Proposal: add forkOnIO and friends to Control.Concurrent:

Bas van Dijk v.dijk.bas at gmail.com
Thu Dec 23 01:12:45 CET 2010


On Wed, Dec 22, 2010 at 4:55 PM, Duncan Coutts
<duncan.coutts at googlemail.com> wrote:
> On 22 December 2010 11:41, Simon Marlow <marlowsd at gmail.com> wrote:
>> Ticket:
>>
>>  http://hackage.haskell.org/trac/ghc/ticket/4859
>>
>> I think these functions are implementation-independent enough to add to the
>> main `Control.Concurrent` API:
>>
>> {{{
>> {- |
>> Like 'forkIO', but lets you specify on which CPU the thread is
>> created.  Unlike a `forkIO` thread, a thread created by `forkOnIO`
>> will stay on the same CPU for its entire lifetime (`forkIO` threads
>> can migrate between CPUs according to the scheduling policy).
>> `forkOnIO` is useful for overriding the scheduling policy when you
>> know in advance how best to distribute the threads.
>>
>> The `Int` argument specifies the CPU number; it is interpreted modulo
>> the value returned by 'getNumCapabilities'.
>
> While Int is clearly enough for a multi-processors single node system,
> I wonder if there is something slightly more general that would give
> us a common API with the Eden / distributed Haskell people. For
> example in MPI the identification of a node is rather more complex
> than a single integer. Perhaps some more abstract identifier for a
> capability would be useful. Closely related of course is how one
> discovers the available capabilities.

Such a design would probably look something like this:

newtype Capability = Capability Int -- abstract

getCapabilities :: IO [Capability]  -- or should this be:
getCapabilities :: IO (Capability, [Capability])
-- to make it clear there's always at least one capability.

forkOnIO :: Capability -> IO () -> IO ThreadId
forkOnIOWithUnmask :: Capability -> ((forall a . IO a -> IO a) -> IO
()) -> IO ThreadId

threadIsPinned :: ThreadId -> IO (Maybe Capability)

I don't really like the name "Capability". "Processing Unit" is more
descriptive but a bit long. Maybe shorten that to "PU".

However I'm not sure yet if such a design provides much benefits over
the current one.

Bas



More information about the Libraries mailing list