Author: Robert Hyatt
Date: 17:13:00 09/03/02
Go up one level in this thread
On September 03, 2002 at 18:29:33, Gian-Carlo Pascutto wrote: >On September 03, 2002 at 18:15:50, Robert Hyatt wrote: > >>On September 03, 2002 at 18:03:14, Gian-Carlo Pascutto wrote: >> >>>However reasonable your explanations may be, the gist of your DTS article >>>and the most important thing for comparison were the speedup numbers. After >>>what we discovered and what you just posted, it is clear that they are >>>based on very shaky foundations. >> >>How so? The speedup numbers were _directly_ computed by dividing times. > >But what times? Certainly not the times you reported. All we have is 5 >speedup numbers that nobody every reproduced and likely will never >reproduce due to the hardware involved. So? What do you have from my dts phd? Times nobody can reproduce since that machine has been dismantled for 10 years. What about Schaeffer's results? If you think I fake data, fine. But be direct about it. 90% of the research results can't be reproduced for that very reason, and I doubt that most of them are "faked" because they can't be repeated... > >>Nodes were impossible to grab in the middle of a search so we computed what >>they "should have been" and we did do some testing to be sure that the >>estimation was very accurate... > >Which wasn't noted anywhere except just now. Had vincent explained his "problem" I could have figured this out. But he never did, he just said 11.1 is faked and can't possibly be true. That I _know_ is a good number. The node counts are well within the normal SMP error margin as well, which makes that discussion extremely moot... > >>University microfilm will sell you a printed copy. It was published in >>1988 at the University of Alabama at Birmingham. So far as I know, they >>microfiche every dissertation published and provide copies for what cost >>I don't know. I certainly bought more than one from them over the years, >>but the last was so long ago I have no idea what they charged me for it. >>(it was berliner's 1970 dissertation). > >Any way to get one of those in Belgium, considering I'm currently no >longer attach to any university? I'm interested in it regardless, since >I'm experimenting with DTS-like algorithms. I went to my local UAB library and asked for a copy of Berliner's thesis. They found it, and said "please pay XX dollars and we will get it for you." I don't know what they did, but a few days later it showed up in my mailbox (back at USM)... All I think you need is my name, and perhaps the title. Which is "A high-performance parallel algorithm to search depth-first game trees" Robert Morgan Hyatt, published in 1988 at the University of Alabama at Birmingham. That should let them find it. But be aware I have no idea about what they charge. It isn't a UAB thing (university microfilm is a company not affiliated with anybody I know of.) The main thing in it is a more detailed study of parallel searching, including special cases like uniform trees, worst-first trees, best-first trees, and finally "normal chess trees." DTS was blinding on worst-first and best-first. :) > >>>I hope you realize that a request from you to trust your numbers isn't >>>very convincing. In fact, with what we know now, I'm pretty sure the >>>article would never have gotten published in the first place. >> >> >>I do not agree. The speedup numbers were verified by several different >>people. > >Apparently, those people completely failed to address all points that >were just brought up, so that's not very convincing, is it. I haven't seen any reasonable points that have been brought up yet, myself. node counts are so variable in an SMP search, as you must know. To worry about a few thousand here and there is pointless when the tree size varies by more than that from one run to the next... Even had I been able to grab real node counts, I could not ever repeat them as I showed in another post using Crafty. > >>I'd be happy to run a short test on Crafty and do the same calculations to >>show you why I am that certain... > >Crafty and Cray Blitz are not comparable. If they were, I would have redone >your tests as soon a I had access to your quad. As it stands, I can only do them >with Crafty, and you know just as well as I that the results of Crafty >don't quite look as spectacular as what you published from Cray Blitz (for >the obvious reasons that they're using different algorithms). Different algorithms, with different hardware. The hardware was as important as the algorithm... zero-cost spinlocks. near-infinite memory bandwidth. Vectors. etc... > >>>If Vincent wanted to discredit your results, then as far as I'm concerned, >>>he's succeeded 100%. >> >>Fine. If that helps him produce a better speedup, good for him. But the >>speedup number was absolutely produced from raw data... > >You were claiming the exact same thing about the time data 2 hours ago. I will say this again. we do _not_ know about the time data. I am 100% positive about the speedup data. Because all of that was checked over and over by several people. We even had long discussions about how to compute the overall speedup. count each position as 1/24 of the total? Or add all the times together and compute one aggregate speedup? (that would favor the speedup values contributed by longer searches, over the results of shorter searches). So the speedups I _know_ were correct. I personally had raw time values because I had the log files. I didn't have node counts as they were not available in the logs for reasons already given. We do not _know_ if the times are real, or if they are computed backward from the speedups, because we had to construct those tables a year after the paper was submitted... Therefore, I won't make any claim about the times. I make a 100% certain claim about the speedups. And I am 100% certain that the calculated node counts are close enough that if we had real data, no one would notice any significant difference between real and calculated... > >I don't know if you faked the results to look better or not. Maybe I don't >want to know. But whatever be of it, there is little scientific ground >to keep them standing, IMHO. That is up to you to believe or not. Doesn't matter to me. Any more than it matters whether vincent thinks I can get a speedup > 1.4 on 2 cpus or not. What he _believes_ simply doesn't matter. The _results_ matter... And those I stand by, completely... > >-- >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.