| Portability | portable |
|---|---|
| Stability | experimental |
| Maintainer | [email protected] |
Control.Parallel.Strategies
Contents
Description
Parallel Evaluation Strategies, or Strategies for short, specify a way to evaluate a structure with components in sequence or in parallel.
Strategies are for expressing deterministic parallelism: the result of the program is unaffected by evaluating in parallel. For non-deterministic parallel programming, see Control.Concurrent.
Strategies let you separate the description of parallelism from the logic of your program, enabling modular parallelism.
Version 1.x
The original Strategies design is described in http://www.macs.hw.ac.uk/~dsg/gph/papers/html/Strategies/strategies.html and the code was written by Phil Trinder, Hans-Wolfgang Loidl, Kevin Hammond et al.
Version 2.x
Later, during work on the shared-memory implementation of parallelism in GHC, we discovered that the original formulation of Strategies had some problems, in particular it lead to space leaks and difficulties expressing speculative parallelism. Details are in the paper "Runtime Support for Multicore Haskell" http://www.haskell.org/~simonmar/papers/multicore-ghc.pdf.
This module has been rewritten in version 2. The main change is to
the 'Strategy a' type synonym, which was previously a -> Done and
is now a -> a. This change helps to fix the space leak described
in "Runtime Support for Multicore Haskell". The problem is that
the runtime will currently retain the memory referenced by all
sparks, until they are evaluated. Hence, we must arrange to
evaluate all the sparks eventually, just in case they aren't
evaluated in parallel, so that they don't cause a space leak. This
is why we must return a "new" value after applying a Strategy,
so that the application can evaluate each spark created by the
Strategy.
The simple rule is this: you must use the result of applying
a Strategy if the strategy creates parallel sparks, and you
should probably discard the the original value. If you don't
do this, currently it may result in a space leak. In the
future (GHC 6.14), it will probably result in lost parallelism
instead, as we plan to change GHC so that unreferenced sparks
are discarded rather than retained (we can't make this change
until most code is switched over to this new version of
Strategies, because code using the old verison of Strategies
would be broken by the change in policy).
The other changes in version 2.x are:
- Strategies can now be defined using a convenient Applicative type Eval. e.g. parList s = unEval $ traverse (Par . s)
-
parListhas been generalised toparTraverse, which works on anyTraversabletype. -
parListandparBufferhave versions specialised torwhnf, and there are transformation rules that automatically translate e.g.parList rwnhfinto a call to the optimised version. -
NFDatahas been moved toControl.DeepSeqin thedeepseqpackage. Note that since theStrategytype changed,rnfis no longer aStrategy: userdeepseqinstead.
- type Strategy a = a -> a
- using :: a -> Strategy a -> a
- withStrategy :: Strategy a -> a -> a
- rwhnf :: Strategy a
- rdeepseq :: NFData a => Strategy a
- seqPair :: Strategy a -> Strategy b -> Strategy (a, b)
- parPair :: Strategy a -> Strategy b -> Strategy (a, b)
- seqTriple :: Strategy a -> Strategy b -> Strategy c -> Strategy (a, b, c)
- parTriple :: Strategy a -> Strategy b -> Strategy c -> Strategy (a, b, c)
- seqTraverse :: Traversable t => Strategy a -> Strategy (t a)
- parTraverse :: Traversable t => Strategy a -> Strategy (t a)
- parList :: Strategy a -> Strategy [a]
- seqList :: Strategy a -> Strategy [a]
- parListN :: Int -> Strategy a -> Strategy [a]
- parListChunk :: Int -> Strategy a -> Strategy [a]
- parMap :: Strategy b -> (a -> b) -> [a] -> [b]
- parBuffer :: Int -> Strategy a -> [a] -> [a]
- parListWHNF :: Strategy [a]
- parBufferWHNF :: Int -> Strategy [a]
- ($|) :: (a -> b) -> Strategy a -> a -> b
- ($||) :: (a -> b) -> Strategy a -> a -> b
- (.|) :: (b -> c) -> Strategy b -> (a -> b) -> a -> c
- (.||) :: (b -> c) -> Strategy b -> (a -> b) -> a -> c
- (-|) :: (a -> b) -> Strategy b -> (b -> c) -> a -> c
- (-||) :: (a -> b) -> Strategy b -> (b -> c) -> a -> c
- data Eval a
- unEval :: Eval a -> a
- class NFData a where
- rnf :: a -> ()
- type Done = ()
- demanding :: a -> Done -> a
- sparking :: a -> Done -> a
- (>|) :: Done -> Done -> Done
- (>||) :: Done -> Done -> Done
- r0 :: a -> ()
Strategy type and basic operations
type Strategy a = a -> aSource
A Strategy is a function that embodies a parallel evaluation strategy.
The function traverses (parts of) its argument, evaluating subexpressions
in parallel or in sequence.
A Strategy may do an arbitrary amount of evaluation of its
argument, but should not return a value different from the one it
was passed.
Parallel computations may be discarded by the runtime system if the
program no longer requires their result, which is why a Strategy
function returns a new value equivalent to the old value. The
intention is that the program applies the Strategy to a
structure, and then uses the returned value, discarding the old
value. This idiom is expressed by the using function.
withStrategy :: Strategy a -> a -> aSource
A Strategy that simply evaluates its argument to Weak Head Normal
Form (i.e. evaluates it as far as the topmost constructor).
rdeepseq :: NFData a => Strategy aSource
A Strategy that fully evaluates its argument
rdeepseq a = rnf a `pseq` a
Tuple strategies
General traversals
seqTraverse :: Traversable t => Strategy a -> Strategy (t a)Source
A strategy that traverses a container data type with an instance
of Traversable, and evaluates each of the elements in left-to-right
sequence using the supplied strategy.
parTraverse :: Traversable t => Strategy a -> Strategy (t a)Source
A strategy that traverses a container data type with an instance
of Traversable, and sparks each of the elements using the supplied
strategy.
List strategies
parList :: Strategy a -> Strategy [a]Source
Spark each of the elements of a list using the given strategy.
Equivalent to parTraverse at the list type.
seqList :: Strategy a -> Strategy [a]Source
Evaluate each of the elements of a list sequentially from left to right
using the given strategy. Equivalent to seqTraverse at the list type.
parListChunk :: Int -> Strategy a -> Strategy [a]Source
parBuffer :: Int -> Strategy a -> [a] -> [a]Source
Applies a strategy to the nth element of list when the head is demanded. More precisely:
- semantics:
parBuffer n s = id :: [a] -> [a] - dynamic behaviour: evalutates the nth element of the list when the head is demanded.
The idea is to provide a `rolling buffer' of length n. It is a
better than parList for a lazy stream, because parList will
evaluate the entire list, whereas parBuffer will only evaluate a
fixed number of elements ahead.
Simple list strategies
parListWHNF :: Strategy [a]Source
version of parList specialised to rwhnf. This version is
much simpler, and may be faster than 'parList rwhnf'. You should
never need to use this directly, since 'parList rwhnf' is
automatically optimised to parListWHNF. It is here for
experimentation purposes only.
parBufferWHNF :: Int -> Strategy [a]Source
version of parBuffer specialised to rwhnf. You should
never need to use this directly, since 'parBuffer rwhnf' is
automatically optimised to parBufferWHNF. It is here for
experimentation purposes only.
Strategy composition operators
($|) :: (a -> b) -> Strategy a -> a -> bSource
Sequential function application. The argument is evaluated using the given strategy before it is given to the function.
($||) :: (a -> b) -> Strategy a -> a -> bSource
Parallel function application. The argument is evaluated using the given strategy, in parallel with the function application.
(.|) :: (b -> c) -> Strategy b -> (a -> b) -> a -> cSource
Sequential function composition. The result of the second function is evaluated using the given strategy, and then given to the first function.
(.||) :: (b -> c) -> Strategy b -> (a -> b) -> a -> cSource
Parallel function composition. The result of the second function is evaluated using the given strategy, in parallel with the application of the first function.
(-|) :: (a -> b) -> Strategy b -> (b -> c) -> a -> cSource
Sequential inverse function composition, for those who read their programs from left to right. The result of the first function is evaluated using the given strategy, and then given to the second function.
(-||) :: (a -> b) -> Strategy b -> (b -> c) -> a -> cSource
Parallel inverse function composition, for those who read their programs from left to right. The result of the first function is evaluated using the given strategy, in parallel with the application of the second function.
Building strategies
Eval is an Applicative Functor that makes it easier to define
parallel strategies that involve traversing structures.
a Seq value will be evaluated strictly in sequence in its context,
whereas a Par value wraps an expression that may be evaluated in
parallel. The Applicative instance allows sequential composition,
making it possible to describe an evaluateion strategy by composing
Par and Seq with <*>.
For example,
parList :: Strategy a -> Strategy [a] parList strat = unEval . traverse (Par . strat)
seqPair :: Strategy a -> Strategy b -> Strategy (a,b) seqPair f g (a,b) = unEval $ (,) <$> Seq (f a) <*> Seq (g b)
Instances
re-exported for backwards compatibility
class NFData a where
Methods
rnf :: a -> ()
rnf should reduce its argument to normal form (that is, fully evaluate all sub-components), and then return '()'.
The default implementation of rnf is
rnf a = a `seq` ()
which may be convenient when defining instances for data types with no unevaluated fields (e.g. enumerations).
Instances