摘要：大家都知道oracle的性能优化，不仅仅是在数据库实例的结构上做文章，同样我们也要在SQL语句上重点着手。下面我们要说的呢，就是分布式SQL的执行机制和一些优化方法。
1. 背景介绍
分布式查询语句对于远程对象的查询在远程库执行，在远程库可以执行的SQL语句会通过优化器的查询转换，执行的是转换后的语句，然后结果集返回到本地，再与本地表运算。当然，
本地操作还是远程操作是相对的，我们可以通过driving_site hint改变主查询计划的执行位置，但是对DML，driving_site是失效的,另外对远程表也可以使用其他hint来控制执行计划。
2. 优化目标
分布式查询语句中可能有不同远程库的表，优化分布式查询要达到3点目标：
1.访问同一个远程库的次数要尽量少,也就是同一远程库的对象应该尽量转为一条SQL运算，一次运算，运算 后将结果返回给本地库。
2.从远程库上返回到本地库的结果集要尽量小，只访问远程对象需要的字段，从而减少网络传输。
3.远程库上执行的语句的执行计划、远程库返回的结果与本地联合查询的计划应该比较高效。
如果能够达到以上3点，一般情况下，分布式查询的效率较高。
3. 优化方法
优化分布式查询需要从以上3个方面着手。
以下样例中，local_tab 7万多条，remote_big_tab百万条，remote_small_tab 7万多条。
3.1 使用Collocated内联视图
也就是说，SQL要引用不同远程库的表，需要组织好语句结构：将相同库的表放一起组成内联视图，这样ORACLE就很容易知道这个内联视图里的表是在同一远程库作完查询，然后再返回给本地库，这样减少了本地库与远程库的交互次数、传输结果集的数量和次数，从而提高效率。比如查询：
SELECT * FROM local_tab a
WHERE EXISTS
(SELECT 1 FROM remote_big_tab@remote b,remote_small_tab@remote c
WHERE b.object_id=c.object_id AND a.object_type=b.object_type);
执行计划
———————————————————-
Plan hash value: 49311412
————————————————————————————————
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | Inst |IN-OUT|
————————————————————————————————
| 0 | SELECT STATEMENT | | 42747 | 4508K| 2152 (4)| 00:00:26 | | |
|* 1 | HASH JOIN | | 42747 | 4508K| 2152 (4)| 00:00:26 | | |
| 2 | VIEW | VW_SQ_1 | 26 | 286 | 1855 (4)| 00:00:23 | | |
| 3 | REMOTE | | | | | | REMOTE | R->S |
| 4 | TABLE ACCESS FULL| LOCAL_TAB | 73985 | 7008K| 296 (1)| 00:00:04 | | |
————————————————————————————————
Predicate Information (identified by operation id):
—————————————————
1 – access(“A”.”OBJECT_TYPE”=”ITEM_0″)
Remote SQL Information (identified by operation id):
—————————————————-
3 – EXPLAIN PLAN SET STATEMENT_ID=’PLUS5801659′ INTO PLAN_TABLE@! FOR SELECT
DISTINCT “A2″.”OBJECT_TYPE” FROM “REMOTE_BIG_TAB” “A2″,”REMOTE_SMALL_TAB” “A1” WHERE
“A2″.”OBJECT_ID”=”A1″.”OBJECT_ID” (accessing ‘REMOTE’ )
可以看出，在远程库remote上执行的语句是两个远程表关联后，并经过查询转换（全转为大写，自己取了别名A1，A2，ORACLE内部自己构造查询语句SELECT DISTINCT…，之后远程查询结果返回给本地，可以去远程库里查询实际的计划，走的是HASH JOIN。通过以上案例看出，对于一些复杂的查询，对象来源于不同远程库，能够通过SQL改写将相同远程库的表先做JOIN，可以提高效率。
3.2 了解CBO优化器对分布式查询的处理
CBO对分布式查询的处理，也是尽量转为Collocated内联视图，CBO会做如下动作：
1 ）所有可mergeable的视图会merge。
2 ) CBO会测试Collocated内联视图的query BLOCK。
3 ) 如果可以将相同库的表组合成SQL一起查询，那么就会一起执行。
当然，CBO对分布式查询的处理，可能是不高效的，这时候得用其他的方法，比如使用HINT,改造SQL，改造分布式查询的方法（远程库用视图）等。
比如下面语句含有子查询：
SELECT * FROM local_tab a,remote_big_tab@remote b,remote_small_tab@remote c
WHERE b.object_id=c.object_id AND a.object_type=b.object_type
AND a.object_id IN (SELECT object_id from sub);
执行计划
———————————————————-
Plan hash value: 252158753
———————————————————————————————————-
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | Inst |IN-OUT|
———————————————————————————————————-
| 0 | SELECT STATEMENT | | 79M| 20G| 3843 (46)| 00:00:47 | | |
|* 1 | HASH JOIN | | 79M| 20G| 3843 (46)| 00:00:47 | | |
