mersenneforum.org > Data Bad LL-D Success Rate
 Register FAQ Search Today's Posts Mark Forums Read

 2014-10-12, 17:12 #1 TheMawn     May 2013 East. Always East. 11×157 Posts Bad LL-D Success Rate I was exploring the data on the pretty new website, looking at the top producers. I was curious to see the success rate of double-checks (which isn't entirely fair given that none of us have any control over the integrity of the first test). A good number of individuals on the top-500 have 100% success and some actually have quite a few attempts as well. Near the bottom, though, we have a number of people with a fairly poor success rate. Code: Rank User Credit Att. Succ. Succ. % 297 Fretard Loic 709.048 19 15 78.9% 376 Judge Hale 538.735 14 11 78.6% 222 hoannis 1046.126 26 20 76.9% 49 George Woltman 7018.413 165 124 75.2% 387 Fedor Liudogovsky 516.796 12 9 75.0% 294 ChrisS94 713.285 11 8 72.7% 169 C.D.B. 1482.304 38 27 71.1% 450 XZT 439.278 10 7 70.0% 466 xiaoqi 422.77 10 7 70.0% 475 Jacques MOLNE 414.447 10 7 70.0% 390 antkh 514.832 13 9 69.2% 195 Chad Funderburg 1213.372 28 19 67.9% 458 pdazzl 426.524 6 4 66.7% 146 Viacheslav Chernetskiy 1731.953 44 27 61.4% 269 enrico 790.912 19 11 57.9% 271 Dusty 788.94 19 11 57.9% 395 crun 507.278 14 8 57.1% 201 Carsten Kossendey 1156.846 40 20 50.0% 368 outofoptions 550.113 14 7 50.0% 480 Robert_SoCal 410.896 2 1 50.0% 208 MikeBerlin 1099.267 30 14 46.7% 95 MrRoboto 2613.878 61 24 39.3% 61 Hinesp 4770.176 121 27 22.3% 451 CJCorley 437.784 10 2 20.0% 366 Eng_A_Morra 557.149 14 2 14.3% 464 technobuffalos 423.015 10 1 10.0% 400 Zach Devlin 499.994 14 0 0.0% 150 Manpowre 1674.447 10 0 0.0% Given that the vast majority of people have a >90% success rate it is fairly clear that anyone significantly below that isn't just picking up badly done LL's but is returning bad results.
2014-10-12, 17:20   #2
retina
Undefined

"The unspeakable one"
Jun 2006
My evil lair

659610 Posts

Quote:
 Originally Posted by TheMawn Code: Rank User Credit Att. Succ. Succ. % ... 49 George Woltman 7018.413 165 124 75.2% ... Given that the vast majority of people have a >90% success rate it is fairly clear that anyone significantly below that isn't just picking up badly done LL's but is returning bad results.
So that user is returning bad results? I wonder if he knows anything about computers and programming? :cheeky:

Yeah, I know that isn't a registered smiley, but it should be!

 2014-10-12, 17:25 #3 kracker     "Mr. Meeseeks" Jan 2012 California, USA 1000011110012 Posts You only get a success IF your result matches the other computer's result. They might have the bad result. EDIT: I may be wrong. But this is how I've seen it.. Last fiddled with by kracker on 2014-10-12 at 17:26
 2014-10-12, 17:36 #4 legendarymudkip     Jun 2014 23×3×5 Posts Manpowre 1674.447 10 This is very high for DCs - looks like almost all of them are LL cat4 sort of size.
 2014-10-12, 18:01 #5 srow7   Jul 2014 22×13 Posts I got this assignment, which will basically be a quadruple check I routinely get triple checks. Exponent Status Data 33371803 No factors below 2^71 33371803 PM1 B1=390000,B2=10335000 33371803 LL Suspect;;Miyahara Shuhei;27C700DC9C27F0__ 33371803 LL Unverified;;Matthias M ller;FB325856C46873__ 33371803 LL Unverified;2014-08-14;Tan Loe Joo;E1DB914F3F6C76__ 33371803 Assigned 2014-01-29;ANONYMOUS;D;expired on 2014-03-30 33371803 Assigned 2014-10-07;srow7;D 33371803 History 2012-09-20;Patrik Johansson;NF;no factor from 2^68 to 2^70 33371803 History 2014-01-24;Chuck;NF;no factor from 2^70 to 2^71 33371803 History 2014-08-14;Tan Loe Joo;C;E1DB914F3F6C76__
2014-10-12, 18:02   #6
Prime95
P90 years forever!

Aug 2002
Yeehaw, FL

175318 Posts

Quote:
 Originally Posted by retina I wonder if he knows anything about computers?
Apparently not

That Haswell was stable (apparently barely) at DDR3-2400 for its first six months. After getting a few roundoff errors, I switched to DC and had several failed attempts at getting it stable again. I'm now running the memory at 2133 and all is well again.

