Author: Robert Hyatt
Date: 18:16:30 05/07/04
Go up one level in this thread
On May 07, 2004 at 19:41:21, Vincent Diepeveen wrote: >On May 07, 2004 at 13:19:09, Robert Hyatt wrote: > >>On May 07, 2004 at 12:11:29, Gian-Carlo Pascutto wrote: >> >>>On May 07, 2004 at 11:37:00, Robert Hyatt wrote: >>> >>>>I later ran the entire CB test set with and without as you well know. >>> >>>I'll repeat, I don't have any logs from you from that (lying around), and I got >>>no reply to my request to (re)send them. The only big test I have from you >>>is the Opteron one that was just recently posted. >> >>Not my problem. The log was sent to you and Vincent at the _same_ time. Then >>you ran the same test on a quad 550 and produced _another_ log... > >you tested 4 positions to disproof everything. No I didn't. You claimed "null-move affects parallel search big-time." (your words). I ran a quick test and responded "nope." Minor changes. You then made one of your famous "Bob didn't even run a test, he just made up an answer." I responded "nope, here is a short test I ran. Is there a test you would prefer?" You wanted the DTS positions so I ran those as well. Continue to lie, fabricate, manipulate, switch the subject, change the topic, etc. Your credibility is dragging the ground at the moment. Where's that ICCA journal citation? What about Martin's speedup numbers? "opteron is better SMP platform + Crafty sucks at NUMA" is a contradiction. You should keep it up. You look incredibly foolish... > >normally you reply within a minute to anything here. just testing those 4 >positions took you more than 24 hours however, in the meantime you did 50 other >posts around the internet. Didn't take any time at all. sel=0/0 turns null off. It took no effort to run. I only wanted a "quick and dirty idea" about how null might affect the parallel search since my gut-feeling was 'not much'. 4 positions were _more_ than enough to say no. The bigger DTS test set showed the same thing. > >when someone else than you tests crafty, we never can repeat the speedups you >claim. your proof is usually based upon running a position or 4. You can't reproduce anything probably because it is not in your best interest to do so. Others often post parallel search numbers from Crafty here. Nobody has _ever_ claimed I get no speedup on a dual, excepting one idiot tester. Guess who? Again, you like to attack my search and results. And always demand log files. Where are yours? Got something to hide? Ever come up with that JICCA paper citation? Ever find that quote you claimed I made where I claimed my simple linear speedup formula worked for _any_ number of processors? All lies. All typical. > >those take you days to create then. > >how many times do you rerun tests? > >5 times? > >100 times? > Depends. For the opteron test I ran each test 4 times. For my dissertation using 16 cpus I ran each test 100 times. Does that matter? How would it matter? >> >>> >>>>You and Vincent are hung up on a >>>>single number. You get 2.8 so 3.0 or 3.1 is "dead wrong" to use Vincent's >>>>words, even though there is enough variability to choke a mule. >>>> >>>>In any case, _I_ understand the issue and it doesn't cause me any grief. I >>>>don't have to get the same number each time myself... >>> >>>Excuse me, but *who exactly* used proper statistics to illustrate how variable >>>the results *were* or *were not*? >> >> >>I didn't use "statistics". I used raw observed data and posted the numbers here >>many times. Ditto in the DTS paper. Position by position, speedup by >>speedup... > >>But you and vincent keep up with the "2.8 is the number". It is't. It is >>certainly "around 2.8." It is closer to 3.1 as I have said... > >all persons who run tests at quads and just calculated the speedup from that, >they always get other speedup results than you post here. please quote a source. I claim you are a liar, as always. It is about time for you to provide real data rather than hand-waving. Anybody with access to a quad opteron can contact me and I'll send them the exact shell script and input to run the tests and logs I posted for Martin. Where is _your_ data? Or is it so bad you'd rather spend all your time trying to make mine look worse rather than improving yours? > >Should they run it 5 times and pick for each position the most positive results >out of it? > I don't do that... >I am simply not trusting you to proof any speedup number. Speedup numbers posted >from you in articles assume a certain speedup and extend upon that. > Fine. Everything you post is a lie anyway so why should I care what _you_ think. You are not _that_ important. To me you are completely unimportant in fact... >> >> >> >> >> >>> >>>-- >>>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.