Author: Gian-Carlo Pascutto
Date: 15:50:44 09/15/05
Go up one level in this thread
On September 15, 2005 at 18:39:47, Dann Corbit wrote: >On September 15, 2005 at 18:16:32, Gian-Carlo Pascutto wrote: > >>On September 15, 2005 at 17:01:37, Dann Corbit wrote: >> >>>On September 15, 2005 at 16:23:13, Gian-Carlo Pascutto wrote: >>> >>>>On September 15, 2005 at 15:17:13, Michael Yee wrote: >>>> >>>>>On September 15, 2005 at 15:13:22, Joshua Shriver wrote: >>>>> >>>>>>Has anyone tried writing an engine that uses PVM, MPI, or Mosix? >>>>>> >>>>>>If you had gigabit, or even fibre wire would the latency still be a problem? >>>>>> >>>>>>Josh >>>>> >>>>>The chessbrain project actually used a distributed beowulf over the internet: >>>>> >>>>>http://www.chessbrain.net/ >>>> >>>>And they never published any serious result or test. >>> >>>This is interesting: >>>http://www.chessbrain.net/docs/chessbrain-discc.pdf >>> >>>Additional overviews: >>>http://www.chessbrain.net/docs/thechessbrainproject.pdf >>>http://www.chessbrain.net/docs/cblinuxjournal0903.pdf >> >>There is nothing interesting in there. In fact these reports don't do much more >>than say "we connected a lot of computers and had them play chess". >> >>No single performance metric. > >5.3 Game Statistics >The total number of useful nodes processed was 84,771,654,525; that is the >number of nodes calculated for work units that were accepted by the central >server. Many more nodes were processed in work units that were never delivered >due to server connection issues, as well as those nodes that were aborted before >a result was returned. ChessBrain used a total of 2 hours, 15 minutes of >thinking time. Using just the useful returned data, this gives an average of >10.5 million nodes per second. By contrast, Beowulf analyses approximately >100,000 nodes per second in an average position on a single P4/2.8GHz machine. >This means, in terms of raw node processing, that ChessBrain performed at the >level of a single 280 GHz CPU. >These figures also tell us that ChessBrain’s efficiency in terms of converting >connected machines to raw processing power was approximately 100 / 2,070 = 5%. >Improving the central server architecture so that the connection problems are >resolved would improve this figure dramatically. We are working on a server >redesign to overcome this challenge before we enter any more high-profile >matches. > >>Why? >> >>Was the speedup on 1000 machines <=1 ? > >They got 5% efficiency. At most. Their measurement doesn't include semi-bad splits for example, nor inefficiencies introduced by worse move ordering, etc... -- 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.