GHC provides two functions for controlling parallel execution, through the Parallel interface:

interface Parallel where
infixr 0 `par`
infixr 1 `seq`

par :: a -> b -> b
seq :: a -> b -> b

The expression (x `par` y) sparks the evaluation of x (to weak head normal form) and returns y. Sparks are queued for execution in FIFO order, but are not executed immediately. At the next heap allocation, the currently executing thread will yield control to the scheduler, and the scheduler will start a new thread (until reaching the active thread limit) for each spark which has not already been evaluated to WHNF.

The expression (x `seq` y) evaluates x to weak head normal form and then returns y. The seq primitive can be used to force evaluation of an expression beyond WHNF, or to impose a desired execution sequence for the evaluation of an expression.

For example, consider the following parallel version of our old nemesis, nfib:

import Parallel

nfib :: Int -> Int
nfib n | n <= 1 = 1
       | otherwise = par n1 (seq n2 (n1 + n2 + 1))
                     where n1 = nfib (n-1)
                           n2 = nfib (n-2)

For values of n greater than 1, we use par to spark a thread to evaluate nfib (n-1), and then we use seq to force the parent thread to evaluate nfib (n-2) before going on to add together these two subexpressions. In this divide-and-conquer approach, we only spark a new thread for one branch of the computation (leaving the parent to evaluate the other branch). Also, we must use seq to ensure that the parent will evaluate n2 before n1 in the expression (n1 + n2 + 1). It is not sufficient to reorder the expression as (n2 + n1 + 1), because the compiler may not generate code to evaluate the addends from left to right.

The functions par and seq are wired into GHC, and unfold into uses of the par# and seq# primitives, respectively. If you'd like to see this with your very own eyes, just run GHC with the -ddump-simpl option. (Anything for a good time…)

Runnable threads are scheduled in round-robin fashion. Context switches are signalled by the generation of new sparks or by the expiry of a virtual timer (the timer interval is configurable with the -C[<num>] RTS option). However, a context switch doesn't really happen until the current heap block is full. You can't get any faster context switching than this.

When a context switch occurs, pending sparks which have not already been reduced to weak head normal form are turned into new threads. However, there is a limit to the number of active threads (runnable or blocked) which are allowed at any given time. This limit can be adjusted with the -t<num> RTS option (the default is 32). Once the thread limit is reached, any remaining sparks are deferred until some of the currently active threads are completed.

In GUM we use an unfair scheduler, which means that a thread continues to perform graph reduction until it blocks on a closure under evaluation, on a remote closure or until the thread finishes.