Author: Robert Hyatt
Date: 09:00:32 09/04/02
Go up one level in this thread
On September 04, 2002 at 10:23:07, Vincent Diepeveen wrote: >On September 04, 2002 at 09:48:14, David Dory wrote: > >If you would have followed the thread you will see i tried to >talk with Bob about it at 13 august for about 3 hours in a calm >way, but it was not possible. Everything was denied. Even >that recursive nullmove (R=3 i use) was not recognized as >something that changes speedup, despite even crafty showed >with asymmetric king safety a 0.2 difference in speedup. > >2.8 versus 3.0 > >Bob denied it and also denies the 2.8 speedup but claims 3.1 instead. Vincent, you are _way_ too full of yourself. I have _not_ denied 2.8. I have not only claimed 3.1... I have _clearly_ _and_ _consistently_ said that the average speedup for Crafty, over a _bunch_ of test positions, fit the curve speedup = 1 + .7 * (NCPUS -1) when I did the test. The parallel search has not changed that much in several years. When I ran the 24 CB positions for you, and gave you the logs, did you see 2.8 or 3.0? Have I not said _many_ times that the parallel speedup is a very dynamic value that can change significantly on the same position run multiple times? have I not given you several examples of such? Do you not understand that you can run a test once and get 2.8, and run it again and get 3.1? If you don't, I can't help you at all. But, since you are in this "bob denies" mode, why didn't you respond to the stuff I posted about the smp vs non-smp crafty? You directly claimed that the version compiled with smp support was "far slower" (your words) than the non-smp version. I posted a test showing that one searched something like 400K nodes per second, the other searched 401K (or whatever the actual numbers are, they are in the post near the bottom of the messages). No response. You said Crafty only gets 1.4 speedup, at best. Eugene at microsoft ran several tests and got 1.9 on a couple of PIV systems, and got 1.4 on a couple of AMD machines. You are using AMD. And somehow, you think it is ok to make AMD "the machine" that must be used to compute a speedup? I posted a one, two and four cpu test in that same post (about the smp vs nonsmp versions). Again, you didn't respond. You like to make wild claims, and when called on them, you run and hide and try to go on a different angle of attack. I can't do much about the old Cray Blitz data. C90s are not available. The source code for CB is long-gone. It would be easy enough to re-run the tests if machine time were available and the '93 version of the program was around. But they aren't. however, for Crafty, the story is different. Several different people have run parallel tests. Some reported here, including Eugene. Others reported in r.g.c.c. And _everybody_ is getting 1.7+ for two processors. Just because _you_ can't get that result using what is apparently problematic hardware doesn't change the fact that _others_ are getting the results. I have, on multiple occasions, run a few test positions with 1,2 and 4 processors and posted the raw data here, usually when someone asks "how much faster is a dual or a quad, really?" I prefer to run the test, show the data, and move on. You prefer to wave your hands. Make false statements. Make wild claims, Produce "proofs" that are nothing more than "it is a proof because I said it is" and continue your ranting and raving. You make statements about how Fritz, or Rebel, or Tiger, or whatever _absolutely must_ be doing something. With no source code to prove it. Only the "this is what they must be doing because I say it is the only way it can be done." You have a serious problem... >His own few testpositions then he emailed to me and he claimed 3.0 >at it, not 3.1. Now he refers again last weeks here at CCC as >3.1 to it. As I said, you are doing the same thing here that some do with Elo numbers. They are _not_ absolute. I can run the same test multiple times and get varying numbers. Bruce used to have a position that, for the longest, caused my program to produce a .1 speedup with four processors. yes, it ran ten times slower on four than one. But not every time. For that problem I could claim .1, 1.0, 2.0, or 3.1 depending on which number came up when I ran it. 3.1 is a good average number over a lot of positions. I have seen much better numbers. And I have seen some down around 2.5-2.8. But 3.1 is a good average, whether you like it or not... > >I showed that every idiot with a parallel search can get a good speedup >in the positions Bob used for his crafty proof. Yes, you moved a lot of air, mainly by waving your arms. I often use the kopec test positions. And "every idiot" has not gotten great speedups using them, as they have been used in almost all of the parallel search result papers written, with the exception of the DTS article. If you don't like the kopec positions, fine. I do. Although those are not the only thing I use to test speedup. Fortunately, you don't get to pick _my_ test positions. I don't try to pick yours. > >If he also misquotes all numbers to me, like not mentionning difference >between analysis mode in crafty NOT using asymmetric king safety and he >does, it is simply *denied* that things like singular extensions are >very bad for speedup, especially when running at a lot of processors. I misquote? That change was made to crafty years ago. Because users requested that it not flip-flop the scores so badly as it changes sides. however, that doesn't do anything statistically to parallel search, so it really doesn't matter... It is just an evaluation change that turns off king safety asymmetry in analysis mode, which means that the scores in analysis mode are different than the scores in a real game. What is the deal here? What did I "misquote". you made an assumption. A bad one. And it is because I "misquoted" something. yet another word you need to look up in your dictionary. To "misquote" I have to first "quote and then get it wrong." I didn't quote _anything_ about analysis mode to you. You chose to use that yourself... > >Even if you show hard proof it is denied. If you show 30 positions >of crafty, not to mention the 100+ i ran here at DIEP, it is denied and >a week later forgotten. What are you talking about now? What "30 positions". Maybe before you make a claim, you ought to do what you said you did in the claim? You said "no way you can get a speedup on the cray blitz DTS positions. I ran them here and your speedup was 1.01." Remember that? I then ran them and gave you the logs. Remember that? Do you recall the speedup? Was it not 3.0? Did I run those positions again with null-move off? Do you recall the speedup? Was it not 3.1? Did you not claim that turning null-move off would make a _huge_ difference? Did I not claim that it would not make any significant difference? Is .1 significant? Again, your rants are meaningless... I