table altered.

$ seq 50 | xargs -I{} echo 'sqlplus -s -l  scott/book <<< "execute do_work(1000000)" & '| bash

SCOTT@book> select count(*),avg(TIME_ELA),sum(TIME_ELA) from job_times ;
  COUNT(*) AVG(TIME_ELA) SUM(TIME_ELA)
---------- ------------- -------------
        50       8573.28        428664
--//可以看到即使这样大并发,采用result cache还是要快许多,没有遇到作者的情况.
--//可以11GR2做了一些改进,不会遇到这样的情况.

SCOTT@book> column name format a30
SCOTT@book> select name, gets, misses, sleeps, wait_time from v$latch where name like 'Result Cache%';
NAME                                 GETS     MISSES     SLEEPS  WAIT_TIME
------------------------------ ---------- ---------- ---------- ----------
Result Cache: RC Latch          103461569    7263987          0          0
Result Cache: SO Latch                302          0          0          0
Result Cache: MB Latch                  0          0          0          0

6.不过当我拿作者的最后的例子做最后的测试发现,使用result cache慢很多.

SCOTT@book> create cluster hc ( n number(*,0)) single table hashkeys 15000 size 230;
Cluster created.

SCOTT@book> create table hc_t ( n number(*,0), v varchar2(200)) cluster hc (n);
Table created.

SCOTT@book> insert into hc_t select level, dbms_random.string('p', 200) from dual connect by level <= 10000;
10000 rows created.

SCOTT@book> commit;
Commit complete.

--//分析表略.

All we need now is two procedures, one with a regular select and another with a cached select:

create or replace procedure do_hc(
 p_iterations in number
) is
 l_rowid  rowid;
 l_n number;
begin
 insert into job_times
  values (sys_context('userenv', 'sid'), dbms_utility.get_time)
  returning rowid into l_rowid;

for i in 1 .. p_iterations
 loop
  l_n:=trunc(dbms_random.value(1, 10000));
  for cur in (select * from hc_t where n=l_n)
  loop
   null;
  end loop;
 end loop;

update job_times set
   time_ela=dbms_utility.get_time-time_ela
  where rowid=l_rowid;
end;
/

Procedure created.

create or replace procedure do_rc(
 p_iterations in number
) is
 l_rowid  rowid;
 l_n number;
begin
 insert into job_times
  values (sys_context('userenv', 'sid'), dbms_utility.get_time)
  returning rowid into l_rowid;

for i in 1 .. p_iterations
 loop
  l_n:=trunc(dbms_random.value(1, 10000));
  for cur in (select /*+ result_cache */ * from hc_t where n=l_n)
  loop
   null;
  end loop;
 end loop;

update job_times set
   time_ela=dbms_utility.get_time-time_ela
  where rowid=l_rowid;
end;
/

Procedure created.

The hash cluster will go first:

SCOTT@book> delete from job_times;
4 rows deleted.

SQL> commit;
Commit complete.

declare
 l_job number;
begin
 for i in 1 .. 4
 loop
  dbms_job.submit(
   job => l_job,
   what => 'do_hc(100000);'
    );
 end loop;
end;
/

PL/SQL procedure successfully completed.

SCOTT@book> commit ;
Commit complete.

--allow jobs to complete

SCOTT@book> select case grouping(sid) when 1 then 'Total:' else to_char(sid) end sid, sum(time_ela) ela from job_times group by rollup((sid, time_ela));
SID      ELA
------- ----
41       446
54       437
80       438
94       437
Total:  1758
--//每个测试仅仅需要4秒.

Now let's see if Result Cache can beat those numbers:

SCOTT@book> delete from job_times;
4 rows deleted.

SCOTT@book> commit ;
Commit complete.

SCOTT@book> select name, gets, misses, sleeps, wait_time from v$latch where name like 'Result Cache%';
NAME                                 GETS     MISSES     SLEEPS  WAIT_TIME
------------------------------ ---------- ---------- ---------- ----------
Result Cache: RC Latch           20385043     535762          5         94
Result Cache: SO Latch                  9          0          0          0
Result Cache: MB Latch                  0          0          0          0

declare
 l_job number;
begin
 for i in 1 .. 4
 loop
  dbms_job.submit(
   job => l_job,
   what => 'do_rc(100000);'
    );
 end loop;
end;
/

PL/SQL procedure successfully completed.

