互联网，请记住我 » php性能 http://www.162cm.com 一米六二的北漂生活 Tue, 18 Nov 2008 14:09:56 +0000 http://wordpress.org/?v=2.6.2 en 用fastCGI协议进行RPC调用 http://www.162cm.com/archives/560.html http://www.162cm.com/archives/560.html#comments Thu, 06 Dec 2007 12:57:13 +0000 xurenlu http://www.162cm.com/archives/560.html 近来由于业务量急剧增长,业务种类越来越多,RPC混乱问题就凸现出来.
具体表现是:
a1,a2,a3,a4 为一组服务器,为用户提供A服务.
b1,b2,b3为一组服务器,为用户提供B服务.
c1,c2,c3为一组服务器,为用户提供C服务.
最初各服务器各自服务是不太相关的.随着业务拓展,用户提出新的需求,我们需要
在A服务中调用B组服务器的API接口.
….
最后我们发现这样一个现象:
用户请求a1 提供服务.a1发现他需要知道另一些数据,于是调用b2的rpc服务.但是b2
上这个数据又需要c3的RPC.结果c3又需要a2来提供一些数据….
我们的远程调用调来调去,最后把自己调晕了.

于是我开始考虑一种解决方案。
首先当然是：对远程调用加以控制和规范，避免不必要的远程调用。
接着我注意到，我们利用CURL来通过http通道进行远程调用其实是没必要的。我可以在做远程调用时，走FastCGI协议。

说干就干:
第一步：建立一个fastCGI:spawn-php

  1. #!/bin/bash
  2.  
  3. ## ABSOLUTE path to the spawn-fcgi binary
  4. SPAWNFCGI="/home/y/opt/lighttpd/bin/spawn-fcgi"
  5.  
  6. ## ABSOLUTE path to the PHP binary
  7. FCGIPROGRAM="/home/y/opt/php/bin/php-cgi"
  8.  
  9. ## TCP port to which to bind on localhost
  10. FCGIPORT="1026"
  11.  
  12. ## number of PHP children to spawn
  13. PHP_FCGI_CHILDREN=10
  14.  
  15. ## maximum number of requests a single PHP process can serve before it is restarted
  16. PHP_FCGI_MAX_REQUESTS=1000
  17.  
  18. ## IP addresses from which PHP should access server connections
  19. FCGI_WEB_SERVER_ADDRS="127.0.0.1,192.168.2.10"
  20.  
  21. # allowed environment variables, separated by spaces
  22. ALLOWED_ENV="ORACLE_HOME PATH USER"
  23.  
  24. ## if this script is run as root, switch to the following user
  25. USERID=y
  26. GROUPID=y
  27.  
  28. ################## no config below this line
  29.  
  30. if test x$PHP_FCGI_CHILDREN = x; then
  31. PHP_FCGI_CHILDREN=5
  32. fi
  33.  
  34. export PHP_FCGI_MAX_REQUESTS
  35. export FCGI_WEB_SERVER_ADDRS
  36.  
  37. ALLOWED_ENV="$ALLOWED_ENV PHP_FCGI_MAX_REQUESTS FCGI_WEB_SERVER_ADDRS"
  38.  
  39. if test x$UID = x0; then
  40. EX="$SPAWNFCGI -p $FCGIPORT -f $FCGIPROGRAM -u $USERID -g $GROUPID -C $PHP_FCGI_CHILDREN"
  41. else
  42. EX="$SPAWNFCGI -p $FCGIPORT -f $FCGIPROGRAM -C $PHP_FCGI_CHILDREN"
  43. fi
  44.  
  45. # copy the allowed environment variables
  46. E=
  47.  
  48. for i in $ALLOWED_ENV; do
  49. E="$E $i=${!i}"
  50. done
  51.  
  52. # clean the environment and set up a new one
  53. env - $E $EX

2.接着我们建立这样的文件：http.txt

  1. for($i=0;$i<100;$i++)
  2. {
  3. echo rand();
  4. echo '....Scene of shooting at Nebraska shopping centre
  5. Eight killed in Nebraska shooting
  6. A gunman opens fire in a shopping centre in the US state of Nebraska, killing at least eight people.
  7. Scene of shooting at Nebraska shopping centre
  8. Eight killed in Nebraska shooting
  9. A gunman opens fire in a shopping centre in the US state of Nebraska, killing at least eight people.
  10. Scene of shooting at Nebraska shopping centre
  11. Eight killed in Nebraska shooting
  12. A gunman opens fire in a shopping centre in the US state of Nebraska, killing at least eight people.
  13. Scene of shooting at Nebraska shopping centre
  14. Eight killed in Nebraska shooting
  15. A gunman opens fire in a shopping centre in the US state of Nebraska, killing at least eight people.
  16. ';
  17. }
  18. print "done!";