will be happy to post all the data here so that everyone can take a look at what you claim, and what is reality... Or someone with a quad can run them for themselves. GCP did. On one of my 550 quads. He got 2.8 on the machine he used. I got 3.0 on the machine I used. Is either number wrong? Is either right? Or is the real answer somewhere close by both??? It certainly isn't down at 1.01 where you claimed it was, is it? > >How can i speak with someone who emails me that crafty copies 44KB >data structure and nalimov finds out it is 3KB? By paying attention? More than once I said "max of 44kb". You could certainly look at CopyToSMP() and CopyFromSMP() to see for yourself. You know what is in my tree structure. You _know_ I would not need to copy the entire move list for all plies, just the one at the split point. I assume a mediocre understanding of what is going on when I talk to you. Sometimes that assumption is really bad, of course. Remember, however, the follow-up. You chastised GCP for copying too much. Then you chastised me for copying too much. Then you went on and on for several emails explaining why my large copy was killing my parallel performance and so forth. Were your conclusions valid? Were they garbage? you jump to conclusions faster than anyone I know. And those conclusions are generally _way_ wrong. This is just another example. > >How can i speak with someone who claims last few weeks >he removed last ply pruning out >of crafty before 1997, where i can show from source code that it >was still in it at 1999? I never claimed that. Don't know where you think I did. It had no relevance to the discussion anyway. > >How can you speak with someone who claims deep blue searched 12 ply >with unimaginable extensions, then later says they searched 17-18, >and even for 6 months doesn't look in hard data written by DB team >themselves, and even then i still didn't hear a statement from bob >they searched 11-12 ply. They claim 12.2 ply, whatever they want, >11 or 12 i don't care. If they say it's 12.2 fine with me. But not >17-18 ply. I don't care what you believe. I asked them directly about the X(Y) notation. Others have as well. We all reported that we got the _same_ answer. You claim it is X plies total, 4 in hardware. Someone here asked last week for you to explain 3(4). How can you do three plies total, four in hardware? So yes, there are some inconsistencies. But not in what _I_ claim. Perhaps in what they have written. the X(Y) explanation they gave me doesn't match some of the results in the paper. The X(Y) explanation you claim _also_ does not match some of the results in the logs. So the truth lies somewhere in that range. It isn't for me to decide... > >How can you speak with someone who claims in conversation to me >that crafty gets beaten by Cray Blitz with 5-0 without ever showing >proofs, games or whatever? :) It wasn't 5 0. It was 7-3. I did it for my own personal interest. I never thought to save any of it as it was not interesting, since I would probably never get another chance to run Cray Blitz again. Do I need to save everything I do for you? I don't think so. You said Cray Blitz was far weaker than anything today, and that you could beat it 10 0 or some such nonsense. I simply reported _my_ result that was not so good, and since you can't beat _my_ program 10 0, I concluded (and told you so) that you probably couldn't beat CB 10 0 either if it had good hardware. > >In fact he quotes that every year. In fact i have email where i ask >for a Cray executable, NWO here has some Crays you know... you never asked for a cray executable. You asked for the source. What did I tell you? "Vincent, I lost _everything_ a couple of years ago due to a hard disk crash, only to discover that all the backups were no good either. I don't have any source so that you can try to decode the DTS stuff, not that you would have made a lot of sense of the Cray assembler stuff anyway." Remember that conversation? It is _still_ true. I might could find an executable if you had a cray it could run on, but that would be a big "if". Cray doesn't have memory mapping hardware so things like shared libraries don't exist. To run the few games I played a year or two ago, I had to get someone to find a very old version of Unicos that was compatible with the MP libraries I used in that particular executable. It wasn't easy. But you didn't ask for an executable, you wanted source code because you couldn't figure out some things in DTS by yourself. > >"i don't have it someone else did the test" Never said that at all. _I_ ran the 10 games as I told you. NO idea where you get the above made-up answer from... If you look back in the CCC archives, you will see that I _clearly_ said that I ran the tests on a whim one afternoon when I was contacted by a friend that had a machine that was being taken down and sold for scrap. He said "hey, interested in running a few tests? all users are off but the machine will be up for the next couple of days..." That wasn't uncommon. Someone sent me a circuit module from the machine we used to win the 1983 WCCC event, because they knew which serial number (201) we used and it was being scrapped. Cray folks are like that. > >I am not someone to spit deeper back than 1997 but i'm sure that if you >really want to hurt you should spit in that thesis and run a statistical >analysis on it. Spit away. But be aware that the data in my thesis is different. It is not raw. It is averaged over _many_ runs. For reasons given in the thesis in fact. I simply didn't have the machine time to do that for the DTS article, as much as I wanted to. > >Not a single logfile from cray blitz exists anymore. I have it in my >email box, but the numbers of the paper get posted here within a few >seconds when he needs them... I have a scanned/OCRed copy of the paper. I have been emailing that to people for several years. I have offered to send it to those interested as a result of this fiasco as well... So yes, I have the paper. But that is _all_ I have, other than some hand-written numbers in a paper file here where I did some of the calculations and wrote down speedups by hand. I really don't think there is much more I can add to the discussion. If you want to continue the bogus claims about Crafty, those I can handle because I can get several people to run tests and report speedups. Eugene has done so twice within the past two weeks. Of course, I am sure you can find something wrong with his results as well, you usually do, when something doesn't agree with your version of reality... I'm out of the DTS discussion, as I don't see any more that I could say that would clarify things further. The speedups are exactly what they claim to be. The nodes were calculated based on the speedups. The times I simply don't recall. have fun...
This page took 0.01 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.