首页 > Linux操作系统 > Linux操作系统 > 关于软解析(soft parse)与硬解析(hard parse),以及session cached cursors (asktom)

关于软解析(soft parse)与硬解析(hard parse),以及session cached cursors (asktom)

原创 Linux操作系统 作者:tolywang 时间:2007-07-10 00:00:00 1 删除 编辑
difference between soft parse and hard parse

-- Thanks for the question regarding "difference between soft parse and hard parse", version
originally submitted on 26-Jan-2002 19:17 Eastern US time, last updated 5-Aug-2004 13:07

Hi Tom

Can you explain briefly the difference between soft and hard parse?


and we said...

Here is a long winded answer, it is extracted in part from a new book oming out soon "beginning Oracle programming" that I collaborated on:


This is the first step in the processing of any statement in Oracle. Parsing is
the act of breaking the submitted statement down into its component parts ? determining what type of statement it is (query, DML, DDL) and performing various checks on it.

The parsing process performs two main functions:

  • Syntax Check: is the statement a valid one? Does it make sense given the SQL grammar documented in the SQL Reference Manual. Does it follow all of the rules for SQL.
    1. 语法检查,这条语句是否合理? 它是否符合SQL语法文档<>的相关语法,是不是符合SQL语句的所有规则.

    o Semantic Analysis: Going beyond the syntax ? is the statement valid in light of the objects in the database (do the tables and columns referenced exist). Do you have access to the objects ? are the proper privileges in place? Are there ambiguities in the statement ? for example if there are two tables T1 and T2 and both have a column X, the query ?select X from T1, T2 where ?? is ambiguous, we don?t know which table to get X from. And so on.
    2.语义分析: 已经处理完语法检查了吗? 这条语句是否正确的引用了数据库中对象
    (是否语句中所有的表和字段都存在). 是否有访问这些对象的权限? 是不是对应的权限都有效(主要指role对应的权限吧,:D)? 比如是否有如下的两个表T1,T2,而这两个表有相同的字段名column X,而查询语句 ? select X from T1,T2 where ??,(没有明确指定列名),我们无法知道从哪个表去取出字段X的值,以及类似的一系列问题.

    So, you can think of parsing as basically a two step process, that of a syntax
    check to check the validity of the statement and that of a semantic check ? to ensure the statement can execute properly. The difference between the two types of checks are hard for you to see ? Oracle does not come back and say ?it failed the syntax check?, rather it returns the statement with a error code and message. So for example, this statement fails with a syntax error:
    因此,你可以认为解析基本上是由这两个步骤组成的: 检查语句有效性的语法检查和确保语句能够正常运行的语义检查.这两种检查的区别就是你所要说的硬解析了.Oracle不会特别的指出这条语句没有通过语法检查,它给你返回这条语句对应的错误代码和错误信息. 比如: 下面这条语句没有通过语法检查.
    SQL> select from where 2;
    select from where 2
    ERROR at line 1:
    ORA-00936: missing expression

    While this statement failed with a semantic error ? if the table NOT_A_TABLE existed and we had permission to access it, this statement would have succeeded:

    而这条语句没有通过语义检查? 如果表NOT_A_TABLE存在,并且我们具有访问这个表的权限,这条语句就会正确执行了.
    SQL> select * from not_a_table;
    select * from not_a_table
    ERROR at line 1:
    ORA-00942: table or view does not exist

    That is the only way to really tell the difference between a semantic and
    syntactic error ? if the statement COULD have executed given the proper objects and privileges, you had a semantic error, otherwise if the statement could not execute under any circumstances, you have a syntax error. Regardless ? Oracle will not execute the statement for you!


    The next step in the parse operation is to see if the statement we are currently parsing has already in fact been processed by some other session. If it has ? we may be in luck here, we can skip the next two steps in the process, that of optimization and row source generation. If we can skip these next two steps in the process, we have done what is known as a Soft Parse ? a shorter process to getting our query going. If we cannot, if we must do all of the steps, we are performing what is known as a Hard Parse ? we must parse, optimize, generate the plan for the query. This distinction is very important. When developing our applications we want a very high percentage of our queries to be Soft Parsed ? to be able to skip the optimize/generate phases ? as they are very CPU intensive as well as a point of contention (serialization). If we have to Hard Parse a large percentage of our queries, our system will function slowly and in some cases ? not at all.

