* Re: Shootout News: 64 Bits of Sweet Pancake
2009-12-15 8:24 ` Martin Krischik
@ 2009-12-15 11:20 ` Georg Bauhaus
2009-12-15 11:25 ` jonathan
1 sibling, 0 replies; 4+ messages in thread
From: Georg Bauhaus @ 2009-12-15 11:20 UTC (permalink / raw)
Martin Krischik schrieb:
> Am 14.12.2009, 18:07 Uhr, schrieb Georg Bauhaus
> <rm.dash-bauhaus@futureapps.de>:
>
>> http://shootout.alioth.debian.org/u64q/benchmark.php?test=fannkuch&lang=all&box=1
>>
>
> Congratulations.
Congratulations need to be sent to Jonathan Parker.
It's his efforts that have once again produced a
program leading in speed.
-- Based on code by Dave Fladebo, Eckehard Berns and Heiner Marxen.
-- Based on the ATS version by Hongwei Xi,
-- and the Java version by The Anh Tran.
-- Contributed by Jonathan Parker.
Many of the leading programs note similar sources; you wonder
how they differ, since they run at different speeds. One thing
that Jonathan has already mentioned is that this is plain Ada,
simply tasking, and no libraries.
The Shootout programs teach, me at least, another thing about language
features and how they contribute to speed; the fundamental
type system, for example, is made to work hand in hand, apparently,
with instruction sets, optimizers and CPUs register allocation.
^ permalink raw reply [flat|nested] 4+ messages in thread
* Re: Shootout News: 64 Bits of Sweet Pancake
2009-12-15 8:24 ` Martin Krischik
2009-12-15 11:20 ` Georg Bauhaus
@ 2009-12-15 11:25 ` jonathan
1 sibling, 0 replies; 4+ messages in thread
From: jonathan @ 2009-12-15 11:25 UTC (permalink / raw)
A few comments on what we've been doing and why.
Rankings are fun (well, at the moment anyway, when all
goes well) but they sure change fast. I'ld expect
someone to submit something faster in the next few weeks.
The important thing in my view was getting some Ada
multi-core programs into the public view.
Here, by the way, are the present overall rankings of
the Ada programs in the 64-bit tests:
http://shootout.alioth.debian.org/u64/benchmark.php?test=all&lang=all&box=1
http://shootout.alioth.debian.org/u64q/benchmark.php?test=all&lang=all&box=1
I remember that after the 6 months we recently spent updating
the Ada suite, the Ada median ranking was at about 1.3
on the above scale (64-bit tests). Three days ago it at
about 1.7. Now (Dec 15) its about 1.25. I won't dare
predict Dec 16.
Here BTW are the number of page-views to the
shootout site over the last 5 or so years (number 2
below, just after Debian Linux):
https://alioth.debian.org/top/toplist.php?type=pageviews_proj
1 Debian Port to AMD64 3938690
2 The Computer Language Benchmarks Game 2179690
3 debian-installer 1402511
That's a lot of hits. I was getting tired of
hearing this cited as definitive evidence that Ada
was 4 times slower than C, or X times slower or whatever.
The test machine has 4 cores available for processing,
and the contest has become a test of language/programmer
ability to exploit these 4 cores in parallel. Most of
the Ada programs weren't multi-tasking/multi-core when
they were originally written, and they were a factor of
4 slower than the competition for that reason alone. For
typical students or programmers, browsing the shootout
site may be the only exposure they'll ever get to Ada
code, Ada runtime performance, or the Ada approach to
exploiting multi-core processors, so these updates were
long overdue.
Finally, the fannkuch numbers - more fun than fractals.
I just invested about 3.5 million processor-seconds in
getting the 16th fannkuch number (139). We had an idle
16-core machine sitting around. I don't suppose there's
a table of these numbers sitting around somewhere. The
17th will take about 70 million processor-seconds to
calculate if I have to do it myself. Is it
a prime number like the 14th, 15th, and 16th?
The more interesting thing is long term trend.
To find out what that is, plot the cubed root
of the fannkuch numbers (the 2nd column) as a function
of the 1st column:
2 1
3 2
4 4
5 7
6 10
7 16
8 22
9 30
10 38
11 51
12 65
13 80
14 101
15 113
16 139
Jonathan
^ permalink raw reply [flat|nested] 4+ messages in thread