上周运营商客户的计费库反应其入库程序很慢,应用方通过监控程序发现主要慢在对于几个表的insert操作上。按照我们的通常理解,insert应该是极快的，为什么会很慢呢？而且反应之前挺好的。这有点让我百思不得其解。通过检查event也并没有发现什么奇怪的地方,于是我通过10046 跟踪了应用的入库程序,如下应用方反应比较慢的表的insert操作，确实非常慢，如下所示：
| 代码如下 |
复制代码 |
|
INSERT INTO XXXX_EVENT_201605C (ROAMING_NBR,.....,OFFER_INSTANCE_ID4)
VALUES
(:ROAMING_NBR,.....:OFFER_INSTANCE_ID4)
call count cpu elapsed disk query current rows
------- ------ -------- ---------- ---------- ---------- ---------- ----------
Parse 17 0.00 0.00 0 0 0 0
Execute 18 1.06 27.41 4534 518 33976 4579
Fetch 0 0.00 0.00 0 0 0 0
------- ------ -------- ---------- ---------- ---------- ---------- ----------
total 35 1.06 27.41 4534 518 33976 4579
Misses in library cache during parse: 1
Misses in library cache during execute: 1
Optimizer mode: ALL_ROWS
Parsing user id: 102
Elapsed times include waiting on following events:
Event waited on Times Max. Wait Total Waited
---------------------------------------- Waited ---------- ------------
db file sequential read 4495 0.03 24.02
gc current grant 2-way 2301 0.00 0.77
SQL*Net more data from client 795 0.00 0.02
......
latch: gcs resource hash 1 0.00 0.00
|
我们可以发现,insert了4579条数据,一共花了27.41秒；其中有24.02秒是处于等待的状态。而且等待事件为顺序读.
很明显这通常是索引的读取操作，实际上检查10046 trace 裸文件，发现等待的对象确实是该表上的2个index。
同时我们从上面10046 trace可以看出，该SQL执行之所以很慢，主要是因为存在了大量的物理读，其中4579条数据的insert，
物理读为4534;这说明什么问题呢？ 这说明,每插入一条数据大概产生一个物理读，而且都是index block的读取。
很明显，通过将该index cache到keep 池可以解决该问题。 实际上也确实如此，通过cache后，应用反馈程序快了很多。
那么对该问题，这里其实有几个疑问，为什么这里的SQL insert时物理读如此之高？ oracle的keep pool对于缓存对象
的清理机制是如何的？
下面我们通过一个简单的实验来进行说明。
首先我们创建2个测试表，并创建好相应的index，如下所示：
| 代码如下 |
复制代码 |
SQL> conn roger/roger
Connected.
SQL> create table t_insert as select * from sys.dba_objects where 1=1;
Table created.
SQL> create index idx_name_t on t_insert(object_name);
Index created.
SQL> analyze table t_insert compute statistics for all indexed columns;
Table analyzed.
SQL> select INDEX_NAME,BLEVEL,LEAF_BLOCKS,DISTINCT_KEYS,CLUSTERING_FACTOR,NUM_ROWS from dba_indexes where table_name='T_INSERT';
INDEX_NAME BLEVEL LEAF_BLOCKS DISTINCT_KEYS CLUSTERING_FACTOR NUM_ROWS
------------- ---------- ----------- ------------- ----------------- ----------
IDX_NAME_T 1 246 29808 24664 49859
SQL> show parameter db_keep
NAME TYPE VALUE
------------------------------------ ----------- --------
db_keep_cache_size big integer 0
SQL> alter system set db_keep_cache_size=4m;
System altered.
SQL> create table t_insert2 as select * from sys.dba_objects where 1=1;
Table created.
SQL> create index idx_name_t2 on t_insert2(object_name);
Index created.
SQL> insert into t_insert select * from sys.dba_objects;
49862 rows created.
SQL> commit;
Commit complete.
