Author: Robert Hyatt
Date: 09:53:45 05/03/04
Go up one level in this thread
On May 03, 2004 at 12:29:37, Gian-Carlo Pascutto wrote: >On May 03, 2004 at 12:12:21, Robert Hyatt wrote: > >>The hardware was _identical_ except for CPU speed. > >I'm talking about Vincent's 512 cpu thing versus the Cray >the DTS paper was run on. OK. You were in on some of the conversations. Did I not tell Vincent "You are running on a horrible architecture compared to the C90 from Cray. It would almost be impossible to produce comparable numbers when the architecture is so inferior." Do you remember that? Do you remember his response? He "waved that off"... NUMA can be great. Etc... I _never_ compare apples and oranges. I've never compared Crafty to DTS for that reason. DTS was better. But I don't want to give up the simpler recursive approach I use today... I don't want to debug DTS for another 2 years as I did when I first got it working... And I don't compare Crafty to those results since I know it is worse. > >>Totally up to you. The main data was not "calculated in a rather funny and >>backward manner". (the speedup data). > >There is nothing left substantiating the single very most important thing >in the entire paper (the speedups). As it is now, the speedups were set >in stone and all the data that were supposed to support them or allow them >to be calculated is based on them, instead of the reverse. No. The speedups were computed by dividing 1cpu time by Ncpu time. There is no other way to calculate that. Later the time info was lost. The node info was never available in the first place. > >For all we know, you determined what looked like reasonable speeups >in advance and invented the 'supporting data' for them. Actually, that >*exactly* how the paper looks without all your storytelling around it. Fine. As I said, it's a pretty free world. You are free to believe what you want. I know what I did... > >And you are surprised people are sceptical about it? > >I'm sorry, but as far as scientific results go it is garbage. The >explanation of the principles behind DTS is interesting, though :) > I don't understand why both you and vincent asked me _so_ many questions about DTS if it is based on garbage? Why would the primciples behind DTS be interesting if the results are garbage? I suppose I miss the point. With you and Vincent I find that happening a lot... >>The actual speedup calculations are around. > >Would be interesting to see what is left of that. It is a sheet of paper in a file. Would a scan of that show anything? It is _exactly_ what you see in the speed-up table in the DTS article, only handwritten over a period of about a year as I slowly ran each test... > >-- >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.