What is the correct shape of a RecursiveTask?

protected Long compute() {
    if (hi - lo <= THRESHOLD) {
        return computeDirectly();
    }
    int mid = (lo + hi) >>> 1;
    SumTask left = new SumTask(array, lo, mid);
    left.fork();                       // hand one half to the pool
    long rightResult = new SumTask(array, mid, hi).compute();   // do the other here
    return rightResult + left.join();  // then collect
}

Forking both halves and joining both wastes the current thread, which sits idle while two other threads work. The order matters too: fork, then compute, then join. Joining before computing the second half serialises the whole thing.

The threshold is the other half of the answer. Too small and the splitting overhead dominates; the usual starting point is a few thousand elements of simple work, tuned by measurement.