Talk:Parallelism vs. Concurrency

From HaskellWiki
Revision as of 21:51, 20 April 2021 by Atravers (talk | contribs) (Using unsafeInterleaveIO where applicable)
Jump to navigation Jump to search

Parallelism vs concurrency: what's the difference?

Visible side effects.

  • Have a look at this ugly eysore "prototype definition" of par:
par     :: a -> b -> b
par x y =  case
             unsafeLocalState (forkIO (evalIO x >> return ()))
           of
             !_ -> y

where:

      evalIO :: a -> IO a
      forkIO :: IO () -> IO ThreadId

Assuming:

  • x is well-defined (it contains no unsafe... calls),
  • x is well-behaved (not throwing exceptions or causing errors);

then:

  1. forkIO attaches a ThreadId to its argument, adds it to the work-queue and returns the identifier;
  2. par then returns y;
  3. Some time later, forkIO's argument is called, causing evalIO to start evaluating x.

If y is still being evaluated when the evaluation of x commences, then we have elementary parallelism: concurrency, but with no visible side-effects.

  • Now have a look at this nearly-as-ugly prototype definition for spawnIO forkIO:
forkIO     :: IO () -> IO ThreadId
forkIO act =  do t <- unsafeInterleaveIO act
                 case par t () of
                   !_ -> do i' <- itsThreadId t
                            case i' of
                              Just i  -> return i
                              Nothing -> ioError "forkIO"

where:

      itsThreadId :: a -> IO (Maybe ThreadId)

Assuming:

  • par and itsThreadId are primitive,
  • itsThreadId would return Nothing if it's argument had not been previously used by par;

then:

  1. Evaluating par t () causes a new ThreadId to be attached to t by the implementation;
  2. itsThreadId retrieves i', the (possible) identifier for t;
  3. forkIO then extracts and returns the identifier.

This looks very much like elementary concurrency: parallelism, but having visible side effects.

Can either of these prototypes ever go mainstream?

  • As shown by it's type signature, par is supposed to be pure: avoiding the use of unsafeLocalState means making it primitive;
  • While the use of unsafeInterleaveIO may annoy some, it being one of the earlier Haskell extensions means it's widely available.

For now, using a primitive par (and others) to define forkIO looks like the simplest option...but if using unsafeInterleaveIO really does annoy you, how about this:

forkIO       :: (OI -> ()) -> OI -> ThreadId
forkIO act u =  let !(u1:u2:u3:_) = parts u in
                let t             = act u1 in
                case par t () of
                  !_ -> case itsThreadId t u2 of
                          Just i  -> i
                          Nothing -> ioError "forkIO" u3

-- Atravers Tue Apr 20 06:04:10 UTC 2021