oracle 都是parallel惹的祸

来源：互联网发布：java字符串concat 编辑：程序博客网时间：2024/05/18 01:46

该项目是中国联通xxxx话务系统，我的架构设计+需求设计，+运维保障+数据库开发，全套服务。

在今天开发完毕后，突然有个模块的需求，用户号码为必须选择，感觉有点郁闷，因为1小时有1000w数据，把所有用户号码显示出来，是不是有点画蛇添足呢。

我的开始设想是查询详单，像中移营业厅，需要输入号码，或者省份证查询模糊查询，没有谓词不能查询。（感觉设计合情合理）

1.但是想了解整个系统用户分布情况，必须输入条件，是不是有点不可用。

2.并且没有谓词过滤，查询会慢，非常慢（1-2分钟出结果），目标是3-5秒内出数据。

注：优化难点是把2秒变成1秒，反之，把2小时变成2分钟非常简单。

第1步：

下面看看语句和执行计划：

SQL>  explain plan for  SELECT /*+ parallel(8)   */  2   starttime starttime,  3   cv.groupid,  4   cs.custmangerid,  5   callercarrier callercarrier,  6   callernum callernum,  7   calledcarrier calledcarrier,  8   callednum callednum,  9   calleenum calleenum, 10   round(duration / 60, 2) CallTimeLen, 11   count(*) over(ORDER BY NULL ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) "@totalrows" 12    FROM CS_xxxx dt, cfg_vipphones cv, cfg_vipusers cs 13   WHERE dt.StartTime >= '2013-05-31 13:00:00' 14     and dt.StartTime < '2013-05-31 14:00:00' 15     AND dt.Callercarrier = 2 16     AND dt.callernum >= cv.beginphone 17     and dt.callernum <= cv.endphone 18     and cv.groupid = cs.groupid; Explained SQL> select * from table(dbms_xplan.display); PLAN_TABLE_OUTPUT--------------------------------------------------------------------------------Plan hash value: 2172492340--------------------------------------------------------------------------------| Id  | Operation                              | Name            | Rows  | Bytes--------------------------------------------------------------------------------|   0 | SELECT STATEMENT                       |                 |   478K|    34|*  1 |  PX COORDINATOR                        |                 |       ||   2 |   PX SEND QC (RANDOM)                  | :TQ10001        |   478K|    34|   3 |    WINDOW BUFFER                       |                 |   478K|    34|*  4 |     FILTER                             |                 |       ||   5 |      MERGE JOIN                        |                 |   478K|    34|   6 |       SORT JOIN                        |                 |    11 |   363|   7 |        BUFFER SORT                     |                 |       ||   8 |         PX RECEIVE                     |                 |       ||   9 |          PX SEND BROADCAST             | :TQ10000        |       ||  10 |           NESTED LOOPS                 |                 |       ||  11 |            NESTED LOOPS                |                 |    11 |   363|  12 |             TABLE ACCESS BY INDEX ROWID| CFG_VIPUSERS    |     3 |    18|  13 |              INDEX FULL SCAN           | PK_CFG_VIPUSERS |     3 ||* 14 |             INDEX RANGE SCAN           | VIPUSERS_FK     |     4 | PLAN_TABLE_OUTPUT--------------------------------------------------------------------------------|  15 |            TABLE ACCESS BY INDEX ROWID | CFG_VIPPHONES   |     4 |   108|* 16 |       FILTER                           |                 |       ||* 17 |        SORT JOIN                       |                 |   516K|    21|  18 |         PX BLOCK ITERATOR              |                 |   516K|    21|* 19 |          TABLE ACCESS FULL             | CS_xxxx          |   516K|    21--------------------------------------------------------------------------------Predicate Information (identified by operation id):---------------------------------------------------   1 - filter(TO_DATE('2013-05-31 13:00:00')<TO_DATE('2013-05-31 14:00:00'))   4 - filter(TO_DATE('2013-05-31 13:00:00')<TO_DATE('2013-05-31 14:00:00'))  14 - access("CV"."GROUPID"="CS"."GROUPID")  16 - filter("DT"."CALLERNUM"<="CV"."ENDPHONE")  17 - access("DT"."CALLERNUM">="CV"."BEGINPHONE")       filter("DT"."CALLERNUM">="CV"."BEGINPHONE")  19 - filter("DT"."CALLERCARRIER"=2 AND "DT"."STARTTIME">='2013-05-31 13:00:00'Note-----   - Degree of Parallelism is 8 because of hint 41 rows selected SQL>

