In the last post I talked about some differences between fibers and continuations. What may not have been clear is that continuations are more primitive and flexible than fibers are. In fact, you can implement fibers using continuations.

Here’s how. The basic idea is that we want to maintain two variables with continuations in them, inside and outside. The first one will transfer execution into the block of code that forms the fiber. The second will transfer control back to the outside world.

When the outside world calls #resume, we save our continuation point as outside, and call the current inside continuation. When, within the block, #yield is called, we save our current continuation point as inside, and transfer code back to the current outside.

There are a few more details in terms of passing values from #yield to #resume, handling the return value of the block, and handling excessive calls to #resume, but that’s the basic story. Here’s the code:

require 'continuation'

class CFiber
  class Error < StandardError; end

  def initialize &block
    @block = block
    callcc do |cc|
      @inside = cc
      return
    end
    @var = @block.call self
    @inside = nil
    @outside.call
  end

  def resume
    raise Error, "dead cfiber called!" unless @inside
    callcc do |cc|
      @outside = cc
      @inside.call
    end
    @var
  end

  def yield var
    callcc do |cc|
      @var = var
      @inside = cc
      @outside.call
    end
  end
end

This is also runnable on Ruby 1.8—just remove the require.

So why does Ruby 1.9 bother to implement fibers, when we can just use continuations? I don’t know what the real answer is, but “speed” is at least a good answer. Let’s do some some benchmarking to compare the two:

require 'benchmark'
n = ARGV.shift.to_i
Benchmark.bm do |bm|
  bm.report " fibers" do
    f = Fiber.new do
      x, y = 0, 1
      loop do
        Fiber.yield y
        x, y = y, x + y
      end
    end

    n.times { |i| f.resume }
  end

  bm.report "cfibers" do
    f = CFiber.new do |c|
      x, y = 0, 1
      loop do
        c.yield y
        x, y = y, x + y
      end
    end

    n.times { |i| f.resume }
  end
end

We’ll start with backporting that code to the Ruby 1.8.7 that Ubuntu provides (ruby 1.8.7 (2008-08-11 patchlevel 72)). For 10000 Fibonacci numbers, we see:

That’s roughly 11.4kfps (that’s thousand Fibonacci numbers per second) that we can produce using continuation-based fibers.

Let’s try the ancient Ruby 1.9.0 that Ubuntu provides (Ruby 1.9.0 (2008-06-20 revision 17482)):

Wow, fibers are fast: 250kfps. But things have gotten significantly worse for cfibers, clocking at a measely 0.489kfps for cfibers.

Finally let’s try the latest and greatest Ruby 1.9.1 (ruby 1.9.1p129 (2009-05-12 revision 23412)):

Fibers are just as fast as before, but continuations have improved dramatically—from 11.4kfps to 66.6kfps. Still, native fibers are more than three times faster.

So perhaps Ruby 1.9.1 is the best of both worlds. When you need fast non-preemptive concurrency, you can use native fibers; when you need to implement your own crazy control structures, you can use continuations and be assured that they’re still pretty darn fast (at least, as far as Ruby operations are concerned).

Ruby 1.9 has both fibers and continuations. The two are often mentioned in the same breath. They do vaguely similar-sounding things, and are implemented in Ruby 1.9 with similar mechanics underneath the hood, much as how continuations and threads were implemented with the same underlying mechanics in Ruby 1.8 [PDF, p. 14].

But implementation similarities aside, continuations and fibers have very different semantics. A fiber behaves as a thread without preemption. Like a thread, you create it, and it eventually dies; unlike a thread, you must manually call yield and resume to transfer control in and out of it, instead of just letting the runtime call them for you whenever it feels like it. Like a thread, when you resume a fiber, you have the same call stack and heap state (local variables) as when you left.

What’s nice about fibers is that, since you keep explicit control of the order of execution, you can get thread-like behavior without all the hassle of mutexes and synchronization. Of course you have to deal with the hassle of ordering all your operations, but you at least have the option of avoiding the fun race-condition game that always seems to crop up in threaded programming.

