Re: Mersenne: GIMPS error rates
George suggested that I refine my method for getting rid of duplicate bad results by considering any results with the same exponent and same residue to be duplicate regardless of the userid. I did this and it reduced the number of unique bad results from 9862 to 9826. This did not significantly affect any of the previous stats I posted, so I won't bother posting the minor adjustments for most of the stats. At 01:54:33, Tuesday, 8/27/02 Daran wrote: results. This is because it requires only 2 LL tests to produce two good results, while it takes 3 or more on the same number to produce bad results ( basically, a disproportionate amount of the bad results have not been uncovered yet ). You could scavenge hrf3.txt for non-matching duplicate entries, which would indicate that one of the results is bad, even though you don't know which one. Yeah, that would probably be fairly accurate. Maybe I'll get around to doing that. However, this information could not be used in the error field analysis since hrf3.txt does not contain that information. Of course, lucas_v.txt didn't contain the error field information until last week. Perhaps George could modify the program that produces hrf3.txt. Can you do a moving average of the error rates within the 1,345,000 to 5,255,000 range to see if they increase with the exponent size? Ok, here it is: For each of these 21 results, I have [error rate]%/[total # results] Columns are ( All results, Error field = , Error field ) 1,345 - 1,991.64%/362380.94%/14426 48.21%/112 1,99 - 2,6552.15%/372161.61%/16870 60.24%/166 2,655 - 3,292.27%/352732.01%/16947 61.33%/75 3,29 - 3,9452.67%/353902.49%/17566 50.00%/122 3,945 - 4,598 3.34%/346162.33%/17057 57.28%/302 4,598 - 5,255 3.57%/343552.22%/16782 54.91%/448 1,345 - 5,255 2.59%/213088 1.96%/99648 55.51%/1225 It is interesting that although the error rate has been increasing, the clean run error rate started decreasing in the 3,945 - 4,598 range, while at the same time the total number of tests with errors indicated by the error field went up. It is likely that this indicates a point where George improved the error checking of the program. -- Nick Glover [EMAIL PROTECTED] It's good to be open-minded, but not so open that your brains fall out. - Jacob Needleman _ Unsubscribe list info -- http://www.ndatech.com/mersenne/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Re: Mersenne: GIMPS error rates
I used MS Access to get some results regarding the bad results file on the status page. These results use the versions of BAD and LUCAS_V.TXT from 8/24/02. I deemed any bad results with the same exponent, user id, and residue to be duplicates and removed them. Oh, and my overall error rate was 2.21% ( pretty much the same as Colin's ). Error rate for a clean run in v19 or later: 2.18% out of 114079 results Error rates of all Woltman versions for all numbers and any value in the error field: ( columns are: version, error rate, total # of results ) v13 1.49% 25928 v14 2.40% 50116 v15 3.62% 18014 v16 3.59% 40954 v17 1.87% 37618 v18 3.05% 65273 v19 3.12% 36428 v20 2.57% 46842 v21 1.84% 30174 v22 0.36% 2800 Some older versions where I don't know what the actual version name is: W? 45.00% 20 W1 2.15% 2323 W2 1.09% 14795 W3 0.97% 9758 W4 0.06% 11090 W5 0.09% 8466 W6 0.01% 8067 W7 0.00% 1843 W8 1.04% 1059 W9 0.00% 961 WA 0.00% 1324 WB 0.65% 3561 WC 0.69% 16155 WL 0.72% 971 Some interesting stats regarding reported errors ( the last 4 digits of the error field ): Error field = : 2.02% error rate out of 213369 results Error field : 22.24% error rate out of 5765 results For bad results with an error field: 77.11% had error field = 22.89% had error field I find this interesting because one time I had a machine that had some overheating problems, but it never got any errors during its double check LL test, so I figured the result would be fine. It turned out the result was incorrect. This stat agrees with this because it indicates that if you are going to return a bad result at some point, it is most likely to occur even when the program gives no errors. This is no reason to doubt your results in general; just doubt them if your computer has problems that might cause bad results even if the program did not record any errors. For good results with an error field: 97.90% had error field = 2.10% had error field The above stats apply to all exponents, but below are some stats that apply to only exponents from 1,345,000 to 5,255,000. This leaves some of the lower exponents out which weren't necessarily using George's program and also are so small that errors are extremely unlikely. It also leaves out exponents above the current limit of what has been fully verified since a disproportionate amount of these exponents will be good results. This is because it