大家看出问题了吗，说实话，执行计划只是一个参考，看看index是否生效，是不是全表scan，nloop，hash，是不是可以增加use_nl, 等hint

OLAP和OLTP 又有很大区别了，包含数据库参数设定，sql写法，hint是否启用等

第2步：

我怀疑是3张表关联，谓词出了问题。

注意看filter，看看是否是分区表搞的鬼。查看后一切正常，因为是我写的，我最清楚。哈哈。。。

在多表关联时，如果有视图，可以考虑视图的合并，关联的优先选择，再hash。都试过了，不行。

第3 步：

怀疑是并行出错了，看看表的并且度，索引并行，

或者我不要并行试试。果然，8-10秒出结果

SQL>  explain plan for  SELECT  2   starttime starttime,  3   cv.groupid,  4   cs.custmangerid,  5   callercarrier callercarrier,  6   callernum callernum,  7   calledcarrier calledcarrier,  8   callednum callednum,  9   calleenum calleenum, 10   round(duration / 60, 2) CallTimeLen, 11   count(*) over(ORDER BY NULL ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) "@totalrows" 12    FROM CS——xx dt, cxg_vippxxx cv, cxg_vipxxx cs 13   WHERE dt.StartTime >= '2013-05-31 13:00:00' 14     and dt.StartTime < '2013-05-31 14:00:00' 15     AND dt.Callercarrier = 2 16     AND dt.callernum >= cv.beginphone 17     and dt.callernum <= cv.endphone 18     and cv.groupid = cs.groupid; Explained SQL> select * from table(dbms_xplan.display); PLAN_TABLE_OUTPUT--------------------------------------------------------------------------------Plan hash value: 1705527799--------------------------------------------------------------------------------| Id  | Operation                         | Name            | Rows  | Bytes |Tem--------------------------------------------------------------------------------|   0 | SELECT STATEMENT                  |                 |   478K|    34M||   1 |  WINDOW BUFFER                    |                 |   478K|    34M||*  2 |   FILTER                          |                 |       |       ||   3 |    MERGE JOIN                     |                 |   478K|    34M||   4 |     SORT JOIN                     |                 |    11 |   363 ||   5 |      NESTED LOOPS                 |                 |       |       ||   6 |       NESTED LOOPS                |                 |    11 |   363 ||   7 |        TABLE ACCESS BY INDEX ROWID| CFGxxUSERS    |     3 |    18 ||   8 |         INDEX FULL SCAN           | PK_CFG_VIPUSERS |     3 |       ||*  9 |        INDEX RANGE SCAN           | VIPUSERS_FK     |     4 |       ||  10 |       TABLE ACCESS BY INDEX ROWID | CFG_xxNES   |     4 |   108 ||* 11 |     FILTER                        |                 |       |       ||* 12 |      SORT JOIN                    |                 |   516K|    21M||  13 |       PARTITION RANGE ITERATOR    |                 |   516K|    21M||* 14 |        TABLE ACCESS FULL          | CS_xxx         |   516K|    21M| PLAN_TABLE_OUTPUT----------------------------------------------------------------------------------------------------------------------------------------------------------------Predicate Information (identified by operation id):---------------------------------------------------   2 - filter(TO_DATE('2013-05-31 13:00:00')<TO_DATE('2013-05-31 14:00:00'))   9 - access("CV"."GROUPID"="CS"."GROUPID")  11 - filter("DT"."CALLERNUM"<="CV"."ENDPHONE")  12 - access("DT"."CALLERNUM">="CV"."BEGINPHONE")       filter("DT"."CALLERNUM">="CV"."BEGINPHONE")  14 - filter("DT"."CALLERCARRIER"=2 AND "DT"."STARTTIME">='2013-05-31 13:00:00'              14:00:00') 32 rows selected SQL>

第4步：

看看并行设置，这个也有很大关系，因为并行的模块太多，造成排队拥塞的情况

<1>如果有并行度低于系统最大并行数的查询在跑,那接下来的并行查询会怎么跑呢?
When you specify parallel degree 4 oracle tries to allocate 4 producer slaves and 4 consumer slaves. The producers can feed any of the consumers.
If there are only 2 slaves available then we use these.
If there is only 1 slave available then we go serial
If there are none available then we use serial.
If parallel_min_percent is set then we error ora 12827 instead of using a lower number of slaves or going serial

<2>设定parallel_max_servers 多大为好?
在多CPU的环境中,一般把CPU-1或CPU的数量做个最大并行数,因为并行查询运行时还需要一个进程协调各并行进程.对于单CPU没什么好说的.

