[
https://issues.apache.org/jira/browse/RNG-50?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16565872#comment-16565872
]
Alex D Herbert commented on RNG-50:
-----------------------------------
I don't see why you have to precompute everything. The cache helps in this
simple example. I've taken the liberty to change the name and have it be
constructed by the PoissonSampler as a suggest API:
{code:java}
PoissonSamplerCache cache = PoissonSampler.createCache(0, 256);
UniformRandomProvider rng = RandomSource.create(RandomSource.SPLIT_MIX_64);
for (int i=0; i<N; i++) {
doSomethingWith(new PoissonSampler(rng, rng.nextDouble() * 256,
cache).sample());
}
{code}
If N=10000000 then pre-computation is fine.
However if N=10 then pre-computation of the cache would be a waste.
With a lazy initialised cache you don't have to know the loop size and make a
decision on whether it should be used. With a cache that has non-synchronised
lazy initialisation then you can even use it across threads with the caveat
that some initialisation may be repeated until the JVM writes local CPU memory
back to main memory. Or just use the {{AtomicReferenceArray}} as above.
So for my use case I would have an idea of the range of values I'd expect for
the Poisson mean. I could then use the cache system to improve performance.
{quote}What I had in mind is a global cache
{quote}
Q. How global is global? Across the simulation and then destroyed (as above)?
In this case any user of the library could keep hold of the cache for use again
if they want so I think this is fine.
> PoissonSampler single use speed improvements
> --------------------------------------------
>
> Key: RNG-50
> URL: https://issues.apache.org/jira/browse/RNG-50
> Project: Commons RNG
> Issue Type: Improvement
> Affects Versions: 1.0
> Reporter: Alex D Herbert
> Priority: Minor
> Attachments: PoissonSamplerTest.java, jmh-result.csv
>
>
> The Sampler architecture of {{org.apache.commons.rng.sampling.distribution}}
> is nicely written for fast sampling of small dataset sizes. The constructors
> for the samplers do not check the input parameters are valid for the
> respective distributions (in contrast to the old
> {{org.apache.commons.math3.random.distribution}} classes). I assume this is a
> design choice for speed. Thus most of the samplers can be used within a loop
> to sample just one value with very little overhead.
> The {{PoissonSampler}} precomputes log factorial numbers upon construction if
> the mean is above 40. This is done using the {{InternalUtils.FactorialLog}}
> class. As of version 1.0 this internal class is currently only used in the
> {{PoissonSampler}}.
> The cache size is limited to 2*PIVOT (where PIVOT=40). But it creates and
> precomputes the cache every time a PoissonSampler is constructed if the mean
> is above the PIVOT value.
> Why not create this once in a static block for the PoissonSampler?
> {code:java}
> /** {@code log(n!)}. */
> private static final FactorialLog factorialLog;
>
> static
> {
> factorialLog = FactorialLog.create().withCache((int) (2 *
> PoissonSampler.PIVOT));
> }
> {code}
> This will make the construction cost of a new {{PoissonSampler}} negligible.
> If the table is computed dynamically as a static construction method then the
> overhead will be in the first use. Thus the following call will be much
> faster:
> {code:java}
> UniformRandomProvider rng = ...;
> int value = new PoissonSampler(rng, 50).sample();
> {code}
> I have tested this modification (see attached file) and the results are:
> {noformat}
> Mean 40 Single construction ( 7330792) vs Loop construction
> (24334724) (3.319522.2x faster)
> Mean 40 Single construction ( 7330792) vs Loop construction with static
> FactorialLog ( 7990656) (1.090013.2x faster)
> Mean 50 Single construction ( 6390303) vs Loop construction
> (19389026) (3.034132.2x faster)
> Mean 50 Single construction ( 6390303) vs Loop construction with static
> FactorialLog ( 6146556) (0.961857.2x faster)
> Mean 60 Single construction ( 6041165) vs Loop construction
> (21337678) (3.532047.2x faster)
> Mean 60 Single construction ( 6041165) vs Loop construction with static
> FactorialLog ( 5329129) (0.882136.2x faster)
> Mean 70 Single construction ( 6064003) vs Loop construction
> (23963516) (3.951765.2x faster)
> Mean 70 Single construction ( 6064003) vs Loop construction with static
> FactorialLog ( 5306081) (0.875013.2x faster)
> Mean 80 Single construction ( 6064772) vs Loop construction
> (26381365) (4.349935.2x faster)
> Mean 80 Single construction ( 6064772) vs Loop construction with static
> FactorialLog ( 6341274) (1.045591.2x faster)
> {noformat}
> Thus the speed improvements would be approximately 3-4 fold for single use
> Poisson sampling.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
