On 2/5/16 12:59 PM, Gilles wrote:
On Fri, 5 Feb 2016 06:50:10 -0700, Phil Steitz wrote:
On 2/4/16 3:59 PM, Gilles wrote:
Hi.
Here is a micro-benchmark report (performed with "PerfTestUtils"):
-----
nextInt() (calls per timed block: 2000000, timed blocks: 100, time
unit: ms)
name time/call std dev total time ratio
cv difference
o.a.c.m.r.JDKRandomGenerator 1.088e-05 2.8e-06 2.1761e+03 1.000
0.26 0.0000e+00
o.a.c.m.r.MersenneTwister 1.024e-05 1.5e-06 2.0471e+03 0.941
0.15 -1.2900e+02
o.a.c.m.r.Well512a 1.193e-05 4.4e-07 2.3864e+03 1.097
0.04 2.1032e+02
o.a.c.m.r.Well1024a 1.348e-05 1.9e-06 2.6955e+03 1.239
0.14 5.1945e+02
o.a.c.m.r.Well19937a 1.495e-05 2.1e-06 2.9906e+03 1.374
0.14 8.1451e+02
o.a.c.m.r.Well19937c 1.577e-05 8.8e-07 3.1542e+03 1.450
0.06 9.7816e+02
o.a.c.m.r.Well44497a 1.918e-05 1.4e-06 3.8363e+03 1.763
0.08 1.6602e+03
o.a.c.m.r.Well44497b 1.953e-05 2.8e-06 3.9062e+03 1.795
0.14 1.7301e+03
o.a.c.m.r.ISAACRandom 1.169e-05 1.9e-06 2.3375e+03 1.074
0.16 1.6139e+02
-----
where "cv" is the ratio of the 3rd to the 2nd column.
Questions are:
* How meaningful are micro-benchmarks when the timed operation has
a very
small duration (wrt e.g. the duration of other machine
instructions that
are required to perform them)?
It is harder to get good benchmarks for shorter duration
activities,
but not impossible. One thing that it would be good to do is to
compare these results with JMH [1].
I was expecting insights based on the benchmark which I did run.
You asked whether or not benchmarks are meaningful when the task
being benchmarked is short duration. I answered that question.
We have a tool in CM; if it's wrong, we should remove it.
How its results compare with JMH is an interesting question,
I will look into this.
I
agree, but I don't have time to make an analysis of benchmarking
tools (on top of what I've been doing since December because
totally innocuous changes in the RNG classes were frowned upon
out of baseless fear).
Please cut the hypberbole.
* In a given environment (HW, OS, JVM), is there a lower limit
(absolute
duration) below which anything will be deemed good enough?
That depends completely on the application.
Sorry, I thought that it was obvious: I don't speak of applications
that don't care about performance. :-)
For those that do, I do not agree with the statement: the question
relates to finding a point below which it is the environment that
overwhelms the other conditions.
A point where there will be _unavoidable_ overhead (transferring
data
from/to memory, JVM book-keeping, ...) and perturbations (context
switches, ...) such that their duration adds a constant time (on
average) that may render most enhancements to an already efficient
algorithm barely noticeable in practice.
Similarly, but in the opposite direction, some language constructs
or design choices might slow down things a bit, but without
endangering any user.
A problem arises when any enhancement to the design is deemed
harmful because it degrades a micro-benchmark, even though that
benchmark may not reflect any real use-cases.
Then, the real harm is against development.
* Can a library like CM admit a trade-off between ultimate
performance and
good design? IOW, is there an acceptable overhead in exchange
for other qualities
(clarity, non-redundancy, extensibility, etc.)?
That is too general a question to be meaningful. We need to look
at specific cases. What exactly are you proposing?
<rant>
It is quite meaningful even if it refers to general principles.
Those could (should, IMO) be taken into account when managing a
project like CM, on a par with "performance" (whose intrinsic value
is never questioned).
</rant>
Rant all you want. Vague generalities and hyperbole have no value.
