Author: Vincent Diepeveen
Date: 11:29:05 08/11/05
Go up one level in this thread
On August 11, 2005 at 13:59:11, Robert Hyatt wrote:
>On August 11, 2005 at 11:36:17, A. Cozzie wrote:
>
>>Well I have taken my holiday from CCC to improve my program, but I'm flying out
>>tomorrow and not going to make any big changes today, so its just a matter of
>>packing up, and I have some free time to speculate. Quite frankly I think
>>guessing who is going to score what in the WCCC might be more fun than playing
>>with it, and unlike boxing programs can't use you ranking them poorly as
>>motivation. So here goes:
>>
>>1. I'm predicting Junior to repeat. No one tests against Junior - they all test
>>against Shredder.
>>
>>2. I'm predicting this is the year that Shredder finally falls into the middle
>>of the pack. Shredder simply hasn't improved since 7.04. 7.04 was a great
>>engine, but the rest of the world has caught up with him. In addition, Fruit is
>>joining, and Fruit is hell for Shredder.
>>
>>3. I'm predicting Crafty and Diep both do pretty well, simply by virtue of
>>hardware. 15M nps is nothing to write off easily, and Peter Berger with his
>>blood contract with the devil will again get Crafty positions where it reacts
>>well. On Diep's side, he might finally get a decent search depth, which has
>>been Diep's achilles heel since time immemorial.
>>
>>4. I'm predicting Fruit only finishes in the middle of the pack. Fruit is a
>>strong program, but it isn't parallel and my prediction is that it gets killed
>>in the opening book at least 3 or 4 games and loses before it computes its first
>>instruction.
>>
>>5. As for myself, my goal is to get 3rd. I'll be disappointed if Zappa finishes
>>less than 5th, but I'm prepared for that :)
>>
>>anthony
>
>
>I'm not predicting anything. Not really much point since the tournament starts
>in a few days.
>
>For those not noticing, looks like the 4 x 875 opteron system is good to go.
>Been testing last two nights on ICC and the parallel search is doing quite well
>on the NUMA architecture, after a bit of modifying the source, and a bit of
>tuning the crafty internal SMP parameters (splitting depth, max processors at
>any single split point, etc.).
Diep's parameters for parallel search:
- maximum allowance of 32 cpu's in 1 position using the below rules:
- minimum splitdepth of 3 ply
SplitInOthers(..)
/* uitvissen hoeveel procs er maximaal opgevraagd mogen worden ivm zettenlijst
*/
nprocs = (int)(rb->zetend-rb->current)-2;
maxhere = MAXPOSPROCS-1; /* maximaal aantal cpu's parallel alhier */
if( maxhere >= 31 ) {
if( rb->ply <= 4 )
maxhere = 7;
else if( rb->ply <= 6 )
maxhere = 15;
else
maxhere = 31;
}
So 8 cpu's are allowed in 1 pos if we have 3-4 ply,
etc.
>
>I'm working on getting the last bit of scaling done so that the 8CPU nps will be
>8.0x the 1 cpu nps. rather than the current 7.5X. I think I can finish this
>today...
Nah, you'll always lose some to the hard fact that hashtable is faster for 1 cpu
than it is for 8.
>Whether we will play well or not is anybody's guess. But I can guaran-damn-tee
>you we will be playing _fast_. :)
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.