| 2 | REMOTE | REMOTE_SMALL_TAB | 5320 | 431K| 8 (0)| 00:00:01 | REMOTE | R->S |
|* 3 | HASH JOIN | | 172M| 31G| 2978 (31)| 00:00:36 | | |
|* 4 | HASH JOIN | | 5260 | 565K| 303 (2)| 00:00:04 | | |
| 5 | SORT UNIQUE | | 5320 | 69160 | 5 (0)| 00:00:01 | | |
| 6 | TABLE ACCESS FULL| SUB | 5320 | 69160 | 5 (0)| 00:00:01 | | |
| 7 | TABLE ACCESS FULL | LOCAL_TAB | 73985 | 7008K| 296 (1)| 00:00:04 | | |
| 8 | REMOTE | REMOTE_BIG_TAB | 1479K| 119M| 1819 (2)| 00:00:22 | REMOTE | R->S |
———————————————————————————————————-
Predicate Information (identified by operation id):
—————————————————
1 – access(“B”.”OBJECT_ID”=”C”.”OBJECT_ID”)
3 – access(“A”.”OBJECT_TYPE”=”B”.”OBJECT_TYPE”)
4 – access(“A”.”OBJECT_ID”=”OBJECT_ID”)
Remote SQL Information (identified by operation id):
—————————————————-
2 – SELECT “OBJECT_NAME”,”SUBOBJECT_NAME”,”OBJECT_ID”,”DATA_OBJECT_ID”,”OBJECT_TYPE”,”CREATED”,
“LAST_DDL_TIME”,”TIMESTAMP”,”STATUS”,”TEMPORARY”,”GENERATED”,”SECONDARY” FROM “REMOTE_SMALL_TAB”
“C” (accessing ‘REMOTE’ )
8 – SELECT “OWNER”,”OBJECT_NAME”,”SUBOBJECT_NAME”,”OBJECT_ID”,”DATA_OBJECT_ID”,”OBJECT_TYPE”,”C
REATED”,”LAST_DDL_TIME”,”TIMESTAMP”,”STATUS”,”TEMPORARY”,”GENERATED”,”SECONDARY” FROM
“REMOTE_BIG_TAB” “B” (accessing ‘REMOTE’ )
通过计划可以看到REMOTE有两条SQL，两张远程表无法做Collocated inline VIEW运算。
再比如下面的语句，有分组运算：
SELECT * FROM local_tab a,remote_big_tab@remote b,(SELECT max(object_id) object_id FROM remote_small_tab@remote c GROUP BY c.object_type) c
WHERE b.object_id=c.object_id AND a.object_type=b.object_type;
执行计划
———————————————————-
Plan hash value: 2122363341
—————————————————————————————————–
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | Inst |IN-OUT|
—————————————————————————————————–
| 0 | SELECT STATEMENT | | 2321K| 431M| 2144 (3)| 00:00:26 | | |
|* 1 | HASH JOIN | | 2321K| 431M| 2144 (3)| 00:00:26 | | |
|* 2 | HASH JOIN | | 1412 | 135K| 1836 (3)| 00:00:23 | | |
| 3 | VIEW | | 11 | 143 | 9 (12)| 00:00:01 | | |
| 4 | REMOTE | | | | | | REMOTE | R->S |
| 5 | REMOTE | REMOTE_BIG_TAB | 1479K| 119M| 1819 (2)| 00:00:22 | REMOTE | R->S |
| 6 | TABLE ACCESS FULL| LOCAL_TAB | 73985 | 7008K| 296 (1)| 00:00:04 | | |
—————————————————————————————————–
Predicate Information (identified by operation id):
—————————————————
1 – access(“A”.”OBJECT_TYPE”=”B”.”OBJECT_TYPE”)
2 – access(“B”.”OBJECT_ID”=”C”.”OBJECT_ID”)
Remote SQL Information (identified by operation id):
—————————————————-
4 – EXPLAIN PLAN SET STATEMENT_ID=’PLUS5801659′ INTO PLAN_TABLE@! FOR SELECT
MAX(“A1″.”OBJECT_ID”) FROM “REMOTE_SMALL_TAB” “A1” GROUP BY “A1″.”OBJECT_TYPE” (accessing
‘REMOTE’ )
5 – SELECT “OWNER”,”OBJECT_NAME”,”SUBOBJECT_NAME”,”OBJECT_ID”,”DATA_OBJECT_ID”,”OBJECT_TYP
E”,”CREATED”,”LAST_DDL_TIME”,”TIMESTAMP”,”STATUS”,”TEMPORARY”,”GENERATED”,”SECONDARY” FROM
