[
https://issues.apache.org/jira/browse/MATH-1154?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14162024#comment-14162024
]
Phil Steitz commented on MATH-1154:
-----------------------------------
I agree that the root issue really is MATH-1124. When we decided to move
sampling into the distributions we created the need for distribution instances
to have access to a PRNG. When we decided we wanted everything to be final we
forced ourselves into the MATH-1124 state, where the only way to avoid
potentially expensive PRNG initialization when creating distribution instances
that may never use sampling is the smelly workaround to null out the (final)
PRNG at instance construction time. Thomas' patch looks fine to me and unless
and until we change one of the decisions above (relax final obsession or pull
sampling back out), we should use the workaround in the unit tests (as the
patch does) and try to reduce the initialization cost of the default or find a
better workaround (reopening MATH-1124). The OPs patch adds complexity, IMO,
without really addressing the core problem, which I think we should address in
MATH-1124. So I am +1 to applying Thomas' patch, resolving this issue and
moving back to MATH-1124.
> Statistical tests in stat.inference package are very slow due to implicit
> RandomGenerator initialization
> --------------------------------------------------------------------------------------------------------
>
> Key: MATH-1154
> URL: https://issues.apache.org/jira/browse/MATH-1154
> Project: Commons Math
> Issue Type: Bug
> Affects Versions: 3.3
> Reporter: Otmar Ertl
> Attachments: MATH-1154.patch, math3.patch
>
>
> Some statistical tests defined in the stat.inference package (e.g.
> BinomialTest or ChiSquareTest) are unnecessarily very slow (up to a factor 20
> slower than necessary). The reason is the implicit slow initialization of a
> default (Well19937c) random generator instance each time a test is performed.
> The affected tests create some distribution instance in order to use some
> methods defined therein. However, they do not use any method for random
> generation. Nevertheless a random number generator instance is automatically
> created when creating a distribution instance, which is the reason for the
> serious slowdown. The problem is related to MATH-1124.
> There are following solutions:
> 1) Fix the affected statistical tests by passing a light-weight
> RandomGenerator implementation (or even null) to the constructor of the
> distribution.
> 2) Or use for all distributions a RandomGenerator implementation that uses
> lazy initialization to generate the Well19937c instance as late as possible.
> This would also solve MATH-1124.
> I will attach a patch proposal together with a performance test, that will
> demonstrate the speed up after a fix.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
