[Haskell-cafe] ANNOUNCE: StrictBench 0.1 - Benchmarking code through strict evaluation

[Kenneth Hoste]

Hi all,

On Jun 8, 2009, at 15:12 , Magnus Therning wrote:

> On Mon, Jun 8, 2009 at 1:56 PM, Martijn van
> Steenbergen<martijn at van.steenbergen.nl> wrote:
>> Magnus Therning wrote:
>>>
>>> Is there no way to force repeated evaluation of a pure value?  (It'd
>>> be nice to be able to perform time measurements on pure code so that
>>> it's possible to compare Haskell implementations of algorithms to
>>> implementations in other languages, without running into confounding
>>> factors.)
>>
>> I'm really curious about this too.
>>
>> My guess is that the answer is "no" because doing so would (among  
>> other
>> things) mean a thunk have to be copied first before it is  
>> evaluated, to
>> preserve the unevaluated version. And what guarantee is there that  
>> values
>> further down the expression haven't been evaluated already?  
>> Efficient lazy
>> evaluation is hard; inefficient lazy evalation is even harder. ;-)
>
> Yes, I guessed as much.  I was hoping that there might be some way of
> tricking GHC into being more inefficient though, something like a
> rollback in evaluation state.

I've been playing around with MicroBench [1], and I believe there is a  
way
to trick GHC (at least the 6.10.2 version) into being inefficient.

Below is a snippet of the code I used to benchmark various  
implementations
of a function. The key is partial function application and the IO monad.
Don't ask me why it works, but I believe it does.

-- benchmark a given function by applying it n times to the given value
benchmark :: (a -> b) -> a -> Int -> IO()
benchmark f x n = do
                         r <- mapM (\y -> return $! f y) (replicate n x)
                         performGC
                         return ()

The performGC might not be 100% necessary, but I see it as a part of the
function evaluation (i.e. make the runtime clean up the mess the  
function made).
Of course this assumes performGC to be called before using "benchmark".

Note: MicroBench was doing something similar, but was using mapM_  
instead,
which no longer seems to fool GHC into evaluating the function n  
times. mapM
does seem to work though.

K.

-- 

Kenneth Hoste
Paris research group - ELIS - Ghent University, Belgium
email: kenneth.hoste at elis.ugent.be
website: http://www.elis.ugent.be/~kehoste
blog: http://boegel.kejo.be