Improve CPU utilization for long running fitness functions

[jenetics]

Jenetics is calculating the fitness values for each generation concurrently. The main assumption is, that the calculation of the fitness function is relatively cheap and the calculation costs are constant for all genotypes. With this assumption in mind, the population is split into sub-populations, which group the fitness function into bigger evaluation units. This should reduce the evaluation overhead. If we have to deal with long-running fitness functions, or with fitness function with unpredictable computation costs, the existing computation strategy no longer works as expected. We get idle CPU cores.

This ticket is about to improve the CPU utilization for long-running fitness functions and for fitness functions with non-constant evaluation costs.

See
#196

Hint
Consider usage of getSurplusQueuedTaskCount() method.

@Override
protected void compute() {
	if ((_high - _low) <= 1 || ForkJoinTask.getSurplusQueuedTaskCount() > 3) {
	//if (_high - _low < threshold) {
		for (int i = _low; i < _high; ++i) {
			_runnables.get(i).run();
		}
	} else {
		final int mid = (_low + _high) >>> 1;
		invokeAll(
			new RunnablesAction(_runnables, _low, mid, threshold),
			new RunnablesAction(_runnables, mid, _high, threshold)
		);
	}
}

[dyorgio]

Hi, where can I get a preview? Do you will create a new branch?
I can test for you :)

[jenetics]

Merged into r3.8.0 branch.