“REMOTE_BIG_TAB” “B” (accessing ‘REMOTE’ )
通过计划看出，将远程表进行分组运算后，传输给本地库，然后大表传输给本地库，之后做HASH JOIN，这是不高效的。运行时间：已用时间: 00: 02: 12.22
可以改造分布式查询，手动组织Collocated inline VIEW,在远程库建立view:
CREATE OR REPLACE VIEW v_remote
AS
SELECT b.* FROM remote_big_tab b,(SELECT max(object_id) object_id FROM remote_small_tab c GROUP BY c.object_type) c
WHERE b.object_id=c.object_id;
查询改为：
SELECT * FROM local_tab a,v_remote@remote v WHERE a.object_type=v.object_type;
SQL> SELECT * FROM local_tab a,v_remote@remote v WHERE a.object_type=v.object_type;
已选择1727104行。
已用时间: 00: 01: 02.81
执行计划
———————————————————-
Plan hash value: 2216230941
————————————————————————————————
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | Inst |IN-OUT|
————————————————————————————————
| 0 | SELECT STATEMENT | | 1 | 274 | 299 (2)| 00:00:04 | | |
|* 1 | HASH JOIN | | 1 | 274 | 299 (2)| 00:00:04 | | |
| 2 | REMOTE | V_REMOTE | 1 | 177 | 2 (0)| 00:00:01 | REMOTE | R->S |
| 3 | TABLE ACCESS FULL| LOCAL_TAB | 73985 | 7008K| 296 (1)| 00:00:04 | | |
————————————————————————————————
Predicate Information (identified by operation id):
—————————————————
1 – access(“A”.”OBJECT_TYPE”=”V”.”OBJECT_TYPE”)
通过计划可以看出，现在是远程表做整体操作之后才返回到本地了，大大减少远程库查询执行次数以及网络传输量。
3.3 使用HINT，特别是driving_site HINT
对远程表可以使用hint,比如parallel,use_nl,use_hash,FULL等。
driving_site hint能够指定执行计划在远程还是本地做，比如下面使用driving_site(b),那么SQL就是放到远程去执行，也就是原来的远程表就相当于本地表，本地表要传输给remote库，主计划在remote库上执行：
SELECT/*+driving_site(b)*/ * FROM local_tab a,remote_big_tab@remote b,(SELECT max(object_id) object_id
FROM remote_small_tab@remote c GROUP BY c.object_type) c
WHERE b.object_id=c.object_id AND a.object_type=b.object_type;
当然，如果是driving_site(a)那么就是本地驱动的，默认的是本地驱动的。
使用driving_site,特别是本地小结果集，远程大结果集的时候，总体结果集较小，希望计划在远程驱动，这样远程执行完毕，将结果集传输到本地，这样避免大结果集的传输。
例1：小表9998条，大表3169376条记录,远程大表sub_id,acc_id上联合索引
SQL> SELECT COUNT(*) FROM small_tab_local a, big_tab_remote@remote b
2 WHERE a.sub_id=b.sub_id AND a.acc_id=b.acc_id;
————
859
已用时间: 00: 00: 50.76
执行计划
———————————————————-
Plan hash value: 1507576754
——————————————————————————————————–
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | Inst |IN-OUT|
——————————————————————————————————–
| 0 | SELECT STATEMENT | | 1 | 41 | 44 (3)| 00:00:01 | | |
| 1 | SORT AGGREGATE | | 1 | 41 | | | | |
| 2 | MERGE JOIN | | 9998 | 400K| 44 (3)| 00:00:01 | | |
| 3 | REMOTE | BIG_TAB_REMOTE | 6771K| 167M| 26 (0)| 00:00:01 | REMOTE | R->S |
|* 4 | SORT JOIN | | 9998 | 146K| 18 (6)| 00:00:01 | | |
| 5 | TABLE ACCESS FULL| SMALL_TAB_LOCAL | 9998 | 146K| 17 (0)| 00:00:01 | | |
——————————————————————————————————–
Predicate Information (identified by operation id):
—————————————————
4 – access(“A”.”SUB_ID”=”B”.”SUB_ID” AND “A”.”ACC_ID”=”B”.”ACC_ID”)
filter(“A”.”ACC_ID”=”B”.”ACC_ID” AND “A”.”SUB_ID”=”B”.”SUB_ID”)
Remote SQL Information (identified by operation id):
—————————————————-
3 – SELECT “SUB_ID”,”ACC_ID” FROM “BIG_TAB_REMOTE” “B” ORDER BY “SUB_ID”,”ACC_ID”
(accessing ‘REMOTE’ )
查询876条数据，耗时50s,显然将大结果集拉到本地做运算是不好的，因为本地表很小，远程大表有索引，如果能在远端执行，并走nl,那么显然效率非常好。使用driving_site hint改造查询如下：
SELECT/*+driving_site(b) ordered use_nl(b)*/ COUNT(*) FROM small_tab_local a, big_tab_remote@remote b
WHERE a.sub_id=b.sub_id AND a.acc_id=b.acc_id;
计划如下：
———————————————————————————————
| Id | Operation | Name | Rows | Bytes | Cost | Inst |IN-OUT|
———————————————————————————————
| 0 | SELECT STATEMENT REMOTE| | 1 | 52 | 10009 | | |
| 1 | SORT AGGREGATE | | 1 | 52 | | | |
| 2 | NESTED LOOPS | | 681 | 35412 | 10009 | | |
| 3 | REMOTE | | 9998 | 253K| 11 | ! | R->S |
|* 4 | INDEX RANGE SCAN | IDX_BIG_TAB_REMOTE | 1 | 26 | 1 | MZT~ | |
———————————————————————————————
Predicate Information (identified by operation id):
—————————————————
4 – access(“A2″.”SUB_ID”=”A1″.”SUB_ID” AND “A2″.”ACC_ID”=”A1″.”ACC_ID”)
Remote SQL Information (identified by operation id):
—————————————————-
3 – SELECT “SUB_ID”,”ACC_ID” FROM “SMALL_TAB_LOCAL” “A2” (accessing ‘!’ )
现在主计划是在远端remote上执行的，本地表small_tab_local变成了远程表，ORACLE会将small_tab_local结果集送到远端，只查询了sub_id,acc_id,然后作为驱动表，与远端表做NESTED LOOPS运算，计划里可以看到远端表走索引了，最后将远端执行结果返回给本地。
CREATE TABLE test_cnt (cnt NUMBER);
INSERT INTO test_cnt
SELECT/*+driving_site(b) ordered use_nl(b)*/ COUNT(*) FROM small_tab_local a, big_tab_remote@remote b
WHERE a.sub_id=b.sub_id AND a.acc_id=b.acc_id;
已用时间: 00: 01: 31.48
执行计划
———————————————————-
Plan hash value: 259989953
————————————————————————————————————
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | Inst |IN-OUT|
————————————————————————————————————
| 0 | INSERT STATEMENT | | 1 | 41 | 10035 (1)| 00:02:01 | | |
| 1 | LOAD TABLE CONVENTIONAL | TEST_CNT | | | | | | |
| 2 | SORT AGGREGATE | | 1 | 41 | | | | |
| 3 | NESTED LOOPS | | 9998 | 400K| 10035 (1)| 00:02:01 | | |
| 4 | TABLE ACCESS FULL | SMALL_TAB_LOCAL | 9998 | 146K| 17 (0)| 00:00:01 | | |
| 5 | REMOTE | BIG_TAB_REMOTE | 1 | 26 | 1 (0)| 00:00:01 | REMOTE | R->S |
————————————————————————————————————
Remote SQL Information (identified by operation id):
—————————————————-
5 – SELECT /*+ OPAQUE_TRANSFORM USE_NL (“B”) */ “SUB_ID”,”ACC_ID” FROM “BIG_TAB_REMOTE” “B”
WHERE :1=”SUB_ID” AND :2=”ACC_ID” (accessing ‘REMOTE’ )
语句执行1分31s,driving_site hint失效，但是后面的NL没有失效，可以从计划中看出类似绑定变量的参数，这实际上是对于每个small_tab_local的结果集的行，将sub_id,acc_id传给远端表big_tab_remote,也就是变量:1，:2，这样本地的表筛选出多少行，远程语句SELECT /*+ OPAQUE_TRANSFORM USE_NL (“B”) */ “SUB_ID”,”ACC_ID” FROM “BIG_TAB_REMOTE” “B”
WHERE :1=”SUB_ID” AND :2=”ACC_ID” 就执行多少次。
这里本地表9998条，无过滤条件，因此远程表语句运行了9998次，虽然远程查询也是走索引的，但是SQL被执行了9998次，是非常影响性能的。可以去远程库查询下：
SQL> SELECT sql_text,executions FROM v$sql WHERE sql_text LIKE ‘%SELECT /*+ USE_NL (“B”) */ “SUB_ID”,”ACC_ID” FROM “BIG_TAB_REMOTE”%’
