[
https://issues.apache.org/jira/browse/RNG-160?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17397988#comment-17397988
]
Alex Herbert commented on RNG-160:
----------------------------------
Note:
The original testing on JDK 11 for the existing Marsaglia Gaussian sampler vs
the new McFarland Gaussian sampler used JDK version 11.0.6 (see RNG-151 comment
17375454). The same performance gap observed in JDK 8 was present in JDK 11.
The newer version of JDK 11 used here has closed this performance gap such that
the Marsaglia sampler is faster than the unchanged McFarland sampler. Updating
the McFarland sampler to allow method inlining makes the two samplers
comparable on the latest LTS JDK, older JDKs will be faster with the McFarland
sampler.
> Performance of modified Ziggurat samplers
> -----------------------------------------
>
> Key: RNG-160
> URL: https://issues.apache.org/jira/browse/RNG-160
> Project: Commons RNG
> Issue Type: Improvement
> Components: sampling
> Affects Versions: 1.4
> Reporter: Alex Herbert
> Priority: Minor
> Fix For: 1.4
>
> Attachments: Exp.jpg, Gaussian.jpg
>
>
> The modified ziggurat algorithm implemented in RNG-151 has a variation for
> sampling from the overhang regions at the ziggurat edges, region A below:
> {noformat}
> \
> ----------+\
> | \
> B |A \
> -------------+\
> | \
> {noformat}
> Note: Region B is the ziggurat layer which is entirely within the
> distribution PDF.
> These overhangs can be convex, concave or an inflection (change from convex
> to concave). Samples from the region A must be below the distribution density
> curve (PDF). When sampling from region A, the sampler will create a random
> point (x,y) uniformly within the rectangle. This is tested to determine if it
> is below the curve. Convex and concave curves can use a fast method that
> knows the largest possible triangle that fits above or below the curve. If a
> sampled point (x,y) is within these triangles then the actual curve position
> (pdf( x)) for the point x does not need to be computed. This can save time if
> the pdf is computationally expensive. In the case of exponential (exp(-x)) or
> Gaussian (exp(-0.5 * x * x)) this involves a call to Math.exp.
> The current sampler implements the fast look-up method. The alternative is to
> always compute pdf( x) and determine if y is below the curve. This is known
> as 'simple overhangs'.
> Investigate the use of simple overhangs on the performance of the sampler.
> Note: The Marsaglia version of the ziggurat sampler uses the 'simple
> overhangs' method.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
