Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Parallel algorithms in chess programming

Author: Robert Hyatt

Date: 06:44:18 04/17/01

Go up one level in this thread


On April 17, 2001 at 08:00:27, Vincent Diepeveen wrote:

>On April 16, 2001 at 22:07:10, Robert Hyatt wrote:
>
>>On April 16, 2001 at 18:15:52, Dieter Buerssner wrote:
>>
>>>In a different discussion, Vincent wrote the following:
>>>
>>>>It is not difficult to implement the form of parallellism as used by
>>>>Rudolf. Invented by a frenchman who couldn't spell a word english and
>>>>who wrote an impossible article for JICCA (did anyone proofread it at
>>>>the time as i'm pretty sure they didn't get his parallel idea?).
>>>>
>>>>At the time when i read the article i was pathetically laughing about it
>>>>actually as i also didn't get the idea of the frenchman. But it appears
>>>>everyone who can make a chessprogram work under win2000 can also get
>>>>within an afternoon his program parallel to work. Then some debugging
>>>>and a day later it works cool.
>>>
>>>I'd be very interested in this algorithm, that can be implemented at an
>>>afternoon :-)
>>>
>>>Could you point elaborate on this.
>>>
>>>BTW. In Paderborn, Roland Pfister also told me, that he knows this from Rudolf
>>>Huber, and he even started to explain it to me. Somhow, we (or me) got
>>>distracted, and I cannot remember the essential things.
>>>
>>>What I remember is, that the time consuming work, of making your
>>>search/evaluation routines free from all those global variables is not needed.
>>>
>>>Regards,
>>>Dieter
>>
>>
>>Global variables will _always_ be a problem.  Unless you avoid threads
>>altogether and use separate processes.  But then you incur other penalties
>>you have to solve...
>
>Multiprocessing is faster anyway of course as multithreading as i
>do not need to read a stupid pointer.
>
>Arrays which are only read and not modified one can put in shared memory
>and read from there without penalties.
>
>The penalty for multithreading is, when using ansi-c conventions,
>way bigger as for multiprocessing.
>
>Multiprocessing is more than ok.
>
>Obviously it's not always so easy to get it to work, because of
>different causes
>  - allocating shared memory in linux isn't simple
>  - windows NT server/tuple server versions have the bad habit
>    to always swap away shared memory to the cache when the allocated
>    memory size is huge.
>  - to share memory in NT in such a way that programs have the same
>    virtual adress space is not evident, as there are 2 functions which
>    one can use and one of them is not going to work for you. Don't need
>    to mention that by accident i picked the wrong function initially :)
>
>Of course all the above problems are solvable


Just use your approach to search trees with 2 nodes.  Then tell me how efficient
it is to use messages and pipes to get the other processes busy on such small
trees, compared to threads which have zero overhead...  :)

The pointer issue is _not_ a big one.  At most it costs a couple of percent
in speed, but it also gives back a lot since threads can look at other thread's
local memory if they know what they are doing...



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.