A million connections and beyond - Node.js at scale

A million connections...
AND BEYOND!!!
Node.js at Scale

A million isn't
very much these
days...

One hundred
billion
connections!!!

What this talk is really
about
Picking meaningful benchmarks

Current state of Node.js benchmarks

Understanding why benchmarks aren't _that_
important

How to pick the right language/framework/lib

What can we measure?

Computation Disk I/O

Net work I/O Throughput

TCP Libraries

UDP URL Parsing

HTTP etc

How do we measure?

Load averages

Memory usage

Average time - Average perf. of each operation

Latency - Actual performance of each operation

Comparison

Erlang vs. Node.js vs. Tornado

3 Erlang libs

Stock Node.js

Stock Tornado

Single core

I pity the fool who
uses Node

Questioning benchmarks

What do average times mean?

Responses on a curve

Who uses in memory servers?

Finding Node's strength
connections: 933990

errors: 0

Read from remote host gqzdj9ab.joyent.us:
Operation timed out

Connection to gqzdj9ab.joyent.us closed.

Enki:~ $

var net = require('net')

var conns = 0

function connectToServer(ip) {
conn = net.createConnection(8000, ip)
conn.on('connect', function() {
conns += 1
})
conn.on('end', function() {
conns -= 1
})
}

for(i=0;i<61000;i++) {
connectToServer('72.2.120.106')
}

var net = require('net')

var s = net.createServer()
var c = 0;
var e = 0;

s.on('connection', function(socket) {
c += 1

socket.on('end', function() {
c -= 1
})
socket.on('error', function(e) {
e += 1
})
})

s.listen(8000)

setInterval(function() {
console.log('connections: ' + c)
console.log('errors: ' + e)
}, 5000)

Finding Node's strength
PID USERNAME SIZE RSS STATE PRI NICE TIME CPU PROCESS/NLWP

22593 root 758M 642M cpu11 25 0 0:17:54 6.2% node/1

22542 croucher 4308K 3084K cpu2 59 0 0:00:08 0.2% prstat/1

...

...

ZONEID NPROC SWAP RSS MEMORY TIME CPU ZONE

27 31 1489M 782M 38% 1:59:54 6.3% gqzdj9ab

Where to use atomic
measurements

Measuring performance improvements over time

Figuring out the hot-path

Proﬁling against your speciﬁc use case

I am a super

stuff
effective
measurement
Because I put stuff on one axis and
things on another and now I'm like things
proper science and you should totally
use me to base all your judgements
about everything on…

no seriously...

Why benchmarks aren't
as important as we think

There is a tradeoff bet ween performance and
productivity

If this weren't true we'd all still be writing in ASM

Fit for purpose
Sandboxed language Event Driven
PHP, Python Node.js

Java, C# Event Machine (Ruby)
Ruby, etc Tornado (Python)

Message Passing
Erlang

Class of languages for the
Internet
New languages/platforms like Node.js designed for
Internet applications

New class of languages are efﬁcient enough, but are
they easy enough?

Node.js designed from scratch for net work server
programming

Reasons to use Node

It's really fast. Very much fast enough.

It's easy to use for Internet/Web applications.

It's JavaScript.

It has an extremely active community.

It has a lot of project velocity.

What we need from
benchmarks
Benchmarks to test real world use cases
Servers that do more than just 'hello world'
Benchmarks to compare real test cases from many
platforms
Benchmarks that reﬂect mobile clients
Node CI on many platforms

Questions?

I'm on Twitter http:/ witter.com/sh1mmer
/t

That's a "one" in my nick

Check out the book on http://ofps.oreilly.com

It's free to read and comment. New release this
week

A million connections and beyond - Node.js at scale

In this document