为何listener要fork两次子进程

freas

wolfop 发表于 2013-6-30 16:13
互联网行业肯定要靠中间件实现连接池的。就算informix支撑互联网应用如果不用连接池也会挂，不过是相对可 ...

1、中间件做流量控制主要是企业应用里的概念，互联网行业相对很少见（我没见到，我特意咨询了下以前的同事求证，也是没有的）。
2、informix来支撑互联网的大并发，无论从技术还是市场来看，基本上都只是个梦想。
3、我以前做的一些不成熟的思考可以参考下
自己写cache server之网络框架处理——对比mysql、Oracle（上）
自己写cache server之网络框架处理——Oracle、Mysql都不靠谱儿（中）

wolfop

freas 发表于 2013-6-30 16:57
1、中间件做流量控制主要是企业应用里的概念，互联网行业相对很少见（我没见到，我特意咨询了下以前的同事 ...

Paypal,支付宝。

freas

wolfop 发表于 2013-7-1 22:22
Paypal,支付宝。

支付属于金融行业，业务类型和典型互联网业务差距比较大。嘎嘎

Yong Huang

Beginning with 12c, you can set the DB parameter threaded_execution to true and listener parameter dedicated_through_broker to on to allow Oracle to spawn threads in an existing server process instead of spawning separate processes. I didn't test. But I think the overhead of spawning a full-fledged process is largely eliminated.

Related documentation:
http://docs.oracle.com/cd/E16655 ... 7615/refrn10335.htm
http://docs.oracle.com/cd/E16655 ... /process.htm#i16690

freas

Yong Huang 发表于 2013-7-3 04:44
Beginning with 12c, you can set the DB parameter threaded_execution to true and listener parameter d ...

12c才出来几天就搞得这么熟，版主威武。Oracle在网络处理上动了大手术，这样的革新让人兴奋。就是不知道会不会带来很多这方面的bug。

Yong Huang

In threaded execution mode (see msg #14 to enable), the number of Oracle processes is drastically reduced:

[oracle@dcdrpcora1a admin]$ ps -fu oracle
UID        PID  PPID  C STIME TTY          TIME CMD
...[a few bash shell processes]...
oracle   12630     1  0 15:45 ?        00:00:00 /u01/app/oracle/product/12.1.0/db/bin/tnslsnr LISTENER -inherit
oracle   15457     1  0 15:47 ?        00:00:00 ora_pmon_test12c
oracle   15459     1  0 15:47 ?        00:00:00 ora_psp0_test12c
oracle   15461     1  0 15:47 ?        00:00:00 ora_vktm_test12c
oracle   15465     1  0 15:47 ?        00:00:03 ora_u004_test12c
oracle   15472     1  5 15:47 ?        00:00:26 ora_u005_test12c
oracle   15478     1  0 15:47 ?        00:00:00 ora_dbw0_test12c

All except the two ora_u* processes are just like in process (non-threaded) execution. The two ora_u* processes each contain multiple threads:

$ ps -fLp 15472
UID        PID  PPID   LWP  C NLWP STIME TTY          TIME CMD
oracle   15472     1 15472  0   40 15:47 ?        00:00:00 ora_u005_test12c
oracle   15472     1 15473  0   40 15:47 ?        00:00:00 ora_u005_test12c
...

Each thread has a specific function:

SQL> select stid, pname, program from v$process where spid = 15472 order by 1;

STID  PNAME PROGRAM
----- ----- ---------------------------------------------
15472 SCMN  oracle@dcdrpcora1a (SCMN)
15474 DIAG  oracle@dcdrpcora1a (DIAG)
15476 DIA0  oracle@dcdrpcora1a (DIA0)
15484 RECO  oracle@dcdrpcora1a (RECO)
15486 MMON  oracle@dcdrpcora1a (MMON)
15487 MMNL  oracle@dcdrpcora1a (MMNL)
...
15712 Q002  oracle@dcdrpcora1a (Q002)
15713 Q003  oracle@dcdrpcora1a (Q003)
21258       oracle@dcdrpcora1a  <-- server thread for my sqlplus connection, no PNAME
29468 W001  oracle@dcdrpcora1a (W001)

SQL> select stid, pname, program from v$process where spid = 15465 order by 1;

STID  PNAME PROGRAM
----- ----- ---------------------------------------------
15465 SCMN  oracle@dcdrpcora1a (SCMN)
15468 GEN0  oracle@dcdrpcora1a (GEN0)
15469 MMAN  oracle@dcdrpcora1a (MMAN)
15475 DBRM  oracle@dcdrpcora1a (DBRM)
15479 LGWR  oracle@dcdrpcora1a (LGWR)
15480 CKPT  oracle@dcdrpcora1a (CKPT)
15481 LG00  oracle@dcdrpcora1a (LG00)
15482 LG01  oracle@dcdrpcora1a (LG01)
15483 SMON  oracle@dcdrpcora1a (SMON)
15485 LREG  oracle@dcdrpcora1a (LREG)

As you can see, the latter one is very critical, as it contains LGWR, CKPT, etc. On Linux/UNIX, you can check a background process environment variable SKGP_HIDDEN_ARGS to see if it's fatal or non-fatal (previously just called background).

$ ps eww 15465
  PID TTY      STAT   TIME COMMAND
15465 ?        Ssl    0:03 ora_u004_test12c... SKGP_HIDDEN_ARGS=<FATAL/...

$ ps eww 15472
...SKGP_HIDDEN_ARGS=<NF/...

If it's fatal, killing the process would crash the instance. It's a quick way to find out whether you can release memory by killing a process. (Of course the real solution is elsewhere, fixing a bug for memory leak, e.g.)

zhu_henry

最近在研究C语言，看到这个帖子，感觉回答的不靠谱啊！特意登上来回复下我的研究结论：

1.listener只fork一个进程，这个进程调exec成为serverprocess
2.listener fork一个子进程，子进程再fork一个孙进程，孙进程调用exec成为serverprocess

1的局限性：listener进程fork子进程后如果不调wait，子进程终止后将成为僵尸进程，严重浪费系统资源；如果listener调用wait势必被阻塞，在子进程退出前，无法响应其它连接请求
2的优势：listener fork子进程,子进程fork孙进程后立马退出，这时候子进程变成僵尸进程,等listener调用Wait后正常退出，由于子进程已经退出变僵尸进程listener进程调用Wait处理不会被阻塞，可以立即响应其它连接请求，孙进程也可以自行终止被init回收不会变为僵尸进程浪费系统资源。

几行C代码就能验证的事。

Yong Huang

> 子进程终止后将成为僵尸进程，严重浪费系统资源

This statement is a little exageration. Defunct or zombie processes do not consume resources, except for occupying process slots. They don't use CPU or memory.

Everything else you said is fine.