A framework for safe, programmable, speculative parallelism, loosely based on:
- Prakash Prabhu, G. Ramalingam, and Kapil Vaswani, "Safe Programmable Speculative Parallelism", In the proceedings of Programming Language Design and Implementation (PLDI) Vol 45, Issue 6 (June 2010) pp 50-61. http://research.microsoft.com/pubs/118795/pldi026-vaswani.pdf
This package provides speculative function application and speculative folds. Speculative
STM transactions take the place
of the transactional rollback machinery from the paper.
You can download it using
cabal install speculation, if you have the Haskell Platform installed.
Various speculative function application combinators are provided. Two fairly canonical samples are described here.
spec :: Eq a => a -> (a -> b) -> a -> b
spec g f a evaluates
f g while forcing
g == a then
f g is returned. Otherwise
f a is evaluated.
Furthermore, if the argument has already been evaluated, we avoid sparking the parallel computation at all.
g is a good guess at the value of
a, this is one way to induce parallelism in an otherwise sequential task.
g isn\'t available more cheaply than
a, then this saves no work, and if
g is wrong, you risk evaluating the function twice.
spec a f a = f $! a
The best-case timeline looks like: [---- f g ----] [----- a -----] [-- spec g f a --]
The worst-case timeline looks like: [---- f g ----] [----- a -----] [---- f a ----] [------- spec g f a -----------]
Compare these to the timeline of
f $! a:
[---- a -----]
[---- f a ----]
specSTM provides a similar compressed timeline for speculated
STM actions, but also rolls back side-effects.
A speculative version of the combinators from
Data.Foldable is provided as
Each combinator therein takes an extra argument that is used to speculate on the value of the list.
foldr :: (Foldable f, Eq b) => (Int -> b) -> (a -> b -> b) -> b -> f a -> b
Given a valid estimator
foldr g f z xs yields the same answer as
Foldable.foldr' f z xs.
g n should supply an estimate of the value returned from folding over the last
n elements of the container.
spec, if the guess
g n is accurate a reasonable percentage of the time and faster to compute than the ensuing fold, then this can provide increased opportunities for parallelism.
foldl :: (Foldable f, Eq b) => (Int -> b) -> (b -> a -> b) -> b -> f a -> b
foldl works similarly to
Foldable.foldl', except that
g n should provide an estimate for the first
Contributions and bug reports are welcome!
I can be reached through the user ekmett on github, as edwardk on irc.freenode.net #haskell channel, or by email at email@example.com.