2 /
SQL_TEXT EXECUTIONS
——————————————————————————– ————————————————————–
SELECT /*+ USE_NL (“B”) */ “SUB_ID”,”ACC_ID” FROM “BIG_TAB_REMOTE” “B” WHERE :1= 9998
这里driving_site失效，但是后面的nl还有效，远程表执行的次数是small_tab_local表的数量（因为这里没有谓词过滤small_tab_local),可以使用其他hint,比如。
INSERT INTO test_cnt
SELECT/*+ordered use_hash(b)*/
COUNT(*) FROM small_tab_local a,big_tab_remote@remote b
WHERE a.sub_id=b.sub_id AND a.acc_id=b.acc_id;
当然效率不一定很好，因为这里由远程驱动效率最好，为了不想driving_site失效，可以使用PL/SQL(这里是只查询数量，如果查询结果集可以使用PL/SQL批处理插入）。
BEGIN
FOR i IN (SELECT/*+driving_site(b) ordered use_nl(b)*/ COUNT(*) cnt FROM small_tab_local a, big_tab_remote@remote b
WHERE a.sub_id=b.sub_id AND a.acc_id=b.acc_id)
LOOP
INSERT INTO test_cnt VALUES(i.cnt);
END LOOP;
COMMIT;
END;
/
已用时间: 00: 00: 00.89
例2：
查询语句：
SELECT * FROM v_remote WHERE object_id IN (
SELECT c.object_id FROM c WHERE c.object_name
IN (SELECT d.object_name FROM d WHERE d.object_id=11)
);
比较慢，返回32行，需要10多秒。其中v_remote是个视图，此视图连接到远程表，其中远程的两张表的object_id都有索引：
CREATE OR REPLACE VIEW v_remote
AS
SELECT object_name,object_id,object_type FROM a@remote
UNION ALL
SELECT object_name,object_id,object_type FROM b@remote;
两表记录数如下：
SQL> SELECT COUNT(*) FROM a;
COUNT(*)
———-
369888
SQL> SELECT COUNT(*) FROM b;
COUNT(*)
———-
5323
c和d是本地表，d.object_id以及c.object_name有索引。单独查询很快，<1s就会返回：
–单独本地语句消耗时间00: 00: 00.01
SQL> SELECT c.object_id FROM c WHERE c.object_name IN (SELECT d.object_name FROM d WHERE d.object_id=11);
已用时间: 00: 00: 00.01
执行计划
———————————————————-
Plan hash value: 2528799293
—————————————————————————————-
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time |
—————————————————————————————-
| 0 | SELECT STATEMENT | | 2 | 94 | 6 (17)| 00:00:01 |
| 1 | NESTED LOOPS | | | | | |
| 2 | NESTED LOOPS | | 2 | 94 | 6 (17)| 00:00:01 |
| 3 | SORT UNIQUE | | 1 | 17 | 2 (0)| 00:00:01 |
| 4 | TABLE ACCESS BY INDEX ROWID| D | 1 | 17 | 2 (0)| 00:00:01 |
|* 5 | INDEX RANGE SCAN | IDX_D | 1 | | 1 (0)| 00:00:01 |
|* 6 | INDEX RANGE SCAN | IDX_C | 2 | | 2 (0)| 00:00:01 |
| 7 | TABLE ACCESS BY INDEX ROWID | C | 2 | 60 | 3 (0)| 00:00:01 |
—————————————————————————————-
Predicate Information (identified by operation id):
—————————————————
5 – access(“D”.”OBJECT_ID”=11)
6 – access(“C”.”OBJECT_NAME”=”D”.”OBJECT_NAME”)
–单独远程语句消耗时间 00: 00: 00.06
SQL> SELECT * FROM v_remote WHERE object_id=11;
已选择32行。
已用时间: 00: 00: 00.06
执行计划
———————————————————-
Plan hash value: 1788691278
————————————————————–
| Id | Operation | Name | Cost (%CPU)| Inst |IN-OUT|
————————————————————–
| 0 | SELECT STATEMENT | | 0 (0)| | |
| 1 | REMOTE | | | REMOTE | R->S |
————————————————————–
Remote SQL Information (identified by operation id):
—————————————————-
1 – EXPLAIN PLAN SET STATEMENT_ID=’PLUS5821518′ INTO PLAN_TABLE@!
