zoukankan      html  css  js  c++  java
  • [20190823]关于CPU成本计算3.txt

    [20190823]关于CPU成本计算3.txt

    --//前几天探究CPU cost时遇到的问题,获取行成本时我的测试查询结果出现跳跃,不知道为什么,感觉有点奇怪,分析看看。
    --//ITPUB原始链接已经不存在,我的日记本还有记录,现在想想当时的记录思路很乱,不过这些都是猜测的过程,以前思路混乱也是正常的。
    --//顺便做一些必要补充唯一索引的相关CPU COST。

    1.环境:
    SCOTT@test01p> @ ver1
    PORT_STRING          VERSION    BANNER                                                                       CON_ID
    -------------------- ---------- ---------------------------------------------------------------------------- ------
    IBMPC/WIN_NT64-9.1.0 12.2.0.1.0 Oracle Database 12c Enterprise Edition Release 12.2.0.1.0 - 64bit Production      0

    2.测试:
    SCOTT@test01p> create table t as select rownum a1 , rownum a2 ,rownum a3 from dual connect by level<=100 ;
    Table created.

    create unique index i_t_a1 on t(a1);
    create unique index i_t_a2 on t(a2);
    create unique index i_t_a3 on t(a3);
    --//分析略.

    column STATEMENT_ID format a10
    column OPTIONS format a20
    explain plan set statement_id='a1' for select a1 from t where a1=100;
    explain plan set statement_id='a2' for select a2 from t where a2=100;
    explain plan set statement_id='a3' for select a3 from t where a3=100;

    SCOTT@test01p> select STATEMENT_ID,OPERATION, OPTIONS, COST, CPU_COST, IO_COST, TIME from plan_table where options='UNIQUE SCAN';
    STATEMENT_ OPERATION OPTIONS     COST CPU_COST IO_COST TIME
    ---------- --------- ----------- ---- -------- ------- ----
    a1         INDEX     UNIQUE SCAN    0     1050       0    1
    a2         INDEX     UNIQUE SCAN    0     1050       0    1
    a3         INDEX     UNIQUE SCAN    0     1050       0    1
    --//通过唯一索引访问cpu cost=1050.这里面包含行成本150,比较50。
    --//通过hack看看变化.

    SCOTT@test01p> select BLEVEL,LEAF_BLOCKS from user_indexes where index_name = 'I_T_A3';
    BLEVEL LEAF_BLOCKS
    ------ -----------
         0           1

    SCOTT@test01p> exec dbms_stats.set_index_stats(ownname=> NULL,indname=>'I_T_A3',indlevel=> 1000);
    PL/SQL procedure successfully completed.

    SCOTT@test01p> select BLEVEL,LEAF_BLOCKS from user_indexes where index_name = 'I_T_A3';
        BLEVEL LEAF_BLOCKS
    ---------- -----------
           255           1

    --//注意:我修改indlevel=> 1000,实际上最大仅仅255.也就是最大blevel=255.
    --//注:我记忆blevel最大64.如果hack blevel最好不要超过6X.不然在11g下会直接导致使用该索引的sql语句挂起!!切记!!
    SCOTT@test01p> exec dbms_stats.set_index_stats(ownname=> NULL,indname=>'I_T_A3',indlevel=> 1000);
    PL/SQL procedure successfully completed.

    SCOTT@test01p> alter system flush shared_pool;
    System altered.

    column STATEMENT_ID format a10
    column OPTIONS format a20

    SCOTT@test01p> explain plan set statement_id='x3' for select /*+ index(t) */ a3 from t where a3=100;
    Explained.

    SCOTT@test01p> select STATEMENT_ID,OPERATION, OPTIONS, COST, CPU_COST, IO_COST, TIME from plan_table where STATEMENT_ID='x3';
    STATEMENT_ OPERATION            OPTIONS                    COST   CPU_COST    IO_COST       TIME
    ---------- -------------------- -------------------- ---------- ---------- ---------- ----------
    x3         SELECT STATEMENT                                   2    1817017          2          1
    x3         INDEX                UNIQUE SCAN                 255    1817017        255          1
    --//访问索引根节点,叶子节点占255块.
    7121.44*255 = 1815967.20
    1817017- 1815967.20 = 1049.80

    --//修改blevel=254看看.
    SCOTT@test01p> exec dbms_stats.set_index_stats(ownname=> NULL,indname=>'I_T_A3',indlevel=> 1000);
    PL/SQL procedure successfully completed.

    SCOTT@test01p> alter system flush shared_pool;
    System altered.

    SCOTT@test01p> explain plan set statement_id='x3' for select /*+ index(t) */ a3 from t where a3=100;
    Explained.

    SCOTT@test01p> select STATEMENT_ID,OPERATION, OPTIONS, COST, CPU_COST, IO_COST, TIME from plan_table where STATEMENT_ID='x3';
    STATEMENT_ OPERATION            OPTIONS                    COST   CPU_COST    IO_COST       TIME
    ---------- -------------------- -------------------- ---------- ---------- ---------- ----------
    x3         SELECT STATEMENT                                   2    1809896          2          1
    x3         INDEX                UNIQUE SCAN                 254    1809896        254          1
    --//访问索引根节点,叶子节点占255块.
    7121.44*254 = 1808845.76
    1809896 - 1808845.76 = 1050.24
    7121.44*254+1050 = 1809895.76
    --//无法通过HACK blevel测试.估计就是1050.
    --//这样扣除行成本150,比较成本50. UNIQUE SCAN block 成本在=850.

    5.看看索引范围的情况呢?
    --//重新分析表T!!

    column STATEMENT_ID format a10
    column OPTIONS format a10
    explain plan set statement_id='x100' for select /*+ index(t) */ 1 from t where a3>=100;
    explain plan set statement_id='x050' for select /*+ index(t) */ 1 from t where a3>=050;

    SCOTT@test01p> select STATEMENT_ID,OPERATION, OPTIONS, COST, CPU_COST, IO_COST, TIME from plan_table ;
    STATEMENT_ OPERATION            OPTIONS                    COST   CPU_COST    IO_COST       TIME
    ---------- -------------------- -------------------- ---------- ---------- ---------- ----------
    x100       SELECT STATEMENT                                   1       7321          1          1
    x100       INDEX                RANGE SCAN                    1       7321          1          1
    x050       SELECT STATEMENT                                   1      17521          1          1
    x050       INDEX                RANGE SCAN                    1      17521          1          1

    --//看看STATEMENT_ID=050的情况.
    SCOTT@test01p> select BLEVEL,LEAF_BLOCKS from user_indexes where index_name = 'I_T_A3';
    BLEVEL LEAF_BLOCKS
    ------ -----------
         0           1
    --//Blevel=0 , LEAF_BLOCKS=1
    --//块1块 7121.44 行1行 150, 比较 1次 50
    --//索引选择率 = (100-50)/99+1/100 = 0.51505050505050505050,也就是card=52.
    --//访问块  0.5150*1 = .5150 = 1 ,块成本= 1*7121.44= 7121
    --//(行成本+比较成本)*card = 200*52 = 10400
    --//10400+7121.44 = 17521.44

    --//仅仅大致确定UNIQUE SCAN 块成本是1050.

  • 相关阅读:
    python爬取北京政府信件信息03
    Python3.7 lxml引入etree
    python爬取北京政府信件信息02
    python爬取北京政府信件信息01
    2020.12.2收获
    2020.12.1收获
    2020.11.30收获
    2020.11.29收获
    2020.11.28收获
    2020.11.27收获
  • 原文地址:https://www.cnblogs.com/lfree/p/11405904.html
Copyright © 2011-2022 走看看