package make-random

Xoshiro

This library includes OCaml implementation of some pseudorandom number generators (PRNGs) designed by David Blackman and Sebastiano Vigna behind an interface that mimmics that of the Random module of the standard library.

The Xoshiro generators (for XOR/shift/rotate) are all-purpose generators (not cryptographically secure). Compared to the standard library, they:

• have a bigger state space: xoshiro256++/xoshiro256** generators have a period of 2²⁵⁶-1.
• pass more tests: xoshiro256++/xoshiro256** pass the whole BigCrush test suite while the Random module of the standard library systematically fails some of the tests.

Usage

Drop-in Replacement

The modules in this library are drop-in replacements of the Random module of the standard library. This means you can use Xoshiro everywhere where you would use Random. For instance:

• use Xoshiro.bits instead of Random.bits
• (same for int, bool, etc. and also for the State submodule)
• use open Xoshiro instead of open Random
• or even write module Random = Xoshiro at the beginning of every file.

Bindings vs. Pure

The library comes in two version: one using C bindings and the other written in pure OCaml. The C bindings are here for performances. They are usually around twice faster as their pure counterpart. The pure implementations come from applications which require it, eg. for programs that compile to JavaScript. The interface is the same, the only thing that changes is whether you depend on the library xoshiro.bindings or xoshiro.pure. By default, xoshiro depends on xoshiro.bindings.

Example

For instance, say you have an executable crazyrandom which uses the standard Random module as source of randomness. You can easily switch to using Xoshiro instead by replacing all occurrences of Random in crazyrandom.ml by Xoshiro (a simple search and replace should do the trick). You can then compile your executable with a Dune file similar to:

(executable
 (name crazyrandom)
 (libraries xoshiro))

Now you realise that you also need to compile for JavaScript. Sadly, C bindings will not work there. You can however keep the same file for crazyrandom.ml and change your Dune file to:

(executable
 (name crazyrandom)
 (libraries xoshiro.pure)
 (modes js))

You're done!

Installation

We recommend installing via OPAM with opam install xoshiro. Otherwise, make followed by make install should do the trick, provided you have the required dependencies.

Documentation

Documentation is available online. It can also be built locally with make doc and by pointing a browser to doc/index.html.

Tests & Benchmarks

The library comes with a set of tests and benchmarks. Tests are ran at every push to the repository. The tests include:

• checking that bindings and pure implementation yield the same bits;
• checking that implementations comply with our claims with respect to the *Crush test batteries (note that, because of time considerations, only SmallCrush is ran in continuous integration).
• checking that the MakeRandom functors build the same interface as the standard library.

It is easy to run other test batteries on generators. For instance, one can run BigCrush on xoshiro256++ by calling:

dune exec xoshiro256plusplus/test/crusher/crusher.exe -- --bigcrush --verbose

The benchmarks allow to compare various PRNGs from this library against each other and, more importantly, against the standard library. They can be ran using dune exec bench/run.exe. For xoshiro256++, we observe that the bindings are slightly slower than the standard library, and almost twice faster than the pure implementations.

Related Word

• Xavier Leroy's PRINGO, “pseudo-random number generators that support splitting and two interfaces: one stateful, one purely functional”. Xavier Leroy's work is very similar to ours: it provides new PRNGs for OCaml (including Split mix which we can also find in the package xoshiro) in an interface similar to that of the standard library (although not exactly the same, contrary to what can be obtained with make-random).
• Mirage's Crypto Library, including in particular the strong, cryptographically secure, PRNG Fortuna. This is further from our work as usage is really not the same.