这个做为是远程接口。
3.下面比较CURL和fastCGI两种访问remote API的方式:
A:走FastCGI:

  1. include "./mod_fcgi.php";
  2. $fcgi=new mod_fcgi();
  3. $args="127.0.0.1:1026";
  4. $filename="/home/y/www/qps/http.php";
  5. $fcgi->parser_open($args,$filename,$rq_err,$cgi_headers);
  6. print($fcgi->parsed_output);
  7. echo "\n=============\n";

B:走CURL通道:

  1. include "./curl.php";
  2. $curl=& new CURL();
  3. echo $curl->get("http://localhost/qps/http.php");
  4. echo "\n=============\n";

4我们比较FCGI方式和CURL方式的差异：
FCGI方式进行远程调用的QPS:

  1. Total transferred:      249306 bytes
  2. HTML transferred:       244410 bytes
  3. Requests per second:    275.77 [#/sec] (mean)
  4. Time per request:       3.626 [ms] (mean)
  5. Time per request:       3.626 [ms] (mean, across all concurrent requests)
  6. Transfer rate:          2233.76 [Kbytes/sec] received

HTTP通道进行远程调用的QPS:

  1. Total transferred:      2244515 bytes
  2. HTML transferred:       2240015 bytes
  3. Requests per second:    204.18 [#/sec] (mean)
  4. Time per request:       4.898 [ms] (mean)
  5. Time per request:       4.898 [ms] (mean, across all concurrent requests)
  6. Transfer rate:          14911.66 [Kbytes/sec] received

经过多次比较，确认用fastCGI通道时，效率比http方式要高。
但是….这个差别并不大，而且,使用lighttpd的spawn-cgi来充当fastcgi的守护进程，还不是特别稳定。
另外，经过我的测试，只有当远程调用的结果有较多的输出时，FastCGI协议与HTTP协议相比才有优势。在通过网络传输的数据不多的情况下（就是http.php几乎没有输出的情况下），用CURL来调远端调用，反而比FastCGI方式慢（显而易见，mod_fcgi是自行封装的用PHP的socket函数来进行网络操作的class,而Curl是一个编译了的php扩展).
因此，我忙活了半天，最后得出的结论是，目前还不能在公司大规模加以运用。
但是对于小型企业，创业型公司来说，这种效率对比还是很有用的，可以加以挖掘。

]]> http://www.162cm.com/archives/560.html/feed 用ab来做性能测评好像不公平吧? http://www.162cm.com/archives/457.html http://www.162cm.com/archives/457.html#comments Tue, 21 Aug 2007 15:47:14 +0000 xurenlu http://www.162cm.com/archives/457.html 我的师兄,从某种程度上说也是我php方面的师父,和他的同事做了测评,分别是针对php4,php5的性能对比和function,class的性能对比的。
这里是师兄的对比结果
师兄的同事做的测评
但是我不认为用ab来做出的测试是合理的.理由:用ab来测试时,结果好像总似是在跳舞一样.做为证据,我将我的notebook依文本方式重启,用ab来测lighttpd的表现.
具体环境:
Haier H40S Notebook
RAM:1.5G
CPU:CY 1.6GHz
硬盘:60G,(具体型号什么的不清楚,列个60G表示不是SCSI硬盘)
OS:Fedora Core 7.0(Moonshine)
web server:lighttpd 1.4.15
测试工具:ab
X环境:未运行X server
命令:ab -n 10000 -c 50 http://localhost:8181/
第一份结果:

  1. This is ApacheBench, Version 2.0.40-dev <$Revision: 1.146 $> apache-2.0
  2. Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
  3. Copyright 2006 The Apache Software Foundation, http://www.apache.org/

Benchmarking localhost (be patient)

Server Software: lighttpd/1.4.15
Server Hostname: localhost
Server Port: 8181

Document Path: /
Document Length: 345 bytes

Concurrency Level: 50
Time taken for tests: 1.786830 seconds
Complete requests: 10000
Failed requests: 0
Write errors: 0
Non-2xx responses: 10000
Total transferred: 4980000 bytes
HTML transferred: 3450000 bytes
Requests per second: 5596.50 [#/sec] (mean)
Time per request: 8.934 [ms] (mean)
Time per request: 0.179 [ms] (mean, across all concurrent requests)
Transfer rate: 2721.58 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 3 2.6 4 8
Processing: 1 4 0.7 5 6
Waiting: 0 1 1.7 2 5
Total: 5 8 2.2 8 13
WARNING: The median and mean for the processing time are not within a normal deviation
These results are probably not that reliable.

