Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: LCT II Fin4, Deep Thought, and Deep Blue (was Re: LCT II results...)

Author: Robert Hyatt

Date: 12:44:13 01/06/98

Go up one level in this thread


On January 06, 1998 at 15:35:18, Stuart Cracraft wrote:

>Maybe you should consider using an extensive just-checkmates
>suite or group of suites to test your parallel program.
>
>I use two combination suites, one mate suite, one endgame suite,
>and two rating tests, plus play multiple games at different time
>controls.
>
>--Stuart

Doesn't help.  The problem, in this case, required the following:

1.  N-1 legal moves where N was the number of processors;

2.  just the right number of interrupts so that the critical timing
hole was activated, allowing that "extra" processor to be delayed just
long enough to report back *after* at least one of the other processors
had completed a search (something normally impossible since it is always
faster to search *nothing* than to search even one node...)

IE, it's not the problem positions, it's the bizarre timing problems
that
occur in a program written on a machine which is built around shared
memory.  IE any processor can write to any memory location whenever it
wants to.  And it's up to the programmer to use semaphores to prevent
the
inevitable race conditions that might happen.

I believe that I found at least one parallel search bug in Cray Blitz
every
time we moved to a different machine, either one with a faster clock or
with
more processors.  *every* time.  Correctness means one thing to a serial
program, something else entirely when discussing parallel programs.
Problems
are very similar to those I handled for 20 years doing operating system
kernel development.  An interrupt at an inopportune time can change
something
in a way you might never expect.  At a time when the whole world can see
it
happen, too.  :)



This page took 0.01 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.