Author: Vincent Diepeveen
Date: 09:30:37 09/03/03
Go up one level in this thread
On September 03, 2003 at 10:57:17, Sune Fischer wrote: >On September 03, 2003 at 10:51:56, Vincent Diepeveen wrote: > >>On September 03, 2003 at 10:44:11, Sune Fischer wrote: >> >>>On September 03, 2003 at 10:35:11, Gian-Carlo Pascutto wrote: >>> >>>>On September 03, 2003 at 10:26:19, Sune Fischer wrote: >>>> >>>>>>What is your point? >>>>> >>>>>Heh :) >>>>> >>>>>You said (remote) checking for fail_high conditions at every node was required, >>>>>and I disagree. >>>> >>>>I never say 'at every node'. >>> >>>You said constantly, then I don't know what you mean by that. >>> >>>>Each time you get a subtree score, you must send out the score update to >>>>all processors, or store it locally and rely on remote processors to check >>>>it in your memory. >>>> >>>>Either way, you need remote accesses. >>>> >>>>Got it now? >>> >>>This is an entirely different matter, you don't get subtree scores "constantly". >>>And I still say you only have to access when there is something to communicate. >>> >>>If you just quietly exit the subtree on a fail low, I see no need for >>>communication. >> >>And how big is your search going to be. 10 ply minimax? >> >>Or do you prefer 10 ply alfabeta + nullmove. >> >>How many times is your processor going to check whether he is doing work for >>nothing? >> >>Each node? >>Each 10 million nodes? >> >>The advantage of each node is that it won't idle too much. >>The advantage of each 10 million nodes is that you don't eat up bandwidth, but >>that your search will be as bad as a single cpu. >> >>So how are you going to do it? > >Ah yes, agreed. >There is a treade off to be made here, some experiments would have to help me >decide on that. > >But since you don't want to spawn new search threads constantly, you don't >terminate search threads constantly either. As a matter of fact the numbers are >identical. I hope you realize that when Huber said his thing would work like the sun on that machine in july 2002, that i was pretty amazed to hear he spawns each move new processes. Thanks to some other 300 processors being busy with something eating 200GB ram (each processor has 1 GB ram) and the i/o being broken (they were upgrading the machine), just that spawning of 59 diep processes took 15 minutes. A few months ago when i spawned 129 diep processes (so 130 processes in total) with just 10GB ram, that took 1.5 hours. They are currently creating a database of 10 TB with extreme weather predictions for the coming centuries, so you'll see the results soon at the discovery channel i guess. Of course they already run for months continuesly at 258 processors (out of 1024). Let's pray they have finished before november ;) Thanks, Vincent >-S. > >> >>>-S. >>>>-- >>>>GCP
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.