requires only 2 LL tests to produce two good results, while it takes 3 or more on the same number to produce bad results ( basically, a disproportionate amount of the bad results have not been uncovered yet ). Total error rate for this range: 2.60% out of 213112 results Error field = : 2.19% out of 99045 results Error field : 25.52% out of 1834 results Does anyone want anything else out of this data? I've gotten to the point where I can get most calculations out of it fairly quickly. At 00:02 24/08/02 -0400, George Woltman wrote: What is the error rate if prime95 reported a clean run (version WVn or later and the last 4 digits of the error field are zero)? Nick Glover [EMAIL PROTECTED] Computer Science, Clemson University It's good to be open-minded, but not so open that your brains fall out. - Jacob Needleman _ Unsubscribe list info -- http://www.ndatech.com/mersenne/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Fwd: Correction -- Re: Mersenne: GIMPS error rates
Uh, sorry, there was an error in the stuff based on the error field. Corrections follow: Error rate for a clean run in v19 or later: 2.05% out of 114550 results Error field = : 1.76% error rate out of 214402 results Error field : 38.50% error rate out of 4732 results For bad results with an error field: 67.47% had error field = 32.53% had error field For good results with an error field: 98.64% had error field = 1.36% had error field 1,345,000 to 5,255,000: Error field = : 1.97% error rate out of 99653 results Error field : 55.55% error rate out of 1226 results Nick Glover [EMAIL PROTECTED] Computer Science, Clemson University It's good to be open-minded, but not so open that your brains fall out. - Jacob Needleman _ Unsubscribe list info -- http://www.ndatech.com/mersenne/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Re: Mersenne: GIMPS error rates
On Tue, 27 Aug 2002 00:14:12 -0400, Nick Glover [EMAIL PROTECTED] wrote: [snip] I find this interesting because one time I had a machine that had some overheating problems, but it never got any errors during its double check LL test, so I figured the result would be fine. It turned out the result was incorrect. [snip] Another statistic that would be interesting, but would be extremely difficult to obtain, would be the number of errors vs. the season of the year where the computer is located. I'll bet you would find a higher error rate in the summer months. Not all of us have our computer spaces air-conditioned 24 hours per day. Gary Edstrom _ Unsubscribe list info -- http://www.ndatech.com/mersenne/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers
Re: Mersenne: GIMPS error rates
- Original Message - From: Nick Glover [EMAIL PROTECTED] To: [EMAIL PROTECTED] Sent: Tuesday, August 27, 2002 5:14 AM Subject: Re: Mersenne: GIMPS error rates Some interesting stats regarding reported errors ( the last 4 digits of the error field ): Error field = : 2.02% error rate out of 213369 results Error field : 22.24% error rate out of 5765 results This suggests to me that the behaviour of the program should be changed upon detection of an error. 22% complete - abandon run 22% complete - restart from last saved file (Or 25% given the statistic quoted below.) Perhaps also the program should immediately revert to trial factorisation, until the operator resets it (presumably after fixing the problem). If it has no TF assignments it could overfactor the LL assignments it already has while until it gets one. [...] The above stats apply to all exponents, but below are some stats that apply to only exponents from 1,345,000 to 5,255,000. This leaves some of the lower exponents out which weren't necessarily using George's program and also are so small that errors are extremely unlikely. It also leaves out exponents above the current limit of what has been fully verified since a disproportionate amount of these exponents will be good results. This is because it requires only 2 LL tests to produce two good results, while it takes 3 or more on the same number to produce bad results ( basically, a disproportionate amount of the bad results have not been uncovered yet ). You could scavenge hrf3.txt for non-matching duplicate entries, which would indicate that one of the results is bad, even though you don't know which one. [...] Error field = : 2.19% out of 99045 results Error field : 25.52% out of 1834 results Does anyone want anything else out of this data? I've gotten to the point where I can get most calculations out of it fairly quickly. Can you do a moving average of the error rates within the 1,345,000 to 5,255,000 range to see if they increase with the exponent size? Nick Glover [EMAIL PROTECTED] Daran G. _ Unsubscribe list info -- http://www.ndatech.com/mersenne/signup.htm Mersenne Prime FAQ -- http://www.tasam.com/~lrwiman/FAQ-mers