【sql调优之执行计划】nested loops join and nested loop join outer

Nested loops（嵌套循环）看一个例子：
SQL> select b.* from scott.emp a,scott.dept b
  2  where a.deptno = b.deptno
  3  and a.empno = 7369
  4  ;
 
    DEPTNO DNAME          LOC
---------- -------------- -------------
        20 RESEARCH       DALLAS
 
 
Execution Plan
----------------------------------------------------------
Plan hash value: 2385808155
 
--------------------------------------------------------------------------------
--------
 
| Id  | Operation                    | Name    | Rows  | Bytes | Cost (%CPU)| Ti
me     |
 
--------------------------------------------------------------------------------
--------
 
|   0 | SELECT STATEMENT             |         |     1 |    27 |     2   (0)| 00
:00:01 |
 
|   1 |  NESTED LOOPS                |         |     1 |    27 |     2   (0)| 00
:00:01 |
 
|   2 |   TABLE ACCESS BY INDEX ROWID| EMP     |     1 |     7 |     1   (0)| 00
:00:01 |
 
|*  3 |    INDEX UNIQUE SCAN         | PK_EMP  |     1 |       |     0   (0)| 00
:00:01 |
 
|   4 |   TABLE ACCESS BY INDEX ROWID| DEPT    |     4 |    80 |     1   (0)| 00
:00:01 |
 
|*  5 |    INDEX UNIQUE SCAN         | PK_DEPT |     1 |       |     0   (0)| 00
:00:01 |
 
--------------------------------------------------------------------------------
--------
 
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
   3 - access("A"."EMPNO"=7369)
   5 - access("A"."DEPTNO"="B"."DEPTNO")
 
 
Statistics
----------------------------------------------------------
          1  recursive calls
          0  db block gets
          4  consistent gets
          0  physical reads
          0  redo size
        533  bytes sent via SQL*Net to client
        400  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
          1  rows processed
 
SQL>
简单的讲就是在表1中每取一行数据，然后从表2中查找匹配的行，然后再回到表1取下一行，如此循环取下去。那么表1中取得的行越少，表2中查找行越容易则嵌套循环的效率越高。
第一个表也通常称为驱动表（或者外部表），第二个表通常称为内部表。
对nested loops join 的选择或者优化，应该考虑驱动表行的选择性和绝对数量大小，并且内部表要能方便的访问到，例如唯一索引，这种结构很多时候用于一个主数据表和一个关联表（或者字典表等）的关联，而关联表很容易使用与主表关联的字段上的索引访问。
 
上述的执行计划可以看出，先访问表emp，emp作为了驱动表，当然也可以使用dept作为驱动表：
 
SQL> select /*+ leading(b) use_nl(a,b)*/b.* from scott.dept b,scott.emp a
  2  where a.deptno = b.deptno
  3  and a.empno = 7369;
 
    DEPTNO DNAME          LOC
---------- -------------- -------------
        20 RESEARCH       DALLAS
 
 
Execution Plan
----------------------------------------------------------
Plan hash value: 3431005640
 
--------------------------------------------------------------------------------
-------
 
| Id  | Operation                    | Name   | Rows  | Bytes | Cost (%CPU)| Tim
e     |
 
--------------------------------------------------------------------------------
-------
 
|   0 | SELECT STATEMENT             |        |     1 |    27 |     7   (0)| 00:
00:01 |
 
|   1 |  NESTED LOOPS                |        |     1 |    27 |     7   (0)| 00:
00:01 |
 
|   2 |   TABLE ACCESS FULL          | DEPT   |     4 |    80 |     3   (0)| 00:
00:01 |
 
|*  3 |   TABLE ACCESS BY INDEX ROWID| EMP    |     1 |     7 |     1   (0)| 00:
00:01 |
 
|*  4 |    INDEX UNIQUE SCAN         | PK_EMP |     1 |       |     0   (0)| 00:
00:01 |
 
--------------------------------------------------------------------------------
-------
 
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
   3 - filter("A"."DEPTNO"="B"."DEPTNO")
   4 - access("A"."EMPNO"=7369)
 
 
Statistics
----------------------------------------------------------
          1  recursive calls
          0  db block gets
         15  consistent gets
          5  physical reads
          0  redo size
        533  bytes sent via SQL*Net to client
        400  bytes received via SQL*Net from client
11034
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
          1  rows processed
 
SQL>
 
上述执行计划先访问dept表（全表扫描），然后通过index访问emp表，可以看到后者的nested loops的cost为7，大于前者的cost2，可以初步判断后者的效率不如前者，或者也可以简单的看看2者的逻辑读数量来判断一下优劣。具体的性能优化结合实例再述。
 
扩展一下，两个数据集做nested loops join和两个表相似。
 
Nested loops join outer 则是由于做了外连接而产生的，例如：
SQL> select b.* from scott.emp a,scott.dept b
  2  where a.deptno = b.deptno(+)
  3  and a.empno = 7369;
 
    DEPTNO DNAME          LOC
---------- -------------- -------------
        20 RESEARCH       DALLAS
 
 
Execution Plan
----------------------------------------------------------
Plan hash value: 1858280091
 
--------------------------------------------------------------------------------
--------
 
| Id  | Operation                    | Name    | Rows  | Bytes | Cost (%CPU)| Ti
me     |
 
--------------------------------------------------------------------------------
--------
 
|   0 | SELECT STATEMENT             |         |     1 |    27 |     2   (0)| 00
:00:01 |
 
|   1 |  NESTED LOOPS OUTER          |         |     1 |    27 |     2   (0)| 00
:00:01 |
 
|   2 |   TABLE ACCESS BY INDEX ROWID| EMP     |     1 |     7 |     1   (0)| 00
:00:01 |
 
|*  3 |    INDEX UNIQUE SCAN         | PK_EMP  |     1 |       |     0   (0)| 00
:00:01 |
 
|   4 |   TABLE ACCESS BY INDEX ROWID| DEPT    |     4 |    80 |     1   (0)| 00
:00:01 |
 
|*  5 |    INDEX UNIQUE SCAN         | PK_DEPT |     1 |       |     0   (0)| 00
:00:01 |
 
--------------------------------------------------------------------------------
--------
 
 
Predicate Information (identified by operation id):
---------------------------------------------------
 
   3 - access("A"."EMPNO"=7369)
   5 - access("A"."DEPTNO"="B"."DEPTNO"(+))
 
 
Statistics
----------------------------------------------------------
          1  recursive calls
          0  db block gets
          4  consistent gets
          0  physical reads
          0  redo size
        533  bytes sent via SQL*Net to client
        400  bytes received via SQL*Net from client
          2  SQL*Net roundtrips to/from client
          0  sorts (memory)
          0  sorts (disk)
          1  rows processed
 
SQL>