Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Parallel search article RBF

Author: Robert Hyatt

Date: 19:09:26 09/11/02

Go up one level in this thread


On September 11, 2002 at 15:27:32, Dann Corbit wrote:

>On September 11, 2002 at 15:10:49, Gian-Carlo Pascutto wrote:
>[snip]
>>>2.  That there is definitely a linear improvement in new CPUs and 64 CPUs is
>>>almost exactly twice as good as 32 CPUs.  I think that is pretty astonishing,
>>>and so however it is that communication happens between nodes should be tried
>>>in other systems.
>>
>>It's a natural consequence of the fact that they have (almost) no
>>synchronisation costs.
>
>Well, that's pretty interesting, isn't it?  Can the same model be copied to
>other algorithms?
>
>Surely they have to coordinate the information from the entire set of processors
>somehow.  The processors cannot be purely independent of one another or no
>resolution would be reached.  If a similar speedup could be obtained with
>another scheme, it might be very valuable.
>
>Also, since they have done a test with 64 CPUs, I assume that they are using
>some sort of NUMA architecture, since there are not many 64 CPU SMP systems
>around.  Hence, it might scale to stupendous CPU counts like the DoD
>supercomputers with thousands of CPUs.  If if the efficiency is 1%, if you have
>10000 CPUs you will be at 100x the speed of a single CPU system.  Maybe no other
>system can match that.


Two comments to _everybody_ fiddling with this thread.

1.  The best serial algorithm is generally _not_ the best parallel
algorithm.  This has been shown over and over and over.  The best parallel
algorithm, by inverse thinking, is not necessarily the best serial algorithm
either.  Just because we use alpha/beta today doesn't mean we will be using
it in 5 years on massively parallel machines.  Quicksort is the best serial
sort today.  It isn't necessarily the best parallel sort for large numbers
of processors.

2.  Best-first is not thought much of today, because it has never worked well
in the past.  But that doesn't mean it _can't_ work well, just that nobody has
really worked with it anywhere near as much as with Shannon's original minimax
approach later augmented with the now infamous alpha/beta modification...

I hesitate to throw rocks at people trying new ideas.  Sometimes they waste
tons of time, but _sometimes_ they find remarkable results...




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.