SQL> insert into t_insert2 select * from sys.dba_objects;
49862 rows created.
SQL> commit;
Commit complete.
|
从前面的信息我们可以看出，object_name上的index其实聚簇因子比较高，说明其数据分布比较离散。
接着我们现在将index都cache 到keep 池中，如下：
| 代码如下 |
复制代码 |
|
SQL> alter index idx_name_t storage (buffer_pool keep);
Index altered.
SQL> alter index idx_name_t2 storage (buffer_pool keep);
Index altered.
SQL> alter system flush buffer_cache;
System altered.
|
这里需要注意的是，仅仅执行alter 命令是不够的，我们还需要手工将index block读取到keep池中，如下：
| 代码如下 |
复制代码 |
SQL> conn /as sysdba
Connected.
SQL> @get_keep_pool_obj.sql
no rows selected
SQL> select /*+ index(idx_name_t,t_insert) */ count(object_name) from roger.t_insert;
COUNT(OBJECT_NAME)
------------------
99721
SQL> @get_keep_pool_obj.sql
SUBCACHE OBJECT_NAME BLOCKS
------------ ------------------------------ ----------
KEEP IDX_NAME_T 499
DEFAULT T_INSERT 431
SQL> select /*+ index(idx_name_t2,t_insert2) */ count(object_name) from roger.t_insert2;
COUNT(OBJECT_NAME)
------------------
99723
SQL> @get_keep_pool_obj.sql
SUBCACHE OBJECT_NAME BLOCKS
------------ ------------------------------ ----------
KEEP IDX_NAME_T 40
KEEP IDX_NAME_T2 459
DEFAULT T_INSERT2 522
DEFAULT T_INSERT 431
SQL> select /*+ index(idx_name_t,t_insert) */ count(object_name) from roger.t_insert;
COUNT(OBJECT_NAME)
------------------
99721
SQL> @get_keep_pool_obj.sql
SUBCACHE OBJECT_NAME BLOCKS
------------ ------------------------------ ----------
KEEP IDX_NAME_T 467
KEEP IDX_NAME_T2 32
DEFAULT T_INSERT2 522
DEFAULT T_INSERT 431
SQL>
|
我们可以大致看出，db keep pool 也是存在LRU的，而且对于block的清除机制是先进先出原则。那么为什么前面的问题中，insert会突然变慢呢？
下面我们来进行3次insert 测试。
| 代码如下 |
复制代码 |
|
#### one
SQL> select /*+ index_ffs(idx_name_t,t_insert) */ count(object_name) from roger.t_insert;
COUNT(OBJECT_NAME)
------------------
149583
SQL> @get_keep_pool_obj.sql
SUBCACHE OBJECT_NAME BLOCKS
------------ ------------------------------ ----------
DEFAULT SQLPLUS_PRODUCT_PROFILE 1
DEFAULT RLM$SCHACTIONORDER 1
DEFAULT RLM$JOINQKEY 1
KEEP IDX_NAME_T 499
DEFAULT T_INSERT2 2113
DEFAULT T_INSERT 2113
6 rows selected.
SQL> oradebug setmypid
Statement processed.
SQL> oradebug event 10046 trace name context forever,level 12
Statement processed.
SQL> set timing on
SQL> insert into roger.t_insert select * from sys.dba_objects;
49862 rows created.
Elapsed: 00:00:03.28
SQL> commit;
Commit complete.
Elapsed: 00:00:00.00
SQL> oradebug tracefile_name
/home/oracle/admin/test/udump/test_ora_11661.trc
++++10046 trace
insert into roger.t_insert select * from sys.dba_objects
call count cpu elapsed disk query current rows
------- ------ -------- ---------- ---------- ---------- ---------- ----------
Parse 1 0.01 0.01 0 0 0 0
Execute 1 0.95 3.07 3289 11592 96374 49862
Fetch 0 0.00 0.00 0 0 0 0
------- ------ -------- ---------- ---------- ---------- ---------- ----------
total 2 0.96 3.08 3289 11592 96374 49862
.....
