CountedCompleter

A {@link ForkJoinTask} with a completion action performed when triggered and there are no remaining pending actions. CountedCompleters are in general more robust in the presence of subtask stalls and blockage than are other forms of ForkJoinTasks, but are less intuitive to program. Uses of CountedCompleter are similar to those of other completion based components (such as {@link java.nio.channels.CompletionHandler}) except that multiple <em>pending</em> completions may be necessary to trigger the completion action {@link #onCompletion(CountedCompleter)}, not just one. Unless initialized otherwise, the {@linkplain #getPendingCount pending count} starts at zero, but may be (atomically) changed using methods {@link #setPendingCount}, {@link #addToPendingCount}, and {@link #compareAndSetPendingCount}. Upon invocation of {@link #tryComplete}, if the pending action count is nonzero, it is decremented; otherwise, the completion action is performed, and if this completer itself has a completer, the process is continued with its completer. As is the case with related synchronization components such as {@link Phaser} and {@link Semaphore}, these methods affect only internal counts; they do not establish any further internal bookkeeping. In particular, the identities of pending tasks are not maintained. As illustrated below, you can create subclasses that do record some or all pending tasks or their results when needed. As illustrated below, utility methods supporting customization of completion traversals are also provided. However, because CountedCompleters provide only basic synchronization mechanisms, it may be useful to create further abstract subclasses that maintain linkages, fields, and additional support methods appropriate for a set of related usages.

<p>A concrete CountedCompleter class must define method {@link #compute}, that should in most cases (as illustrated below), invoke {@code tryComplete()} once before returning. The class may also optionally override method {@link #onCompletion(CountedCompleter)} to perform an action upon normal completion, and method {@link #onExceptionalCompletion(Throwable, CountedCompleter)} to perform an action upon any exception.

<p>CountedCompleters most often do not bear results, in which case they are normally declared as {@code CountedCompleter!(void)}, and will always return {@code null} as a result value. In other cases, you should override method {@link #getRawResult} to provide a result from {@code join(), invoke()}, and related methods. In general, this method should return the value of a field (or a function of one or more fields) of the CountedCompleter object that holds the result upon completion. Method {@link #setRawResult} by default plays no role in CountedCompleters. It is possible, but rarely applicable, to override this method to maintain other objects or fields holding result data.

<p>A CountedCompleter that does not itself have a completer (i.e., one for which {@link #getCompleter} returns {@code null}) can be used as a regular ForkJoinTask with this added functionality. However, any completer that in turn has another completer serves only as an internal helper for other computations, so its own task status (as reported in methods such as {@link ForkJoinTask#isDone}) is arbitrary; this status changes only upon explicit invocations of {@link #complete}, {@link ForkJoinTask#cancel}, {@link ForkJoinTask#completeExceptionally(Throwable)} or upon exceptional completion of method {@code compute}. Upon any exceptional completion, the exception may be relayed to a task's completer (and its completer, and so on), if one exists and it has not otherwise already completed. Similarly, cancelling an internal CountedCompleter has only a local effect on that completer, so is not often useful.

<p><b>Sample Usages.</b>

<p><b>Parallel recursive decomposition.</b> CountedCompleters may be arranged in trees similar to those often used with {@link RecursiveAction}s, although the constructions involved in setting them up typically vary. Here, the completer of each task is its parent in the computation tree. Even though they entail a bit more bookkeeping, CountedCompleters may be better choices when applying a possibly time-consuming operation (that cannot be further subdivided) to each element of an array or collection; especially when the operation takes a significantly different amount of time to complete for some elements than others, either because of intrinsic variation (for example I/O) or auxiliary effects such as garbage collection. Because CountedCompleters provide their own continuations, other tasks need not block waiting to perform them.

<p>For example, here is an initial version of a utility method that uses divide-by-two recursive decomposition to divide work into single pieces (leaf tasks). Even when work is split into individual calls, tree-based techniques are usually preferable to directly forking leaf tasks, because they reduce inter-thread communication and improve load balancing. In the recursive case, the second of each pair of subtasks to finish triggers completion of their parent (because no result combination is performed, the default no-op implementation of method {@code onCompletion} is not overridden). The utility method sets up the root task and invokes it (here, implicitly using the {@link ForkJoinPool#commonPool()}). It is straightforward and reliable (but not optimal) to always set the pending count to the number of child tasks and call {@code tryComplete()} immediately before returning.

