HTTP压测工具之wrk

前言

wrk是一款简单的HTTP压测工具,托管在Github上,https://github.com/wg/wrk.
wrk 的一个很好的特性就是能用很少的线程压出很大的并发量. 原因是它使用了一些操作系统特定的高性能 io 机制, 比如 select, epoll, kqueue 等. 其实它是复用了 redis 的 ae 异步事件驱动框架. 确切的说 ae 事件驱动框架并不是 redis 发明的, 它来至于 Tcl的解释器 jim, 这个小巧高效的框架, 因为被 redis 采用而更多的被大家所熟知.

安装

git clone https://github.com/wg/wrk.git  
cd wrk  
make  

如果报错：

src/wrk.h:11:25: fatal error: openssl/ssl.h: No such file or directory  
 #include <openssl/ssl.h>  

则需要安装openssl,使用sudo apt-get install libssl-dev或 sudo yum install openssl-devel安装即可,最后编辑etc/profile配置环境变量。由于笔者使用的是阿里云centos7,相关依赖都已经存在了,所以可以直接使用。

测试一下

wrk -t12 -c100 -d30s http://www.baidu.com  
[root@jerrik /]# wrk -t12 -c100 -d30s http://www.baidu.com  
Running 30s test @ http://www.baidu.com
  12 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   211.76ms  304.92ms   1.97s    88.17%
    Req/Sec    72.93     68.72   797.00     90.97%
  23725 requests in 30.05s, 347.47MB read
  Socket errors: connect 0, read 48, write 0, timeout 50
Requests/sec:    789.57
Transfer/sec:     11.56MB
[root@jerrik /]# 

一般线程数不宜过多. 核数的2到4倍足够了. 多了反而因为线程切换过多造成效率降低. 因为 wrk 不是使用每个连接一个线程的模型, 而是通过异步网络 io 提升并发量. 所以网络通信不会阻塞线程执行. 这也是 wrk 可以用很少的线程模拟大量网路连接的原因. 而现在很多性能工具并没有采用这种方式, 而是采用提高线程数来实现高并发. 所以并发量一旦设的很高, 测试机自身压力就很大. 测试效果反而下降.

参数解析

• 12 threads and 100 connections:
  总共是12个线程,100个连接(不是一个线程对应一个连接)

• latency和Req/Sec:
  代表单个线程的统计数据,latency代表延迟时间,Req/Sec代表单个线程每秒完成的请求数，他们都具有平均值, 标准偏差, 最大值, 正负一个标准差占比。一般我们来说我们主要关注平均值和最大值. 标准差如果太大说明样本本身离散程度比较高. 有可能系统性能波动很大.

• 23725 requests in 30.05s, 347.47MB read
  在30秒之内总共有23725个请求,总共读取347.47MB的数据

• Socket errors: connect 0, read 48, write 0, timeout 50
  总共有48个读错误,50个超时.

• Requests/sec和Transfer/sec
  所有线程平均每秒钟完成了789.57个请求,每秒钟读取11.56MB数据量

如果想看看响应时间的分布,可以增加–latency:

wrk -t12 -c100 -d30s --latency http://www.baidu.com 
[root@jerrik ~]# wrk -t12 -c100 -d30s --latency http://www.baidu.com 
Running 30s test @ http://www.baidu.com
  12 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   204.30ms  287.90ms   1.97s    88.61%
    Req/Sec    71.43     67.59   810.00     89.77%
  Latency Distribution
     50%   14.76ms
     75%  296.79ms
     90%  545.03ms
     99%    1.40s 
  23676 requests in 30.03s, 346.84MB read
  Socket errors: connect 0, read 42, write 0, timeout 46
Requests/sec:    788.29
Transfer/sec:     11.55MB

说明有50%的请求在14.76ms之内,90%在545.03ms之内。

高级用法

wrk可以结合lua来做,通过wrk提供的几个lua函数来对请求进行修改,结果输出、设置延迟等操作。下面来看看wrk提供的几个lua函数:

• setup 函数
  这个函数在目标 IP 地址已经解析完, 并且所有 thread 已经生成, 但是还没有开始时被调用. 每个线程执行一次这个函数.
  可以通过thread:get(name), thread:set(name, value)设置线程级别的变量.
• init 函数
  每次请求发送之前被调用.
  可以接受 wrk 命令行的额外参数. 通过 – 指定.
• delay函数
  这个函数返回一个数值, 在这次请求执行完以后延迟多长时间执行下一个请求. 可以对应 thinking time 的场景.
• request函数
  通过这个函数可以每次请求之前修改本次请求的属性. 返回一个字符串. 这个函数要慎用, 会影响测试端性能.
• response函数
  每次请求返回以后被调用. 可以根据响应内容做特殊处理, 比如遇到特殊响应停止执行测试, 或输出到控制台等等.

