Iodine - HTTP / Websocket Server & EventMachine alternative: C kqueue/epoll extension

Logo

Build Status Gem Version Inline docs GitHub

Iodine is a fast concurrent web server for real-time Ruby applications, with native support for Websockets, Pub/Sub, static file service, HTTP/1.1 and Redis Pub/Sub scaling (2 connections per Iodine process).

Iodine also supports custom protocol authoring, making Object Oriented Network Services easy to write.

Iodine is an evented framework with a simple API that builds off the low level C code library facil.io with support for epoll and kqueue - this means that:

  • Iodine can handle thousands of concurrent connections (tested with more then 20K connections).

    That's right, Iodine isn't subject to the 1024 connection limit imposed by native Ruby and select/poll based applications.

    This makes Iodine ideal for writing HTTP/2 and Websocket servers (which is what started this whole thing).

  • Iodine supports only Linux/Unix based systems (i.e. OS X, Ubuntu, FreeBSD etc'), which are ideal for evented IO (while Windows and Solaris are better at IO completion events, which are totally different).

Iodine is a C extension for Ruby, developed for Ruby MRI 2.2.2 and up... it should support the whole Ruby 2.0 MRI family, but Rack requires Ruby 2.2.2, and so Iodine matches this requirement.

Iodine::Rack == a fast and powerful HTTP + Websockets server with native Pub/Sub

Iodine includes a light and fast HTTP and Websocket server written in C that was written according to the Rack interface specifications and the Websocket draft extension.

With Iodine.listen2http it's possible to run multiple HTTP applications in addition to (or instead of) the default Iodine::Rack HTTP service.

Iodine also supports native process cluster Pub/Sub and a native RedisEngins to easily scale Iodine's Pub/Sub horizontally.

Running the web server

Using the Iodine server is easy, simply add Iodine as a gem to your Rack application:

gem 'iodine', '~>0.4'

Iodine will calculate, when possible, a good enough default concurrency model for lightweight applications... this might not fit your application if you use heavier database access or other blocking calls.

To get the most out of Iodine, consider the amount of CPU cores available and the concurrency level the application requires.

The common model of 16 threads and 4 processes can be easily adopted:

bundler exec iodine -p $PORT -t 16 -w 4

Static file serving support

Iodine supports an internal static file service that bypasses the Ruby layer and serves static files directly from "C-land".

This means that Iodine won't lock Ruby's GVL when sending static files. The files will be sent directly, allowing for true native concurrency.

Since the Ruby layer is unaware of these requests, logging can be performed by turning iodine's logger on.

To use native static file service, setup the public folder's address before starting the server.

This can be done when starting the server from the command line:

bundler exec iodine -p $PORT -t 16 -w 4 -www /my/public/folder

Or by adding a single line to the application. i.e. (a config.ru example):

require 'iodine'
Iodine::Rack.public = '/my/public/folder'
out = [404, {"Content-Length" => "10".freeze}.freeze, ["Not Found.".freeze].freeze].freeze
app = Proc.new { out }
run app

To enable logging from the command line, use the -v (verbose) option:

bundler exec iodine -p $PORT -t 16 -w 4 -www /my/public/folder -v

X-Sendfile

Ruby can leverage static file support (if enabled) by using the X-Sendfile header in the Ruby application response.

This allows Ruby to send very large files using a very small memory footprint, as well as (when possible) leveraging the sendfile system call.

i.e. (example config.ru for Iodine):

app = proc do |env|
  request = Rack::Request.new(env)
  if request.path_info == '/source'.freeze
    [200, { 'X-Sendfile' => File.expand_path(__FILE__) }, []]
  elsif request.path_info == '/file'.freeze
    [200, { 'X-Header' => 'This was a Rack::Sendfile response sent as text.' }, File.open(__FILE__)]
  else
    [200, { 'Content-Type'.freeze => 'text/html'.freeze,
            'Content-Length'.freeze => request.path_info.length.to_s },
     [request.path_info]]
 end
end
# # optional:
# use Rack::Sendfile
run app

Go to localhost:3000/source to download the config.ru file using the X-Sendfile extension.

