Author: Robert Hyatt
Date: 19:10:22 09/03/02
Go up one level in this thread
On September 03, 2002 at 20:00:56, Vincent Diepeveen wrote: >On September 03, 2002 at 19:54:41, Robert Hyatt wrote: > >How can you calculate speedup based upon node counts >instead of time? I don't, and I didn't. As I clearly said. I calculated the speedups directly from the log file times. Then I was asked to supply node counts which are _impossible_ to produce due to the way the test was run and the way Cray Blitz (and crafty btw) counts nodes during a parallel search. So. Speedups were computed normally. In the log file, for anyone that has ever seen Cray Blitz run, we produced a continual display of how busy each processor was. I could tell at any point I wanted "how much of each processor has been devoted to searching nodes only? (as opposed to spinning or whatever). Given that data, which in the case of Crafty is the cpu% that it displays in the log files, I can easily compute a node count that will be _very_ close to what is searched. IE in crafty, the cpu% is computed by _me_ and it is not just total cpu time divided by wall clock time. I'll let you look at the source to confirm that it is only time spent actually _searching_ that is counted... That is how the node counts were produced, since I was not doing searches to a fixed depth, because I used the actual game log for the 16 cpu numbers, so that I could just use an 8cpu C90 for the rest which was easier to get access to. Clear enough??? > >Where is the accurate time table if your latest statement now >is that the time table in journal of icca is not correct? The speedup numbers in the JICCA paper are _exactly_ correct. As I have told you all along. I had no idea you were on a node count tangent, even though I have certainly told you often enough "I can't display node counts when the pv changes because I don't count nodes that way... I couldn't do it in Cray Blitz for the same reason." So you should have known what I was talking about, but aparently not... There are three tables in the JICCA paper. 1. speedup. This is correct and was computed by eating the logs and computing speedups by dividing raw times. 2. nodes. Computed from the speedups computed above. I can guarantee that they would be within less than 1% of the real numbers, because I did some fixed- depth testing to make sure that calculation was reasonable. 3. Time. I don't know whether time was computed from 1 above, or whether it was raw. The paper first only had speedup. I added time/nodes a year later after one (or more) referees requested it to be compatible with other papers that did provide them. After talking to my "helper" this afternoon, I realized that I should have remembered the node issue, because I have only explained that to you a dozen times in the past 2-3 weeks as you continue to try to make crafty display data that I have explained can not be displayed. And then you use that flawed data to draw silly conclusions such as "if you didn't have that smp lock overhead, and the 44KB copy overhead, you could have gotten a speedup of > 2. Of course, the fact that your 2 cpu node count was way "understated" (because of not listening to me about trying to do that in Crafty) didn't deter you one bit... > >You seem to have you roriginal text of your paper still at home, >so you must have the logfiles still too? I don't have anything "original". Fortunately I had sent several ASCII copies out before "the crash" and was able to get an ascii version back. I don't have the log files. I don't have any source except for one version Harry found, well prior to the singular extension additions. I don't have any of the carefully saved log files from _every_ game Cray Blitz ever played. I don't have any versions of Crafty prior to 13.x except for a few sent to me after I posted a request on r.g.c.c a few years back. As I said, I lost _everything_ because the disk crashed and we discovered that the backups were unusable... Not a big deal although there was a lot of nostalgia in some of the games that I would have loved to keep "just because"... But of course, I had already told you that I had lost the SE source code when you asked last year. I had already told you that the log files were gone when you asked last week. The only reason I have a fair number of old crafty versions is that I used to keep 2-3 major versions on my ftp site, and _everything_ on my office machine. Luckily when it went down, the 13.x and 14.x versions were still on the ftp... but it was a big loss. Lots of old code dealing with null-move threat extensions, early singular extensions attempts... lots of things I would have liked to have kept... > >Best regards, That rings a tad hollow, IMHO. >Vincent > >>On September 03, 2002 at 18:22:04, Dave Gomboc 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. >>>> >>>>What's far worse, until you were directly accused, there was no indication >>>>whatsoever for all the fiddling that was done with the auxiliary data. When >>>>you were accused, you denied again, until other people supported Vincent's >>>>point of view, when you suddenly got an email from an unknown person you're >>>>not willing to disclose that 'refreshed your memory'. >>>> >>>>Additionally, the only other thing to support DTS, you PhD thesis, appears >>>>to be basically totally unfindable for third parties. >>>> >>>>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. >>>> >>>>If Vincent wanted to discredit your results, then as far as I'm concerned, >>>>he's succeeded 100%. >>>> >>>>-- >>>>GCP >>> >>> >>>I don't agree: the main result is the speedup, which was directly measured -- >>>though I'm certainly not a fan of the estimated node counts being there (at >>>least without something saying "estimated", which the article may or may not >>>have done). >> >>In this case, Vincent's "your bad memory" would actually be right. >> >>The only thing that would have been better would be had he simply asked me >>specifically what he stated in his first post here. He mentioned a week ago >>that he thought my data was "faked". I had no idea what he was talking about >>and I told him _exactly_ what we had done for the speedup stuff.. How I ran >>the tests, taking pondering into account, etc. Never for a minute thought >>he was talking about the nodes or times, because he kept talking about the >>11.1 is fake... >> >>He was right and wrong. The 11.1 was _not_ a faked result. But the node >>counts were definitely calculated because that was the only way I could think >>of to come up with them. >> >> >> >>> >>>But in the main, the least reliable source for information at CCC is Vincent -- >>>indeed, I'd go so far as to say that whenever he claims something, I tend to >>>believe that the opposite is true, unless there are some other, more credible >>>people who agree with him. >>> >>>Dave
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.