<pre> {@code static <E> void forEach(E[] array, Consumer<E> action) { class Task extends CountedCompleter!(void) { final int lo, hi; Task(Task parent, int lo, int hi) { super(parent); this.lo = lo; this.hi = hi; }

void compute() { if (hi - lo >= 2) { int mid = (lo + hi) >>> 1; // must set pending count before fork setPendingCount(2); new Task(this, mid, hi).fork(); // right child new Task(this, lo, mid).fork(); // left child } else if (hi > lo) action.accept(arraylo); tryComplete(); } } new Task(null, 0, array.length).invoke(); }}</pre>

This design can be improved by noticing that in the recursive case, the task has nothing to do after forking its right task, so can directly invoke its left task before returning. (This is an analog of tail recursion removal.) Also, when the last action in a task is to fork or invoke a subtask (a "tail call"), the call to {@code tryComplete()} can be optimized away, at the cost of making the pending count look "off by one".

<pre> {@code void compute() { if (hi - lo >= 2) { int mid = (lo + hi) >>> 1; setPendingCount(1); // looks off by one, but correct! new Task(this, mid, hi).fork(); // right child new Task(this, lo, mid).compute(); // direct invoke } else { if (hi > lo) action.accept(arraylo); tryComplete(); } }}</pre>

As a further optimization, notice that the left task need not even exist. Instead of creating a new one, we can continue using the original task, and add a pending count for each fork. Additionally, because no task in this tree implements an {@link #onCompletion(CountedCompleter)} method, {@code tryComplete} can be replaced with {@link #propagateCompletion}.

<pre> {@code void compute() { int n = hi - lo; for (; n >= 2; n /= 2) { addToPendingCount(1); new Task(this, lo + n/2, lo + n).fork(); } if (n > 0) action.accept(arraylo); propagateCompletion(); }}</pre>

When pending counts can be precomputed, they can be established in the constructor:

<pre> {@code static <E> void forEach(E[] array, Consumer<E> action) { class Task extends CountedCompleter!(void) { final int lo, hi; Task(Task parent, int lo, int hi) { super(parent, 31 - Integer.numberOfLeadingZeros(hi - lo)); this.lo = lo; this.hi = hi; }

void compute() { for (int n = hi - lo; n >= 2; n /= 2) new Task(this, lo + n/2, lo + n).fork(); action.accept(arraylo); propagateCompletion(); } } if (array.length > 0) new Task(null, 0, array.length).invoke(); }}</pre>

Additional optimizations of such classes might entail specializing classes for leaf steps, subdividing by say, four, instead of two per iteration, and using an adaptive threshold instead of always subdividing down to single elements.

<p><b>Searching.</b> A tree of CountedCompleters can search for a value or property in different parts of a data structure, and report a result in an {@link hunt.concurrency.atomic.AtomicReference AtomicReference} as soon as one is found. The others can poll the result to avoid unnecessary work. (You could additionally {@linkplain #cancel cancel} other tasks, but it is usually simpler and more efficient to just let them notice that the result is set and if so skip further processing.) Illustrating again with an array using full partitioning (again, in practice, leaf tasks will almost always process more than one element):

<pre> {@code class Searcher<E> extends CountedCompleter<E> { final E[] array; final AtomicReference<E> result; final int lo, hi; Searcher(ICountedCompleter p, E[] array, AtomicReference<E> result, int lo, int hi) { super(p); this.array = array; this.result = result; this.lo = lo; this.hi = hi; } E getRawResult() { return result.get(); } void compute() { // similar to ForEach version 3 int l = lo, h = hi; while (result.get() is null && h >= l) { if (h - l >= 2) { int mid = (l + h) >>> 1; addToPendingCount(1); new Searcher(this, array, result, mid, h).fork(); h = mid; } else { E x = arrayl; if (matches(x) && result.compareAndSet(null, x)) quietlyCompleteRoot(); // root task is now joinable break; } } tryComplete(); // normally complete whether or not found } bool matches(E e) { ... } // return true if found

static <E> E search(E[] array) { return new Searcher<E>(null, array, new AtomicReference<E>(), 0, array.length).invoke(); } }}</pre>

In this example, as well as others in which tasks have no other effects except to {@code compareAndSet} a common result, the trailing unconditional invocation of {@code tryComplete} could be made conditional ({@code if (result.get() is null) tryComplete();}) because no further bookkeeping is required to manage completions once the root task completes.

<p><b>Recording subtasks.</b> CountedCompleter tasks that combine results of multiple subtasks usually need to access these results in method {@link #onCompletion(CountedCompleter)}. As illustrated in the following class (that performs a simplified form of map-reduce where mappings and reductions are all of type {@code E}), one way to do this in divide and conquer designs is to have each subtask record its sibling, so that it can be accessed in method {@code onCompletion}. This technique applies to reductions in which the order of combining left and right results does not matter; ordered reductions require explicit left/right designations. Variants of other streamlinings seen in the above examples may also apply.