Special HTTP Upgrade support

Iodine's HTTP server includes special support for the Upgrade directive using Rack's env Hash, allowing the application to focus on services and data while Iodine takes care of the network layer.

Upgrading an HTTP connection can be performed either using Iodine's Websocket Protocol support with env['upgrade.websocket'] or by implementing your own protocol directly over the TCP/IP layer - be it a websocket flavor or something completely different - using env['upgrade.tcp'].

Websockets

When an HTTP Upgrade request is received, Iodine will set the Rack Hash's upgrade property to true, so that: env[upgrade.websocket?] == true

To "upgrade" the HTTP request to the Websockets protocol, simply provide Iodine with a Websocket Callback Object instance or class: env['upgrade.websocket'] = MyWebsocketClass or env['upgrade.websocket'] = MyWebsocketClass.new(args)

Iodine will adopt the object, providing it with network functionality (methods such as write, each, defer and close will become available) and invoke it's callbacks on network events.

Here is a simple chatroom example we can run in the terminal (irb) or easily paste into a config.ru file:

require 'iodine'
class WebsocketEcho
  def on_open
    # Pub/Sub directly to the client (or use a block to process the messages)
    subscribe channel: :chat
    # Writing directly to the socket
    write "You're now in the chatroom."
  end
  def on_message data
    # Strings and symbol channel names are equivalent.
    publish channel: "chat", message: data
  end
end
Iodine::Rack.app= Proc.new do |env|
  if env['upgrade.websocket?'.freeze] && env["HTTP_UPGRADE".freeze] =~ /websocket/i.freeze
    env['upgrade.websocket'.freeze] = WebsocketEcho # or: WebsocketEcho.new
    [0,{}, []] # It's possible to set cookies for the response.
  else
    [200, {"Content-Length" => "12"}, ["Welcome Home"] ]
  end
end
# Pus/Sub can be server oriented as well as connection bound
root_pid = Process.pid
Iodine.subscribe(channel: :chat) {|ch, msg| puts msg if Process.pid == root_pid }
# By default, Pub/Sub performs in process cluster mode.
Iodine.processes = 4
# static file serving can be set manually as well as using the command line:
Iodine::Rack.public = "www/public"
#
Iodine.start

TCP/IP (raw) sockets

Upgrading to a custom protocol (i.e., in order to implement your own Websocket protocol with special extensions) is performed almost the same way, using env['upgrade.tcp']. In the following (terminal) example, we'll use an echo server without direct socket echo:

require 'iodine'
class MyProtocol
  def on_message data
    # regular socket echo - NOT websockets - notice the upgrade code
    write data
  end
end
Iodine::Rack.app = Proc.new do |env|
  if env['upgrade.tcp?'.freeze] && env["HTTP_UPGRADE".freeze] =~ /echo/i.freeze
    env['upgrade.tcp'.freeze] = MyProtocol
    # no HTTP response will be sent when the status code is 0 (or less).
    # to upgrade AFTER a response, set a valid response status code.
    [1000,{}, []]
  else
    [200, {"Content-Length" => "12"}, ["Welcome Home"] ]
  end
end
Iodine.start

A few notes

This design has a number of benefits, some of them related to better IO handling, resource optimization (no need for two IO polling systems), etc. This also allows us to use middleware without interfering with connection upgrades and provides backwards compatibility.

Iodine::Rack imposes a few restrictions for performance and security reasons, such as that the headers (both sending and receiving) must be less than 8Kb in size. These restrictions shouldn't be an issue and are similar to limitations imposed by Apache.

Of course, if you still want to use Rack's hijack API, Iodine will support you - but be aware that you will need to implement your own reactor and thread pool for any sockets you hijack, as well as a socket buffer for non-blocking write operations (why do that when you can write a protocol object and have the main reactor manage the socket?).

