Newer milestone thread

[axn]

Originally Posted by Brian-E

But if there was ever a question seriously asked about the integrity of the double-checked residues in the database...

...one need only fire up one's favorite LL testing program and check it again. _That_ is the only thing that matters. The rest is just conspiracy-mongering and no attention should be paid to it.

[axn]

Originally Posted by Brian-E

doing the first-time test and the double-check as a close-knit team of two is a bad idea.

I'll also present a scenario where the opposite is true -- running really big tests (say 100 million digits). Here, it is very irresponsible if you don't simultaneously run doublechecks and frequently compare residues (either by yourself or coordinating with another).

[chalsall]

Originally Posted by Brian-E

I have no idea if it actually occurs amongst the high-throughput community of GPS272, but if it does I would like to see it strongly discouraged by the administrator.

I don't believe any of the group would ever tamper with the results. But if there was ever a question seriously asked about the integrity of the double-checked residues in the database, the whole project should be able to assert that users do not generally have access to the full residue until there have been two matching ones.

To the best of my knowledge, the only two GPU72ers who ever coordinate double-checks / triple checks are LaurV and Dubslow, and they are so fanatical about correctness that they will often run tests individually on their own in parallel or in series to ensure they have the correct results. As LaurV said, for 40902149 he ran the test on two separate cards at the same time to ensure his residues matched.

I fail to see how someone who lowers their own rankings by such behavior could be suspected of falsifying results.

On the other hand, the question of the same user submitting LL and DC results using different UIDs should be taken seriously. Here is an example of a candidate where this appears to be the case. Interestingly, Primenet won't let GPU72 reserve it for Trial Factoring, which suggests to me that it considers it proven.

[LaurV]

Originally Posted by chalsall

so fanatical about correctness

well, say better that I am driven by the fear of missing that prime... (and therefore losing the fame and the money)

[Dubslow]

Originally Posted by chalsall

To the best of my knowledge, the only two GPU72ers who ever coordinate double-checks / triple checks are LaurV and Dubslow, and they are so fanatical about correctness that they will often run tests individually on their own in parallel or in series to ensure they have the correct results. As LaurV said, for 40902149 he ran the test on two separate cards at the same time to ensure his residues matched.

That's only LaurV who's that fanatical. I only ever run an exponent once on my system before passing it off if it doesn't match the first test.

[Brian-E]

Originally Posted by axn

I'll also present a scenario where the opposite is true -- running really big tests (say 100 million digits). Here, it is very irresponsible if you don't simultaneously run doublechecks and frequently compare residues (either by yourself or coordinating with another).

Yes, but the status of the exponent when you finish will (I hope) be "one-LL", not "double checked".

[rcv]

I know I'm blunt. And I believe GPU272'ers are hard-working, dedicated, technically advanced users who want what's best for GIMPS. But you GPU272'ers just can't take criticism.

The scientific method wasn't my main concern, but I appreciate that some users, apparently including Brian-E and LaurV understand that it could be a concern. It's really George's decision, as he is the one to make announcements, and it's his credibility on the line. [I suspect there's a little cross-checking that goes on behind the scenes just prior to an announcement that M(x) is the nth Mersenne Prime.] And the language that is used in the announcement can vary with the circumstances. "All candidates below x have been independently verified by at least two independent users at least two years apart" versus "All candidates below x have been tested twice" versus "At least two reports have been received for all candidates below x".

@Dubslow: If you understood my points, asking me to TC doesn't achieve an independent result. There would still be an excessive number of LL/DC/TC results highly correlated between by some combination of Dubslow, LaurV, and rcv.

@flash: I am certainly capable of writing my own bot. But we both know that's not good for GIMPS. [In fact, I understand the one extant bot already slams the server, and work has been done between George and chalsall to mitigate some of those effects.] Your second post about who should get the "plum" assignments is well taken! That's exactly the kind of thing we should be discussing (see below).

