Sunday, June 29, 2014

Optimisation with Continuations

Summary: Continuations are confusing. Here we solve a simple problem (that is at the heart of the Shake build system) using continuations.

Imagine we are given two IO a computations, and want to run them both to completion, returning the first a value as soon as it is produced (let's ignore exceptions). Writing that in Haskell isn't too hard:

parallel :: IO a -> IO a -> IO a
parallel t1 t2 = do
    once <- newOnce
    var <- newEmptyMVar
    forkIO $ t1 >>= once . putMVar var
    forkIO $ t2 >>= once . putMVar var
    readMVar var

We create an empty variable var with newEmptyMVar, fire off two threads with forkIO to run the computations which write their results to var, and finish by reading as soon as a value is available with readMVar. We use a utility newOnce to ensure that only one of the threads calls putMVar, defined as:

newOnce :: IO (IO () -> IO ())
newOnce = do
    run <- newMVar True
    return $ \act -> do
        b <- modifyMVar run $ \b -> return (False, b)
        when b act

Calling newOnce produces a function that given an action will either run it (the first time) or ignore it (every time after). Using newOnce we only call putMVar for the first thread to complete.

This solution works, and Shake does something roughly equivalent (but much more complex) in it's main scheduler. However, this solution has a drawback - it uses two additional threads. Can we use only one additional thread?

For the problem above, running the computations to completion without retrying, you can't avoid two additional threads. To use only one additional thread and run in parallel you must run one of the operations on the calling thread - but if whatever you run on the additional thread finishes first, there's no way to move the other computation off the the calling thread and return immediately. However, we can define:

type C a = (a -> IO ()) -> IO ()

Comparing IO a to C a, instead of returning an a, we get given a function to pass the a to (known as a continuation). We still "give back" the a, but not as a return value, instead we pass it onwards to a function. We assume that the continuation is called exactly once. We can define parallel on C:

parallel :: C a -> C a -> C a
parallel t1 t2 k = do
    once <- newOnce
    forkIO $ t1 (once . k)
    t2 (once . k)

This definition takes the two computations to run (t1 and t2), plus the continuation k. We fork a separate thread to run t1, but run t2 on the calling thread, using only one additional thread. While the parallel function won't return until after t2 completes, subsequent processing using the a value will continue as soon as either finishes.

Looking at the transformers package, we see Control.Monad.Trans.Cont contains ContT, which is defined as:

newtype ContT r m a = ContT {runContT :: (a -> m r) -> m r}

If we use r for () and IO for m then we get the same type as C. We can redefine C as:

type C a = ContT () IO a

The changes to parallel just involve wrapping with ContT and unwrapping with runContT:

parallel :: C a -> C a -> C a
parallel t1 t2 = ContT $ \k -> do
    once <- newOnce
    forkIO $ runContT t1 (once . k)
    runContT t2 (once . k)

Now we've defined our parallel function in terms of C, it is useful to convert between C and IO:

toC :: IO a -> C a
toC = liftIO

fromC :: C a -> IO a
fromC c = do
    var <- newEmptyMVar
    forkIO $ runContT c $ putMVar var
    readMVar var

The toC function is already defined by ContT as liftIO. The fromC function needs to change from calling a callback on any thread, to returning a value on this thread, which we can do with a forkIO and MVar. Given parallel on IO takes two additional threads, and parallel on C takes only one, it's not too surprising that converting IO to C requires an additional thread.

Aren't threads cheap?

Threads in Haskell are very cheap, and many people won't care about one additional thread. However, each thread comes with a stack, which takes memory. The stack starts off small (1Kb) and grows/shrinks in 32Kb chunks, but if it ever exceeds 1Kb, it never goes below 32Kb. For certain tasks (e.g. Shake build rules) often some operation will take a little over 1Kb in stack. Since each active rule (started but not finished) needs to maintain a stack, and for huge build systems there can be 30K active rules, you can get over 1Gb of stack memory. While stacks and threads are cheap, they aren't free.

The plan for Shake

Shake currently has one thread per active rule, and blocks that thread until all dependencies have rebuilt. The plan is to switch to continuations and only have one thread per rule executing in parallel. This change will not require any code changes to Shake-based build systems, hopefully just reduce memory usage. Until then, huge build systems may wish to pass +RTS -kc8K, which can save several 100Mb of memory.


Anonymous said...

Is this type signature messed up? It looks like it's missing the type for k to me:

parallel :: C a -> C a -> C a
parallel t1 t2 k = do ...

Neil Mitchell said...

C a is a type alias for (a -> IO ()) -> IO (), so you can rewrite it as:

parallel :: C a -> C a -> (a -> IO ()) -> IO ()

Now you can see where the k comes from.

Thomas Schilling said...

When I built my clone of Shake (that was shortly before you published your personal rewrite) I ended up using an explicit data type to represent the current state of a rule. Some of those did include a continuation.

Neil Mitchell said...

Thomas: Very interesting to see. Did it work in practice? How did you cope with exceptions? Any general hardwon advice?

Thomas Schilling said...

It's been too long ago to remember the details.

I mostly did it because that was the most understandable way to design it. Actually, I looked at Max's Openshake and couldn't really make sense of it's inner workings. So, after thinking about how it should work conceptually, explicitly enumerating the states each Action could be in seemed to be the way that seemed to be easiest to follow (also for potential contributors).

I don't think I ever finished the code. Looking through the code, exceptions handling was never fully implemented, so I don't know if it can be made to work correctly.

The key driver is

Maybe you can see some fatal flaw in it :)

J-W Maessen said...

This observation – that running the continuation on the first thread to complete lets you use one less thread – is very similar to how the work-stealing scheduler in Cilk works. There we want to run both pieces of work to completion, but the first thread to finish wants to go off and find other work to do. So we actually steal the continuation of pending work.

S├ębastien Bocq said...

I thought also forkIO would be so cheap already that you wouldn't have to resort to CPS in Haskell.

This reminds me of a project I did in Scala a few years ago to support user-level green threads. Scala being impure, CPS with effects can be directly expressed as:
(a => ()) => ()

To cope with user-level threads and exceptions, the signature had to be changed to:
((t, a) => ()) => ()
Where t is a user-level thread. It is used to schedule back continuations sequentially (only for processing results!) and it acts also as an escape hatch in case of exception by implementing the function to handle them:
Exception => ().

After, I went on adding automatic resource control, and so on... If you are curious about it, the details are explained in the paper here:
(see section 3: Monadic User-Level Threads, code is on github too)

Getting the details rights was tricky and I would love to see if it can be expressed in a better way.

Side remark: I would call the operator something else because a parallel operator should really return both results: IO a -> IO b -> IO (a, b)

Neil Mitchell said...

S├ębastien: forkIO is cheap in terms of computation time, and thread stacks are small if you never do any meaningful work on them. Shake is in the rare position of having a lot of queued threads, each of which did a reasonable amount of work, and will again in the future - a worst case for the Haskell RTS.

Exceptions did turn out to be the hardest bit of the project, and I did end up including an exception handling bit, and some resource finalisation code. Thanks for the link, I'll have a read, and I'm also intending to post my notes on what ended up with in the end. My finalisers were quite neat (but not very advanced). My exception handlers on the continuation threads were complex, powerful, and seem a bit inelegant.