Percentage of the requests served within a certain time (ms)
50% 8
66% 10
75% 10
80% 11
90% 11
95% 12
98% 12
99% 12
100% 13 (longest request)

这里的结果是Requests per second: 5596.50 [#/sec] (mean)(插一句,lighttpd的性能真是好,用apache从来没上过3000)
第二份ab -n 10000 -c 50的结果:

  1. This is ApacheBench, Version 2.0.40-dev <$Revision: 1.146 $> apache-2.0
  2. Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
  3. Copyright 2006 The Apache Software Foundation, http://www.apache.org/

Benchmarking localhost (be patient)

Server Software: lighttpd/1.4.15
Server Hostname: localhost
Server Port: 8181

Document Path: /
Document Length: 345 bytes

Concurrency Level: 50
Time taken for tests: 1.313719 seconds
Complete requests: 10000
Failed requests: 0
Write errors: 0
Non-2xx responses: 10011
Total transferred: 4985478 bytes
HTML transferred: 3453795 bytes
Requests per second: 7611.98 [#/sec] (mean)
Time per request: 6.569 [ms] (mean)
Time per request: 0.131 [ms] (mean, across all concurrent requests)
Transfer rate: 3705.51 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 1 1.0 1 5
Processing: 1 5 2.0 5 32
Waiting: 0 3 1.8 3 30
Total: 1 6 2.1 6 35

Percentage of the requests served within a certain time (ms)
50% 6
66% 6
75% 6
80% 6
90% 7
95% 8
98% 10
99% 11
100% 35 (longest request)

结果表明RPS是Requests per second: 7611.98 [#/sec] (mean).
第三份结果:

  1. This is ApacheBench, Version 2.0.40-dev <$Revision: 1.146 $> apache-2.0
  2. Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
  3. Copyright 2006 The Apache Software Foundation, http://www.apache.org/

Benchmarking localhost (be patient)

Server Software: lighttpd/1.4.15
Server Hostname: localhost
Server Port: 8181

Document Path: /
Document Length: 345 bytes

Concurrency Level: 50
Time taken for tests: 1.300247 seconds
Complete requests: 10000
Failed requests: 0
Write errors: 0
Non-2xx responses: 10009
Total transferred: 4984482 bytes
HTML transferred: 3453105 bytes
Requests per second: 7690.85 [#/sec] (mean)
Time per request: 6.501 [ms] (mean)
Time per request: 0.130 [ms] (mean, across all concurrent requests)
Transfer rate: 3743.14 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 1 1.0 1 5
Processing: 1 4 1.4 5 11
Waiting: 0 2 1.1 3 8
Total: 1 6 1.2 6 14

Percentage of the requests served within a certain time (ms)
50% 6
66% 6
75% 6
80% 6
90% 7
95% 9
98% 10
99% 11
100% 14 (longest request)

RPS结果是Requests per second: 7690.85 [#/sec] (mean).
第四份结果是:

  1. This is ApacheBench, Version 2.0.40-dev <$Revision: 1.146 $> apache-2.0
  2. Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
  3. Copyright 2006 The Apache Software Foundation, http://www.apache.org/

Benchmarking localhost (be patient)

Server Software: lighttpd/1.4.15
Server Hostname: localhost
Server Port: 8181

Document Path: /
Document Length: 345 bytes

Concurrency Level: 50
Time taken for tests: 1.315784 seconds
Complete requests: 10000
Failed requests: 0
Write errors: 0
Non-2xx responses: 10033
Total transferred: 4996434 bytes
HTML transferred: 3461385 bytes
Requests per second: 7600.03 [#/sec] (mean)
Time per request: 6.579 [ms] (mean)
Time per request: 0.132 [ms] (mean, across all concurrent requests)
Transfer rate: 3708.06 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 1 1.2 1 8
Processing: 1 4 1.4 5 11
Waiting: 0 2 1.1 3 8
Total: 1 6 1.4 6 15

Percentage of the requests served within a certain time (ms)
50% 6
66% 6
75% 6
80% 6
90% 7
95% 9
98% 11
99% 12
100% 15 (longest request)

结果是Requests per second: 7600.03 [#/sec] (mean).
第五份结果是:

