JRuby vs. Java speed (re: Stochastic Simulation with SSJ)

Just read Ali Rizvi's blog post about "JRuby: Stochastic Simulation with SSJ" and decided to take a look at the implementation real quick... of course the Java version will be much faster, but the numbers given seemed a little strange.  After setting up the environment and running the tests, I saw the major slowdown as expected, then I decided to take a look at the Ruby code to see if I could speed things up a little.  This was done rather quickly, if you have better result... please share!

First, the Java version:

real    0m7.960s
user    0m7.839s
sys     0m0.056s

The JRuby version took too long to run (I'm pressed for time) so I added a few options to the runtime:

jruby -J-Djruby.compile.fastest=true --server collision.rb

real    3m55.834s

user    3m53.684s

sys     0m2.355s

Not as bad as reported by the blog ref'd above... but still pretty bad. I had a hunch the loops were what was killing the time and made the following changes to the collision.rb file:

@k.times { |i| @used[i] = false}

changed to:

@used.fill(false)

That brought the time down a bit:

real    1m1.430s

user    1m1.344s

sys     0m0.538s

It would seem loops are a sore spot. (kinda expected? the loop for setting happens in the Java impl) Decided to make one last change to see what happened. I condensed the "generate_c" method back into "simulate_runs" method to get the loops in one place. Some (not much) benefit was shown in my tests:

real    0m59.081s

user    0m59.027s

sys     0m0.502s

CONCLUSION: Yes, straight java is faster. Yes, there are ways to tune your Ruby to make things a bit more efficient. Yes, I'm sure there are other ways to make this run faster. I think prototyping this sort of thing in JRuby is just fine... then when you need more perf (if this stuff will be called from say a JRuby on Rails app) you can code it up in Java (Collision.java) and just call that from JRuby layer.

Here is the final collision.rb I ended up with :

require 'java'

require 'ssj.jar'



import 'umontreal.iro.lecuyer.rng.RandomStream'
import 'umontreal.iro.lecuyer.stat.Tally'
import 'umontreal.iro.lecuyer.rng.MRG32k3a'

class Collision
  def initialize(k, m)
    @k = k
    @m = m
    @lambda = m * m / (2.0 * k)
    @used = Array.new(k, false)
  end


def simulate_runs(n, stream, stat_c)

stat_c.init

n.times do

c = 0

@used.fill(false)

@m.times do

loc = stream.nextInt(0, @k-1)

if @used[loc]

c += 1

else

@used[loc] = true

end

end

stat_c.add(c)

end

stat_c.setConfidenceIntervalStudent()

puts stat_c.report(0.95, 3)

puts " Theoretical mean: #{@lambda} "

end



def self.run

stat_c = Tally.new("Statistics on collision")

col = Collision.new(10000,500)

col.simulate_runs(100000, MRG32k3a.new, stat_c)

end



end



Collision.run