mersenneforum.org > Data Project Unsportsmanlike: gunning for curtisc
 User Name Remember Me? Password
 Register FAQ Search Today's Posts Mark Forums Read

 2003-09-27, 19:02 #1 GP2     Sep 2003 2×5×7×37 Posts Project Unsportsmanlike: gunning for curtisc The following are exponents that: 1) have been LL tested once 2) have been LL tested by curtisc! 3) have been tested on machines that may be more error-prone than average, based on past historical results 4) are not currently assigned by Primenet Doing an early double-check of these exponents could: a) Find a mismatch with curtisc's result, which means he loses credit when a triple-check is done later to confirm that his result was bad. b) Find a Mersenne prime that curtisc missed (OK, highly unlikely). Warning: the error-proneness of these machines has been determined based on very low stats. For instance the machine that returned the first 3 exponents below has one known-good result and two known-bad results for an error rate of 66%. Warning: curtisc actually has a fairly low overall error rate, whereas Team_Prime_Rib has a number of error-prone (presumably overclocked) machines. So he could turn around and test some of TPR's error-prone first-time tests and cause TPR to lose credit. Ordered in decreasing order of likeliness to be erroneous. The top exponents were for a machine with 66% error rate, the bottom one for a machine with 6% error rate (all based on very low stats, however). DoubleCheck=13695317,65,1 DoubleCheck=17514929,65,1 DoubleCheck=18500561,66,1 DoubleCheck=16760893,65,1 DoubleCheck=16000987,65,0 DoubleCheck=16164389,65,1 DoubleCheck=16749487,65,1 DoubleCheck=17723819,65,1 DoubleCheck=16179931,65,1 DoubleCheck=16480921,65,1 DoubleCheck=17218867,65,1 DoubleCheck=18124811,66,1 DoubleCheck=14605973,65,1 DoubleCheck=16665697,65,1 DoubleCheck=12562483,64,1 DoubleCheck=16853293,65,1 Last fiddled with by GP2 on 2003-09-30 at 02:02
 2003-09-27, 19:09 #2 GP2     Sep 2003 A1E16 Posts There are a number of cases where we might want to do early double-checks of exponents. George is already doing this for first-time tests that return most kinds of nonzero error code. It could also be done for results returned by machines that are suspected of being error-prone based on previous stats. The problem is, the server only hands out the currently lowest available exponent for double-checking. So if the leading edge of double-checking is sweeping through the 10.2M range, it will be a long time (maybe years) before an exponent in the 17M range gets assigned for automatic double-checking. Ideally, there should be some mechanism where the server occasionally randomly tosses out a non-lowest-available exponent for early double-checking. Then these could be done for Primenet credit. But if the only way to do such early double checks is by manual testing, then it could be a project for volunteers. George how do you currently handle early double checks of nonzero-error-code results? Do you just do them yourself, or is there a more formal mechanism?
 2003-09-27, 21:47 #3 garo     Aug 2002 Termonfeckin, IE 276410 Posts GP2, George has a system for this. He tosses all those exponents that have non-zero error codes back in the pile as LL tests. He does it manuall so it happens once every few months. Since they are assigned as LL tests they are on the top of the pile and get assigned quickly. However, this method does not deal with exponents that are from suspect machines but with zero error codes. You may want to send him a list of such exponents and he will toss them in the next time he does a manual release. All these exponents including the ones you listed above should easily work with Primenet and for Primenet credit. There is no reason to run them as manual tests. You wonn't be breaking anything by running them with Primenet.
 2003-09-28, 04:40 #4 dswanson     Aug 2002 110010002 Posts For the past several months shaneamy and I have been running doublechecks of all numbers in the range 10.95M - 11.0138M (see threads Doublechecking 10950053 - 10987343 and Doublechecking 10987349 - 11013853). Among these we've doublechecked 64 curtisc exponents without finding a single error. But, in the spirit of this thread, we have eliminated four curtisc exponents via P-1: Code: Exponent Factor 10987841 3602179992480329067961 10953673 1197987487463370010447 10956797 2035213159147168949089 10974919 111036620818926603303137
 2003-09-28, 06:06 #5 garo     Aug 2002 Termonfeckin, IE 1010110011002 Posts GP2, Why are you gunning for curtisc? He is #2. If anything, you should be going after #1. Especially since TPR has many "latent" bad tests due to bad overclocks.
2003-09-28, 12:39   #6
GP2