Two specific cases are:
* inheritance vs delegation (a.k.a. composition)
* generics (that could require runtime casts)
This is getting closer to meaningful. Where exactly in the code are
you wanting to use something and seeing benchmark damage?
* Does ultimate performance for the base functionality (generation
of a
random number) trump any consideration of use-cases that would
need an
extension (of the base functionality, such as computation to
match another
distribution) that will unavoidably degrades the performance
(hence the
micro-benchmark will be completely misleading for those users)?
Again, this is vague and the answer depends on what exactly you are
talking about. Significantly damaging performance of PRNG
implementations is a bad idea,
Now, *this* is vague: what do you mean by "significantly"?
That was actually my question in the first place.
If you are talking about PRNG performance, I would say a 1% hit is
significant.
Referring to the
benchmark above, people who'd know why they require ultimate
performance
should be able to tell what range of numbers they'd find
acceptable in
that table.
<rant>
Actually my questions are very precise, but the answers would
require
some decent analysis, rather than the usual "bad idea" dismissal.
</rant>
In the Javadoc of the "random" package, there is information about
performance but no reference as to the benchmarking procedure.
It would be great to repeat these using JMH, which is emerging as a
de facto standard for java benchmarking. I will look into this.
I can consistently observe a totally different behaviour (using
"PerfTestUtils"):
1. "MersenneTwister" is *always* faster than all of the WELL RNGs;
2. moreover, the ratio *grows* with each of the longer periods
members of the WELL family (see the above table).
This makes me wonder how someone who purports to need "ultimate"
performance can have any objective basis to determine what is good
or bad for his own applications.
unless there are actual practical use
cases you can point to that whatever changes you are proposing
enable.
As I've explained in very much details in another thread, I've
reviewed (from a design POV) the RNG code in "random" and IMHO,
there
is room for improvement (cf. above for what I mean by that term).
<rant>
I have some code ready for review but I had to resort to what I
considered sub-optimal design (preemptively renouncing to propose a
"delegation"-based design) solely because of the destructive
community
process that takes place here.[1]
</rant>
More vague hyperbole that serves no purpose. Please focus on actual
code or design issues.
The practical use-cases is anything that needs further processing of
the numbers produced according to a uniform distribution:
Isn't that what the samplers in the distributions package do? What
we need from the PRNG implementations is just blocks of bits. Since
we wanted a pluggable replacement for j.u.Random, we added uniform
ints, longs and floats and gaussian floats. The samplers just need
uniform doubles. The practical use case we need is well-supported
in the code we have. What is missing, exactly?
I agree that
there would be little sense to code that latter part in a "pure" OO
way[2]. And Luc made it indeed quite efficient, I think, in the
various
concrete classes.
What I want to reconsider is how those concrete low-level
algorithms can
be plugged in a higher-level function that just requires a "source
of
randomness", as I'd call a provider of "int" (or "long") values,
where
the high level functionality does not care at all about the
provider's
inner working (a.o. how it's seeded!).
This is why many higher-level samplers and other things that require
random data inside [math] have a pluggable RandomGenerator.
A case in point is the sampling of other distributions (namely the
Normal distribution).
Or any of the others. We have a default, inversion-based method
that the abstract distribution classes provide and some pretty good
specialized implementations within individual distributions. Most
of these just require uniform random doubles as source.
Here is the benchmark report:
-----
nextGaussian() (calls per timed block: 2000000, timed blocks: 100,
time unit: ms)
name time/call std dev total time ratio
cv difference
o.a.c.m.r.JDKRandomGenerator 1.200e-05 1.7e-06 2.4001e+03 1.000
0.14 0.0000e+00
o.a.c.m.r.JDKRandomGenerator 7.646e-05 5.1e-06 1.5292e+04 6.371
0.07 1.2892e+04
o.a.c.m.r.MersenneTwister 6.396e-05 3.6e-06 1.2793e+04 5.330
0.06 1.0393e+04
o.a.c.m.r.Well512a 6.880e-05 5.0e-06 1.3760e+04 5.733
0.07 1.1360e+04
o.a.c.m.r.Well1024a 6.956e-05 3.0e-06 1.3913e+04 5.797
0.04 1.1513e+04
o.a.c.m.r.Well19937a 7.262e-05 2.0e-06 1.4525e+04 6.052
0.03 1.2125e+04
o.a.c.m.r.Well19937c 7.164e-05 4.3e-06 1.4329e+04 5.970
0.06 1.1928e+04
o.a.c.m.r.Well44497a 8.166e-05 3.2e-06 1.6332e+04 6.804
0.04 1.3931e+04
o.a.c.m.r.Well44497b 8.259e-05 4.6e-06 1.6518e+04 6.882
0.06 1.4118e+04
o.a.c.m.r.ISAACRandom 6.724e-05 5.4e-06 1.3449e+04 5.603
0.08 1.1049e+04
-----
where the first line is JDK's "nextInt()" and the remaining are
"nextGaussian()".