Elapsed times include waiting on following events:
Event waited on Times Max. Wait Total Waited
---------------------------------------- Waited ---------- ------------
db file sequential read 3168 0.00 0.50
db file scattered read 1 0.00 0.00
#### two
SQL> oradebug setmypid
Statement processed.
SQL> oradebug event 10046 trace name context forever,level 12
Statement processed.
SQL> oradebug tracefile_name
/home/oracle/admin/test/udump/test_ora_13163.trc
SQL> set timing on
SQL> insert into roger.t_insert select * from sys.dba_objects;
49825 rows created.
++++10046 trace
insert into roger.t_insert select * from sys.dba_objects
call count cpu elapsed disk query current rows
------- ------ -------- ---------- ---------- ---------- ---------- ----------
Parse 1 0.01 0.01 0 0 0 0
Execute 1 0.87 3.10 3817 8134 87352 49825
Fetch 0 0.00 0.00 0 0 0 0
------- ------ -------- ---------- ---------- ---------- ---------- ----------
total 2 0.88 3.11 3817 8134 87352 49825
.....
Elapsed times include waiting on following events:
Event waited on Times Max. Wait Total Waited
---------------------------------------- Waited ---------- ------------
db file sequential read 3827 0.00 0.56
#### three
SQL> oradebug setmypid
Statement processed.
SQL> oradebug event 10046 trace name context forever,level 12
Statement processed.
SQL> set timing on
SQL> insert into roger.t_insert select * from sys.dba_objects;
49825 rows created.
Elapsed: 00:00:03.94
SQL> commit;
Commit complete.
Elapsed: 00:00:00.01
SQL> oradebug tracefile_name
/home/oracle/admin/test/udump/test_ora_13286.trc
SQL> select /*+ index_ffs(idx_name_t,t_insert) */ count(object_name) from roger.t_insert;
COUNT(OBJECT_NAME)
------------------
249233
SQL> @get_keep_pool_obj.sql
SUBCACHE OBJECT_NAME BLOCKS
------------ ------------------------------ ----------
DEFAULT SQLPLUS_PRODUCT_PROFILE 1
......
DEFAULT RLM$JOINQKEY 1
KEEP IDX_NAME_T 499
++++10046 trace
insert into roger.t_insert select * from sys.dba_objects
call count cpu elapsed disk query current rows
------- ------ -------- ---------- ---------- ---------- ---------- ----------
Parse 1 0.00 0.00 0 0 0 0
Execute 1 1.60 3.84 7598 13208 104820 49825
Fetch 0 0.00 0.00 0 0 0 0
------- ------ -------- ---------- ---------- ---------- ---------- ----------
total 2 1.60 3.84 7598 13208 104820 49825
.....
Elapsed times include waiting on following events:
Event waited on Times Max. Wait Total Waited
---------------------------------------- Waited ---------- ------------
db file sequential read 7618 0.00 1.07
free buffer waits 224 0.02 1.30
|
从测试来看,随着表的数据越来越大，insert的效率会越来越低，也其实主要在于index的问题。
我们可以发现,3次测试过程中,物理读越来越大，而且db file sequential read的等待时间分别从0.5秒，增加到0.56秒，最后增加到1.07秒。 为什么会出现这样的情况呢？
随着表数据的日益增加,导致表上的index也不断增大,同时index的离散度比较高，这样就导致每次insert时，oracle在进行index block读取时，可能在buffer cache中都无法命中相应的block；这样就会导致每次读取需要的index block时，可能都要进行物理读，这势必会导致性能问题的出现。
同时默认的default buffer cache pool虽然也可以缓存index 块，但是也要同时缓存其他的数据块，这样很容易导致
相关的index block被从buffer cache pool中移走。所以这也是前面为什么需要将index cache到keep 池的原因。
