Author: Robert Hyatt
Date: 07:09:16 05/26/04
Go up one level in this thread
On May 26, 2004 at 09:51:43, Anthony Cozzie wrote: >On May 25, 2004 at 21:33:32, Robert Hyatt wrote: > >>On May 25, 2004 at 19:52:56, Will Singleton wrote: >> >>>On May 25, 2004 at 17:33:12, Sean Empey wrote: >>> >>>>On May 25, 2004 at 17:09:02, Matthew Hull wrote: >>>> >>>>>On May 25, 2004 at 05:48:02, Sean Empey wrote: >>>>> >>>>>>No, it is not a clone of crafty. I will tell you that it does use Dan Corbit's >>>>>>big book and crafty's book code, and the same egtb code that is used in crafty >>>>>>18.15. Other than that not much else is from crafty. Oh some of the Print >>>>>>functions I referenced are from crafty 16.x if memory serves correct. Storm is >>>>>a >SMP engine and uses a completely different algorithm for its search eval. It >>>>>>uses ABDADA.I have offered portions of code to Professor Hyatt but on the >>>>>terms >it not be published or used by him or others. My move generator >>>>>functions are >completely different and I can not go into detail regarding them >>>>>as some of the >code came from a professional programmer (at least the >>>>>approach) on the >condition I not release that code ever. Because a program has >>>>>similar evals does >not make it a clone. If I wanted to clone crafty I would >>>>>not have been working >on Storm for years; actually starting about one year >>>>>before CCT1. As I have told >Professor Hyatt. I am not cloning crafty. We have >>>>>talked and he is fine. Any >other questions someone may have; I'm willing to >>>>>answer them to the best of my >ability. I would appreciate ad hominems and >>>>>other unnecessary comments not be >made as I have stated I'm willing to back-up >>>>>my claim. >>>>> >>>>>Hi Sean, >>>>> >>>>>A question about SMP. Does Storm use threads or processes when splitting the >>>>>search? Which do you think is most effective? >>>>>Thanks! >>>>> >>>> >>>>I use ABDADA, which I learned from an ICCA publication. You can find info here: >>>>http://www.recherche.enac.fr/~weill/publications.html >>>> >>>>There is no communication between threads. Just a shared hash table and it's >>>>pretty easy to implement. I have been happy with it. It may not be the best for >>>>SMP machines. Professor Hyatt says nothing beats a perfectly implemented PVS. He >>>>may be right. >>> >>>Hi Sean, >>> >>>I'm like someday to do multiprocessing, but I haven't studied it. You indicate >>>there's no communciation between threads, how does that work? I assume there's >>>one thread that does i/o, and that thread must somehow inform the other what the >>>position is and get search results. >>> >>>Also, I think I understand how two threads can share the same hash table, but >>>how do you handle locking? Is it automatic or do you need to explicitly lock >>>the memory while writing? And what kind of speedup do you get? Does the ICCA >>>article use the same approach (shared hash, multiple processes, no comm)? >>> >>>Will >> >> >>This is a _very_ primitive approach to parallel search. It will do fairly >>reasonably with 2 processors. Go beyond that and performance becomes very >>poor... >> >>It is one of those "easy to implement" approaches, but in this case you really >>do get what you "pay for" and since you don't "pay very much in terms of >>effort..." > >In the paper they quoted a 64X speedup on a 128 processor machine, which is >pretty good I would think. It doesn't seem like it would perform that well >after looking at the algorithm though. > >anthony Yes but you have to watch the search depths. They also reported that as they went deeper, the search did worse, IIRC... I'll try to re-read it as I have not looked at it since it was first published.
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.