Java 5 has introduced support for asynchronous task execution by a thread pool in the form of the Executor framework, whose heart is the thread pool implemented by java.util.concurrent.ThreadPoolExecutor. Java 7 has added an alternative thread pool in the form of java.util.concurrent.ForkJoinPool.
Looking at their respective API, ForkJoinPool provides a superset of ThreadPoolExecutor’s functionality in standard scenarios (though strictly speaking ThreadPoolExecutor offers more opportunities for tuning than ForkJoinPool). Adding to this the observation that
fork/join tasks seem to be faster (possibly due to the work stealing scheduler), need definitely fewer threads (due to the non-blocking join operation), one might get the impression that ThreadPoolExecutor has been superseded by ForkJoinPool.
But is this really correct? All the material I have read seems to sum up to a rather vague distinction between the two types of thread pools:
- ForkJoinPool is for many, dependent, task-generated, short, hardly ever blocking (i.e. compute-intensive) tasks
- ThreadPoolExecutor is for few, independent, externally-generated, long, sometimes blocking tasks
Is this distinction correct at all? Can we say anything more specific about this?
ThreadPool (TP) and ForkJoinPool (FJ) are targeted towards different use cases. The main difference is in the number of queues employed by the different executors which decide what type of problems are better suited to either executor.
The FJ executor has n (aka parallelism level) separate concurrent queues (deques) while the TP executor has only one concurrent queue (these queues/deques maybe custom implementations not following the JDK Collections API). As a result, in scenarios where you have a large number of (usually relatively short running) tasks generated, the FJ executor will perform better as the independent queues will minimize concurrent operations and infrequent steals will help with load balancing. In TP due to the single queue, there will be concurrent operations every time work is dequeued and it will act as a relative bottleneck and limit performance.
In contrast, if there are relatively fewer long-running tasks the single queue in TP is no longer a bottleneck for performance. However, the n-independent queues and relatively frequent work-stealing attempts will now become a bottleneck in FJ as there can be possibly many futile attempts to steal work which add to overhead.
In addition, the work-stealing algorithm in FJ assumes that (older) tasks stolen from the deque will produce enough parallel tasks to reduce the number of steals. E.g. in quicksort or mergesort where older tasks equate to larger arrays, these tasks will generate more tasks and keep the queue non-empty and reduce the number of overall steals. If this is not the case in a given application then the frequent steal attempts again become a bottleneck. This is also noted in the javadoc for ForkJoinPool: