Just so we are clear, Random123 is not a library of ours. Read their website.
Use mt19937 even with the same seed cannot guarantee you the same results as your old simulation. I don't know what do mean by "same". If you mean exactly the same results, you will never get it. Reordering floating points, change of language and library will certainly change the results. If you mean asymptotically the same results. What matters is the quality of RNG, there is no reason you need to set the seed the same. Setting seed to reproduce the results is only possible when using exactly the same programming environment Yan Zhou On Feb 10, 2013, at 8:50 AM, Simon Zehnder <[email protected]> wrote: > Dear Yan, dear Dirk, > > thank you for these very precise answers! I think at the beginning I'll fall > back to the R RNG. I will test if it is faster to create an Armadillo > vector/matrix and fill it via R::rnorm, or if it is faster to use Rcpp sugar > and change the Rcpp::NumericVector to an arma::vec via arma::conv_to. > > Btw Dirk, this is a nice overview given at the Rcpp gallery for timing random > number generators. > > Your HPC library Random123 looks pretty interesting to me. I am very into HPC > and use openMP and OpenMPI a lot. In this case however, I have to use > specifically a Mersenne Twister RNG, as I want to compare results from my > older simulation package, using Scythe Statistical Library, and my newer one > using RcppArmadillo. As Scythe uses MT, I want to use the same RNG and same > seed to have results comparable to each other. > > Later on, when I know all results are plausible, I will concentrate on speed. > > Best > > > Simon > > > > On Feb 9, 2013, at 5:42 PM, Yan Zhou <[email protected]> wrote: > >>> >>> This really is a BIG topic and worth a few more comments. Note that I wrote >>> a >>> few related posts on RNGs at the Rcpp Gallery, see for example >>> >>> http://gallery.rcpp.org/articles/timing-normal-rngs/ >>> >>> which compares the RNGs from R, C++11 and Boost. Simon just added Armadillo >>> to >>> the list, we can add even more RNGs fromn other packages. >> >> If it is of interest to anyone, I once timed Boost, C++11 and Random123 (A >> high performance parallel RNG, >> http://www.thesalmons.org/john/random123/releases/latest/docs/, It come with >> a C++11 compatible RNG engine, can be used just like std::mt19937) once for >> different compilers on Linux. I just uploaded them to >> https://github.com/zhouyan/vSMC/wiki/RNG-performance-comparison >> >> There are two benchmark, one for the performance of URNG (mt19937 etc). >> These include those in C++11 <random> and Boost.Random, which are almost >> identical in functionality (C++11 <random> is based on Boost.Random after >> all). Also they include two URNG from Random123, threefry and philox (both >> come with four basic configurations) >> >> Another benchmark is the performance of generating distribution random >> numbers (such as normal). The Random123 threefry2x64 was used for all >> distribution and compilers, since it is the one with least performance >> difference between compilers. >> >> Compilers include, >> clang SVN with libstdc++ 4.7 >> clang SVN with libc++ SVN >> gcc 4.7 >> intel icpc 13 >> >> clang and gcc version also come with results when using AMD libm instead of >> glibc. However, the benchmark are not Rcpp specific. They are compiled to >> standalone C++ programs. But all these URNGs can be used Rcpp. As >> demonstrated in Dirk's example. >> >> Best, >> >> Yan Zhou >
_______________________________________________ Rcpp-devel mailing list [email protected] https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
