[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [computer-go] Hardware-Instruction.
"Despite your belief, our speedups for going to multiple processors were
impressive."
Please show some logfiles of cilkchess which it had in tournament games,
including especially nodes a second and number of processors and at which
processor you ran it at. In fact i remember the hardware very well you used
the tournaments you were there.
Don't tell the same bullshit like hyatt that the logfiles are away. At
supercomputer *nothing ever* gets away. Things are 10 times backed up if
not more.
The diep logfiles, despite that i just ran once at it in a single
tournament, and cilkchess tens of times, will be even there when i'm dead
and buried in a year or 60.
The inefficiency of cilk is very well known of course. No need to defend a
hopeless cause.
By the way, in 1996 diep was barely 2 years old and i joined at 3 times
slower hardware first weekend than any of the other PC-participants (all
single cpu).
The operator of cilkchess first year was Aske Plaat. He didn't know how to
setup a chess position in cilkchess when Cilkchess had crashed. Aske Plaat
was given each game something like 15 minutes extra to compensate for
remote operation.
You were allowed to crash 3 times in those days before a game was resigned.
The nonsense about cilkchess is that you guys did speedclaims: "cilk is
doing real great".
Never any hard data. NEVER. Not a logfile. NOTHING.
I just saw it was 40 times slower than it could be.
The program was a simple bitboards program. "Even simpler than crafty" to
quote your own words in 1999.
In short you could on paper get the same nps like crafty times 512 at the
same hardware (1999 world champs).
Did you get that?
Remember i played you a few times. I know exactly how many nps a cpu you got.
Dare you say it publicly?
Vincent
At 18:29 5-11-2004 -0500, Don Dailey wrote:
>
>> Cilkchess doesn't show any speedup but clearly first slowed down their
>> program 40 times with 'cilk' from MiT, in order to show a better speedup.
>
>This shows that you have no clue about what you are talking about.
>
>In 1996 we won the Dutch championship with a score of 10 out of 11
>beating 9 programs including YOUR program and drawing only two. Not a
>single loss. Not bad for a program that was "slowed down 40X" in
>order to make the numbers look good.
>
>The following year we came in a very strong second, two points ahead
>of every one behind us and only 1/2 point behind the winner (Nimzo,
>Chrilly's program.) Not bad for a program that was "slowed down 40X"
>in order to make the numbers look good.
>
>Cilk is a superset of C, and in fact is just a frontend to C. I
>literally wrote a serial chess program in C, and added a few lines of
>"cilk" code to make it parallel. When the program is run as a 1
>processor "parallel" program, it slows down less than 5 percent over
>the purely serial program. Cilk is a fantastic language for high
>performance parallel programming. Despite your belief, our speedups
>for going to multiple processors were impressive.
>
>I won the 1996 tournament and came in second in the 1997 tournament
>because we had an excellent parallel implementation of a good but not
>great program. My program had a lot of evaluation problems and quite
>a few other bugs but none related to the parallelism, thanks to Cilk.
>Both programs were developed in just weeks (yes, the 1997 program was
>rewritten from scratch because I felt the winning 1996 program needed
>too much revamping to be what it should be.) Not like Diep which had
>evolved over many years but won less than half it's games in those 2
>tournaments.
>
>I think we did pretty good for a program that "showed no speedup" due
>to parallelism and I also think you don't know what you are talking
>about.
>
>> How much more nonsense will we keep seeing in the search world?
>
>It's ironic that you ask this question.
>
>
>- Don
>
>_______________________________________________
>computer-go mailing list
>computer-go@xxxxxxxxxxxxxxxxx
>http://www.computer-go.org/mailman/listinfo/computer-go/
>
>
_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/