One of the fun exercises in The Complete Guide to Rails Performance is to profile the dalli gem with ruby-prof.

After cloning the repository and running ruby-prof test/benchmark_test.rb, we get the following summary:

 %self      total      self      wait     child     calls  name
 13.56      9.181     9.181     0.000     0.000   193917   <Class::IO>#select
  8.12      9.604     5.498     0.000     4.106   275009   Dalli::Ring#binary_search
  5.01      3.390     3.390     0.000     0.000   272536   Kgio::SocketMethods#kgio_write

That sparked my curiosity. Why is dalli spending so much time in binary_search and can it be improved?

As it turns out, dalli already takes care of it in a pretty interesting way. Here’s a simplified version of binary_search method definition (full source here):

  require 'inline'
  inline do |builder|
    builder.c <<-EOM
      int binary_search(VALUE ary, unsigned int r) {
        // C implementation
rescue LoadError
  def binary_search(ary, value)
    # Ruby implementation

We’re making use of rubyinline gem which lets you embed small snippets of C into your code. They get automatically compiled and loaded into the class/module that defines them. Definitely handy if all you need is a few lines of C to speed things up.

Why not go the more common native extension route, you might ask? This comment from the source provides an indication:

optional for performance and only necessary if you are using multiple memcached servers.

Clearly, the typical user won’t need the native extension. More often than not, they’re a source of frustration.

I like the flexibility of this simple-by-default design. It allows you to optimize only if your metrics tell you so. If you need more speed, all you need to do is include rubyinline in your bundle, without further configuration in your code. Here’s the kind of performance you’ll get:

 %self      total      self      wait     child     calls  name
 13.57      7.796     7.796     0.000     0.000   197184   <Class::IO>#select
  6.39      3.672     3.672     0.000     0.000   272536   Kgio::SocketMethods#kgio_write
  0.77      0.440     0.440     0.000     0.000   275009   Dalli::Ring#binary_search

Out of the box, fallback to Ruby implementation by means of LoadError has you covered.