实例演示使用并行和非并行选项的性能差异

比较direct方式使用并行和非并行选项的性能差异。

并行测试

创建测试表

create table sqlldr_paral
nologging
PARTITION BY HASH (sfzh)
(partition p1 tablespace users , partition p2tablespace zdry_tbs2 )
AS SELECT *FROM sfxx2 where 1=2 ;

sqlldr_paral.ctl

LOAD DATA
INFILE 'd:\flatfile\flatfile.dat'
APPEND
INTO TABLE sqlldr_paral
fields terminated by X'23'
TRAILING NULLCOLS
(zj,
…字段太多省略
cjdwdm)
sqlldr_paral.par

USERID=zdrygk/zdrygk
CONTROL='d:\flatfile\sqlldr_paral.ctl'
DATA='d:\flatfile\flatfile.dat'
LOG='d:\flatfile\sqlldr_paral.log'
DIRECT=true
parallel=true
ERRORS=10000

执行

D:\flatfile>sqlldrparfile=d:\flatfile\sqlldr_paral.par

SQL*Loader: Release 11.2.0.1.0 - Production onMon Jan 7 23:21:42 2013


Copyright (c) 1982, 2009, Oracle and/or itsaffiliates. All rights reserved.


Load completed - logical record count 3022375.
sqlldr_paral.log

Control File: d:\flatfile\sqlldr_paral.ctl
Data File: d:\flatfile\flatfile.dat
BadFile: d:\flatfile\flatfile.bad
DiscardFile: none specified

(Allowall discards)

Number to load: ALL
Number to skip: 0
Errors allowed: 10000
Continuation: none specified
Path used: Direct- with parallel option.
-----这里指定了并行选项

Table SQLLDR_PARAL, loaded from every logicalrecord.
Insert option in effect for this table: APPEND
TRAILING NULLCOLS option in effect




Table SQLLDR_PARAL:
3017264Rows successfully loaded.
5111Rows not loaded due to data errors.
0 Rows not loaded because all WHEN clauseswere failed.
0 Rowsnot loaded because all fields were null.

Dateconversion cache disabled due to overflow (default size: 1000)

Partition P1: 1509091 Rows loaded.
Partition P2: 1508173 Rows loaded.

Bind array size not used in direct path.
Column array rows : 5000
Stream buffer bytes: 256000
Read buffer bytes: 1048576

Total logical records skipped: 0
Total logical records read: 3022375
Total logical records rejected: 5111
Total logical records discarded: 0
Total stream buffers loaded by SQL*Loader mainthread: 1301
Total stream buffers loaded by SQL*Loader loadthread: 3901

Run began on Mon Jan 07 23:21:42 2013
Run ended on Mon Jan 07 23:22:18 2013

Elapsed time was: 00:00:35.91
CPU time was: 00:00:25.79

非并行测试

创建测试表

create table sqlldr_nopal
nologging
PARTITION BY HASH (sfzh)
(partition p1 tablespace users , partition p2tablespace zdry_tbs2 )
AS SELECT *FROM sfxx2 where 1=2 ;

sqlldr_nopar.ctl

LOAD DATA
INFILE 'd:\flatfile\flatfile.dat'
APPEND
INTO TABLEsqlldr_nopar
fields terminated by X'23'
TRAILING NULLCOLS
(zj,
…字段太多省略
cjdwdm)
sqlldr_nopar.par

这里删除了并行选项

USERID=zdrygk/zdrygk
CONTROL='d:\flatfile\sqlldr_nopar.ctl'
DATA='d:\flatfile\flatfile.dat'
LOG='d:\flatfile\sqlldr_nopar.log'
DIRECT=true
ERRORS=10000

执行

D:\flatfile>sqlldrparfile=d:\flatfile\sqlldr_nopar.par

SQL*Loader: Release 11.2.0.1.0 - Production onMon Jan 7 23:20:54 2013

Copyright (c) 1982, 2009, Oracle and/or itsaffiliates. All rights reserved.


Load completed - logical record count 3022375.
sqlldr_nopar.log

Control File: d:\flatfile\sqlldr_nopar.ctl
Data File: d:\flatfile\flatfile.dat
BadFile: d:\flatfile\flatfile.bad
DiscardFile: none specified

(Allowall discards)

Number to load: ALL
Number to skip: 0
Errors allowed: 10000
Continuation: none specified
Path used: Direct
----这里只是指定了直接路径加载

Table SQLLDR_NOPAR, loaded from every logicalrecord.
Insert option in effect for this table: APPEND
TRAILING NULLCOLS option in effect

Table SQLLDR_NOPAR:
3017264Rows successfully loaded.
5111Rows not loaded due to data errors.
0 Rows not loaded because all WHEN clauseswere failed.
0 Rowsnot loaded because all fields were null.

Dateconversion cache disabled due to overflow (default size: 1000)

Partition P1: 1509091 Rows loaded.
Partition P2: 1508173 Rows loaded.

