Le 20/09/2010 15:28, Mikkel Meyer Andersen a écrit : > I would like to see/try a comparison between the default and new PRNG > under different conditions. Just to get an idea of how they perform.
I gave it a try. The version from yesterday was really slow (about 3 times slower than other generators). It seems the JVM optimizer (at least on my computer) was not able to inline everything. So I inlined myself and also tried several different tricks to avoid modulo operations in indices. I ended up with a solution different from the one of the authors: I used precomputed indices arrays. This increase data size by a factor 6 but really speeds up generation. Here are my results, using a 64 bits JVM on a linux computer with a 2 years old processor (AMD phenom Quad 9550 at 2.2 GHz): generation rate for MersenneTwister: about 27 millions doubles per second (remember I generate two 32 bits integers for each double). generation rates for other PRNG, relative to MersenneTwister: MersenneTwister 1.0 JDKRandomGenerator 1.16 Well512a 0.85 Well1024a 0.63 Well19937a 0.71 Well19937c 0.57 Well44497a 0.69 Well44497b 0.65 So JDK is 16% faster than MersenneTwister (but very poor) and the new generators are between 15% slower and 43% slower. The order of magnitude are the same. I could perhaps optimize a little further for example by generating several numbers and caching them but am not sure this is worth the effort. I have sent a mail to the paper authors with our questions. Luc --------------------------------------------------------------------- To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org For additional commands, e-mail: dev-h...@commons.apache.org
