Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Non-deterministic behaviour of Deep Shredder - REALLY interesting

Author: Uri Blass

Date: 10:09:16 01/01/06

Go up one level in this thread


On January 01, 2006 at 11:17:26, Robert Hyatt wrote:

>On December 31, 2005 at 17:41:27, George Sobala wrote:
>
>>I was playing around with seeing how Deep Shredder performs with different
>>numbers of threads, and was fascinated to discover that the program behaviour
>>becomes completely non-repeatable / non-deterministic once more than one thread
>>is running.
>
>This is perfectly normal, and every parallel search program (that is any good)
>exhibits this same behavior.  It's just something one lives with to obtain the
>improved performance.

I wonder if it is not possible to have deterministic parallel search that is
only slightly slower.

speed is not everything and bugs that you cannot reproduce can be also a
problem.

It may be better to be 10% slower with deterministic search for being able to
reproduce errors and the question is if it is not impossible to have rules that
enable it.

Note that I understand nothing about parallel search and if my idea cannot work
then I am interested to know what is the problem with it.

My idea is the following:

Suppose you have 4 processors that have big jobs to do.

You can divide the jobs to small parts of searching of exactly 10,000 nodes or
searching less than it in case that the processor finish.

You can decide that processors do not continue after finishing their part and
simply wait for other processors(hopefully they do not wait big time because
searching 10,000 nodes takes almost the same time for all processors and cases
when they search less than 10,000 nodes are not common).

You can also decide that processors do not store information in the global hash
during the time that they did not finish their part and simply store the
information in small tables.

After all the processors finish their parts they store the information in the
global hash and the computer decide based on deterministic rules how to
continue.

Uri



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.