Are P-1 errors detected?
 

Or am I just on a string of bad luck.

I have a couple dozen cores doing P-1 in the 5xM Range for the last 6 months
... for exponents where the prior P-1 has B1=B2.

Overall I am averaging 2.82% (as expected);
but I have 1 PC (the fastest) at about half that success rate.
- 14 factors out of 907 attempts; about 1.54%.
- All in the 50.2 and 50.3M range.
- Current TF level is 73 bits for them.
- I am using B1=1,000,000 B2=20,000,000
- And expect about 2.5% success rate (Because the previous B1=B2 had about a 1.9% and my Bounds are about 4.4)

============================

So is it possible that hardware errors are causing it to miss some?
i.e. is there such thing as an error code or any kind of error checking for P-1?
None of the 14 successes (Factors found) were marked BAD; i.e. not-factors.

Or is it more likely I am just on a bad stretch?

Prior to this work this PC did over 500 DC/LL tests prior to this of which 2 were marked as BAD - Mismatch; the error code was 000000.
0.4% doesn't seem like a bad error rate.


Thanks

Just do another, single DC test (or one per core or whatever). It's certainly the simplest way to verify stability

[QUOTE=Dubslow;484228]Just do another, single DC test (or one per core or whatever). It's certainly the simplest way to verify stability[/QUOTE]

Yeah, I'm doing the same after one of my computers completed over 90 P-1 assignments without finding a factor. Granted, the chance was lower because I'm mostly redoing exponents with only stage 1 done, but I requested a double-check assignment anyway in order to have peace of mind. Funnily enough, that computer found a factor [I]just[/I] after the double check began.

[QUOTE=petrw1;484225]

I have a couple dozen cores doing P-1 in the 5xM Range for the last 6 months
... for exponents where the prior P-1 has B1=B2.
[/QUOTE]

If one of my P-1 machines is suspect I feed it 20 exponents that have been successfully factored by other P-1 boxes. If it scores 20 out of 20 I assume it to be safe. Otherwise the entire range done by that machine is being redone by another machine up from the last point where I am (almost) certain it was functioning correctly.

[QUOTE=petrw1;484225]
- 14 factors out of 907 attempts; about 1.54%.
[/QUOTE]

I suspect it's a string of bad luck. 907 attempts is not a large enough sample size.

If you are still worried, I would try the 20 known factors test.

[QUOTE=tha;484247]If one of my P-1 machines is suspect I feed it 20 exponents that have been successfully factored by other P-1 boxes. If it scores 20 out of 20 I assume it to be safe. Otherwise the entire range done by that machine is being redone by another machine up from the last point where I am (almost) certain it was functioning correctly.[/QUOTE]

Now that's a good idea.

Better than my thought of having another PC redo these 900 no-factor tests at least until a factor is found.

I found a case in which the computer missed a P-1 factor but got a correct LL result: [url]https://mersenne.org/m40470509[/url]

I wonder how often this happens.

[QUOTE=ixfd64;485406]I found a case in which the computer missed a P-1 factor but got a correct LL result: [url]https://mersenne.org/m40470509[/url]

I wonder how often this happens.[/QUOTE]

Often. I find p-1 misses regularly and once tried to redo P-1 work on exponents that had their later first time LL test proven faulty. It appears to be a better case when one does a redo of P-1 work done with a low B1/B2 setting.

[QUOTE=masser;484276]I suspect it's a string of bad luck. 907 attempts is not a large enough sample size.

If you are still worried, I would try the 20 known factors test.[/QUOTE]

I got a little impatient and did 16 and got 100% success. How much more relieved should I be?

IN OTHER WORDS:
On the work it is doing in the 50M range; the P-1 Probability Calculator it tells me I should get about a 2.45% success rate;
the difference between the expected Factor Ratio on the current P-1 (B1=B2=800K)... AND ... the work I am doing (B1=1M; B2=20M).

I am seeing about a 1.57% success rate on this PC; 955 done; 15 Factors
If I expect 2.45% that would be 23 factors.

If statistically it missed 8 factors what would be the statistical odds it would be 16/16 (in other words miss 0/16) in this known factors test?

Thanks for everyone thoughts on this.

I think you are ok, first because there was P-1 done in the range, so the probability is lower than calculated by the tool, and second, even if not, you are still "in the confidence interval" in any case (you found over 65% of the expected factors, i.e. 15/23). You should not be worried at all.

I've got two more questions:

1. Assuming that the same amount of work is done, would a P-1 run be more likely to be bad than an LL test, given that P-1 uses much more memory? For example, consider an LL test of M80614843 and a P-1 run on M332194529 with B1=5,000,000 and B2=130,000,000, both of which require around 250 GHz-days. Would the P-1 run on M332194529 be more likely to have errors?

2. If a computer has flaky hardware, are factors with smaller bounds less likely to be missed?