Performance oriented design - but safety first

Iodine is an evened server, similar in it's architecture to nginx and puma. It's different than the simple "thread-per-client" design that is often taught when we begin to learn about network programming.

By leveraging epoll (on Linux) and kqueue (on BSD), iodine can listen to multiple network events on multiple sockets using a single thread.

All these events go into a task queue, together with the application events and any user generated tasks, such as ones scheduled by Iodine.run.

In pseudo-code, this might look like this

QUEUE = Queue.new

def server_cycle
    QUEUE << get_next_32_socket_events # these events schedule the proper user code to run
    QUEUE << [server]
end

def run_server
      while ((event = QUEUE.pop))
            event.shift.call(*event)
      end
end

In pure Ruby (without using C extensions or Java), it's possible to do the same by using select... and although select has some issues, it works well for smaller concurrency levels.

The server events are fairly fast and fragmented (longer code is fragmented across multiple events), so one thread is enough to run the server including it's static file service and everything... but single threaded mode should probably be avoided.

The thread pool is there to help slow user code. It's very common that the application's code will run slower and require external resources (i.e., databases, a pub/sub service, etc'). This slow code could "starve" the server (that is patiently waiting to run it's tasks on the same thread) - which is why a thread pool is often necessary.

The slower your application code, the more threads you will need to keep the server running smoothly.

How does it compare to other servers?

Personally, after looking around, the only comparable servers are Puma and Passenger, which Iodine significantly outperformed on my tests (I didn't test Passenger's enterprise version).

Since the HTTP and Websocket parsers are written in C (with no RegExp), they're fairly fast.

Also, Iodine's core and parsers are running outside of Ruby's global lock, meaning that they enjoy true concurrency before entering the Ruby layer (your application) - this offers Iodine a big advantage over other Ruby servers.

Another assumption Iodine makes is that it is behind a load balancer / proxy (which is the normal way Ruby applications are deployed) - this allows Iodine to disregard header validity checks (we're not checking for invalid characters) and focus it's resources on other security and performance concerns.

I recommend benchmarking the performance for yourself using wrk or ab:

$ wrk -c200 -d4 -t12 http://localhost:3000/
# or
$ ab -n 100000 -c 200 -k http://127.0.0.1:3000/

Create a simple config.ru file with a hello world app:

App = Proc.new do |env|
   [200,
     {   "Content-Type" => "text/html".freeze,
         "Content-Length" => "16".freeze },
     ['Hello from Rack!'.freeze]  ]
end

run App

Then start comparing servers. Here are the settings I used to compare Iodine and Puma (4 processes, 16 threads):

$ RACK_ENV=production iodine -p 3000 -t 16 -w 4
# vs.
$ RACK_ENV=production puma -p 3000 -t 16 -w 4
# Review the `iodine -?` help for more command line options.

When benchmarking with wrk, Iodine performed significantly better, (~62K req/sec vs. ~44K req/sec) while keeping a lower memory foot print (~60Mb vs. ~111Mb).

When benchmarking with ab, I got different results, where Iodine still performed significantly better, (~72K req/sec vs. ~36K req/sec and ~61Mb vs. ~81.6Mb). I suspect the difference between the two benchmarks has to do with system calls to write, but I have no real proof.

Remember to compare the memory footprint after running some requests - it's not just speed that C is helping with, it's also memory management and object pooling (i.e., Iodine uses a buffer packet pool management).

Can I try before I buy?

Well, it is free and open source, no need to buy.. and of course you can try it out.

It's installable just like any other gem on MRI, run:

$ gem install iodine

If building the native C extension fails, please note that some Ruby installations, such as on Ubuntu, require that you separately install the development headers (ruby.h and friends). I have no idea why they do that, as you will need the development headers for any native gems you want to install - so hurry up and get them.

If you have the development headers but still can't compile the Iodine extension, open an issue with any messages you're getting and I'll be happy to look into it.

