id	summary	reporter	owner	description	type	status	milestone	component	version	severity	resolution	keywords	cc
6135	Significant speed-up in gamma rng by switching from Knuth method to Marsaglia and Tsang (2000)	Jan Drugowitsch <jdrugo@…>	No-Maintainer	"When compared to the GNU Scientific Library (GSL), the generation of Gamma-distributed random numbers is very slow in Boost. A simple test, generating 100000000 random numbers, reveals

GSL   : 22.3233 MSamples/s[[BR]]
Boost : 5.41275 MSamples/s

where MSamples/s is million samples per second. The difference is due to Boost using the Knuth method, whereas the GSL implements a newer method by Marsaglia and Tsang (2000).

With the attached patch, which implements the newer method, the performance improves to

GSL   : 22.7472 MSamples/s[[BR]]
Boost : 14.5018 MSamples/s

The remaining difference between Boost and GSL can be traced back to GSL using Zigguraut rather than Box-Muller to generate Normal random numbers. At some point I might submit a patch for normal_distribution.hpp that implements the Zigguraut algorithm.

The attached patch exactly follows gsl_ran_gamma of GSL v1.15 (gsl-1.15/randdist/gamma.c)"	Patches	new	To Be Determined	random	Boost Development Trunk	Optimization		gamma, rng