目的
测试个人经常使用的不同 App server（主要用来做 API）的吞吐率。
测试主机
- 系统： MacBook Air (11-inch, Early 2015)
- CPU: 1.6 GHz Intel Core i5 x4 (4 核)
- 内存： 4 GB 1600 MHz DDR3
参数对比
1. Golang Gin 框架（协程模式）
- Go 1.6
- Gin v1.0rc1
2. Node Koa 框架 (多进程的异步IO模式)
- Node v6.1.0
- Koa 1.2.0
3. Ruby Sinatra 框架 (多进程多线程模式)
- Ruby 2.3.1p112
- Sinatra 1.4.7
- Puma 3.4.0
测试方法
- 使用 wrk 测试工具
- 压力参数参考 http://www.littlelines.com/blog/2014/07/08/elixir-vs-ruby-showdown-phoenix-vs-rails/
- 分别测试
/users(有 DB 查询) 和/static（无 DB 查询）
数据准备
采用 mongodb 作为测试，并选用了各自最为流行的 mongo 驱动。
进入 mongo console 执行以下命令：
use ruby2go2node2elixir
db.users.insert([
{"_id": ObjectId(), "username": "Josh Williams", "created_at": new Date(), "updated_at": new Date()},
{"_id": ObjectId(), "username": "Josh Williams", "created_at": new Date(), "updated_at": new Date()},
{"_id": ObjectId(), "username": "Josh Williams", "created_at": new Date(), "updated_at": new Date()},
{"_id": ObjectId(), "username": "Josh Williams", "created_at": new Date(), "updated_at": new Date()},
{"_id": ObjectId(), "username": "Josh Williams", "created_at": new Date(), "updated_at": new Date()}
])
启动服务：
所有命令都是在项目根目录前提下执行：
Gin (golang)（绑定在3000端口）
cd go
source env.sh
make
Koa (nodejs)（绑定在3001端口）
cd nodejs
npm install
npm start
Sinatra (ruby)（绑定在3002端口）
cd ruby
bundle
puma -C puma.rb
Wrk 本地测试结果
Gin (golang)
wrk -t4 -c100 -d30S --timeout 2000 "http://127.0.0.1:3000/static"
Running 30s test @ http://127.0.0.1:3000/static
4 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 14.63ms 59.10ms 779.91ms 96.68%
Req/Sec 5.42k 1.37k 20.49k 86.36%
634806 requests in 30.10s, 167.09MB read
Requests/sec: 21092.78
Transfer/sec: 5.55MB
wrk -t4 -c100 -d30S --timeout 2000 "http://127.0.0.1:3000/users"
Running 30s test @ http://127.0.0.1:3000/users
4 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 17.63ms 5.65ms 125.03ms 80.56%
Req/Sec 1.43k 317.13 2.19k 59.50%
171245 requests in 30.03s, 143.06MB read
Requests/sec: 5702.07
Transfer/sec: 4.76MB
Koa (nodejs)
#### 多进程模式
wrk -t4 -c100 -d30S --timeout 2000 "http://127.0.0.1:3001/static"
Running 30s test @ http://127.0.0.1:3001/static
4 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 10.77ms 8.97ms 199.54ms 94.74%
Req/Sec 2.56k 374.25 4.50k 87.25%
305517 requests in 30.03s, 84.20MB read
Requests/sec: 10172.22
Transfer/sec: 2.80MB
wrk -t4 -c100 -d30S --timeout 2000 "http://127.0.0.1:3001/users"
Running 30s test @ http://127.0.0.1:3001/users
4 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 44.25ms 18.20ms 460.36ms 85.57%
Req/Sec 576.93 83.73 0.85k 76.90%
68859 requests in 30.01s, 56.08MB read
Requests/sec: 2294.72
Transfer/sec: 1.87MB
##### 单进程模式
wrk -t4 -c100 -d30S --timeout 2000 "http://127.0.0.1:3001/static"
Running 30s test @ http://127.0.0.1:3001/static
4 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 21.21ms 1.45ms 47.63ms 93.30%
Req/Sec 1.18k 103.09 1.88k 86.25%
141520 requests in 30.06s, 39.00MB read
Requests/sec: 4708.48
Transfer/sec: 1.30MB
wrk -t4 -c100 -d30S --timeout 2000 "http://127.0.0.1:3001/users"
Running 30s test @ http://127.0.0.1:3001/users
4 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 73.37ms 10.07ms 212.29ms 87.93%
Req/Sec 342.42 66.59 505.00 63.45%
40923 requests in 30.08s, 33.33MB read
Requests/sec: 1360.40
Transfer/sec: 1.11MB
Sinatra (ruby)
wrk -t4 -c100 -d30S --timeout 2000 "http://127.0.0.1:3002/static"
Running 30s test @ http://127.0.0.1:3002/static
4 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 22.83ms 28.96ms 392.78ms 88.99%
Req/Sec 1.23k 334.43 2.07k 60.92%
147405 requests in 30.06s, 35.85MB read
Requests/sec: 4903.72
Transfer/sec: 1.19MB
wrk -t4 -c100 -d30S --timeout 2000 "http://127.0.0.1:3002/users"
Running 30s test @ http://127.0.0.1:3002/users
4 threads and 100 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 153.62ms 121.39ms 1.09s 73.08%
Req/Sec 159.87 36.28 287.00 69.44%
19111 requests in 30.05s, 16.51MB read
Requests/sec: 635.96
Transfer/sec: 562.70KB
吞吐率倍数关系
- 无 DB 查询:
gin > koa(多进程) > koa(单进程) ~= sinatra
(2x) (2x)
- 有 DB 查询:
gin > koa(多进程) > koa(单进程) > sinatra (2x) (2x) (2x)
结论：
- 在 4 核 CPU 条件下， golang 并发大概是 Nodejs 的两倍，Ruby 8 倍。因为 golang 有更好的并发模型，所以核数越多，这个差距越大 。
- 多进程的并发模型，代价其实挺大的。 比如 Node 多进程模式并发能力并没有随核数成倍增加，大致是其核数／2 的提升。
- Ruby GIL 缘故，其并发能力最差不足为奇。但是现在主流的都是微服务，RESTful API，一些性能要求较高的 业务完全可以考虑其他性能更好的语言。在其他语言领域，有些 ruby 棘手的问题，或许根本就不存在。