Moonbase

Since Thread.critical will be going away with Ruby 1.9 (not to mention the pain it’s been for other Ruby implementors), I decided to remove the main reason most programs use it: the fact that stdlib’s Mutex is slow.

Most of Mutex’s performance problems come from two sources:

1. It involves a lot of Ruby method calls
2. It uses a dynamically sized array to maintain its wait list

To that end, I:

1. Rewrote Mutex in pure C (big savings)
2. Switched to using a double-headed linked list (O(1) get and put)
3. Used a per-mutex memory pool for list entries (so malloc overhead doesn’t kill us)

The end result is actually faster than using Thread.critical directly! Even Mutex#synchronize, with its block call, is faster than the typical:

saved = Thread.critical
Thread.critical = true
begin
  # ...
ensure
  Thread.critical = saved
end

Here’s a git repository where you can check out my work:

http://moonbase.rydia.net/software/optimized-locking/optimized-locking.git

I’ve also put up a snapshot tarball for the git-disinclined.

I’ll be working on condition variables next. If they also turn out fast, then purging most explicit uses of Thread.critical from stdlib may become a real possibility.