[摘]将“相关值”导致的Nested Loop优化成Hash Join

      原文地址:http://www.itpub.net/thread-1376537-1-2.html

      “相关值”是我剽窃的一个词,与之相反的是“不相关值”,我借用来描述以下这种情况:在SQL中有两个表A和B要join,join条件是A.name=B.first_name,判定符“=”左右两边的值是“不相关”的,而substr(A.name,1,length(B.first_name))=B.first_name,判定符“=”左右两边的值是“相关”的,因为要通过“=”右边的值才能求出左边的值(相反亦然)。

      由于这种“相关值”的情况,A、B表join的算法注定只能走Nested Loop了,因为每一个不同的驱动值(驱动表中的值,也就是B表中的值)将改变A表负责join的值,执行计划如下:

select a.*,b.object_name from TEST_OBJECT a, TEST_OBJ1 b
where substr(a.object_name,1,length(b.object_name))=b.object_name;
已选择108612行。

已用时间:  00: 00: 21.54

---------------------------------------------------------------------
| Id  | Operation            |  Name        | Rows  | Bytes | Cost  |
---------------------------------------------------------------------
|   0 | SELECT STATEMENT     |              |       |       |       |
|   1 |  NESTED LOOPS        |              |       |       |       |
|   2 |   TABLE ACCESS FULL  | TEST_OBJ1    |       |       |       |
|*  3 |   TABLE ACCESS FULL  | TEST_OBJECT  |       |       |       |
---------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   3 - filter("B"."OBJECT_NAME"=SUBSTR("A"."OBJECT_NAME",1,LENGTH("B"."OBJ
              ECT_NAME")))

统计信息
----------------------------------------------------------
          1  recursive calls
          0  db block gets
    9262734  consistent gets

      这里展开一下,Nested Loop这个算法复杂度是 O(B*A) (B表示驱动表的记录数,A表示内部表),如果A、B都很大,那就杯具了,关于Nested Loop驱动表的选取就不展开了。

      网友anlinew提供了一种非常精妙的,决的优化方法,将语句改写成:

select a.*,b.object_name from TEST_OBJECT a, TEST_OBJ1 b
where substr(a.object_name,1,length(b.object_name))=b.object_name
and substr(a.object_name,1,4)=substr(b.object_name,1,4)  
and length(b.object_name)>3
union all
select a.*,b.object_name from TEST_OBJECT a, TEST_OBJ1 b
where substr(a.object_name,1,length(b.object_name))=b.object_name
and length(b.object_name)<4;
已选择108612行。

已用时间:  00: 06: 51.24

-----------------------------------------------------------------------------------
| Id  | Operation           | Name        | Rows  | Bytes | Cost (%CPU)| Time     |
-----------------------------------------------------------------------------------
|   0 | SELECT STATEMENT    |             |  3468 |   365K| 90997 (100)| 00:18:12 |
|   1 |  UNION-ALL          |             |       |       |            |          |
|*  2 |   HASH JOIN         |             |    34 |  3672 |   176   (2)| 00:00:03 |
|*  3 |    TABLE ACCESS FULL| TEST_OBJ1   |   659 | 12521 |    35   (0)| 00:00:01 |
|   4 |    TABLE ACCESS FULL| TEST_OBJECT | 52544 |  4566K|   139   (1)| 00:00:02 |
|   5 |   NESTED LOOPS      |             |  3434 |   362K| 90821   (1)| 00:18:10 |
|*  6 |    TABLE ACCESS FULL| TEST_OBJ1   |   659 | 12521 |    35   (0)| 00:00:01 |
|*  7 |    TABLE ACCESS FULL| TEST_OBJECT |     5 |   445 |   138   (1)| 00:00:02 |
-----------------------------------------------------------------------------------

Predicate Information (identified by operation id):
---------------------------------------------------

   2 - access(SUBSTR("A"."OBJECT_NAME",1,4)=SUBSTR("B"."OBJECT_NAME",1,4))
       filter("B"."OBJECT_NAME"=SUBSTR("A"."OBJECT_NAME",1,LENGTH("B"."OBJE
              CT_NAME")))
   3 - filter(LENGTH("B"."OBJECT_NAME")>3)
   6 - filter(LENGTH("B"."OBJECT_NAME")<4)
   7 - filter("B"."OBJECT_NAME"=SUBSTR("A"."OBJECT_NAME",1,LENGTH("B"."OBJE
              CT_NAME")))

统计信息
----------------------------------------------------------
          1  recursive calls
          0  db block gets
      33946  consistent gets

      执行计划虽然变复杂了,但是耗时大幅减少,consistent gets也大幅降低,作出巨大贡献的是Hash Join的引入。

      这里再展开一下,Hash Join的复杂度是O(A+B),简单来说就是对A、B表各扫描一次,如果A、B都比较大的情况来看,无疑Hash Join要比Nested Loop 优越很多。

      扯远了,回到anlinew的具体方法上吧,导致Hash Join出现的关键因素是一个谓词的引入:

and substr(a.object_name,1,4)=substr(b.object_name,1,4)

      套用我剽窃的那个词来说,这是“不相关值”的对比!

      anlinew的核心思想是将数据“分片”,该例子中分片的依据是多少位首字母(这里是4),其中“大头”由Hash Join处理,而“小头”走Nested Loop,这种“抓大放小”的做法直接就从复杂度上进行了优化。

      这种“分片”的思想非常值得借鉴,将“相关值”判断转化成“不相关值”的判断也是处理问题的一种有效手法。

posted @ 2011-04-11 22:46 killkill 阅读(...) 评论(...) 编辑 收藏