& Kubernetes 
Lighttpd, Mongrel, Ruby on Rails, and Lighttpd Proxy backend Performance Comparison
Mar 18, 2009 - 6 minutesFirst I was kind of surprised at the lack of performance comparisons people have made between the different lighttpd proxies with a rails/mongrel cluster backend and which ones would be best.
Lets start off the the specs of the box: Intel Atom 330 CPU Standard 7,200RPM HDD CentOS 5.2 2GB Ram Lighttpd with 3 Mongrel Clusters running an online food ordering system I wrote (MySQL + Rails)
Other than that there was nothing special. I’ve done no kernel tweaking or any other performance optimizations with Lighttpd or any other part of the system.
Result Data / Proxy Type / Result
Hash 21.41 Round-Robin 39.73 First 40.07
Hash is the loser in the proxy test over all. The requests per second capabilities of it were nearly half of what Round-Robin and First were able to achieve with the 3 node mongrel cluster running in the background. This is to be expected though since the Hash algorithm (so far as I understand it) will send the same url requests to the same proxy to help caching, whereas round-robin and first proxy methods balance it over the 3 nodes (or however many you have running).
Hash 467.20772 Round-Robin 251.725273 First 249.578092
Round-Robin are nearly neck and neck with the total time taken to return 10,000 hits with 250 concurrency, while the hash algorithm still lags far behind.
Hash 59.74 Round-Robin 110.84 First 111.79
Since only one mongrel cluster is able to serve all the requests in Hash, where as 3 are able to in the Round Robin and First tests, the transfer rate is much higher. I’m sure this could have been improved even further had I put all static files like style sheets and images on a separate instance so Mongrel could just worry about the rails files instead of rails + static, but since this site only has about 5 or so images, I did not think it necessary.
Hash 22795 Round-Robin 17132 First 11936
This is the result data where First comes through, beating Round Robin by a good portion (~6 seconds) and proving to be the quickest to serve up the website.
Hash 1.15 Round-Robin 3.15 First 3.28
Since Hash is only using one mongrel cluster instance, instead of the 3, the CPU load was much lower. However, Round Robin and First were very close, but I believe that when comparing the request times for First and Round Robin, First is worth the very miniscule increase in server load.
While Phusion makes deploying rails apps a bit simpler / easier to add new apps, in the long run I view it as detrimental should your application ever grow. The reason I chose to work with mongrel and Lighttpd (although Nginx is another viable option) is because I looked at sites like GitHub and EngineYard (a large rails host) and checked a few of their web server headers and found them to be running Nginx (with a Mongrel backend I’m assuming).
If you’re worried about memory I did not notice much of a difference between what my application used while in Phusion vs Mongrel, but one of the HUGE differences I did notice was when running ab -n 1000 -c 10 the load would shoot up to 5, while Lighttpd’s would barely hit .89. I didn’t bother attempting ab -n 10000 -c 250 with apache since I didn’t want the box to melt/halt/be-killed-by-apache. If load isn’t an issue and you’re a die-hard Apache fan, stick with Phusion, but my tests put it way behind Lighttpd+Mongrel in comparison.
1
2If anyone has any suggestions for other analytic's to test let me know! I'd love to do some more research regarding this and post my findings since there seems to be a lack of them on google.
3
4<strong>Proxy: Hash Raw Data</strong>
5
6<pre>
7Server Port: 80
8
9Document Path: /
10Document Length: 2495 bytes
11
12Concurrency Level: 250
13Time taken for tests: 467.20772 seconds
14Complete requests: 10000
15Failed requests: 0
16Write errors: 0
17Total transferred: 28570261 bytes
18HTML transferred: 24950000 bytes
19Requests per second: 21.41 [#/sec] (mean)
20Time per request: 11675.520 [ms] (mean)
21Time per request: 46.702 [ms] (mean, across all concurrent requests)
22Transfer rate: 59.74 [Kbytes/sec] received
23
24Connection Times (ms)
25 min mean[+/-sd] median max
26Connect: 0 0 3.2 0 35
27Processing: 72 11547 1056.2 11694 22795
28Waiting: 72 11547 1056.2 11694 22795
29Total: 72 11548 1055.2 11694 22795
30
31Percentage of the requests served within a certain time (ms)
32 50% 11694
33 66% 11717
34 75% 11739
35 80% 11794
36 90% 11887
37 95% 11907
38 98% 11928
39 99% 11942
40 100% 22795 (longest request)
41</pre>
42
43<strong>Proxy Type: Round Robin Raw Data</strong>
44<pre>
45Server Port: 80
46
47Document Path: /
48Document Length: 2495 bytes
49
50Concurrency Level: 250
51Time taken for tests: 251.725273 seconds
52Complete requests: 10000
53Failed requests: 0
54Write errors: 0
55Total transferred: 28570213 bytes
56HTML transferred: 24950000 bytes
57Requests per second: 39.73 [#/sec] (mean)
58Time per request: 6293.132 [ms] (mean)
59Time per request: 25.173 [ms] (mean, across all concurrent requests)
60Transfer rate: 110.84 [Kbytes/sec] received
61
62Connection Times (ms)
63 min mean[+/-sd] median max
64Connect: 0 0 3.2 0 37
65Processing: 103 6192 2439.0 6653 17132
66Waiting: 103 6192 2439.0 6653 17132
67Total: 103 6192 2438.5 6653 17132
68
69Percentage of the requests served within a certain time (ms)
70 50% 6653
71 66% 7665
72 75% 8169
73 80% 8495
74 90% 9038
75 95% 10066
76 98% 10377
77 99% 10468
78 100% 17132 (longest request)
79</pre>
80
81<strong>Proxy Type: First Raw Data</strong>
82<pre>
83--- Proxy: First 3.28 ---
84Server Port: 80
85
86Document Path: /
87Document Length: 2495 bytes
88
89Concurrency Level: 250
90Time taken for tests: 249.578092 seconds
91Complete requests: 10000
92Failed requests: 0
93Write errors: 0
94Total transferred: 28570190 bytes
95HTML transferred: 24950000 bytes
96Requests per second: 40.07 [#/sec] (mean)
97Time per request: 6239.453 [ms] (mean)
98Time per request: 24.958 [ms] (mean, across all concurrent requests)
99Transfer rate: 111.79 [Kbytes/sec] received
100
101Connection Times (ms)
102 min mean[+/-sd] median max
103Connect: 0 0 3.3 0 37
104Processing: 307 6169 621.6 6256 11936
105Waiting: 306 6168 621.7 6256 11936
106Total: 307 6169 620.0 6256 11936
107
108Percentage of the requests served within a certain time (ms)
109 50% 6256
110 66% 6450
111 75% 6550
112 80% 6598
113 90% 6734
114 95% 6795
115 98% 6858
116 99% 6897
117 100% 11936 (longest request)
118</pre>
GitHub