turbo-http

This is a fork that can be installed via @kore/http. It brings the SO_REUSEPORT cluster hack for Linux, resulting in up to 2x more r/s numbers in benchmarks (can't be tested properly with autocannon in some cases, might need to use wrk).

A low level http library for Node.js based on turbo-net

npm i turbo-http

WIP, this module is already really fast but there are some HTTP features missing and easy performance gains to be had. :D :D :D

On my laptop I can serve simple hello world payloads at around 100k requests/seconds compared to 10k requests/second using node core.

Usage

const turbo = require('turbo-http')

const server = turbo.createServer(function (req, res) {
  res.setHeader('Content-Length', '11')
  res.write(Buffer.from('hello world'))
})

server.listen(8080)

API

server = turbo.createServer([onrequest])

Create a new http server. Inherits from the turbo-net tcp server

server.on('request', req, res)

Emitted when a new http request is received.

res.statusCode = code

Set the http status

res.setHeader(name, value)

Set a http header

res.write(buf, [length], [callback])

Write a buffer. When the callback is called, the buffer has been completely flushed to the underlying socket and is safe to reuse for other purposes

res.writev(buffers, [lengths], [callback])

Write more that one buffer at once.

res.end([buf], [length], [callback])

End the request. Only needed if you do not provide a Content-Length.

req.url

Request url

req.method

Request method

req.socket

Request turbo-net socket

value = req.getHeader(name)

Get a request header.

headers = req.getAllHeaders()

Get all request headers as a map.

req.ondata(buffer, start, length)

Called when there is data read. If you use the buffer outside of this function you should copy it.

req.onend()

Called when the request is fully read.

Benchmarks

Comparing turbo-http to other frameworks is like comparing oranges to apples. turbo-http could be thought of as a replacement of Node's native http module, while all available frameworks actually use it.

Benchmark it:

• clone this repo,
• npm i
• npm run bench

Benchmark averages are taken after one warm-up round.

  | Requests/s | Latency | Throughput/Mb ------------- | ---------- | ------- | -------------- turbo-http.js | 32592 | 3.03 | 2.43 bare-node.js | 18396 | 5.32 | 1.98 rayo.js | 16249.6 | 6.03 | 1.77 polka.js | 15802.4 | 6.2 | 1.71 fastify.js | 15141.6 | 6.47 | 2.26 express.js | 13408.8 | 7.31 | 1.46 hapi.js | 9675.6 | 10.15 | 1.42

Note: Nevermind these numbers, this benchmark was run on a slow computer and the above table is for illustrative purposes only.

Optionally, you may also define your test's parameters:

$> npm run bench -- -u http://localhost:5050 -c 100 -p 10 -d 5

• -u (url) -Defaults to http://localhost:5050
• -c (connections) -Defaults to 100
• -p (pipelines) -Defaults to 10
• -d (duration) -Defaults to 5 (seconds)

Acknowledgements

This project was kindly sponsored by nearForm.

License

MIT