I performed a similar test yesterday: https://www.reddit.com/r/lolphp/comments/46fxi8/typofixing_commit_in_mersenne_twister_rng_code_is/d05zwgg
There is some discussion happening on the PR: https://github.com/php/php-src/pull/1681 On Sat, 20 Feb 2016 at 04:04 Yasuo Ohgaki <yohg...@ohgaki.net> wrote: > Hi all, > > On Sat, Feb 20, 2016 at 10:39 AM, Andrea Faulds <a...@ajf.me> wrote: > > > > Yasuo Ohgaki wrote: > >> > >> This should be discussed in internals. > >> MT rand being not MT rand is unacceptable. > >> This fix must included in released versions. IMHO. > >> > >> In any case we should provide broken MT rand, we must use INI switch > >> or other options. > > > > > > PHP's implementation of the Mersenne Twister algorithm is buggy, so it > > doesn't produce the same output as in other languages. But the buggy > > algorithm produces sufficiently random sequences of apparently the same > > quality as the proper algorithm. So we *could* simply consider this as a > > documentation issue if we wanted to. I'm not saying that's the right > course > > of action, though. > > It seems good enough for quick analysis via ent. > Note: Ent tests per "byte", not php_uint32. > http://www.fourmilab.ch/random/ > > WITH FIX > [yohgaki@dev rand]$ ~/git/oss/php.net/php-src/php-bin -r '$i = > 1000000;while($i--) echo pack("I", mt_rand());' | ./ent > Entropy = 7.954269 bits per byte. > > Optimum compression would reduce the size > of this 4000000 byte file by 0 percent. > > Chi square distribution for 4000000 samples is 250901.85, and randomly > would exceed this value less than 0.01 percent of the times. > > Arithmetic mean value of data bytes is 111.4725 (127.5 = random). > Monte Carlo value for Pi is 3.486975487 (error 10.99 percent). > Serial correlation coefficient is -0.049294 (totally uncorrelated = 0.0). > > > WITHOUT FIX > [yohgaki@dev rand]$ php -r '$i = 1000000;while($i--) echo pack("I", > mt_rand());' | ./ent > Entropy = 7.954382 bits per byte. > > Optimum compression would reduce the size > of this 4000000 byte file by 0 percent. > > Chi square distribution for 4000000 samples is 250280.94, and randomly > would exceed this value less than 0.01 percent of the times. > > Arithmetic mean value of data bytes is 111.5000 (127.5 = random). > Monte Carlo value for Pi is 3.483663484 (error 10.89 percent). > Serial correlation coefficient is -0.049192 (totally uncorrelated = 0.0). > > > I'm not sure if this result is good enough to say "It's ok". > MT rand has very long cycle and this quick test does not prove it. > > Did anyone do better analysis? > Otherwise, I'm uncomfortable with basic algorithm change. > > Regards, > > -- > Yasuo Ohgaki > yohg...@ohgaki.net > > -- > PHP Internals - PHP Runtime Development Mailing List > To unsubscribe, visit: http://www.php.net/unsub.php > >
