Looking for a faster RNG

Questions / Help
17

I realize the BIF option may not be preferred, but out of curiosity I put together a proof-of-concept similar to unique_integer where the random state is stored on each scheduler’s ErtsSchedulerData. There’s also a global state which jumps for each scheduler’s initial state so each scheduler has its own non-overlapping subsequence.

It uses the Xoroshiro116+ code in rand.erl as well as SplitMix64 for establishing the initial global state on boot.

Here’s the code (based on master at 86fc3db): WIP - erlang:random_integer/0,1 BIF · potatosalad/otp@e3427a8 · GitHub

Single process test:

erlperf -c 1 'erlang:random_integer(1000).' 'erlang:phash2(erlang:unique_integer(), 1000).' 'rand:uniform(1000).'
Code                                                  ||        QPS     Rel
erlang:random_integer(1000).                           1   33956 Ki    100%
erlang:phash2(erlang:unique_integer(), 1000).          1   22099 Ki     65%
rand:uniform(1000).                                    1   10102 Ki     29%

64 process test (matches the number of online schedulers):

erlperf -c 64 'erlang:random_integer(1000).' 'erlang:phash2(erlang:unique_integer(), 1000).' 'rand:uniform(1000).'
Code                                                  ||        QPS     Rel
erlang:random_integer(1000).                          64   37147 Ki    100%
erlang:phash2(erlang:unique_integer(), 1000).         64   27644 Ki     74%
rand:uniform(1000).                                   64   14981 Ki     40%
5 Likes