Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: I played 4 games against Zappa 2.0b tonight if anyone wants to see

Author: Robert Hyatt

Date: 13:02:32 08/23/05

Go up one level in this thread


On August 23, 2005 at 10:25:36, Joachim Rang wrote:

>On August 23, 2005 at 06:18:31, A. Cozzie wrote:
>
>>Zappa is not really that great at blitz.  In fact before the WCCC I did a test
>>at 5+0 vs Fruit and Fruit won 75% - and I was using 2 processors for Zappa.
>>
>>However at 15+5 that score was reversed (of course I was still cheating there :)
>>
>>The reason is that Zappa concentrates on getting a deep search by an efficient
>>branching factor, not pruning/big nps, so it needs some time before it begins to
>>outsearch opponents.  It is a serious chess engine, not a blitzer.
>>
>>anthony
>
>Bob stated in some posts that a parallel search needs some time to "stabilize"
>which means it can be weaker in blitz. Do you think this is related to your
>phenomenon? Was it a joke that Fruit got 75% in 5+0 but only 5% in 15+5? I can't
>imagine such a difference.
>
>regards Joachim


My weaker blitz performance this year was a different issue.  I play blitz all
the time on ICC.  But for the new dual-core opteron, I had to tune the parallel
search to split less often (greater distance from the leaf positions) to reduce
memory bandwidth.  The dual-cores have a bit of a bottleneck there compared to
the single-core opterons I have used in the past.  Each dual core chip has two
L1 and two L2 caches, but just one hypertransport interface and one memory
controller.  The hypertranspost handles a _lot_ of cache-to-cache traffic, and
with two cpus, it sees 2x the traffic a single-core processor sees.  Ditto for
the memory controller.  As a result, I had to tune things to do fewer splits
(which is a high bandwidth operation).  But that hurts blitz trees.  For
example, at the WCCC, I didn't split within 6 plies of a leaf position.
Searching 15-16 plies, that worked extremely well and parallel performance was
good.  But in very fast blitz games, where I might only get to search 10 plies
deep in a .1 second search, suddenly the extra processors are not helping that
much...




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.