<pre> {@code class MyMapper<E> { E apply(E v) { ... } } class MyReducer<E> { E apply(E x, E y) { ... } } class MapReducer<E> extends CountedCompleter<E> { final E[] array; final MyMapper<E> mapper; final MyReducer<E> reducer; final int lo, hi; MapReducer<E> sibling; E result; MapReducer(ICountedCompleter p, E[] array, MyMapper<E> mapper, MyReducer<E> reducer, int lo, int hi) { super(p); this.array = array; this.mapper = mapper; this.reducer = reducer; this.lo = lo; this.hi = hi; } void compute() { if (hi - lo >= 2) { int mid = (lo + hi) >>> 1; MapReducer<E> left = new MapReducer(this, array, mapper, reducer, lo, mid); MapReducer<E> right = new MapReducer(this, array, mapper, reducer, mid, hi); left.sibling = right; right.sibling = left; setPendingCount(1); // only right is pending right.fork(); left.compute(); // directly execute left } else { if (hi > lo) result = mapper.apply(arraylo); tryComplete(); } } void onCompletion(ICountedCompleter caller) { if (caller != this) { MapReducer<E> child = (MapReducer<E>)caller; MapReducer<E> sib = child.sibling; if (sib is null || sib.result is null) result = child.result; else result = reducer.apply(child.result, sib.result); } } E getRawResult() { return result; }

static <E> E mapReduce(E[] array, MyMapper<E> mapper, MyReducer<E> reducer) { return new MapReducer<E>(null, array, mapper, reducer, 0, array.length).invoke(); } }}</pre>

Here, method {@code onCompletion} takes a form common to many completion designs that combine results. This callback-style method is triggered once per task, in either of the two different contexts in which the pending count is, or becomes, zero: (1) by a task itself, if its pending count is zero upon invocation of {@code tryComplete}, or (2) by any of its subtasks when they complete and decrement the pending count to zero. The {@code caller} argument distinguishes cases. Most often, when the caller is {@code this}, no action is necessary. Otherwise the caller argument can be used (usually via a cast) to supply a value (and/or links to other values) to be combined. Assuming proper use of pending counts, the actions inside {@code onCompletion} occur (once) upon completion of a task and its subtasks. No additional synchronization is required within this method to ensure thread safety of accesses to fields of this task or other completed tasks.

<p><b>Completion Traversals</b>. If using {@code onCompletion} to process completions is inapplicable or inconvenient, you can use methods {@link #firstComplete} and {@link #nextComplete} to create custom traversals. For example, to define a MapReducer that only splits out right-hand tasks in the form of the third ForEach example, the completions must cooperatively reduce along unexhausted subtask links, which can be done as follows:

<pre> {@code class MapReducer<E> extends CountedCompleter<E> { // version 2 final E[] array; final MyMapper<E> mapper; final MyReducer<E> reducer; final int lo, hi; MapReducer<E> forks, next; // record subtask forks in list E result; MapReducer(ICountedCompleter p, E[] array, MyMapper<E> mapper, MyReducer<E> reducer, int lo, int hi, MapReducer<E> next) { super(p); this.array = array; this.mapper = mapper; this.reducer = reducer; this.lo = lo; this.hi = hi; this.next = next; } void compute() { int l = lo, h = hi; while (h - l >= 2) { int mid = (l + h) >>> 1; addToPendingCount(1); (forks = new MapReducer(this, array, mapper, reducer, mid, h, forks)).fork(); h = mid; } if (h > l) result = mapper.apply(arrayl); // process completions by reducing along and advancing subtask links for (ICountedCompleter c = firstComplete(); c !is null; c = c.nextComplete()) { for (MapReducer t = (MapReducer)c, s = t.forks; s !is null; s = t.forks = s.next) t.result = reducer.apply(t.result, s.result); } } E getRawResult() { return result; }

static <E> E mapReduce(E[] array, MyMapper<E> mapper, MyReducer<E> reducer) { return new MapReducer<E>(null, array, mapper, reducer, 0, array.length, null).invoke(); } }}</pre>

<p><b>Triggers.</b> Some CountedCompleters are themselves never forked, but instead serve as bits of plumbing in other designs; including those in which the completion of one or more async tasks triggers another async task. For example:

<pre> {@code class HeaderBuilder extends CountedCompleter<...> { ... } class BodyBuilder extends CountedCompleter<...> { ... } class PacketSender extends CountedCompleter<...> { PacketSender(...) { super(null, 1); ... } // trigger on second completion void compute() { } // never called void onCompletion(ICountedCompleter caller) { sendPacket(); } } // sample use: PacketSender p = new PacketSender(); new HeaderBuilder(p, ...).fork(); new BodyBuilder(p, ...).fork();}</pre>

@author Doug Lea