    解析中的下一个步骤就是检查我们现在正在解析的语句是否已经被其他的session执行过,在这一步我们可能很幸运,我们可以跳过下面的两个步骤: 语句的优化(生成执行计划) 和 生成执行编码. 如果我们能够跳过这两个步骤,我们就实现了一般意义上的软解析, 我们必须解析,优化,并为这条语句生成执行计划.这个特性(distinction)非常重要,当我们开发应用程序的时候,我们会希望大部分的语句都是软解析的,以跳过优化和生成编码的步骤,因为他们和争用(串行化,)(contention)一样,都是非常消耗CPU的操作.如果我们必须硬解析我们语句的大部分的话,在某些情况下我们的系统就会运行的很慢,并不是任何时候(在olap,dss系统应该是个例外吧:))

    The way this sharing of SQL in Oracle is accomplished is via the shared pool, a piece of memory in the SGA maintained by Oracle. We covered this topic in chapter 5 but will revisit it again in the context of processing a query. After Oracle parses the query and it passes the syntax and semantic checks ? it will look in the shared pool component of the SGA to see if that same exact query has already been processed by another session. Since it has performed the semantic check it has already figured out:

    在Oracle中共享SQL的方式是通过共享池--SGA中的一块内存(由Oracle系统来维护)--实现的.我们在第五章中讨论了这个主题,并且会在查询处理的章节再次讨论这个主题.当Oracle已经解析了这条语句,并且已经通过语法和语义检查的时候,Oracle就会在SGA的shared pool(共享池)组件中查询,看看,是否已经有一个完全一样的语句已经被另外一个session执行过.因为语句已经通过了oracle的语义检查,并且oracle已经计算出:

    o Exactly what tables are involved

    o That we have access to the tables (the proper privileges are there)

    And so on. Now, it can look at all of the queries in the shared pool that have already been parsed/optimized and generated to see if the work has already been done.
    Soft Parse and session_cashed_cursor parameter January 28, 2002
    Reviewer: A reader


    can you explain what means the default value (0) of session_ casched_cursor parameter? Oracle always has a cache of sqls in sga and size of this cache is determinated through init.ora parameter shared_pool_size.

    tom :

    The default value of ZERO means that by default, ZERO cursors will be cached for your session.


    They will be cached in the shared pool -- but your session will have to find
    them there. Session Cached Cursors can remove the need to have to "find" them. Consider this example that shows the difference spent when soft parsing without session cached cursors vs soft parsing WITH session cached cursors:
    他们会被缓存在共享池(Shared Pool)中,但是你的session必须在共享池中去查找他们,session cached cursors可以省略掉去共享池中查找这一步.下面这个例子
    ops$[email]tkyte@ORA817DEV.US.ORACLE.COM[/email]> create table emp as select * from scott.emp;
    Table created.

    ops$[email]tkyte@ORA817DEV.US.ORACLE.COM[/email]> create table run_stats ( runid varchar2(15),
    name varchar2(80), value int );
    Table created.

    ops$[email]tkyte@ORA817DEV.US.ORACLE.COM[/email]> create or replace view stats
    2 as select 'STAT...' || name, b.value
    3 from v$statname a, v$mystat b
    4 where a.statistic# = b.statistic#
    5 union all
    6 select 'LATCH.' || name, gets
    7 from v$latch;

    View created.

