sampling distributions

[ewitwer]

Right now, every time a Dist is sampled, a new rand_distr distribution is created and subsequently discarded. Since sampling distributions is a major factor in Maybenot's runtime, it may be desirable to change this.

With one exception, all distributions currently included in the framework have a constant-time new function - some simply check parameter validity and return an instance of the relevant struct (e.g. Normal), and others require some additional logic (e.g. Beta and Poisson). The Geometric distribution has a new function whose runtime depends on the shape parameter, but only a few loop iterations should occur, even with very small probabilities.

From the above, there may be minimal benefit to storing a persistent rand_distr distribution in each Dist, also because this would make the Dist type's size dynamic. Benchmark tests with a prototype implementation may shed more light on this.