php5-cgi+Nginx: which is faster 127.0.0.1:9000 or Unix sockets?

[Joe Aston]

Thanks for sharing - this is very interesting.

How did you create the socket for fastcgi?

On Thu, Nov 6, 2008 at 2:07 PM, Mark Alan <varia at e-healthexpert.org> wrote:
> We intended to test server performance using either:
>  /usr/bin/php-cgi -q -b 127.0.0.1:9000  # i.e., the classic IP:port
> or:
>  /usr/bin/php-cgi -q -b /tmp/php-fastcgi.socket  # i.e. using Unix domain
> sockets (and having taken care to change the relevant lines in nginx config
> files from: fastcgi_pass 127.0.0.1:9000 to fastcgi_pass
> unix:/tmp/php-fastcgi.socket)
>
> In order to verify the impact of the differences described at:
> <http://lists.freebsd.org/pipermail/freebsd-performance/2005-February/001143.html>
>
> *Material*
> - Ubuntu 8.04.1 LTS server
> - Nginx 0.6.32 installed with:
>  wget
> http://http.us.debian.org/debian/pool/main/n/nginx/nginx_0.6.32-3_i386.deb
>  sudo gdebi nginx_0.6.32-3_i386.deb
> - PH5 5.2.4 installed with:
>  sudo aptitude -R install php5-cli php5-cgi php5-mysql php5-imap
> - Apache Bench 2.0.40-dev installed with:
>  sudo aptitude -R install apache2-utils
>
> *Methods*
> - Create php test script with:
> echo -e '<?php\n;\n?>' | sudo -u www-data tee
> /var/www/nginx-default/benchtest.php
> - Start php-cgi using the Ubuntu's UPSTART script posted as "20081106 - How
> to enable Nginx to serve PHP code/pages in UBUNTU" in this Nginx mailing
> list.
> - Start Apache Bench script with:
>  sudo ab -n 10000 -c 100 http://127.0.0.1/benchtest.php | tee bench-ip.log #
> after configuring php-cgi to listen at 127.0.0.1:9000
>  sudo ab -n 10000 -c 100 http://127.0.0.1/benchtest.php | tee
> bench-socket.log #after configuring php-cgi a unix socket at
> /tmp/php-fastcgi.socket
>
> *Results and Discussion*
> We made a delta file (bench-diff.txt) using the linux diff utility to show
> the differences between the two approaches:
> - diff -Naur bench-ip.log bench-socket.log > bench-diff.txt
>
> Using the described testing procedure shows that using using Unix domain
> sockets, besides the elsewhere described advantage of being  more secure
> (see external link above), significantly improves server cgi performance.
>
>
> Regards,
>
> M.
>
>
> File bench-diff.txt follows:
> --- cut here ---
> --- bench-ip.log        2008-11-06 13:22:30.000000000 +0000
> +++ bench-socket.log    2008-11-06 13:18:15.000000000 +0000
> @@ -13,31 +13,31 @@
>  Document Length:        0 bytes
>
>  Concurrency Level:      100
> -Time taken for tests:   5.708770 seconds
> +Time taken for tests:   4.980691 seconds
>  Complete requests:      10000
>  Failed requests:        0
>  Write errors:           0
>  Total transferred:      1510000 bytes
>  HTML transferred:       0 bytes
> -Requests per second:    1751.69 [#/sec] (mean)
> -Time per request:       57.088 [ms] (mean)
> -Time per request:       0.571 [ms] (mean, across all concurrent requests)
> -Transfer rate:          258.20 [Kbytes/sec] received
> +Requests per second:    2007.75 [#/sec] (mean)
> +Time per request:       49.807 [ms] (mean)
> +Time per request:       0.498 [ms] (mean, across all concurrent requests)
> +Transfer rate:          295.94 [Kbytes/sec] received
>
>  Connection Times (ms)
>               min  mean[+/-sd] median   max
> -Connect:        0    0   1.4      0      27
> -Processing:     9   55   9.5     55     156
> -Waiting:        9   55   9.3     55     154
> -Total:         15   56   9.0     55     156
> +Connect:        0    1   4.1      0      35
> +Processing:     4   47  12.8     48     145
> +Waiting:        4   46  12.6     48     129
> +Total:         12   48  11.1     49     145
>
>  Percentage of the requests served within a certain time (ms)
> -  50%     55
> -  66%     57
> -  75%     59
> -  80%     61
> -  90%     67
> -  95%     71
> -  98%     81
> -  99%     91
> - 100%    156 (longest request)
> +  50%     49
> +  66%     51
> +  75%     53
> +  80%     55
> +  90%     60
> +  95%     69
> +  98%     78
> +  99%     83
> + 100%    145 (longest request)
> --- cut here ---
>
>