I now permanently have one core of both my Haswells running DC so I can monitor for mismatches (that's why you can now click on the exponent in your account results page to quickly see if the result matched).

EDIT: Do you think I should mark the first 6 months of LL tests as suspect so that they get double-checked right now?

Last fiddled with by Prime95 on 2014-10-12 at 18:06

 2014-10-12, 18:16 #7 legendarymudkip     Jun 2014 23·3·5 Posts You can't tell whether they were bad or not. If the machines have a "history" afterwards of being unreliable, I would, but I don't know if that would be the best thing to do.
2014-10-12, 18:40   #8
chalsall
If I May

"Chris Halsall"
Sep 2002

24×5×7×19 Posts

Quote:
 Originally Posted by Prime95 EDIT: Do you think I should mark the first 6 months of LL tests as suspect so that they get double-checked right now?
I would argue it would be better if effort was put into Primenet considered the situation where the DC was actually correct, proven by the TC (or QC), and then correct the percentages for the DC'er (and, potentially, TC'er).

Otherwise, IMO, the "meta stats" as presented by Primenet aren't terribly valuable.

(Personally, I use DCs to ensure the machines I'm responsible for are sane. Such monitoring once determined that a critical machine of mine was close to failing about two months before it actually did. FWTW.)

 2014-10-12, 21:10 #9 Xyzzy     Aug 2002 204748 Posts Vaguely related: We wish there was a work option that consisted of double checks and first time work. You would have the option of setting the ratio. Say 1 double check for every 5 first time tests, or whatever you wanted. Taken further, it would be nice if the client checked to see if the double check you turned in matched the original, and then if it didn't, it would somehow alert the user, maybe with a pop-up message, or an email, or something like that.
 2014-10-13, 05:16 #10 TheMawn     May 2013 East. Always East. 11·157 Posts I saw George's name on that list but I didn't think much of it since I know how much trouble he was having with the Haswell. If all is well now then that is good, but the question of whether the other tests by that machine are suspect is maybe valid. To be honest I wasn't really trying to hint at anything. I tried fairly hard to write another paragraph to make something of the findings but I couldn't really think of anything. I was wondering if there was a mechanism in place to say "Hey, this computer isn't particularly reliable" but it twigged right away that there is. It's called "reliability"
2014-10-13, 05:31   #11
TheMawn

May 2013
East. Always East.

11×157 Posts

Quote:
 Originally Posted by kracker You only get a success IF your result matches the other computer's result. They might have the bad result. EDIT: I may be wrong. But this is how I've seen it..
Quote:
 Originally Posted by chalsall I would argue it would be better if effort was put into Primenet considered the situation where the DC was actually correct, proven by the TC (or QC), and then correct the percentages for the DC'er (and, potentially, TC'er). Otherwise, IMO, the "meta stats" as presented by Primenet aren't terribly valuable.
I like the phrasing "meta stats" because that's exactly what I'm doing. I too question the value of some of the numbers because I've essentially given bias to the people lucky enough to only pick up good first-LL results for DC.

1. How can I look up the exponents for which I have returned a mismatched DC?

2. Does Primenet internally distinguish between "mismatch" and "incorrect" for residues?

3. Would it be valuable to change the headers in the top-500 reports to contain "Attempts" "Matches" "Mismatches" and "Incorrect" to help us see what's going on? I can think of no way to check which of George's 41 mismatches ended up being errors on his end, errors on someone else's end, or still unknown (i.e. a TC is pending).

Kracker brings up a point I tried to address. Primenet only currently gives a "success" if you get a matched residue, not if you get the correct one. Your 100% stable computer can get a "failure" because the original LL test returned a bad residue.

However, because of the overwhelming mass of 90% success rates and higher, it seems that there aren't all that many bad first LL's and it would be very unlucky for George to get 41 bad ones yet for me to only get 2 and for a bunch to have none at all.

Quote:
 Originally Posted by legendarymudkip Manpowre 1674.447 10 This is very high for DCs - looks like almost all of them are LL cat4 sort of size.
You know, I was too concerned over his 0 / 10 to pay any attention to the size of the exponents. He's been quiet on the forums lately but he was pretty big into his GPU's. I wonder if he was using them to double check some of his own big LL's and ran into some issues.

 Similar Threads Thread Thread Starter Forum Replies Last Post paul0 Factoring 2 2015-02-23 03:55 ATH Data 133 2012-03-09 17:19 ATH Data 80 2011-11-15 11:40 lycorn News 631 2010-05-26 10:33 edorajh Lounge 5 2004-08-15 16:51

All times are UTC. The time now is 16:09.

Mon Sep 26 16:09:33 UTC 2022 up 39 days, 13:38, 1 user, load averages: 2.24, 2.14, 2.22