function response(status, headers, body)  
   if status ~= 200 then  
      print(body)  
      wrk.thread:stop()  
   end  
end  

• done函数
  在所有请求执行完以后调用, 一般用于自定义统计结果.

done = function(summary, latency, requests)  
   io.write("------------------------------\n")  
   for _, p in pairs({ 50, 90, 99, 99.999 }) do  
      n = latency:percentile(p)  
      io.write(string.format("%g%%,%d\n", p, n))  
   end  
end  

wrk官网提供的setup.lua实例:
-- example script that demonstrates use of setup() to pass
-- data to and from the threads

local counter = 1
local threads = {}

function setup(thread)
   thread:set("id", counter)
   table.insert(threads, thread)
   counter = counter + 1
end

function init(args)
   requests  = 0
   responses = 0

   local msg = "thread %d created"
   print(msg:format(id))
end

function request()
   requests = requests + 1
   return wrk.request()
end

function response(status, headers, body)
   responses = responses + 1
end

function done(summary, latency, requests)
   for index, thread in ipairs(threads) do
      local id        = thread:get("id")
      local requests  = thread:get("requests")
      local responses = thread:get("responses")
      local msg = "thread %d made %d requests and got %d responses"
      print(msg:format(id, requests, responses))
   end
end

[root@jerrik wrk]# wrk -t 4 -c 100 -d 20s --latency -s scripts/setup.lua https://www.baidu.com
thread 1 created
thread 2 created
thread 3 created
thread 4 created
Running 20s test @ https://www.baidu.com
  4 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   251.75ms  336.19ms   2.00s    86.89%
    Req/Sec   138.51     69.90   690.00     71.23%
  Latency Distribution
     50%  215.74ms
     75%  401.87ms
     90%  664.84ms
     99%    1.54s 
  11021 requests in 20.02s, 162.82MB read
  Socket errors: connect 0, read 3, write 0, timeout 50
Requests/sec:    550.62
Transfer/sec:      8.13MB
thread 1 made 2945 requests and got 2919 responses
thread 2 made 2831 requests and got 2807 responses
thread 3 made 2772 requests and got 2747 responses
thread 4 made 2573 requests and got 2548 responses
[root@jerrik wrk]# 

将每个线程的请求数和响应数输出来了。其它更多使用可以参考github script目录下的lua脚本。

总结

wrk作为http压测还是非常简便的,但是要想应对更多复杂场景,就需要多熟悉lua的使用,深入了解wrk提供的那几个函数。其它http压测工具,jmeter,apache ab,siege也可以了解一下。

附件

  function done(summary, latency, requests)

  The done() function receives a table containing result data, and two
  statistics objects representing the per-request latency and per-thread
  request rate. Duration and latency are microsecond values and rate is
  measured in requests per second.

  latency.min              -- minimum value seen
  latency.max              -- maximum value seen
  latency.mean             -- average value seen
  latency.stdev            -- standard deviation
  latency:percentile(99.0) -- 99th percentile value
  latency(i)               -- raw value and count

  summary = {
    duration = N,  -- run duration in microseconds
    requests = N,  -- total completed requests
    bytes    = N,  -- total bytes received
    errors   = {
      connect = N, -- total socket connection errors
      read    = N, -- total socket read errors
      write   = N, -- total socket write errors
      status  = N, -- total HTTP status codes > 399
      timeout = N  -- total request timeouts
    }
  }
-----------------------
print("latency.min:%s",latency.min)
print("latency.max:%s",latency.max)
print("latency.mean:%s",latency.mean)
print("latency.stdev:%s",latency.stdev)
print("latency:percentile(99.0):%s",latency:percentile(99.0))
print("requests.min:%s",requests.min)
print("requests.max:%s",requests.max)
print("requests.mean:%s",requests.mean)
print("requests.stdev:%s",requests.stdev)
print("requests:percentile(99.0):%s",requests:percentile(99.0))

print("summary['duration']:%s",summary["duration"])
print("summary['requests']:%s",summary["requests"])
print("summary['bytes']:%s",summary["bytes"])
print("summary['errors'].connect:%s",summary['errors'].connect])
print("summary['errors'].read:%s",summary['errors'].read])
print("summary['errors'].write:%s",summary['errors'].write])
print("summary['errors'].status:%s",summary['errors'].status])
print("summary['errors'].timeout:%s",summary['errors'].timeout])