    ops$[email]tkyte@ORA817DEV.US.ORACLE.COM[/email]> column name format a40
    ops$[email]tkyte@ORA817DEV.US.ORACLE.COM[/email]> declare
    2 l_start number;
    3 l_cnt number;
    4 begin
    5 execute immediate 'alter session set session_cached_cursors=0';
    6 insert into run_stats select 'before', stats.* from stats;
    8 l_start := dbms_utility.get_time;
    9 for i in 1 .. 1000
    10 loop
    11 execute immediate 'select count(*) from emp' into l_cnt;
    12 end loop;
    13 dbms_output.put_line( (dbms_utility.get_time-l_start) || ' hsecs' );
    15 execute immediate 'alter session set session_cached_cursors=100';
    16 insert into run_stats select 'after 1', stats.* from stats;
    18 l_start := dbms_utility.get_time;
    19 for i in 1 .. 1000
    20 loop
    21 execute immediate 'select count(*) from emp' into l_cnt;
    22 end loop;
    23 dbms_output.put_line( (dbms_utility.get_time-l_start) || ' hsecs' );
    25 insert into run_stats select 'after 2', stats.* from stats;
    26 end;
    27 /
    45 hsecs
    35 hsecs

    PL/SQL procedure successfully completed.

    so, session cached cursors RAN faster (i ran this a couple of times, there
    were no hard parses going on. But the real good news is:

    ops$[email]tkyte@ORA817DEV.US.ORACLE.COM[/email]> select, b.value-a.value run1,
    c.value-b.value run2,
    2 ( (c.value-b.value)-(b.value-a.value)) diff
    3 from run_stats a, run_stats b, run_stats c
    4 where =
    5 and =
    6 and a.runid = 'before'
    7 and b.runid = 'after 1'
    8 and c.runid = 'after 2'
    9 and (c.value-a.value) > 0
    10 and (c.value-b.value) <> (b.value-a.value)
    11 order by abs( (c.value-b.value)-(b.value-a.value))
    12 /

    ---------------------------------------- ---------- ---------- ----------
    LATCH.checkpoint queue latch 3 4 1
    LATCH.redo allocation 30 31 1
    STAT...consistent gets 5088 5089 1
    STAT...deferred (CURRENT) block cleanout 2 3 1

    STAT...calls to get snapshot scn: kcmgss 5019 5018 -1
    STAT...enqueue releases 10 9 -1
    STAT...execute count 1015 1014 -1
    STAT...opened cursors cumulative 1015 1014 -1
    STAT...parse count (total) 1015 1014 -1
    STAT...session cursor cache count 0 1 1
    STAT...redo entries 28 27 -1
    STAT...recursive calls 1180 1179 -1
    STAT...physical reads 1 0 -1 msg latch 2 0 -2
    LATCH.session queue latch 2 0 -2
    LATCH.done queue latch 2 0 -2 buffer requested 8 6 -2
    STAT...enqueue requests 11 9 -2
    LATCH.messages 3 0 -3
    STAT...db block changes 47 44 -3
    LATCH.redo writing 3 0 -3
    LATCH.ksfv messages 4 0 -4
    STAT...session logical reads 17128 17123 -5
    LATCH.row cache objects 184 178 -6
    STAT...db block gets 12040 12034 -6
    STAT...parse time elapsed 9 3 -6
    STAT...parse time cpu 13 4 -9
    STAT...recursive cpu usage 51 38 -13
    LATCH.cache buffers chains 34315 34335 20
    STAT...redo size 23900 24000 100
    STAT...session cursor cache hits 3 1002 999
    LATCH.shared pool 2142 1097 -1045
    LATCH.library cache 17361 2388 -14973

    34 rows selected.


    see the significantly REDUCED number of LATCH counts on the library and shared pool. Since a latch is a lock, a lock is a serialization device, serialization
    implies WAITS -- using the session cached cursors will increase scalability and performance as you add more and more users. Its not only FASTER, but more scalable as well...


    来自 “ ITPUB博客 ” ,链接:,如需转载,请注明出处,否则将追究法律责任。

  • 全部评论
    Oracle , MySQL, SAP IQ, SAP HANA, PostgreSQL, Tableau 技术讨论,希望在这里一起分享知识,讨论技术,畅谈人生 。


    • 博文量
    • 访问量