Sep 2003

2×5×7×37 Posts

Quote:
 Originally posted by garo GP2, Why are you gunning for curtisc? He is #2. If anything, you should be going after #1. Especially since TPR has many "latent" bad tests due to bad overclocks.
Actually I'm not, I'm just getting into the spirit of this "mercenary" forum by suggesting some exponents that some of you might find interesting to work on, since there's a lot of TPR folks here.

But you're right, it's better to avoid manual tests.

After the next release of data files (which ought to be in a day or so) I should have a fairly complete list of "error-prone" first-time exponents, so we can take it from there.

 2003-09-28, 19:03 #7 outlnder     Aug 2002 2×3×53 Posts You will find a lot of bad results and triple checks needed from outlnder boxes. Do not worry, all those boxes are gone. I no longer crunch with them.
 2003-09-28, 23:34 #8 garo     Aug 2002 Termonfeckin, IE 276410 Posts GP2, that was meant to be tongue in cheeck since I'm from TPR too!! BTW, our folks were ambitious with their overclocking but more importantly relied on the torture test to see if an overclock is stable or not. Well, turns out the stress/torture test was not reliable and machines that passed it still returned bad results. If anything, our experience has shown that the best way to check the stability of the machine is to run a few DCs and wait a week to see the results show up in the LUCAS_V.TXT file. DO NOT rely on the torture test, Anyway, I think your efforts are wonderful, I had many of the things you have done on my wishlist for a long time and time was the only thing that had prevented me for so long. Regardless it is really good of you to do all this work. Much appreciated!
2003-09-29, 15:49   #9
GP2

Sep 2003

2·5·7·37 Posts

Quote:
 Originally posted by garo GP2, George has a system for this. He tosses all those exponents that have non-zero error codes back in the pile as LL tests. He does it manuall so it happens once every few months. Since they are assigned as LL tests they are on the top of the pile and get assigned quickly. However, this method does not deal with exponents that are from suspect machines but with zero error codes. You may want to send him a list of such exponents and he will toss them in the next time he does a manual release. All these exponents including the ones you listed above should easily work with Primenet and for Primenet credit. There is no reason to run them as manual tests. You wonn't be breaking anything by running them with Primenet.

I think you're right. If George is willing I'll just send him a batch of such exponents after the next release of the data files (which will hopefully include the computer-id in HRF3.TXT).

I think it would be legitimate to release exponents from error-prone machines as "first time" LL tests because there's a much higher than average probability of getting a different residue. So for people who request only first-time tests because they hope to be the next Mersenne prime discoverer, I don't think they'd mind getting these exponents to work on.

All machines could be classified as either "error-prone", "not error-prone", or "unknown". The unknown machines would be the ones that have returned nothing but large exponents and none of them have ever been verified as good or bad. So Mersenne-aries could do manual double-checks of small random samples of exponents from such "unknown" machines, to definitively classify every machine as either "error-prone" or "not error-prone". In this case, it might not be legitimate to release such exponents to Primenet as "first time" LL tests, because they really would be just routine double-checks with no more than average (low) chance of getting a different result, and people who request true first-time LL tests because they hope to find a prime might be annoyed to get assigned such exponents.

So to summarize: the "curtisc" exponents above probably should just get sent to Primenet for reassignment. But there will be some other manual LL tests for Mersenne-aries to do...

 Similar Threads Thread Thread Starter Forum Replies Last Post NBtarheel_33 GPU to 72 11 2013-08-05 07:43 NBtarheel_33 PrimeNet 70 2012-08-07 11:06 schickel Aliquot Sequences 307 2011-10-28 01:29 schickel Aliquot Sequences 29 2011-08-12 17:45 ATH Miscellaneous Math 4 2006-08-30 17:59

All times are UTC. The time now is 15:18.

Wed May 12 15:18:31 UTC 2021 up 34 days, 9:59, 0 users, load averages: 2.80, 2.76, 2.81