FOR SELECT “A1″.”OBJECT_NAME”,”A1″.”OBJECT_ID”,”A1″.”OBJECT_TYPE” FROM
( (SELECT “A4″.”OBJECT_NAME” “OBJECT_NAME”,”A4″.”OBJECT_ID”
“OBJECT_ID”,”A4″.”OBJECT_TYPE” “OBJECT_TYPE” FROM “A” “A4” WHERE
“A4”.”OBJECT_ID”=11) UNION ALL (SELECT “A3″.”OBJECT_NAME”
“OBJECT_NAME”,”A3″.”OBJECT_ID” “OBJECT_ID”,”A3″.”OBJECT_TYPE”
“OBJECT_TYPE” FROM “B” “A3” WHERE “A3”.”OBJECT_ID”=11)) “A1” (accessing
‘REMOTE’ )
–联合查询消耗时间00: 00: 10.95
SQL> SELECT * FROM v_remote WHERE object_id IN (
2 SELECT c.object_id FROM c WHERE c.object_name IN (SELECT d.object_name FROM d WHERE d.object_id=11)
3 );
已选择32行。
已用时间: 00: 00: 10.95
执行计划
———————————————————-
Plan hash value: 2118901120
————————————————————————————————————-
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | Inst |IN-OUT|
————————————————————————————————————-
| 0 | SELECT STATEMENT | | 65 | 6695 | 471 (3)| 00:00:06 | | |
|* 1 | HASH JOIN | | 65 | 6695 | 471 (3)| 00:00:06 | | |
| 2 | VIEW | VW_NSO_1 | 2 | 26 | 6 (17)| 00:00:01 | | |
| 3 | HASH UNIQUE | | 2 | 94 | 6 (17)| 00:00:01 | | |
| 4 | NESTED LOOPS | | | | | | | |
| 5 | NESTED LOOPS | | 2 | 94 | 5 (0)| 00:00:01 | | |
| 6 | TABLE ACCESS BY INDEX ROWID| D | 1 | 17 | 2 (0)| 00:00:01 | | |
|* 7 | INDEX RANGE SCAN | IDX_D | 1 | | 1 (0)| 00:00:01 | | |
|* 8 | INDEX RANGE SCAN | IDX_C | 2 | | 2 (0)| 00:00:01 | | |
| 9 | TABLE ACCESS BY INDEX ROWID | C | 2 | 60 | 3 (0)| 00:00:01 | | |
| 10 | VIEW | V_REMOTE | 375K| 32M| 462 (2)| 00:00:06 | | |
| 11 | UNION-ALL | | | | | | | |
| 12 | REMOTE | A | 369K| 29M| 454 (2)| 00:00:06 | REMOTE | R->S |
| 13 | REMOTE | B | 5323 | 431K| 8 (0)| 00:00:01 | REMOTE | R->S |
————————————————————————————————————-
Predicate Information (identified by operation id):
—————————————————
1 – access(“OBJECT_ID”=”OBJECT_ID”)
7 – access(“D”.”OBJECT_ID”=11)
8 – access(“C”.”OBJECT_NAME”=”D”.”OBJECT_NAME”)
Remote SQL Information (identified by operation id):
—————————————————-
12 – SELECT “OBJECT_NAME”,”OBJECT_ID”,”OBJECT_TYPE” FROM “A” “A” (accessing ‘REMOTE’ )
13 – SELECT “OBJECT_NAME”,”OBJECT_ID”,”OBJECT_TYPE” FROM “B” “B” (accessing ‘REMOTE’ )
单独查询很快，为什么联合查询就慢了呢？原因在于：
单独执行远程查询 本地与远程混合查询
直接执行视图，并将OBJECT_ID=11谓词推入到视图中，走索引，最后只将32行结果返回给本地 从计划中可以看到，本地查询与远程查询做HASH JOIN，但是访问远程的SQL是没有谓词的，这样必然全表从远程将结果传输到本地，结果集大，所以慢
因此，优化此混合查询的语句有多种办法：可以使用PL/SQL拆分本地与远程查询，或者可以使用driving_site hint,将主计划推到远程库去执行，本地的结果集少，推到远程，远程视图走索引，效率高。如下：
–耗时已用时间: 00: 00: 00.08
SQL> SELECT/*+ driving_site(v_remote.a) */ * FROM v_remote WHERE object_id IN (