SCOTT@book> commit ;
Commit complete.

--allow jobs to complete

SCOTT@book> select case grouping(sid) when 1 then 'Total:' else to_char(sid) end sid, sum(time_ela) ela from job_times group by rollup((sid, time_ela));
SID       ELA
------ ------
41       3850
54       3853
80       3860
94       3863
Total:  15426
--//我的测试使用Result Cache 更加糟糕!!每个测试需要38秒.而作者的测试两者几乎差不多.作者用 Nothing (almost) 来表达.

SCOTT@book> select name, gets, misses, sleeps, wait_time from v$latch where name like 'Result Cache%';
NAME                                 GETS     MISSES     SLEEPS  WAIT_TIME
------------------------------ ---------- ---------- ---------- ----------
Result Cache: RC Latch           21768802    1045691     663187   64314325
Result Cache: SO Latch                 17          0          0          0
Result Cache: MB Latch                  0          0          0          0

--//我开始以为这里有1个将结果集放入共享池的过程,每一次执行都需要放入共享池.再次调用应该会快一些.
create or replace procedure do_rc(
 p_iterations in number
) is
 l_rowid  rowid;
 l_n number;
begin
 insert into job_times
  values (sys_context('userenv', 'sid'), dbms_utility.get_time)
  returning rowid into l_rowid;

for i in 1 .. p_iterations
 loop
  l_n:=trunc(dbms_random.value(1, 10000));
  for cur in (select /*+ result_cache */ * from hc_t where n=l_n)
  loop
   null;
  end loop;
 end loop;

update job_times set
   time_ela=dbms_utility.get_time-time_ela
  where rowid=l_rowid;
end;
/

--//再次执行:
declare
 l_job number;
begin
 for i in 1 .. 4
 loop
  dbms_job.submit(
   job => l_job,
   what => 'do_rc(100000);'
    );
 end loop;
end;
/

PL/SQL procedure successfully completed.
SCOTT@book> commit ;
Commit complete.

SCOTT@book> select case grouping(sid) when 1 then 'Total:' else to_char(sid) end sid, sum(time_ela) ela from job_times group by rollup((sid, time_ela));
SID     ELA
----- -----
72     3980
81     3900
96     3936
108    3922
Total 15738

--//问题依旧.我估计不同查询存在select /*+ result_cache */ * from hc_t where n=l_n的情况下,探查Result Cache: RC Latch持有
--//时间很长,导致使用result cache更慢,这样看来result_cache更加适合统计类结果不变的语句.而且绑定变量不要变化很多的情况.

--//换成普通表测试看看:
SCOTT@book> rename  hc_t to hc_tx;
Table renamed.

SCOTT@book> create table hc_t as select * from hc_tx ;
Table created.

SCOTT@book> create unique index i_hc_t on hc_t(n);
Index created.

--//分析表略.
--//调用do_hc的情况如下:
SCOTT@book> select count(*),avg(TIME_ELA),sum(TIME_ELA) from job_times ;
  COUNT(*) AVG(TIME_ELA) SUM(TIME_ELA)
---------- ------------- -------------
         4         431.5          1726

--//调用do_rc的情况如下:
SCOTT@book> select count(*),avg(TIME_ELA),sum(TIME_ELA) from job_times ;
  COUNT(*) AVG(TIME_ELA) SUM(TIME_ELA)
---------- ------------- -------------
         4       4027.75         16111

--//结果一样.删除索引在测试看看.
SCOTT@book> drop index i_hc_t ;
Index dropped.

--//调用do_hc的情况如下:
--//delete from job_times;
--//commit ;
SCOTT@book> select count(*),avg(TIME_ELA),sum(TIME_ELA) from job_times ;
  COUNT(*) AVG(TIME_ELA) SUM(TIME_ELA)
---------- ------------- -------------
         4          4160         16640

--//调用do_rc的情况如下:
--//delete from job_times;
--//commit ;
SCOTT@book> select count(*),avg(TIME_ELA),sum(TIME_ELA) from job_times ;
  COUNT(*) AVG(TIME_ELA) SUM(TIME_ELA)
---------- ------------- -------------
         4          3828         15312

--//这个时候result cache优势才显示出来.总之在生产系统使用要注意这个细节,一般result cahe仅仅只读表(dml很少的静态表)外.
--//如果经常使用不同变量查询表,能使用索引的情况,使用result cache毫无优势可言.