@tarheel: The masthead says there are 88000+ users and 640000+ computers registered. [I'm sure those numbers are inflated.] You can't be serious about asking all who care to change their procedures. Asking people to make significant changes is equivalent to asking them to reevaluate whether GIMPS is their first choice in computing projects.

@axn (with respect to "BS" and "conspiracy mongering"): If you were the editor of a prestigious mathematical journal, what you publish would depend, among a thousand other things, on credibility and independence of the sources. (I think George has earned high credibility, although my recollection of the early years is that he and his project were viewed with some skepticism.)


My biggest concern was that long-time, dedicated, trustworthy PrimeNet users no longer have an equal chance to obtain some of the best assignments. (I hope I'm included in that set or users. My first LL, in 1997, was a little over 2,000,000.) No matter how much GPU272 team members can contribute to LL/DC, you are only a drop in the bucket compared to the traditional Prime95/PrimeNet users.

Now, can we sanely discuss the merits and the alternatives and the pros and cons? Can we attempt to reach some consensus and/or compromise? [Before we engage in wide-spread implementation of a policy that might be detrimental to the overall well-being of GIMPS.]

[chalsall]

Originally Posted by rcv

But you GPU272'ers just can't take criticism.

I don't think that is true.

Could you please provide evidence which supports your claim?

[Dubslow]

Originally Posted by rcv

@Dubslow: If you understood my points, asking me to TC doesn't achieve an independent result. There would still be an excessive number of LL/DC/TC results highly correlated between by some combination of Dubslow, LaurV, and rcv.

My point was that anyone who suspects collusion is absolutely free to run a TC themselves.

Originally Posted by rcv

My biggest concern was that long-time, dedicated, trustworthy PrimeNet users no longer have an equal chance to obtain some of the best assignments.

The problem is determining who exactly qualifies for that, excepting people we know who post here (and thus are at least a little motivated to finish an assignment quickly). You would certainly be trusted, if only because you're raising such a fuss.

PrimeNet's first (and so far only) attempt is the "preferred" assignment system, which notes bounds below which exponents are assigned only to trusted users. (Btw, as an aside to some earlier discussion of those bounds, they have been raised to 48M and 28M repsectively.)

However, you'll note that all that is required to get preferred exponents is two error-free tests (a sufficient but not necessary condition). You'll also note that, by the very existence of this conversation, many "preferred" workers still take a year or more to finish an assignment. What's also been discussed is that a really slow worker gets an exponent before it's designated as "preferred"; then the range becomes preferred, and the rest get cleared out before the slow/non-preferred worker finishes his/her assignment, also holding up milestones.

GPUto73 is the only other method for "trusted" workers to reliably get exponents. Since many "preferred" workers are still really slow, and since when an assignment is released is largely random (unless it's expired by PrimeNet, and even then it's still something of a crapshoot), no worker who is really motivated has any way to get his hands on an exponent without going through GPUto73. The reason we allow it is that anyone who registers for GPUto73 is generally more motivated than the average GIMPSter.

Now, as you've made a large point of, ideally we shouldn't be abusing GPUto73's spiders to get the preferred assignments. However, as I've pointed out, that requires fixing/extending PrimeNet, which is notoriously difficult. If you search back through this thread (and others) there have been many proposals to make a more robust (and typically more layered) "trust" system on PrimeNet. However, until such time as that's implemented, GPUto73 is the best method we have to get such assignments. Anyone who asks chalsall for "plum" assignments (and who can show quick/reliable work) will get them.

[chalsall]

Originally Posted by rcv

@flash: I am certainly capable of writing my own bot. But we both know that's not good for GIMPS.

I always find it very interesting when the first party tells the second party what the second party knows....

[Dubslow]

Originally Posted by chalsall

I always find it very interesting when the first party tells the second party what the second party knows....

Originally Posted by chalsall

I don't think that is true.

Could you please provide evidence which supports your claim?

To play the devil's advocate , neither of these two posts is really responding to the criticism he presents. (The second post responds to the meta-criticism, which doesn't really count.)

@rcv: flash is a large GPUto73 contributor (axn and tarheel are, AFAIK, observers to the project).