Ten times slower than gzip?

[watson]

I performed a few benchmarks to see how much faster lz4 would be compared to gzip when decompressing an incoming HTTP payload. I'm not sure if I'm doing something wrong, but it seems that this library is a lot slower than just using plain gzip.

I had 3 test JSON docs of different sizes, that I compressed with both LZ4 and gzip:

Uncompressed	LZ4	gzip
7,369 bytes	3,975 bytes	2,772 bytes
73,723 bytes	33,028 bytes	21,790 bytes
716,995 bytes	311,365 bytes	202,697 bytes

The LZ4 version was compressed using default options:

uncompressed.pipe(lz4.createEncoderStream()).pipe(compressed)

The gzip version was compressed from my macOS command line:

gzip uncompressed.json

I used autocannon to hammer a test HTTP server with the compressed document. The server would decompress the payload but otherwise discard it afterwards.

Here's an example of how autocannon was configured:

autocannon -i body.json.lz4 -H 'Content-Encoding: lz4' localhost:3000

And here's my test server running on localhost:

'use strict'

const http = require('http')
const zlib = require('zlib')
const lz4 = require('lz4')

const server = http.createServer(function (req, res) {
  const enc = req.headers['content-encoding'] || ''

  let decompressed
  if (/\bgzip\b/.test(enc)) {
    decompressed = req.pipe(zlib.createGunzip())
  } else if (/\blz4\b/.test(enc)) {
    decompressed = req.pipe(lz4.createDecoderStream())
  } else {
    decompressed = req
  }

  const buffers = []
  decompressed.on('data', buffers.push.bind(buffers))
  decompressed.on('end', function () {
    const data = Buffer.concat(buffers)
    res.end()
  })
})

server.listen(3000, function () {
  console.log('Server listening on http://localhost:3000')
})

Test 1 - Decompressing a 7,369 byte JSON document

LZ4 (3,975 bytes):

Stat         Avg     Stdev   Max
Latency (ms) 23.29   9.31    61.38
Req/Sec      419.7   13.36   434
Bytes/Sec    41.4 kB 1.31 kB 43 kB

4k requests in 10s, 416 kB read

Gzip (2,772 bytes):

Stat         Avg    Stdev   Max
Latency (ms) 1.07   0.67    13.48
Req/Sec      7064.4 704.93  7733
Bytes/Sec    699 kB 67.8 kB 766 kB

71k requests in 10s, 6.99 MB read

Test 2 - Decompressing a 73,723 byte JSON document

LZ4 (33,028 bytes):

Stat         Avg     Stdev  Max
Latency (ms) 23.28   8.94   55.9
Req/Sec      419.8   11.45  435
Bytes/Sec    41.8 kB 1.1 kB 43.1 kB

4k requests in 10s, 416 kB read

Gzip (21,790 bytes):

Stat         Avg    Stdev   Max
Latency (ms) 2.7    1.61    21.23
Req/Sec      3131   105.16  3342
Bytes/Sec    313 kB 13.1 kB 331 kB

31k requests in 10s, 3.1 MB read

Test 3 - Decompressing a 716,995 byte JSON document

On a large document like this the difference between gzip and lz4 is much smaller, but gzip still wins:

LZ4 (311,365 bytes):

Stat         Avg     Stdev Max
Latency (ms) 41.56   13.21 102
Req/Sec      237.6   6.95  250
Bytes/Sec    23.7 kB 819 B 24.8 kB

2k requests in 10s, 235 kB read

Gzip (202,697 bytes):

Stat         Avg     Stdev Max
Latency (ms) 26.11   6.51  137.09
Req/Sec      375.4   7.61  381
Bytes/Sec    37.5 kB 819 B 37.7 kB

4k requests in 10s, 372 kB read

[simonh1000]

I'm seeing a big difference between node and the browser for a 17mb text. 16s in browser vs 6.5 in NodeJs 10

[pierrec]

The browser implementation uses pure JS to decode/encode, whereas the node one binds to the C implementation.