The generation time is thus about 4-fold that of "nextInt()".
Thus, degrading the performance of "nextInt()" by 10% would
degrade the
performance of "nextGaussian()" by half that.
For a performance discussion to be meaningful, I think that we'd
need
to know how that fact would affect, even modestly, any moderately
complex
post-processing of the generated values.
Another case, for modularity, would be to consider that other
algorithms could
be implemented to provide the required distribution.[3]
In the current design (inheritance-based), that can only be done
by creating
a subclass, even though the core functionality ("nextDouble()") is
not
overridden.
* What are usages of the CM RNGs?
Do those use-cases strictly forbid "loosing" a dozen
milliseconds per
million calls?
There are many different use cases. My own applications use them
in
simulations to generate random deviates, to generate random hex
strings as identifiers and in stochastic algorithms like some of
our
internal uses. The last case is definitely sensitive to PRNG
performance.
Thanks for giving examples, but since we talk about performance, I
was hoping for some real flesh, like the relative duration of
numbers
generation (e.g. the total duration of calls to the
"RandomGenerator"
instances wrt to the total duration of the application).
I don't know if by "last case", you are referring to code that is
inside CM. I didn't spot anything that makes "heavy" usage of a
RNG (in the sense that generation would count as a sizable part of
the whole processing).
monteCarloP in KolmogorovSmirnovTest is one to check.
As I pointed out many times: if an application is severely dependent
on the performance of RNG, the user probably will turn to specific
tools (e.g. GPUs? [4]) rather than use CM.
That is a bogus argument. We should make our PRNGs simple and fast
so their use can extend to performance-sensitive applications.
Conversely, using Java might be preferred for its flexibility, which
is destroyed by a search for ultimate performance (which nobody
seems
able to define reasonably).
Performance is not a goal in itself; it should not be a trophy which
sits uselessly on a shelf.
Nor should "beautiful design" in the eyes of one person.
My goal is not to deliberately slow things down; it is to allow some
leeway so that designs which are deemed better (on all counts
except,
perhaps, performance) are given a chance to show their strengths, in
particular in areas where performance in absolute terms is "good
enough" for all use-cases which CM should care about (hence the need
of actual data points[5]).
I see no reason that we can't have it both ways - good design and
good performance. What we have now, modulo maybe some small changes
to reduce code duplication, works fine. If you want to play with
64-bit generators and can find reference implementations and verify
that they do in fact perform better, great. If not, I don't see the
point. You can rant and complain all you want; but I am not going
to let us trash performance or correctness of code in the random
class or anywhere else just because you think it is somehow "better
designed" unless you can show specific, practical use cases
demonstrating the value of the changes.
Phil
Gilles
[1] "Is it faster?"
"No."
"Then, no."
[2] Although that is in some sense what you indirectly defend by
wanting
to stick with a meaningless "next(int bits)" method.
[3] http://www.doornik.com/research/ziggurat.pdf
[4] http://http.developer.nvidia.com/GPUGems3/gpugems3_ch37.html
[5] Hence the need to agree on a methodology/policy for
benchmarking.
Phil
[1] http://openjdk.java.net/projects/code-tools/jmh/
IOW, would those users for which such a difference matters use
CM at all?
Thanks,
Gilles
