iteratee-0.8.9.0: src/Data/Iteratee/Parallel.hs
{-# LANGUAGE NoMonomorphismRestriction, BangPatterns, ScopedTypeVariables #-}
module Data.Iteratee.Parallel (
psequence_
-- ,psequence
,parE
,parI
,liftParI
,mapReduce
)
where
import Control.Monad.IO.Class
import Control.Monad.Trans.Class
import Data.Iteratee as I hiding (mapM_, zip, filter)
import qualified Data.ListLike as LL
import Data.Monoid
import Control.Concurrent
import Control.Parallel
import Control.Monad
-- | Transform usual Iteratee into parallel composable one, introducing
-- one step extra delay.
--
-- Ex - time spent in Enumerator working on x'th packet
-- Ix - time spent in Iteratee working on x'th packet
-- z - last packet, y = (z-1)'th packet
--
-- regular Iteratee: E0 - I0, E1 - I1, E2 - I2 .. Ez -> Iz
-- parallel Iteratee: E0, E1, E2, .. Ez
-- \_ I0\_ I1\_ .. Iy\__ Iz
--
parI :: forall s a. (Nullable s, Monoid s) => Iteratee s IO a -> Iteratee s IO a
parI = icont . firstStep
where
-- first step, here we fork separate thread for the next chain and at the
-- same time ask for more date from the previous chain
firstStep iter chunk = do
var <- liftIO newEmptyMVar
_ <- sideStep var chunk iter
return (icont $ go var, mempty)
-- somewhere in the middle, we are getting iteratee from previous step,
-- feeding it with some new data, asking for more data and starting
-- more processing in separete thread
go var chunk@(Chunk _) = do
iter <- liftIO $ takeMVar var
_ <- sideStep var chunk iter
return (icont $ go var, mempty)
-- final step - no more data, so we need to inform our consumer about it
go var e = do
iter <- liftIO $ takeMVar var
return (join . lift $ enumChunk e iter, e)
-- forks away from the main computation, return results via MVar
sideStep var chunk iter = liftIO . forkIO
$ runIter iter onDone onCont onErr onReq
where
onDone a = putMVar var $ idone a
onCont k = k chunk >>= \(i',_) ->
runIter i' onDone onFina onErr onReq
onErr i e = putMVar var $ ierr i e
onReq :: IO x -> (x -> Iteratee s IO a) -> IO ()
onReq mb doB = mb >>= \b ->
runIter (doB b) onDone onCont onErr onReq
onFina k = putMVar var $ icont k
-- | Transform an Enumeratee into a parallel composable one, introducing
-- one step extra delay, see 'parI'.
parE ::
(Nullable s1, Nullable s2, Monoid s1)
=> Enumeratee s1 s2 IO r
-> Enumeratee s1 s2 IO r
parE outer inner = parI (outer inner)
-- | Enumerate a list of iteratees over a single stream simultaneously
-- and discard the results. Each iteratee runs in a separate forkIO thread,
-- passes all errors from iteratees up.
psequence_ ::
(LL.ListLike s el, Nullable s)
=> [Iteratee s IO a]
-> Iteratee s IO ()
psequence_ = I.sequence_ . map parI
{-
-- | Enumerate a list of iteratees over a single stream simultaneously
-- and keeps the results. Each iteratee runs in a separete forkIO thread, passes all
-- errors from iteratees up.
psequence = I.sequence . map parI
-}
-- | A variant of 'parI' with the parallelized iteratee lifted into an
-- arbitrary MonadIO.
liftParI ::
(Nullable s, Monoid s, MonadIO m)
=> Iteratee s IO a
-> Iteratee s m a
liftParI = ilift liftIO . parI
-- | Perform a parallel map/reduce. The `bufsize` parameter controls
-- the maximum number of chunks to read at one time. A larger bufsize
-- allows for greater parallelism, but will require more memory.
--
-- Implementation of `sum`
--
-- > sum :: (Monad m, LL.ListLike s, Nullable s) => Iteratee s m Int64
-- > sum = getSum <$> mapReduce 4 (Sum . LL.sum)
mapReduce ::
(Monad m, Nullable s, Monoid b)
=> Int -- ^ maximum number of chunks to read
-> (s -> b) -- ^ map function
-> Iteratee s m b
mapReduce bufsize f = icontP (step (0, []))
where
step a@(!buf,acc) (Chunk xs)
| nullC xs = (icontP (step a), Chunk xs)
| buf >= bufsize =
let acc' = mconcat acc
b' = f xs
in b' `par` acc' `pseq` (icontP (step (0,[b' `mappend` acc'])), Chunk empty)
| otherwise =
let b' = f xs
in b' `par` (icontP (step (succ buf,b':acc)), Chunk empty)
step (_,acc) s@(EOF Nothing) =
(idone (mconcat acc), s)
step acc s@(EOF (Just err)) =
(throwRecoverableErr err (icontP $ step acc), s)