Strategic Double Clicking

[Mark Rose]

Originally Posted by chalsall

The point being?

Thought about it. Haven't decided to. There isn't much point to doing it :)

[Madpoo]

Originally Posted by chalsall

The point being?

Probably the same reason I was doing it...just checking up on old code/old results/old people.

Quixotic effort, in all likelihood, but since the computing cost was relatively low (for the tests I did, under 2M) it wasn't a big deal either way. And hey, now we know: those old results were actually decent.

[chalsall]

Originally Posted by Madpoo

Quixotic effort, in all likelihood, but since the computing cost was relatively low (for the tests I did, under 2M) it wasn't a big deal either way. And hey, now we know: those old results were actually decent.

Hey, your and Mark's (et al) kit/time/money. Rock your boat!

In the future compute will be so fast we can triple check everything we've already done.

Perhaps the future is now....

[Mark Rose]

It will probably be fastest to triple check everything after we've found all the primes.

[Madpoo]

Originally Posted by Mark Rose

...Has anything interesting come from the mandatory DC work change, by the way?

Probably too soon to tell. I'll wait and look for changes to these graphs:
http://www.mersenne.org/primenet/graphs.php

I run a daily query to look for exponents that need a triple-check... I guess in theory when more machines are double-checking, I should start seeing more that need a triple-check (about 5% of double-checks should mismatch, if the past is any guide).

The trouble there is that AirSquirrels and I are actively attacking the triple-check list, so it's also going down as a result. Currently ~ 2800 unassigned exponents needing triple-checks, so we'll see.

The other way to know is when we start finding more, previously unknown, bad machines, as a result of those increased double-checks. Even when it's a match, that means *two* machines just got one more "good" result notched up for them, which I use to help guess the winner/loser in a mismatch, so that helps too.

[Dubslow]

Originally Posted by Mark Rose

It will probably be fastest to triple check everything after we've found all the primes.

[Mark Rose]

Originally Posted by Madpoo

The other way to know is when we start finding more, previously unknown, bad machines, as a result of those increased double-checks. Even when it's a match, that means *two* machines just got one more "good" result notched up for them, which I use to help guess the winner/loser in a mismatch, so that helps too.

It would be nice to prioritize DC assignments to give out DC for machines that have never had a result DC'ed. That would expose the historically bad machines quicker. Do know how many machines have returned an LL result without having any of its LL results verified?

[chalsall]

Originally Posted by Mark Rose

It would be nice to prioritize DC assignments to give out DC for machines that have never had a result DC'ed. That would expose the historically bad machines quicker. Do know how many machines have returned an LL result without having any of its LL results verified?

I support that idea.

I don't have the time at the moment to manually reserve and then place onto my machines specific assignments.

But if there was a button somewhere I could click which said "Use my machines to check possibly bad machines" I would click it.

[Madpoo]

Originally Posted by Mark Rose

It would be nice to prioritize DC assignments to give out DC for machines that have never had a result DC'ed. That would expose the historically bad machines quicker. Do know how many machines have returned an LL result without having any of its LL results verified?

In overall terms, just looking at each CPU (both v4/legacy and v5 data) there are 21,162 systems that have never had a verified or bad result... it's just kinda hanging out, waiting to see which way it goes.

Of those, 932 have a mismatch and/or suspect result, but it's pending a triple-check.

6,133 of the 21,162 have just a single unknown exponent, so it's not really worth a preemptive check on them because one way or another, when their solo exponent is checked, that's it, end of the line. Another 4,013 have just 2 unknowns, and the case could be made it's not worth it for those either... if you do test one of them and it's bad, yeah, there's one more you can check to see if it's bad too, but there are bigger fish to fry.

So it's really the 10,153 that have 3+ unknowns, and even then I'd probably start at the top of list (by # of unknowns) and work down, because if someone has 20 unknowns and they started turning up bad, well, that's more interesting than someone with only 5.

It's surprising... there are machines out there that have 100-200+ unknown results and zero bad, zero good. Those are "newer", as in the exponents they did are 50M+ in size, so they're ahead of the DC group.

It just dawned on me that another way to see how the extra DC assignments impact that count of "machines with no good/no bad". When those active systems start doing DC work, we should *hopefully* see "good" results being tallied by them. Or, they mismatch and at least I can see "oh, they have zero good/zero bad, but they do have a mismatch or two...that's curious".

I noted with some amusement that in 74 instances, a machine with zero bad/zero good had a mismatch, but my "guess" for which one was correct (based in this case on the other machine's history) managed to give those 74 systems at least one bad/good result. I really should do my own triple-check on those guesses (if I wasn't the one doing the double-check) to confirm my guess.

For example, in one case a CPU had 5 mismatches, and my algorithm guessed that 3 of them were bad. Two other mismatches were unable to make a prediction, and there's one solo-checked exponent. (no wonder, those 3 "guesses" were because I did double-checks on those and mismatched... I assume my results are always awesome)

The solo is checked out to an anonymous user, of course, so who knows if we'll get an answer to that anytime soon: M41218607

The other two that have been double-checked and await a triple-check (but I couldn't make a good guess) are: M39787061 and M46849639, which I think are both available.

Anyway, that's the kind of fun and in-depth analysis we can really dig into, if we have enough data to get going.

[UBR47K]

Requesting DC on

DoubleCheck=40935859,74,1

I have suspicions that I have a machine going bad.

[Mark Rose]

Originally Posted by UBR47K

Requesting DC on

DoubleCheck=40935859,74,1

I have suspicions that I have a machine going bad.

Will do. I should have it done in a few days.