Mr. Sandman, write me a server

Girls love flowers, or so my ex used to keep telling me... but I think code is the way to really show that something is hot!

I mean, look at this short and sweet echo server - No HTTP, just use telnet... but it's so elegant I could cry:


require 'iodine'

# an echo protocol with asynchronous notifications.
class EchoProtocol
  # `on_message` is an optional alternative to the `on_data` callback.
  # `on_message` has a 1Kb buffer that recycles itself for memory optimization.
  def on_message buffer
    # writing will never block and will use a buffer written in C when needed.
    write buffer
    # close will be performed only once all the data in the write buffer
    # was sent. use `force_close` to close early.
    close if buffer =~ /^bye[\r\n]/i
    # use buffer.dup to save the data from being recycled once we return.
    data = buffer.dup
    # run asynchronous tasks with ease
    run do
      sleep 1
      puts "Echoed data: #{data}"
    end
  end
end

# listen on port 3000 for the echo protocol.
Iodine.listen 3000, EchoProtocol
Iodine.threads = 1
Iodine.processes = 1
Iodine.start

Why not EventMachine?

You can go ahead and use EventMachine if you like. They're doing amazing work on that one and it's been used a lot in Ruby-land... really, tons of good developers and people on that project, I'm sure...

But me, I prefer to make sure my development software runs the exact same code as my production software. So here we are.

Also, I don't really understand all the minute details of EventMachine's API, it kept crashing my system every time I reached 1K-2K active connections... I'm sure I just don't know how to use EventMachine, but that's just that.

Besides, you're here - why not take Iodine out for a spin and see for yourself?

Can I contribute?

Yes, please, here are some thoughts:

  • I'm really not good at writing automated tests and benchmarks, any help would be appreciated. I keep testing manually and that's less then ideal (and it's mistake prone).

  • If we can write a Java wrapper for the C libraries, it would be nice... but it could be as big a project as the whole gem, as a lot of minor details are implemented within the bridge between these two languages.

  • Bug reports and pull requests are welcome on GitHub at https://github.com/boazsegev/iodine.

  • If you love the project or thought the code was nice, maybe helped you in your own project, drop me a line. I'd love to know.

License

The gem is available as open source under the terms of the MIT License.

"I'm also writing a Ruby extension in C"

Really?! That's great!

We could all use some more documentation around the subject and having an eco-system for extension tidbits would be nice.

Here's a few things you can use from this project and they seem to be handy to have (and easy to port):

  • Iodine is using a Registry to keep dynamic Ruby objects that are owned by C-land from being collected by the garbage collector in Ruby-land...

    Some people use global Ruby arrays, adding and removing Ruby objects to the array, but that sounds like a performance hog to me.

    This one is a simple binary tree with a Ruby GC callback. Remember to initialize the Registry (Registry.init(owner)) so it's "owned" by some Ruby-land object, this allows it to bridge the two worlds for the GC's mark and sweep.

    I'm attaching it to one of Iodine's library classes, just in-case someone adopts my code and decides the registry should be owned by the global Object class.

  • I was using a POSIX thread pool library (libasync.h) until I realized how many issues Ruby has with non-Ruby threads... So now there's a Ruby-thread port for this library at (rb-libasync.h).

    Notice that all the new threads are free from the GVL - this allows true concurrency... but, you can't make Ruby API calls in that state.

    To perform Ruby API calls you need to re-enter the global lock (GVL), albeit temporarily, using rb_thread_call_with_gvl and rv_protect (gotta watch out from Ruby longjmp exceptions).

  • Since I needed to call Ruby methods while multi-threading and running outside the GVL, I wrote RubyCaller which let's me call an object's method and wraps all the rb_thread_call_with_gvl and rb_protect details in a secret hidden place I never have to see again. It also keeps track of the thread's state, so if we're already within the GVL, we won't enter it "twice" (which will crash Ruby sporadically).

These are nice code snippets that can be easily used in other extensions. They're easy enough to write, I guess, but I already did the legwork, so enjoy.