Haywire

Haywire is an asynchronous HTTP server framework written in C. The goal of Haywire is to learn how to create a server with a minimal feature set that can handle a high rate of requests and connections with as low of latency and resource usage as possible.

Haywire uses the event loop based libuv platform layer that node.js is built on top of (also written in C). libuv abstracts IOCP on Windows and epoll/kqueue/event ports/etc. on Unix systems to provide efficient asynchronous I/O on all supported platforms.

Haywire isn’t very useful yet but I wanted to open source it from the very beginning. I started experimenting with a HTTP server for Dazzle queries and I decided I would split it into its own project.

Features

• Cross platform (Windows, Linux, OSX)
• HTTP keep-alive
• HTTP pipelining
• Non-blocking I/O
• SO_REUSEPORT multi-process and multi-threaded load balancing across CPU cores.

Plans or Ideas

• HTTP handler routing
• SPDY support
• Export an API that is easily consumable by other languages like on the JVM (Java, etc) and the CLR (C#, etc)

Contributions

Feel free to join in if you feel like helping progress Haywire. I’m open to new ideas and would love to work with some people instead of by myself!

Dependencies

apt-get install git gcc make cmake automake autoconf libssl-dev libtool

Compiling on Linux

./make.sh

Compiling on Mac OSX

brew install automake
brew install libtool
brew install openssl
./make.sh

Compiling on Windows

These instructions are old but may still work. We need to make Windows support in make.sh.

Open the Developer Command Prompt for Visual Studio

build.bat

Benchmarks

Bare metal Rackspace instance
Intel® Xeon® CPU E5-2680 v2 @ 2.80GHz 20 physical cores

Throughput

../bin/wrk/wrk --script ./pipelined_get.lua --latency -d 5m -t 40 -c 760 http://server:8000 -- 32

Running 5m test @ http://server:8000
  40 threads and 760 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     2.04ms    5.38ms 389.00ms   92.75%
    Req/Sec   233.38k    48.72k  458.99k    86.19%
  Latency Distribution
     50%    1.26ms
     75%    1.96ms
     90%    4.09ms
     99%    0.00us
  2781077938 requests in 5.00m, 409.23GB read
Requests/sec: 9,267,161.41
Transfer/sec:      1.36GB

Latency distribution with coordinated omission at 3.5 million requests/second

../bin/wrk2/wrk --script ./pipelined_get.lua --latency -d 10s -t 80 -c 512 -R 3500000 http://server:8000 -- 32

Running 10s test @ http://server:8000
  80 threads and 512 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency     1.82ms    3.00ms  64.86ms   94.17%
    Req/Sec       -nan      -nan   0.00      0.00%
  Latency Distribution (HdrHistogram - Recorded Latency)
 50.000%    1.18ms
 75.000%    1.62ms
 90.000%    2.34ms
 99.000%   16.75ms
 99.900%   33.12ms
 99.990%   44.70ms
 99.999%   52.99ms
100.000%   64.89ms

----system---- ----total-cpu-usage---- ------memory-usage----- --io/total- -dsk/total- -net/total- ---system-- -pkt/total- ----tcp-sockets----
     time     |usr sys idl wai hiq siq| used  buff  cach  free| read  writ| read  writ| recv  send| int   csw |#recv #send|lis act syn tim clo
03-12 18:58:10|  0   0  99   0   0   0| 712M 95.3M 1203M  124G|0.05  0.22 | 684B   10k|   0     0 |5066  1368 |   0     0 |  3   2   0   0   0
03-12 18:58:11|  0   0 100   0   0   0| 711M 95.3M 1203M  124G|   0     0 |   0     0 |5538B 1588B| 425   975 |92.0  2.00 |  3   2   0   0   0
03-12 18:58:12|  5   2  92   0   0   1| 725M 95.3M 1203M  124G|   0     0 |   0     0 |  85M  285M| 220k  112k| 127k  226k|  3 482   0   0   1
03-12 18:58:13| 10   4  85   0   0   1| 724M 95.3M 1203M  124G|   0  2.00 |   0    32k| 166M  556M| 427k  173k| 243k  438k|  3 482   0   0   1
03-12 18:58:14| 10   4  85   0   0   1| 725M 95.3M 1203M  124G|   0     0 |   0     0 | 165M  555M| 435k  172k| 243k  438k|  3 482   0   0   1
03-12 18:58:15| 10   4  85   0   0   1| 723M 95.3M 1203M  124G|   0     0 |   0     0 | 166M  555M| 440k  172k| 243k  438k|  3 482   0   0   1
03-12 18:58:16| 10   4  86   0   0   1| 724M 95.3M 1203M  124G|   0     0 |   0     0 | 166M  555M| 415k  172k| 243k  438k|  3 482   0   0   1
03-12 18:58:17| 10   4  85   0   0   1| 723M 95.3M 1203M  124G|   0     0 |   0     0 | 165M  555M| 404k  172k| 243k  438k|  3 482   0   0   1
03-12 18:58:18| 10   4  85   0   0   1| 724M 95.3M 1203M  124G|   0  5.00 |   0    24k| 165M  555M| 404k  171k| 243k  438k|  3 482   0   0   1
03-12 18:58:19| 10   4  85   0   0   1| 724M 95.3M 1203M  124G|   0     0 |   0     0 | 166M  555M| 411k  171k| 243k  438k|  3 482   0   0   1
03-12 18:58:20| 10   4  85   0   0   1| 723M 95.3M 1203M  124G|   0     0 |   0     0 | 166M  555M| 412k  170k| 244k  438k|  3 482   0   0   1
03-12 18:58:21| 10   4  85   0   0   1| 722M 95.3M 1203M  124G|   0     0 |   0     0 | 166M  555M| 411k  171k| 244k  438k|  3 482   0   0   1
03-12 18:58:22|  5   2  93   0   0   1| 724M 95.3M 1203M  124G|   0     0 |   0     0 |  76M  256M| 190k   81k| 113k  202k|  3   2   0   0 391

Users of Haywire

There’s a production system running Haywire that serves 700 million requests a day from an Aerospike cluster with 2ms or lower response times.

pyrs Python handlers for Haywire.