Some Experimental Observations on the Behavior of Composite Random Number Generators

We describe a series of experiments with composite random number generators using shuffling tables. The research explores the issues of improvement in statistical behavior claimed for composite generators and the overhead incurred by shuffling. On the basis of the limited experimental work, three observations are warranted: (1) Shuffling does not effect significant improvement in the statistical behavior of good simple generators implemented on large-word-length machines (with large periods). (2) Shuffling can significantly improve the behavior of generators on small-word-length machines, at least partly by increasing the maximum period obtained. (3) A table size of two leads to results comparable with larger tables and incurs less overhead.