最近實在好
不過我放上面那張圖並不是說我一秒鐘幾十萬上下所以很忙，而是我新接觸的產業對於系統效能是很講究，因為真的是一秒鐘幾十萬上下~:P
以前可以偷懶只要求功能有做出來就好，效能不要太差就好，但是現在的產品對於系統反應速度(Response time )可以說是錙銖必較，所以想盡辦法都得從每個地方榨出效能。
通常一個系統效能可以調教的部分主要有三個部分：OS, Web Server, DB，最近我就先從最基本的Web端著手。
最近看到這個投影片，裡面也講到，不要偷懶上網看人家stackoverflow 的文章就直接下結論什麼東西比較好，也不要直接抄別人的參數，因為每個應用的情境不同，可以調教的參數和選擇的點也會差很多，而且網路上很多測試評比也有很多可議之處，還是自己來比較準，參考 Lies, Damned Lies and Benchmarks
所以那就只好捲起袖子來自己動手做實驗，我得實驗方法也很間單，以Spring boot 為基礎，然後抽換Embedded Web Container來測試，主要測試對象Jetty和Tomcat。
我的測試專案：Servlet Container Test
Jetty 是以NIO為基礎的Container ，從第九版之後做了大幅度的架構重整，號稱效能很威，然後調教的部分可以參考他官網wiki：Optimizing Jetty (雖然沒啥內容....)
可是Tomcat 就沒那麼簡單了，他光是Connector就有好幾種，對於系統效能影響有很大，主要分為以下幾種：
1. [BIO] Http11Protocol (BIO 就是 Blocking I/O 縮寫)
2. [NIO] Http11NioProtocol
3. [NIO2] Http11Nio2Protocol (tomcat 8以後推出)
4. [APR] Http11AprProtocol (Native implement)
詳細資料可以參考官方文件：Connector Comparison
如果要使用APR 就必須先安裝Native Lib 比較麻煩，可以參考這篇 CentOS: Installing Apache Portable Runtime (APR) for Tomcat
測試系統Azure A2 機器
簡單的測試指令： ab -n 1000000 -c 300
Tomcat 8 BIO
Requests per second: 2602.83 [#/sec] (mean)
Time per request: 115.259 [ms] (mean)
Time per request: 0.384 [ms] (mean, across all concurrent requests)
Transfer rate: 538.87 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 1 65 316.2 1 15196
Processing: 2 50 54.2 36 8608
Waiting: 2 47 51.6 35 8608
Total: 3 115 324.1 38 15316
Percentage of the requests served within a certain time (ms)
50% 38
66% 48
75% 58
80% 66
90% 117
95% 1032
98% 1066
99% 1115
100% 15316 (longest request)
tomcat 8 apr
Requests per second: 2657.37 [#/sec] (mean)
Time per request: 112.893 [ms] (mean)
Time per request: 0.376 [ms] (mean, across all concurrent requests)
Transfer rate: 550.16 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 1 55 240.8 3 7123
Processing: 2 58 145.5 40 8087
Waiting: 1 38 53.5 27 3752
Total: 3 113 287.3 45 8986
Percentage of the requests served within a certain time (ms)
50% 45
66% 56
75% 67
80% 77
90% 142
95% 570
98% 1067
99% 1102
100% 8986 (longest request)
tomcat 8 nio2
Requests per second: 2658.73 [#/sec] (mean)
Time per request: 112.836 [ms] (mean)
Time per request: 0.376 [ms] (mean, across all concurrent requests)
Transfer rate: 550.44 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 1 63 300.5 1 15148
Processing: 1 49 50.0 35 3198
Waiting: 1 46 46.0 34 3197
Total: 2 113 307.6 37 15200
Percentage of the requests served within a certain time (ms)
50% 37
66% 47
75% 56
80% 65
90% 122
95% 1030
98% 1064
99% 1107
100% 15200 (longest request)
jetty 9
Requests per second: 2829.32 [#/sec] (mean)
Time per request: 106.033 [ms] (mean)
Time per request: 0.353 [ms] (mean, across all concurrent requests)
Transfer rate: 549.84 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 1 72 328.7 1 15158
Processing: 2 34 37.9 24 4125
Waiting: 1 34 37.8 24 4125
Total: 3 106 333.1 27 15388
Percentage of the requests served within a certain time (ms)
50% 27
66% 35
75% 42
80% 49
90% 86
95% 1026
98% 1055
99% 1091
100% 15388 (longest request)
從RPS來看 jetty 9 > tomcat NIO2 ~ tomcat APR < tomcat BIO
不過說實在也差不了多少(何苦這樣比較呢~~開始自暴自棄)
此外nio 的在worst case 都會炸裂，打回跟BIO差不多程度?
這就結束了嗎? 不...
在許多Production 環境的架構通常會在Servlet Container 前面放台Proxy(Apache or Nginx)，這時候故事又可能不一樣了...
將會有以下案例：
1. Apache (mod_proxy) <---->Jetty 9
2. Nginx (mod_proxy) <---> Jetty 9 [註1]
3. Apache (mod_proxy_ajp) <----> tomcat 8 + AjpNioProtocol (NIO)
4. Apache (mod_proxy_ajp) <----> tomcat 8 + AjpAprProtocol (APR)
5. Nginx (mod_proxy_ajp) <----> tomcat 8 + AjpNioProtocol (NIO)
6. Nignx (mod_proxy_ajp) <----> tomcat 8 + AjpAprProtocol (APR)[註2]
5. Nginx (mod_proxy) <----> tomcat 8 + NioProtocol (NIO)
6. Nignx (mod_proxy) <----> tomcat 8 + AprProtocol (APR)
[註1] Jetty 本身不建議使用AJP
[註2] Nginx 本身沒有提供AJP module 只有3rd-party 所以效能和穩定度待確認
這麼多排列組合看了就累了....XDrz....
Referenc:
[1] NIO 2 in Apache Tomcat 8
[2] Tomcat 8（十）HTTP/AJP Connector、Bio/Nio/Apr性能对比
[3] How to Optimize Tomcat Performance
[4] Planning for High Concurrency Load Tests with JMeter
沒有留言 :
張貼留言