Comments on: Parallel Random Number Generation using TRNG

By: Helmingstay

Helmingstay — Fri, 17 Oct 2014 20:42:01 +0000

Thanks, nice writeup.
Here’s the best theoretical discussion of threads and rng streams, with a clear discussion of problems and justification for TRNG’s methodology, FWIW.
http://arxiv.org/abs/0905.4238

By: Matteo Fasiolo

Matteo Fasiolo — Wed, 07 May 2014 15:21:19 +0000

Actually using one seed per thread is what I generally do too… but sometimes I do wonder about the hidden risks!

By: admin

admin — Wed, 07 May 2014 14:54:30 +0000

Hey Matteo, it would be nice to have a package for parallel random number generation for sure, but it might be possible to achieve this with the C++11 pseudo random number generators instead of TRNG (I wrote this post before I was aware of C++11 PRNG). A lot of people would be happy to have one C++11 PRNG object per thread and seed them differently (i.e. seed according to thread id), this is what I always used to do, but the purists out there I think would say that seeding according to thread id is simply not enough as it may still be possible to get overlapping streams. Having said that, I think in practice the chances of getting an overlapping stream are rare, so you might be willing to accept that risk. I know enough to be cautious, but I don’t really know enough how best to proceed.

By: Matteo Fasiolo

Matteo Fasiolo — Tue, 06 May 2014 21:52:14 +0000

Hi Michael, many thanks for sharing this! I wonder how difficult would it be to build a R package that makes all the distributions in TRNG available at R level, by using Rcpp. I guess that would be quite useful, as I don’t know about any package that provides such functionality.