Author: Robert Hyatt
Date: 09:08:31 09/20/00
Go up one level in this thread
On September 20, 2000 at 11:52:17, Mogens Larsen wrote: >On September 20, 2000 at 08:52:03, Robert Hyatt wrote: > >>Jonathan Schaeffer tried this a good while back with his distributed version >>of Phoenix. > >Has he published anything on the net about this subject? I don't know about "on the net" but I am sure something was in an ICCA journal several years back. The tactical searcher was called "minix" and it searched deeper than the normal positional searcher, with the idea that it would either find a winning tactical shot the positional searcher could not find, or it would find the positional move proposed by the positional searcher was a blunder. > >>The problem is, suppose your "positional" search says play "X" >>while the tactical search says "X loses"? > >I guess that would depend on the actual evaluation. Maybe in very rough terms >something like: if the tactical evaluation is very poor then go with positional, >unless the number of pieces on the board favours a tactical approach or the >position isn't at least semiclosed. Maybe a cpu for comparative evaluation would >be ideal as well :o). You've got four. What should the last one do? :o)) > It is difficult. If there is one positional best move you find near the end of the search, do you give the tactical searcher a chance to search it for sanity? Suppose that fails? >>It is not easy to coordinate a pair >>of search results like this. And remember both would be searching the same >>part of the tree, twice. > >That wouldn't be a major problem IMO as the actual result would still >(hopefully) be superior to the single cpu approach. The SMP approach is just a >remedy to the lack of single cpu clock speed as far as I can tell, not an >advanced computer chess concept. Am I wrong about that? > I am not quite sure what you mean. In my case, I consider my search to be "optimal" in the sense that it is searching the tree that I want it to search. Using the current threaded approach simply searches that same tree significantly faster. I don't particularly have a requirement that all the CPUs run at the same clock speed, although SMP motherboards generally do have that requirement to make interrupt delivery 'sane' as well as handling a shared bus. I will add that a parallel search is a significant contributor to overall program complexity, particularly for someone that has never written any parallel code or done operating system development (interrupt handling presents many of the same bugs as multiple threads sharing a memory address space). >>a standard SMP approach is way more efficient... and also easier to understand >>when you try to address the above problem. > >I think it's too simple an approach, but I know too little about it to put >weight behind that opinion. Is there any available papers for technical morons >about SMP? > >Best wishes... >Mogens I've written up the Cray Blitz parallel search in the JICCA. I think I may have an ASCII version if you will email me. As far as simple, I don't agree. IE I want to search the _optimal_ tree that the single-cpu version of the tree searches. I only want to search in N times faster. other approaches (ie Schaeffer's minix idea) is definitely simpler to implement, but it presents some issues that are difficult to handle as well. I'd rather simply search the usual tree, but faster, so that I can go deeper. Otherwise on a single cpu I would _still_ use a minix-approach and allocate half the total time to the positional search and the rest to the minix search...
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.