<3>并行查询能提高系统的性能吗?
并行查询运行时,很容易会使机器运行在高负荷下,令系统对其它事务的处理时间大大加长.并行查询一般适合在非业务高峰值人工执行,并不适合在程序中指定运行并行查询.
PINNER:
并行不等于快速，仅仅是适合在数据仓库环境，低业务请求与低并发操作的时候
典型的OLTP系统，如果我们的系统，是绝对不允许并行查询出现的。

(引荐哈)

第5步：

问题解决，注意看看问题，paralle的写法，当一个表时，用parallel(8) , 表示当前表并行8个进程

当有多个表是，请指定某一个表，否则会默认3个表，当然执行计划上看不出来，可以trace一把看看

SQL>  explain plan for  SELECT /*+ parallel(dt,8)   */  2   starttime starttime,  3   cv.groupid,  4   cs.custmangerid,  5   callercarrier callercarrier,  6   callernum callernum,  7   calledcarrier calledcarrier,  8   callednum callednum,  9   calleenum calleenum, 10   round(duration / 60, 2) CallTimeLen, 11   count(*) over(ORDER BY NULL ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING) "@totalrows" 12    FROM CS_CDR dt, cfg_vipphones cv, cfg_vipusers cs 13   WHERE dt.StartTime >= '2013-05-31 13:00:00' 14     and dt.StartTime < '2013-05-31 14:00:00' 15     AND dt.Callercarrier = 2 16     AND dt.callernum >= cv.beginphone 17     and dt.callernum <= cv.endphone 18     and cv.groupid = cs.groupid; Explained SQL> select * from table(dbms_xplan.display); PLAN_TABLE_OUTPUT--------------------------------------------------------------------------------Plan hash value: 2172492340--------------------------------------------------------------------------------| Id  | Operation                              | Name            | Rows  | Bytes--------------------------------------------------------------------------------|   0 | SELECT STATEMENT                       |                 |   478K|    34|*  1 |  PX COORDINATOR                        |                 |       ||   2 |   PX SEND QC (RANDOM)                  | :TQ10001        |   478K|    34|   3 |    WINDOW BUFFER                       |                 |   478K|    34|*  4 |     FILTER                             |                 |       ||   5 |      MERGE JOIN                        |                 |   478K|    34|   6 |       SORT JOIN                        |                 |    11 |   363|   7 |        BUFFER SORT                     |                 |       ||   8 |         PX RECEIVE                     |                 |       ||   9 |          PX SEND BROADCAST             | :TQ10000        |       ||  10 |           NESTED LOOPS                 |                 |       ||  11 |            NESTED LOOPS                |                 |    11 |   363|  12 |             TABLE ACCESS BY INDEX ROWID| CFG_VIPUSERS    |     3 |    18|  13 |              INDEX FULL SCAN           | PK_CFG_VIPUSERS |     3 ||* 14 |             INDEX RANGE SCAN           | VIPUSERS_FK     |     4 | PLAN_TABLE_OUTPUT--------------------------------------------------------------------------------|  15 |            TABLE ACCESS BY INDEX ROWID | CFG_VIPPHONES   |     4 |   108|* 16 |       FILTER                           |                 |       ||* 17 |        SORT JOIN                       |                 |   516K|    21|  18 |         PX BLOCK ITERATOR              |                 |   516K|    21|* 19 |          TABLE ACCESS FULL             | CS_xxxx         |   516K|    21--------------------------------------------------------------------------------Predicate Information (identified by operation id):---------------------------------------------------   1 - filter(TO_DATE('2013-05-31 13:00:00')<TO_DATE('2013-05-31 14:00:00'))   4 - filter(TO_DATE('2013-05-31 13:00:00')<TO_DATE('2013-05-31 14:00:00'))  14 - access("CV"."GROUPID"="CS"."GROUPID")  16 - filter("DT"."CALLERNUM"<="CV"."ENDPHONE")  17 - access("DT"."CALLERNUM">="CV"."BEGINPHONE")       filter("DT"."CALLERNUM">="CV"."BEGINPHONE")  19 - filter("DT"."CALLERCARRIER"=2 AND "DT"."STARTTIME">='2013-05-31 13:00:00' 37 rows selected SQL>

目前是3秒出结果，已经达到预期，当然谓词为1小时，或者有号码过滤绝对是1秒内响应速度。