C[omp]ute

From: andrew cooke <andrew@...>

Date: Sun, 10 Jun 2012 14:28:20 -0400

I just wrote my first "worthwhile" macro in Clojure (by which I mean, does
something non-trivial I expected the language to do for me).

I have some code I want to run multiple times.  The only things that changes
each time is an integer, but each run takes 10 hours.  So I had a loop that
took 10 hours per iteration.  If you know Clojure then you'll know that my
code looked like this:

  (doseq [i (range 4)]
    (do-stuff-with i))

which is basically a "for loop" that changes "i".

To save time I wanted to run those things in parallel.  So I wanted something
like Python's multiprocessing - a simple way to make "do-stuff-with i" run on
multiple cores.

Solving this went through three phases:

 - Looking for an obvious existing solution.  Couldn't find one.

 - Trying to write an implementation using basic concepts from Java like
   semaphores or ThreadPoolExecutor.  This was complicated and "felt wrong".

 - Working out how to do it the "Clojure way".  Understanding that I needed
   agents, and then writing a macro to manage the agents.

At first I was confused by agents - I thought they would be something like
processes.  But really they are just values that a process can access.  So in
my code above.  So my original idea to put "do-stuff-with" in an agent was
wrong - it is the "i" that "goes inside" the agent.

So the basic way to solve this problem in Clojure is to:

 - Create an agent for each "i"

 - Use "send" to send the function "do-stuff-with" to each agent.

 - Wait for all the agents to finish.

Behind the scenes, Clojure is careful to only let a few agents run at a time
(this is what "send" takes care of).

So a solution would look something like:

  (apply await
    (for [i (range 4)]
      (let [a (agent i)]
        (send a do-stuff-with)
        a)))

which:

 - Generates an agent for each i

 - Sends the work to the agent (which is queued to run in a thread from a 
   thread pool)

 - Creates a list of agents (the result of the "for")

 - Waits for all the agents to finish.

OK, so where does the macro come into this?  Well, this code:

  (defmacro dopar [seq-expr & body]
    (assert (= 2 (count seq-expr)) "single pair of forms in sequence
    expression")
    (let [[k v] seq-expr]
      `(apply await
	 (for [k# ~v]
	   (let [a# (agent k#)]
	     (send a# (fn [~k] ~@body))
	     a#)))))

lets me write the (apply....) above as:

  (dopar [i (range 4)]
    (do-stuff-with i))

which is, I think, pretty awesome.

Andrew