[Haskell-cafe] performance difference for binary-0.4.3.1 with ghc-6.8.3 and ghc-6.10

[Don Stewart]

jwlato:
> On Tue, Oct 28, 2008 at 5:43 PM, Don Stewart <dons at galois.com> wrote:
> > jwlato:
> >> Hello,
> >>
> >> I was experimenting with using ghc-6.10.0.20081007 on a project, and
> >> it seems that binary-0.4.3.1 has markedly worse performance in certain
> >> cases.  With the following simple test:
> >>
> >> > import qualified Data.ByteString.Lazy as L
> >> > import Data.Binary
> >> > import Data.Binary.Get
> >> > import Control.Monad
> >> >
> >> > main :: IO ()
> >> > main = do
> >> >     b <- L.readFile "some_binary_file"
> >> >     putStrLn $ show $ runGet getter b
> >>
> >> > getter :: Get [Word16]
> >> > getter = replicateM 1000000 getWord16le
> >>
> >> running this program compiled with ghc-6.10 takes about 4 times as
> >> long (and consumes much more memory) as when compiled with ghc-6.8.3.
> >> The extra time appears to be proportional to the number of elements
> >> processed in the Get.  Running the programs with -hT shows a clear
> >> memory difference, which I think is the source of the problem.  I've
> >> placed pdfs of that output at https://webspace.utexas.edu/latojw/data/
> >>
> >> The difference seems to manifest itself only when the elements are
> >> actually processed; changing "show $ runGet " to "show $ length $
> >> runGet " is slightly faster in 6.10.
> >>
> >> I was working on an Intel Mac with OS 10.4, binary-0.4.3.1, and
> >> bytestring-0.9.1.4.  Can anyone confirm this, or suggest what might be
> >> the difference?
> >>
> >
> > Is this the sole test case?
> >
> > I can investigate. Though perhaps using a newer GHC release candidate is
> > also a good idea.
> >
> > -- Don
> >
> 
> I'll try a newer release candidate, although at the time I first ran
> this it was the latest.
> 
> Last night I tried creating the bytestring using "repeat 1" to remove
> the file I/O, and the result was the same.  There are a few other
> things I want to try as well (removing replicateM, folding the list to
> force values rather than printing them all), but my time is extremely
> limited at the moment.


Could you send me a minimal test case not involving IO?

-- Don