Newer milestone thread

[ric]

Originally Posted by rudy235

So what is the Percentage of Failure? (meaning by that that the first and the second result are not the same)

On my (quite limited) sample of Madpoo's Strategic Double Checks, which is naturally a very biased sample, over 60%. YMMV.

[VictordeHolland]

Originally Posted by rudy235

So what is the Percentage of Failure? (meaning by that that the first and the second result are not the same)

Historically 3-4% of the tests are bad, with a few spikes in ranges with >10% failure. Usually these are at FFT crossover points and the first 10,000,000 digits test (M33,000,000+)
http://mersenneforum.org/showpost.ph...&postcount=102

[Dubslow]

Originally Posted by rudy235

So what is the Percentage of Failure? (meaning by that that the first and the second result are not the same)

This thread is devoted to the topic: http://mersenneforum.org/showthread....883#post449883

[GP2]

Originally Posted by VictordeHolland

Historically 3-4% of the tests are bad

It's worth mentioning that the 3–4% error rate is just an average over all machines. Many machines have a perfect or near-perfect record, while a few have very high error rates.

Originally Posted by VictordeHolland

with a few spikes in ranges with >10% failure.

Actually, from the graph, the most salient feature is the set of regular downward spikes, starting at around 20M and repeating four more times at even intervals just under 5M apart.

A bit of data mining might shed some light on this, but it's more fun to speculate idly. If this was a user or group of users with highly reliable machines doing LL tests ahead of time in order to test the software ahead of the wavefront, then you'd expect the intervals to be exactly 5M apart and exactly aligned with human-friendly thresholds like 20M, 25M, 30M, 35M and 40M; instead, they're just a bit off. But you wouldn't really expect these little islands of hyper-reliability to be an artifact of the algorithms either, so it seems likely to be some kind of selection effect.

[VictordeHolland]

Originally Posted by GP2

Actually, from the graph, the most salient feature is the set of regular downward spikes, starting at around 20M and repeating four more times at even intervals just under 5M apart.

A bit of data mining might shed some light on this, but it's more fun to speculate idly. If this was a user or group of users with highly reliable machines doing LL tests ahead of time in order to test the software ahead of the wavefront, then you'd expect the intervals to be exactly 5M apart and exactly aligned with human-friendly thresholds like 20M, 25M, 30M, 35M and 40M; instead, they're just a bit off. But you wouldn't really expect these little islands of hyper-reliability to be an artifact of the algorithms either, so it seems likely to be some kind of selection effect.

They roughly correspond with the underlined?
range - FFT
29.69M-34.56M - 1792K
34.56M-39.50M - 2048K
39.50M-49.10M - 2560K
49.10M-58.52M - 3072K

[GP2]

Originally Posted by VictordeHolland

They roughly correspond with the underlined?
range - FFT
29.69M-34.56M - 1792K
34.56M-39.50M - 2048K
39.50M-49.10M - 2560K
49.10M-58.52M - 3072K

Hmmm, mprime does some extra checking around crossover points, but I think this is mostly to see if it needs to go to a higher FFT size. Can it really be that this extra checking has the side effect of turning unreliable machines (overheated, overclocked, cheap parts, etc) into nearly perfectly reliable machines? If this were true, then the server and software should cooperate to turn on this extra checking all the time for machines that have a track record of being unreliable. A priori it would seem more likely that George or others do exponents near the crossover points before the wavefront reaches them, as a way of doing quality assurance of the software.

[flashjh]

Originally Posted by rudy235

In some way the issue of the 71'000,000 milestone is not resolved.

The Prime Minister Residue is 42A93C867B552E__
while Albert Pettersson is 9C157058B13DB9__

Thus one of the two is not correct (or both )

So I hope someone is able to take care of that by doing a triple check

https://www.mersenne.org/report_expo...0723879&full=1

Verified C from Albert Pettersson's run

[rudy235]

Originally Posted by flashjh

Verified C from Albert Pettersson's run

That is good. The poaching was 100% not worth it.

[LaurV]

Originally Posted by rudy235

Thats the good thing of having a Forum. We get closer to the truth, one post at a time.

I love this line!

[Madpoo]

Originally Posted by rudy235

Thats the good thing of having a Forum. We get closer to the truth, one post at a time.

True dat.

I had a bizarre experience at another forum (IPCamTalk) with one of the admins of the forum freaking out at me. I wish I'd paid more attention to the mood of the place before signing up; some of them (including the guy who runs the place) are borderline insane as it turns out.

I wound up getting banned, but seems I'm in good company... I think he just bans everyone he disagrees with (and then brags about his user base of 30K people). It was amusing to be accused of not knowing anything about networking. Entire threads exist where nearly everyone who posted has a big "Banned" note.

On the bright side, it made me appreciate the people here. Yeah, we have some characters (ahem... RD Silverman... ahem...) but even they're mostly harmless and give the place character. When new folks show up and ask the same questions, Mike doesn't tear them a new one if they're confused about anything or ask a question that's in a FAQ somewhere.

Point being, we're better than them. LOL

[rudy235]

Originally Posted by Madpoo

On the bright side, it made me appreciate the people here. Yeah, we have some characters (ahem... RD Silverman... ahem...) but even they're mostly harmless and give the place character. When new folks show up and ask the same questions, Mike doesn't tear them a new one if they're confused about anything or ask a question that's in a FAQ somewhere.

Point being, we're better than them. LOL

Yes I do love this forum. To make things clear I am no mathematician and I no longer contribute to GIMPS, however I am very enthusiastic about very large prime numbers. I find it amazing that with methods like ECPP we can prove primality of any form of random number under 30,000 digits. We now have about 120,000 primes over 1000 digits. Up from about 170 known only 33 years ago. At that time the largest Mersenne (now M30) had less than 40,000 digits.

What will the future bring? The first Sierpinsky prime over 10,000,000 digits should be coming soon, as well as the first Mersenne over the Classic limit of 79.3 million digits.

Yes everyone here is very helpful and I for one I'm not afraid of asking stupid questions, although of course I try to make them sound intelligent.