Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Bob....how strong Crafty on Blue Pacific (@3.9 trillion p/s)?

Author: Robert Hyatt

Date: 07:40:58 10/31/98

Go up one level in this thread


On October 30, 1998 at 14:43:19, José de Jesús García Ruvalcaba wrote:

>On October 29, 1998 at 21:58:24, Robert Hyatt wrote:
>
>>On October 29, 1998 at 18:27:10, Terry Presgrove wrote:
>>
>>>Bob,Ed programers everywhere....just awe struck by the lastest
>>>super computer just unveiled by V.P. Gore. At 3.9 Trilion calculations
>>>per second just how strong would Crafty , Rebel, any of the comercialy
>>>available programs be? Unless I misunderstood its Over 15,000 times faster
>>>the the fastest PC's currrently available.  More than 5,800
>>>processors....wouldn't it be nice to play around with this baby for just a day?
>>
>>
>>first warning is a large number of processors...  ie it isn't a shared memory
>>machine, which means that it is not easy or efficient to distribute the search
>>on such machines.  If you have seen crafty on a 4 processor machine, and expect
>>that to scale so such a machine, you'll be disappointed...
>>
>>However Sequent makes a 32 processor xeon-based machine that features a
>>shared memory.  *that* would scream...  IE a NPS of something around 10M
>>would be easy...
>
>What would be the expected performance of craty on a 64-processor Sun Enterprise
>10000?
>As far as I understand, it has symmetric multiprocessing.



Assuming no "bottlenecks" it should be roughly .7*64 times faster.  However,
I wouldn't be certain that the .7 holds...  IE I have a single lock for each
of the two hash tables.  With 64 cpus, that is likely going to produce some
blocking conditions.  I carefully only lock while I read an entry then unlock,
but it is possible that this will start to add up.  Not having a 64 processor
machine to work on, I can only speculate.  In Cray Blitz, with 32 processors,
I didn't see significant blocking... but there I did everything in assembly
and the lock/load/unlock and lock/store/unlock for hashing was *very* quick.

Also I don't know how sun has done their 64 processor architecture.  A non-
blocking crossbar is very expensive for 64 processors.  Also, what about memory?
Cray uses a bunch of memory banks (things that can do 2 reads and 1 write at
the same time) to avoid processor stalls.  What does Sun do?  If the memory
bandwidth isn't there, then the .7 is going to shrink...



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.