  1. This is ApacheBench, Version 2.0.40-dev <$Revision: 1.146 $> apache-2.0
  2. Copyright 1996 Adam Twiss, Zeus Technology Ltd, http://www.zeustech.net/
  3. Copyright 2006 The Apache Software Foundation, http://www.apache.org/

Benchmarking localhost (be patient)

Server Software: lighttpd/1.4.15
Server Hostname: localhost
Server Port: 8181

Document Path: /
Document Length: 345 bytes

Concurrency Level: 50
Time taken for tests: 1.812989 seconds
Complete requests: 10000
Failed requests: 0
Write errors: 0
Non-2xx responses: 10000
Total transferred: 4980000 bytes
HTML transferred: 3450000 bytes
Requests per second: 5515.75 [#/sec] (mean)
Time per request: 9.065 [ms] (mean)
Time per request: 0.181 [ms] (mean, across all concurrent requests)
Transfer rate: 2682.31 [Kbytes/sec] received

Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 3 2.7 4 9
Processing: 1 4 0.8 5 6
Waiting: 0 1 1.7 2 5
Total: 5 8 2.3 9 13
WARNING: The median and mean for the processing time are not within a normal deviation
These results are probably not that reliable.

Percentage of the requests served within a certain time (ms)
50% 9
66% 10
75% 10
80% 11
90% 12
95% 12
98% 12
99% 12
100% 13 (longest request)

RPS结果是Requests per second: 5515.75 [#/sec] (mean);

可以看到,同是ab -n 10000 -c 50 http://localhost:8181/,结果大不一样,不仅分为7600RPS和5500RPS两档,而各档中具体数据也不一样.这个结果可以理解,OS中总是会有各种各样的消息要处理,影响具体数据结果也很正常.
因此,谢华亮(sir?mm?)的测试用ab的结果来进行php4,php5的性能比较是不太对的.至少目前看来,我没有从他的文章中看到他进行了多次ab测评后取结果的平均值.但是他列出来这样一个结果:

  1. 版本 函数测试 不实例化类 实例化类 类的继承
  2. PHP 4.4.2 1047.23/rps 1034.98/rps 1006.14/rps 992.95/rps
  3. PHP 5.2.1 1176.06/rps 1197.17/rps 1187.93/rps 1128.54/rps

这大致可以说明,php5.2确实是比php4.4.2要快.但是从1197.17>1176.06得出php5的类跑得比functiong还快,我不这么认为.只是21个request的差别,太小了,ab测试的随便一个风吹草吹就行了(而且说class比function要快,严重地伤害了我的感情,哈哈)…
嗯,先整这么多.有空好好研究一下,如何对web程序进行性能测评.
另外,有空见到这两位兄台,好好争论一番.

]]> http://www.162cm.com/archives/457.html/feed 老覃的性能测试 http://www.162cm.com/archives/455.html http://www.162cm.com/archives/455.html#comments Mon, 20 Aug 2007 17:33:14 +0000 xurenlu http://www.162cm.com/archives/455.html 老覃同学做了一个测试，来比较PHP5下利用function,class来编程时的效率差异。
我一向喜欢钻牛角尖：
有两个疑问：
1.ab本身的文档说了,ab自身是需要CPU资源的,有时候要注意你可能不是在测apache的性能，而是在测ab这个软件的性能。老覃同志如何去除ab的影响?
2.php编译时也是花费CPU时间的。我做了另一组测试:

  1. //t1.php

function signin()

{

echo “signin”;

}

signin();

?>

ab -n 10000 -c 50 的结果是:
1534.15 [#/sec]
Time per request: 32.591 [ms] (mean)

,再另外一个:

function signin()

{

echo “signin”;

}

//signin();

?>

ab -n 10000 -c 50的结果是:
1748.48 [#/sec], Time per request: 28.596 。
就是说，算上ab的花费,apache的花费,php编译的花费,这几个的时间消费可能已经比php代码运行的时间花费要大了。加上运行的代码后,只相关几个ms。
如果要进行精确的比较,老覃的做法可能很不可靠。
PHP4与PHP5.20的效率目前还只有看官方的数据。
另外有个疑问，我的php比老覃同志的似乎快了不少,相差好几倍。可能是我的是Fedora 下运行的结果。

附注：

我后来做了一个实验,ab -n 10000 -c 50 http://localhost/test/23442.php ，返回的结果是:

  1. Requests per second:    2192.03 [#/sec] (mean)
  2. Time per request:       22.810 [ms] (mean)

这是在请求的文件不存在下,apache简单返回一个静态面的情况下的。由此看来，ab和apache的开销就已经占了大约20ms.