2 SELECT c.object_id FROM c WHERE c.object_name IN (SELECT d.object_name FROM d WHERE d.object_id=11
3 );
已选择32行。
已用时间: 00: 00: 00.08
—————————————————————————————
| Id | Operation | Name | Rows | Bytes | Cost | Inst |IN-OUT|
—————————————————————————————
| 0 | SELECT STATEMENT REMOTE | | 15 | 1425 | 109 | | |
| 1 | NESTED LOOPS | | 15 | 1425 | 109 | | |
| 2 | SORT UNIQUE | | | | | | |
| 3 | VIEW | VW_NSO_1 | 3 | 39 | 8 | MZT~ | |
| 4 | REMOTE | | | | | ! | R->S |
| 5 | VIEW | | 5 | 410 | 33 | | |
| 6 | UNION-ALL PARTITION | | | | | | |
| 7 | TABLE ACCESS BY INDEX ROWID| A | 32 | 960 | 35 |MZT~ | |
|* 8 | INDEX RANGE SCAN | IDX_A | 32 | | 3 |MZT~ | |
| 9 | TABLE ACCESS BY INDEX ROWID| B | 1 | 32 | 2 | MZT~ | |
|* 10 | INDEX RANGE SCAN | IDX_B | 1 | | 1 | MZT~ | |
—————————————————————————————
Predicate Information (identified by operation id):
—————————————————
8 – access(“A6″.”OBJECT_ID”=”VW_NSO_1″.”OBJECT_ID”)
10 – access(“A5″.”OBJECT_ID”=”VW_NSO_1″.”OBJECT_ID”)
Remote SQL Information (identified by operation id):
—————————————————-
4 – SELECT /*+ */ “A1″.”OBJECT_ID” FROM “D” “A2″,”C” “A1” WHERE
“A1″.”OBJECT_NAME”=”A2″.”OBJECT_NAME” AND “A2”.”OBJECT_ID”=11 (accessing ‘!’ )
现在的SQL效率很高，从计划中可以看出，现在计划在远程库上执行，本地的查询传输给远程，并驱动视图查询，视图走索引，然后再将结果返回给本地。另外注意hint 视图的用法是hint(视图名.表名)。
再次说说driving_site失效的情况：
DML,DDL让driving_site失效,driving_site hint会自动被oracle忽略掉，因为此时以目标表所在的库为主计划驱动，相当于driving_site(目标表库）
1）DML，DDL如果是对本地表操作，主计划是在本地做的，远程数据拉到本地,driving_site(remote)失效。
2）DML如果是对远程表DML，主计划是在远程做的，本地数据送到远程，相当于自动driving_site(remote)
–本地建表和操作表test,driving_site失效
CREATE TABLE test AS
SELECT/*+driving_site(v_remote.a)*/ * FROM v_remote WHERE object_id IN (
SELECT c.object_id FROM c WHERE c.object_name IN (SELECT d.object_name FROM d WHERE d.object_id=11)
);
————————————————————————————————————–
| Id | Operation | Name | Rows | Bytes | Cost (%CPU)| Time | Inst |IN-OUT|
————————————————————————————————————–
| 0 | CREATE TABLE STATEMENT | | 65 | 6695 | 1267 (1)| 00:00:16 | | |
| 1 | LOAD AS SELECT | TEST | | | | | | |
|* 2 | HASH JOIN | | 65 | 6695 | 1266 (1)| 00:00:16 | | |
| 3 | VIEW | VW_NSO_1 | 2 | 26 | 7 (15)| 00:00:01 | | |
| 4 | HASH UNIQUE | | 2 | 94 | 6 (17)| 00:00:01 | | |
| 5 | NESTED LOOPS | | | | | | | |
| 6 | NESTED LOOPS | | 2 | 94 | 5 (0)| 00:00:01 | | |
| 7 | TABLE ACCESS BY INDEX ROWID| D | 1 | 17 | 2 (0)| 00:00:01 | | |
|* 8 | INDEX RANGE SCAN | IDX_D | 1 | | 1 (0)| 00:00:01 | | |
|* 9 | INDEX RANGE SCAN | IDX_C | 2 | | 2 (0)| 00:00:01 | | |
