[Haskell-cafe] Re: True Random Numbers

[James Andrew Cook]

Thanks for the clues, I'll try and make some time this weekend to track it down.  I do have some gentoo x64 systems to play with.  My first impulse is actually that it is likely due to differences in inlining and/or rewrite rule processing between the GHC versions, but we'll see what turns up.

-- James

On Apr 9, 2010, at 6:51 AM, Gökhan San wrote:

> mokus at deepbondi.net writes:
> 
>> are you using the hackage-released version of random-fu or the darcs
>> one?
> 
> I was using the hackage version, but since then I switched to the darcs
> version. (Btw, began using it in some of my projects and I'm really
> happy about it.)
> 
>> In the above case, I was using IO in the first bgroup and State StdGen
>> in the second.
> 
> I'm running it on a x86_64 Gentoo Linux box with GHC 6.10.4 and was
> unable to install Criterion (apparently, impossible is happening while
> compiling vector-algorithms) so I used 'time' to come up with some
> results.
> 
> Below doesn't include IO tests (randomRIO, etc.), since they turned out
> to be spectacularly slow anyway. Results using ghc -O2.
> 
>> module Main (main) where 
> 
>> import Data.Random
>> import Data.List
>> import Control.Monad.State
>> import Control.Monad.Random
>> import System.Random
> 
>> test       = p1 `fmap` getStdGen
>> type RType = Double
> 
> /usr/bin/time results for (test, RType):
> 
> (p1, Double) : ~3.3 secs
> (p2, Double) : ~1.7 secs
> (p3, Double) : ~1.0 sec
> (p1, Int)    : ~1.9 secs
> (p2, Int)    : ~1.0 sec
> (p3, Int)    : ~0.5 sec
> 
>> count  = 10 ^ 6
>> range  = (-10, 10)
>> type P = StdGen -> [RType]
> 
>> p1 = evalState (sample (replicateM count (uncurry uniform range))) :: P
> 
>> p2 = evalRand (replicateM count (getRandomR range)) :: P
> 
>> p3 = take count . evalRand (getRandomRs range) :: P
> 
>> main = test >>= (print . foldl' (+) 0)
> 
> Using 'sum' turned to be rather misleading (took up to a minute to sum
> up 'Double's; this problem was less apparent for p1), so I had to use
> foldl' here to get consistent results between 'Int's and
> 'Double's. '`using` rnf' produced similar results.
> 
> Also, using DevURandom for random-fu produces almost the same results.
> 
> -- 
> 
> Gökhan San
>