The Mersenne Twister in Go
An implementation of Takuji Nishimura's and Makoto Matsumoto's Mersenne Twister pseudo random number generator in Go.
Copyright (C) 2013 Jochen Voss
This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
The Mersenne Twister is a pseudo random number generator (PRNG), developed by Takuji Nishimura and Makoto Matsumoto. The Mersenne Twister is, for example, commonly used in Monte Carlo simulations and is the default random number generator for many programming languages, e.g. Python and R. This package implements the 64bit version of the algorithm.
This package can be installed using the
go get command:
go get github.com/seehuhn/mt19937
Detailed usage instructions are available via the package's online help, either on godoc.org or on the command line:
go doc github.com/seehuhn/mt19937
MT19937 represents instances of the Mersenne Twister.
New instances can be allocated using the
A seed can be set using the .Seed() or .SeedFromSlice() methods.
MT19937 implements the
rand.Source interface from the
math/rand package. Typically the PRNG is wrapped in a rand.Rand
object as in the following example:
rng := rand.New(mt19937.New()) rng.Seed(time.Now().UnixNano())
Comparison to the Go Default PRNG
Go has a built-in PRNG provided by the math/rand package. I did not find any information about this built-in PRNG except for a comment in the source code which says "algorithm by DP Mitchell and JA Reeds". In contrast, the MT19737 generator provided in this package is a well-understood random number generator. Relevant references include [Ni2000] and [MatNi1998].
|[Ni2000]||T. Nishimura, Tables of 64-bit Mersenne Twisters, ACM Transactions on Modeling and Computer Simulation 10, 2000, pages 348-357.|
|[MatNi1998]||M. Matsumoto and T. Nishimura, Mersenne Twister: a 623-dimensionally equidistributed uniform pseudorandom number generator, ACM Transactions on Modeling and Computer Simulation 8, 1998, pages 3--30.|
The unit tests for the mt19937 package verify that the output of the Go implementation coincides with the output of the references implementation.
The mt19937 generator is slightly slower than the Go default PRNG. A speed comparison can be performed using the following command:
go test -bench=. github.com/seehuhn/mt19937
On my (64bit) system I get the following results:
method time per call thoughput MT19937.Uint64 14.9 ns/op 537.48 MB/s MT19937.Int63 15.0 ns/op 533.54 MB/s builtin Int63 11.9 ns/op 674.07 MB/s
This shows that, on my system, a call to the
Int63() method of the
built-in PRNG takes about 80% of the time that MT19937.Int63() takes.