Bind array size not used in direct path.
Column array rows : 5000
Stream buffer bytes: 256000
Read buffer bytes: 1048576

Total logical records skipped: 0
Total logical records read: 3022375
Total logical records rejected: 5111
Total logical records discarded: 0
Total stream buffers loaded by SQL*Loader mainthread: 1301
Total stream buffers loaded by SQL*Loader loadthread: 3901

Run began on Mon Jan 07 23:20:54 2013
Run ended on Mon Jan 07 23:21:28 2013

Elapsed time was: 00:00:33.81
CPU time was: 00:00:25.32

分析

Load Mode	Elapsed time(ss.99)
Direct	33.81
Direct + parallel	35.91

笔者做了多次试验，除了第一次消耗时间较长（因为表空间自动拓展），后面的实验中两种模式的用时大致相同。自己猜测在使用Direct 直接路径加载的情况下，多线程并不能提高效率。

在Util中找到关于sql*loader command line的描述

PARALLEL(parallel load)

Default: false

PARALLELspecifies whether direct loads canoperate in multiple concurrent sessions

to load data into the same table.

parallel这个参数用来设定使用direct loads的时候是否使用并发的session去加载数据到相同的表中。

于是测试了多文件，多session加载数据的情况。

D:\flatfile>sqluldr2 user=zdrygk/zdrygk@orclquery=sfxx2 degree=8 file=d:\flatfile\ldrfiles%t.dat field=0x23 size=300MB

0rows exported at 2013-01-08 00:07:26, size 0 MB.

output file d:\flatfile\ldrfiles1357574846.dat closed at 707972 rows,size 304 MB.

output file d:\flatfile\ldrfiles1357574861.dat closed at 697498 rows,size 604 MB.

output file d:\flatfile\ldrfiles1357574876.dat closed at 701379 rows,size 904 MB.

output file d:\flatfile\ldrfiles1357574891.dat closed at 692556 rows,size 1204 MB.

222868rows exported at 2013-01-08 00:08:31, size 1300 MB.

output file d:\flatfile\ldrfiles1357574906.dat closed at 222868 rows,size 1300 MB.



一共5个文件。

多文件并行测试

如果指定了parallel选项，sqlldr运行的时候会对表加4级锁，允许多个sqlldr session同时对同一个表执行加载作用。在我们执行了truncate 操作后对表进行并行加载操作。

parallelpar1文件

USERID=zdrygk/zdrygk

CONTROL='D:\flatfile\multifile\paral\sqlldr_paral_1.ctl'

DATA='D:\flatfile\multifile\ldrfiles1357574846.dat'

LOG='D:\flatfile\multifile\paral\sqlldr_paral_1.log'

DIRECT=true

parallel=true

ERRORS=10000

parallel par2、3、4、5的内容大致相同。

在windows下只能打开多个窗口，通过复制粘贴的方法保证并行了。

执行

sqlldrparfile=D:\flatfile\multifile\paral\sqlldr_paral_1.par

sqlldrparfile=D:\flatfile\multifile\paral\sqlldr_paral_2.par

sqlldrparfile=D:\flatfile\multifile\paral\sqlldr_paral_3.par

sqlldr parfile=D:\flatfile\multifile\paral\sqlldr_paral_4.par

sqlldrparfile=D:\flatfile\multifile\paral\sqlldr_paral_5.par

最后得到了5个log文件关键内容：

log1

Run began on Tue Jan 08 01:51:11 2013

Run ended on Tue Jan 08 01:51:21 2013



Elapsed time was: 00:00:09.42

CPU time was: 00:00:06.43

…

中间的几个log文件省略

log5

Run began on Tue Jan 08 01:51:34 2013

Run ended on Tue Jan 08 01:51:38 2013



Elapsed time was: 00:00:03.27

CPU time was: 00:00:02.03



因为最后一个sqlldr session 是最后一个执行结束的，所以需要计算并行sqlldr的执行消耗时间，只需要使用log5中的结束时间减去log1中的开始时间即可。

01:51:38-01:51:11=22秒

因为不是使用程序执行这几个sqlldr程序，复制、粘贴、执行之间会有空隙。真实用时应该比22秒这个数字要小的多。

多文件非并行测试

因为如果没有指定parallel选项，sqlldr运行的时候会对表加6级锁，其他sqlldr session会无法工作，所以只能串行的执行sqlldr操作。在我们执行了truncate操作之后对表进行串行加载操作。

noparallelpar1文件

USERID=zdrygk/zdrygk

CONTROL='D:\flatfile\multifile\nopar\sqlldr_nopar_1.ctl'

DATA='D:\flatfile\multifile\ldrfiles1357574846.dat'

LOG='D:\flatfile\multifile\nopar\sqlldr_nopar_1.log'

DIRECT=true

ERRORS=10000

noparallel par2、3、4、5的内容大致相同。

执行

sqlldrparfile=D:\flatfile\multifile\nopar\sqlldr_nopar_1.par

sqlldr parfile=D:\flatfile\multifile\nopar\sqlldr_nopar_2.par

sqlldrparfile=D:\flatfile\multifile\nopar\sqlldr_nopar_3.par

sqlldrparfile=D:\flatfile\multifile\nopar\sqlldr_nopar_4.par

sqlldrparfile=D:\flatfile\multifile\nopar\sqlldr_nopar_5.par



最后得到了5个log文件关键内容：

log1

Run began on Tue Jan 08 01:47:35 2013

Run ended on Tue Jan 08 01:47:48 2013



Elapsed time was: 00:00:12.61

CPU time was: 00:00:06.92

…

中间的几个log文件省略

log5

Run began on Tue Jan 08 01:48:26 2013

Run ended on Tue Jan 08 01:48:30 2013



Elapsed time was: 00:00:03.68

CPU time was: 00:00:02.01



最后执行时间相加

12.61+ 12.54+ 12.09+ 11.40+ 03.68=52.32秒

总结

Load Mode	Elapsed time(ss.99)
Direct	52.32
Direct + parallel	22

sqlldr是cpu密集型的操作，并且多线程，只有在使用多个数据文件，并且开启多个sqlldr session的时候才会有效果。并且效果非常明显。在消耗更多cpu和内存的同时能够给予非常高的效率，在真实情况下与单文件加载相比效率相差有1倍以上。