What about continuations? Instead of fibers’ create, kill, yield, and resume operations, a continuation only really has two operations: capture and resume. A continuation is captured once, and may be resumed multiple times. When you resume a continuation, the call stack is reverted to what it looked like when it was captured, but the heap state stays the same. There’s no exit point or death for a continuation (at least until Ruby gets bounded continuations); execution simply continues from the capture point.

What’s nice about continuations is that you can use them to implement control structures. Loops, exceptions, cross-procedure gotos… almost every control structure you can come up with can be implemented with continuations. In fact, you can implement fibers using continuations!

Let’s look at an example. Here’s the fiber-based Fibonacci computation from the InfoQ article on Fibers in Ruby 1.9:

fib = Fiber.new do  
  x, y = 0, 1 
  loop do  
    Fiber.yield y 
    x, y = y, x + y 
  end 
end 
20.times { puts fib.resume }

Here we call yield from within the fiber once we’ve computed a number, which transfers control to the main function, and which prints out the number yielded and then calls resume to transfer control back to the fiber. A thread version looks very similar:

require 'thread'
q = SizedQueue.new 1
fib = Thread.new do  
  x, y = 0, 1 
  loop do  
    q.push y 
    x, y = y, x + y 
  end 
end 

20.times { puts q.pop }

Since we don’t have explicit control over the scheduling, we implicitly scheduled the order of operations by using a synchronized SizedQueue data structure, which blocks the computation thread from computing a new number until the printing thread is ready to receive it. (There are many ways we could’ve accomplished this.)

Here’s the version using continuations:

require 'continuation'
c, x, y, i = callcc { |cc| [cc, 0, 1, 1] }
puts y
c.call c, y, x + y, i + 1 if i < 20

You’ll notice there are no loops, and variables are never changed after assignment. In fact the code is starting to look suspiciously like an inductive proof, with one line that like a base case and another line that looks like a recursive case. You can see why continuations make functional-programming enthusiasts get excited!

This implementation works because resuming the continuation (the call to c.call) replaces the call stack and point of execution with what they were at the point it was captured (the call to callcc). In contrast, resume-ing the fiber moved us back to the point we were when the fiber called yield, and so the outer loop in the fiber implementation was necessary.

Beyond call stacks, another major difference between fibers and continuations is the way the heap is treated. Multiple fibers on the same section of code do not share local variables. Multiple continuations on the same section of code do. Here’s a brief example. First, the fibers version:

fib = (0 ... 5).map do |i|
  Fiber.new do
    x = 0
    Fiber.yield x
    x += 1
  end
end

fib.each { |f| puts f.resume }

We create five fibers, and call resume on them once each. As you’d expect, this prints out a series of 0’s. The variable x is not shared between the multiple fibers. Of course, the fiber constructor here is a block, and blocks are closures, so we could make them share state by moving the x = 0 line outside the map line. But that’s a result of having closures, not of fibers per se.

Let’s try an example with multiple continuations, all jumping into the same point in the code:

require 'continuation'

x = 0
c = callcc { |cc| cc }
d = callcc { |cc| cc } if c
e = callcc { |cc| cc } if c && d
f = callcc { |cc| cc } if c && d && e
x += 1

puts x
c.call if c
d.call if d
e.call if e
f.call if f

We initialize x to 0, create 4 separate continuations, add one to x, and call the continuations in order. (The postfix if statements ensure that the continuations variables aren’t set or called more than once. Calling c.call without arguments will jump back to the c = callcc line and set c to nil.)

Silly, but it illustrates the point: the output is “1 2 3 4 5”, meaning that the four continuations all share the same heap. When d is called, its x is the same as the x of c, and even though it was 0 when d was captured, it has since been modified by the resumption of c. When e is called, its x is also the same x, and so on. (In fact this whole example depends on this behavior—each of the continuation variables are only set once, and must “retain” their value across all rentries to continuations above them.)

In additon to multiple continuations being able to share state, the converse is true too: multiple resumes on the same continuation will share state:

require 'continuation'

x = 0
c = callcc { |cc| cc }
x += 1
puts x
c.call c while x < 5

This outputs the same thing as the examples above.

Hopefully that clears up some of the confusion. Here’s the summary: