Re: [Haskell-cafe] Question about the Monad instance for Iteratee (from the enumerator package)
Joining slightly late...
From: John Millikin
John Lato's "iteratee" package is based on IterateeMCPS.hs[1]. I used IterateeM.hs for "enumerator", because when I benchmarked them the non-CPS version was something like 10% faster on most operations.
Based on tests I did before iteratee-0.4, the fastest implementation was Oleg's alternate design which is in the comments of IterateeM.hs. I don't think any current packages use that however, with good reason. I'd be interested to see the results of a shootout between iteratee and enumerator. I would expect them to be basically equivalent most of the time, with maybe two or three operations with a small (but consistent) difference one way or the other. John L.
On Tuesday, April 26, 2011 7:19:25 AM UTC-7, John Lato wrote:
I'd be interested to see the results of a shootout between iteratee and enumerator. I would expect them to be basically equivalent most of the time, with maybe two or three operations with a small (but consistent) difference one way or the other.
I did some basic benchmarks a few months ago; if I remember correctly, it depends almost entirely on how well GHC optimizes CPS on a particular platform. The relative performace was very similar to Lennart Kolmodin's benchmarks of "binary" at < http://lennartkolmodin.blogspot.com/2011/02/binary-by-numbers.html >. In particular, CPS/"iteratee" is faster on 32-bit, while state passing/"enumerator" is faster on 64-bit. This difference exists for almost all operations, and was on the order of 5-15% depending on the shape of the input. I couldn't figure out a good way to benchmark the libraries themselves when there's so much noise from the compiler. I'm waiting for GHC7 to stabilise a bit before doing another benchmark -- the LLVM backend and new CPS handling should make both implementations faster, once they're a bit more mature. I really hope CPS becomes as fast on 64-bit as it is on 32-bit, because it's a much cleaner implementation.
On Tue, Apr 26, 2011 at 6:32 PM, John Millikin
On Tuesday, April 26, 2011 7:19:25 AM UTC-7, John Lato wrote:
I'd be interested to see the results of a shootout between iteratee and enumerator. I would expect them to be basically equivalent most of the time, with maybe two or three operations with a small (but consistent) difference one way or the other.
I did some basic benchmarks a few months ago; if I remember correctly, it depends almost entirely on how well GHC optimizes CPS on a particular platform. The relative performace was very similar to Lennart Kolmodin's benchmarks of "binary" at < http://lennartkolmodin.blogspot.com/2011/02/binary-by-numbers.html >. In particular, CPS/"iteratee" is faster on 32-bit, while state passing/"enumerator" is faster on 64-bit.
This difference exists for almost all operations, and was on the order of 5-15% depending on the shape of the input. I couldn't figure out a good way to benchmark the libraries themselves when there's so much noise from the compiler.
That figures. Thanks for sharing this. John L.
participants (2)
-
John Lato -
John Millikin