On 12/20/11 1:29 AM, Russel Winder wrote:
The system as set out is biased though, systematically so. This is not
a problem per se since all the micro-benchmarks are about
computationally intensive activity. Native code versions are therefore
always going to appear better. But then this is fine the Shootout is
about computationally intensive comparison.
This is fine, so no bias so far. It's a speed benchmark, so it's
supposed to measure speed. It says as much. If native code comes usually
in top places, the word is "expected", not "biased".
Actually I am surprised
that Java does so well in this comparison due to its start-up time
issues.
I suppose this is because the run time of the tests is long enough to
bury VM startup time. Alternatively, the benchmark may only measure the
effective execution time.
Part of the "problem" I alluded to was people using the numbers without
thinking. No amount of words on pages affect these people, they take
the numbers as is and make decisions based solely on them.
Well, how is that a bias of the benchmark?
C, C++ and
Fortran win on most of them and so are the only choice of language.
The benchmark measures speed. If one is looking for speed wouldn't the
choice of language be in keeping with these results? I'd be much more
suspicious of the quality and/or good will of the benchmark if other
languages would frequently come to the top.
As I understand it, Isaac ruins this basically single handed, relying of
folk providing versions of the code. This means there is a highly
restricted resource issue in managing the Shootout. Hence a definite
set of problems and a restricted set of languages to make management
feasible. This leads to interesting situation such as D is not part of
the set but Clean and Mozart/Oz are. But then Isaac is the final
arbiter here, as it is his project, and what he says goes.
If I recall things correctly, Isaac dropped the D code because it was
32-bit only, which was too much trouble for his setup. Now we have good
64 bit generation, so it may be a good time to redo D implementations of
the benchmarks and submit it again to Isaac for inclusion in the shootout.
Quite frankly, however, your remark (which I must agree, for all respect
I hold for you, is baseless) is a PR faux pas - and unfortunately not
the only one of our community. I'd find it difficult to go now and say,
"by the way, Isaac, we're that community that insulted you on a couple
of occasions. Now that we got to talk again, how about putting D back in
the shootout?"
I looked at the Java code and the Groovy code a couple of years back (I
haven't re-checked the Java code recently), and it was more or less a
transliteration of the C code.
That is contributed code. In order to demonstrate bias you'd need to
show that faster code was submitted and refused.
This meant that the programming
languages were not being shown off at their best. I started a project
with the Groovy community to provide reasonable version of Groovy codes
and was getting some take up. Groovy was always going to be with Python
and Ruby and nowhere near C, C++, and Fortran, or Java, but the results
being displayed at the time were orders of magnitude slower than Groovy
could be, as shown by the Java results. The most obvious problem was
that the original Groovy code was written so as to avoid any parallelism
at all.
Who wrote the code? Is the owner of the shootout site responsible for
those poor results?
Of course Groovy (like Python) would never be used directly for this
sort of computation, a mixed Groovy/Java or Python/C (or Python/C++,
Python/Fortran) would be -- the "tight loop" being coded in the static
language, the rest in the dynamic language. Isaac said though that
this was not permitted, that only pure single language versions were
allowed. Entirely reasonable in one sense, unfair in another: fair
because it is about language performance in the abstract, unfair because
it is comparing languages out of real world use context.
I'd find it a stretch to label that as unfair, for multiple reasons. The
shootout measures speed of programming languages, not speed of systems
languages wrapped in shells of other languages. The simpler reason is
that it's the decision of the site owner to choose the rules. I happen
to find them reasonable, but I get your point too (particularly if the
optimized routines are part of the language's standard library).
(It is worth noting that the Python is represented by CPython, and I
suspect PyPy would be a lot faster for these micro-benchmarks. But only
when PyPy is Python 3 compliant since Python 3 and not Python 2 is the
representative in the Shootout. A comparison here is between using
Erlang and Erlang HiPE.)
In the event, Isaac took Groovy out of the Shootout, so the Groovy
rewrite effort was disbanded. I know Isaac says run your own site, but
that rather misses the point, and leads directly to the sort of hassles
Walter had when providing a benchmark site.
That actually hits the point so hard, the point is blown into so little
pieces, you'd think it wasn't there in the first place. It's a website.
If it doesn't do what you want, at the worst case that would be "a
bummer". But it's not "unfair" as the whole notion of fairness is
inappropriate here. Asking for anything including fairness _does_ miss
the point.
There is no point in a
language development team running a benchmark. The issues are
perceived, if not real, bias in the numbers. Benchmarks have to be run
by an independent even if the contributions are from language
development teams.
But I'm sure Russel had something in mind. Russel, would you want to
expand a bit?
Hopefully the above does what you ask.
The summary is that Isaac is running this in good faith, but there are
systematic biases in the whole thing, which is entirely fine as long as
you appreciate that.
Well, to me your elaboration seems like one of those delicious
monologues Ricky Gervais gets into in the show "Extras". He makes some
remark, figures it's a faux pas, and then tries to mend it but instead
it all gets worse and worse.
Andrei