再注,在这里
是老覃的同事做的测试.我google得到了这个结果.

]]> http://www.162cm.com/archives/455.html/feed 聪明人的算法 http://www.162cm.com/archives/452.html http://www.162cm.com/archives/452.html#comments Tue, 14 Aug 2007 16:48:40 +0000 xurenlu http://www.162cm.com/archives/452.html 两段代码,都是用来1到100之前计算三次方的.
同样的代码,第一个其实只需要一行(当然要考虑美观,分成了四行);
第二个估计要用上100来行;但是哪个更快呢?….
读者说哪个程序更好?

  1.  

function getPower($n)
{
return $n*$n*$n;
}

global $data;
$data=array(
1=>1,
2=>8,
3=>27,
4=>64,
5=>125…
);
function getPower($n)
{
global $n;
return $data[$n];
}

]]> http://www.162cm.com/archives/452.html/feed php-APC介绍 http://www.162cm.com/archives/451.html http://www.162cm.com/archives/451.html#comments Tue, 14 Aug 2007 14:20:40 +0000 xurenlu http://www.162cm.com/archives/451.html 前面几篇文章提到了APC.很多人问我:APC是什么?这里做个比较全面的介绍:
APC,全称是Alternative PHP Cache,官方翻译叫”可选PHP缓存”,但我个人觉得应该叫”另一个PHP缓存”.因为这个东西如果叫”可选PHP缓存”,容易给人一种可要可不要的,不怎么有用的错觉.
APC的主页是:<a href=”http://pecl.php.net/package/apc” mce_href=”http://pecl.php.net/package/apc”>http://pecl.php.net/package/apc</a>.目前的版本是3.0.14,通过PECL安装,需要php4.3.0或更高版本.
<b>APC的安装:</b>
一般是下载源代码然后phpize来编译安装,安装完以后在加上php.ini里加上
<coolcode>
extension=apc.so
</coolcode>
这么一行就行了.
<b>APC的使用</b>
APC的使用其实倒说不上.APC是个优化器,自安装之日起,就默默地在后台为您的PHP应用服务了.您的所有PHP代码会被缓存起来.

另外,APC可提供一定的内存缓存功能.但是这个功能并不是十分完美,有报告说如果频繁使用APC缓存的写入功能,会导致不可预料的错误.如果想使用这个功能,可以看看apc_fetch,apc_store等几个与apc缓存相关的函数.
从PHP5.2开始,APC引入了一个小甜饼,解决了困扰大家已久的大文件上传的进度条问题.具体请看看我这篇<a href=”http://www.162cm.com/archives/406.html” mce_href=”http://www.162cm.com/archives/406.html”>blog</a>.
<b>APC的高级使用</b>
1.缓存期限:
APC的缓存分两部分:系统缓存和用户数据缓存.
系统缓存是自动使用的,是指APC把PHP文件源码的编译结果缓存起来,然后在再次调用时先对比时间标记。如果未过期,则使用缓存代码运行。默认缓存3600s(一小时).但是这样仍会浪费大量CPU时间.因此可以在php.ini中设置system缓存为永不过期(apc.ttl=0).不过如果这样设置,改运php代码后需要restart一下您的web服务器(比如apache…).目前对APC的性能测试一般指的是这一层cache;
用户数据缓存由用户在编写php代码时用apc_store和apc_fetch函数操作读取、写入的.如果量不大的话我建议可以使用一下.如果量大,我建议使用memcache会更好.
如果要享受APC带来的缓存大文件上传进度的特性,需要在php.ini中将apc.rfc1867设为1,并且在表单中加一个隐藏域APC_UPLOAD_PROGRESS,这个域的值可以随机生成一个hash,以确何唯一.具体例子请参见前面给出的链接.
2.状态控制和分析:
APC的源码包自带了一个apc.php;您可以将这个文件上传到web服务器的某个目录下,用浏览器访问,这会显示当前的状态.我们可以从这里的表格分析当前的缓存状况,作出进一步优化.
apc-info-clublocalhost2.png
这是某test站点的状态.您可以慢慢分析,这个工具会提供很多有用的工具.比如您可以看到哪些文件经常被包含(访问),您缓存的哪个变量经常被读取,或经常被更新等.
最后顺便提一句,有独立报告说,APC的代码缓存、优化效果要高出zend优化器.就算不是真的,他开源而又免费,实在是一个相当不错的选择.

]]> http://www.162cm.com/archives/451.html/feed