| 10 | TABLE ACCESS BY INDEX ROWID | C | 2 | 60 | 3 (0)| 00:00:01 | | |
| 11 | VIEW | V_REMOTE | 375K| 32M| 1257 (1)| 00:00:16 | | |
| 12 | UNION-ALL | | | | | | | |
| 13 | REMOTE | A | 369K| 29M| 454 (2)| 00:00:06 | REMOTE | R->S |
| 14 | REMOTE | B | 5323 | 431K| 8 (0)| 00:00:01 | REMOTE | R->S |
————————————————————————————————————–
Predicate Information (identified by operation id):
—————————————————
2 – access(“OBJECT_ID”=”OBJECT_ID”)
8 – access(“D”.”OBJECT_ID”=11)
9 – access(“C”.”OBJECT_NAME”=”D”.”OBJECT_NAME”)
Remote SQL Information (identified by operation id):
—————————————————-
13 – SELECT /*+ */ “OBJECT_NAME”,”OBJECT_ID”,”OBJECT_TYPE” FROM “A” “A” (accessing ‘REMOTE’ )
14 – SELECT “OBJECT_NAME”,”OBJECT_ID”,”OBJECT_TYPE” FROM “B” “B” (accessing ‘REMOTE’ )
–远程表操作，自动driving_site(remote)
INSERT INTO xm@remote( object_name,object_id,object_type)
SELECT * FROM v_remote WHERE object_id IN (
SELECT c.object_id FROM c WHERE c.object_name IN (SELECT d.object_name FROM d WHERE d.object_id=11)
);
—————————————————————————————
| Id | Operation | Name | Rows | Bytes | Cost | Inst |IN-OUT|
—————————————————————————————
| 0 | INSERT STATEMENT REMOTE | | 15 | 1425 | 109 | | |
| 1 | NESTED LOOPS | | 15 | 1425 | 109 | | |
| 2 | SORT UNIQUE | | | | | | |
| 3 | VIEW | VW_NSO_1 | 3 | 39 | 8 | MZT~ | |
| 4 | REMOTE | | | | | ! | R->S |
| 5 | VIEW | | 5 | 410 | 33 | | |
| 6 | UNION-ALL PARTITION | | | | | | |
| 7 | TABLE ACCESS BY INDEX ROWID| A | 32 | 960 | 35 | MZT~ | |
|* 8 | INDEX RANGE SCAN | IDX_A | 32 | | 3 | MZT~ | |
| 9 | TABLE ACCESS BY INDEX ROWID| B | 1 | 32 | 2 | MZT~ | |
|* 10 | INDEX RANGE SCAN | IDX_B | 1 | | 1 | MZT~ | |
—————————————————————————————
Predicate Information (identified by operation id):
—————————————————
8 – access(“A7″.”OBJECT_ID”=”VW_NSO_1″.”OBJECT_ID”)
10 – access(“A6″.”OBJECT_ID”=”VW_NSO_1″.”OBJECT_ID”)
Remote SQL Information (identified by operation id):
—————————————————-
4 – SELECT /*+ */ “A1″.”OBJECT_ID” FROM “D” “A2″,”C” “A1” WHERE
“A1″.”OBJECT_NAME”=”A2″.”OBJECT_NAME” AND “A2”.”OBJECT_ID”=11 (accessing ‘!’ )
4. 总结
含有dblink的SQL语句，特别是DML SQL，优化是很复杂的，特别是远程表与本地表结果集都很大，或含有多个不同的远程对象，这时更加复杂。很多含有dblink的SQL语句可以通过Collocated inline view，定义视图，driving_site hint,PL/SQL程序等进行优化，当然，在业务允许的情况下也可以通过MV等技术，减少dblink使用，可以在一定程度上优化含有DBLINK的分布式操作语句。
5. 作者简介
丁俊，网名：dingjun123
ITPUB开发版资深版主、ITPUB社区专家、ITPUB名人堂成员、ChinaUnix BLOG专家，
ITPUB 2010-2013连续4届最佳精华获得者、2011-2014连续4届最佳版主。
电子工业出版社终身荣誉作者，《剑破冰山-Oracle开发艺术》副主编。
曾多次参与和ORACLE相关的公共活动：ITPUB 2011演讲嘉宾、OOW 2013 上海 weibo特使等.
