6.1.0.Beta1 will support alternative Pseudo Random Number Generator implementations
(such as Mersenne Twister, Isaac and Well), which you can easily change with a single line in the solverConfig XML.
These alternative Random implementations have a more uniform distribution than the JDK and most of them do this at no noticeable performance cost.
Below is the result of the initial benchmark over 6 use cases over 11 datasets (about 2 datasets each), over the out of the box supported random implementations. The winners are green, the losers are red. For each run, the score difference with the winner is shown, follow by it's ranking number.
Conclusion: Generally speaking, the choice of the Random implementation doesn't matter much, the JDK random remains a good default. But for some use cases (machine reassignment), I still believe it does matter and I 'll (dis)prove that with a benchmark run tonight.
For more information, see https://issues.jboss.org/browse/PLANNER-210