Strategic Double Clicking

[Madpoo]

Originally Posted by Brain

Sorry if I'm in the wrong thread but is 48000619 really verified as it says?
I did 2 tests on my own, Titan 1 and Titan 2. 1x with assignment, 1x without.
Feel free to do a triple check...

Yeah, that one popped up in my query on self-verified results, so I've already scheduled it for an independent triple-check.

[Madpoo]

Originally Posted by Prime95

I propose we simply toss all 70+M exponents, meeting some criteria that we agree on, back in the first-time assignment pool. We could run a nightly or weekly stored procedure to do this.

Maybe, although I'd be more comfortable, personally, setting the assignment type back to 'first time check' manually, since the queries and thresholds I use to look for potentially bad stuff is an ever evolving and somewhat fluid concept.

For those 70+M exponents, AirSquirrels expressed an interest in doing some strategic DC in the 4M FFT size which would be in that general range, and I've been tackling a few of them myself here and there.

Right now my goal is to finish doing at least one double-check of all his exponents for each month of each year, just to get an idea of badness/goodness. And then ideally, if I find a bad one, do a second one to determine if there's a trend or not, or if the bad result was a fluke.

There's another user with a similar track record, although not as many turned in each month, but also in that 65-75M range, but it's a little weirder because there's no discernible pattern to his results. Might be good for a few months, bad for a few months, good again, etc. Over the course of several years. Weird.

In a case like that I'd feel bad marking his past results as reassigned as first-time checks, because maybe half of his are good and half are bad.

When I started the strategic double-checking, it was easy to find systems with 20 or even 30 bad for every good result (if they had any at all), and now it's down to systems with 2:1 or maybe just even odds on bad/good, with a few exceptions like the above.

What would help more in the long run is probably to try and do at least one double-check of every CPU that's turned in a result, and then maybe one double-check for each year of that CPU's history.

To give you an idea of what's involved, there are 21,413 machines without any good or bad results at all. 976 of those have at least one mismatched or suspect result, but nothing proven (or guessed) either way.

Breaking it down by cpu *and* year, it's 46,250 total, and 2,046 with at least one mismatch/suspect result.

If we did something to make it a priority of some kind to verify at least one result from each of those, given a historical 5% average of bad results, we'd discover quite a few previously unknown bad apples out there with who knows how many collective results to go through.

That's why AirSquirrels and I are having some fun going through all the exponents needing triple-checks and trying to clear a bunch. We've currently assigned (between the two of us) just about all of those under 40M, then there's another 2500 or so higher than that. Roughly 5% of those will result in a quad check needed, and quite a few results are going to push one of the older runs or the other into the "bad" category.

So that might be another fun work option... "request triple checks".

[airsquirrels]

With some combined effort we could make a dent in ensuring every machine has had at least one result checked. The GPU72 project is about 150+/- days from reaching the caught up point, at which point some more of at least my energy can turn towards LL.

Right now at current levels it would probably take Madpoo and I 2-3 years to independently give every machine at least one check, accounting for those exponents being larger than current DC work. In the process we probably discover a good number of new bad actors to triple check. With the Strategic Double group all aiding in that effort we could catch up much faster. It seems like worthwhile effort to increase the confidence that there aren't primes missing in the DC haystack.

[Madpoo]

Originally Posted by airsquirrels

...
Right now at current levels it would probably take Madpoo and I 2-3 years to independently give every machine at least one check, accounting for those exponents being larger than current DC work. In the process we probably discover a good number of new bad actors to triple check. With the Strategic Double group all aiding in that effort we could catch up much faster. It seems like worthwhile effort to increase the confidence that there aren't primes missing in the DC haystack.

Maybe Curtis would be interested in devoting a wee smidgen of time for that, although in fairness he'd be limited to doing DC work on machines besides his own... still a lot to choose from though. :)

Right now, Curtis' systems have a ratio of 3430 first time to 193 DCs, and he's just using "whatever makes sense" if I recall correctly, leaving the assignment choice up to Primenet. If he (or anyone else) was on board with doing a couple months of "strategic DC" work, I think we'd make some awesome progress.

I'm just thinking of other heavy LL workers out there that might be interested in a short term boost of DC effort. Maybe George and/or others wouldn't be keen to miss out on a couple months of first-time LL work from Dr. Cooper, or maybe it would be hard to implement... I'm just thinking some pie-in-the-sky wishlist stuff.

[Mark Rose]

Well, with the default setting of getting one DC per year, I imagine most of his machines will grab a DC assignment soon, no? So it would only be his retired machines we'd have to adjust.

[Madpoo]

Originally Posted by Mark Rose

Well, with the default setting of getting one DC per year, I imagine most of his machines will grab a DC assignment soon, no? So it would only be his retired machines we'd have to adjust.

Yeah, I expect we'll get more DC assignments in general, but it'd be nice to have a selected pool of exponents being handed out that specifically advance our knowledge of good/bad machines... a guy can dream, right?

[Mark Rose]

Originally Posted by Madpoo

Yeah, I expect we'll get more DC assignments in general, but it'd be nice to have a selected pool of exponents being handed out that specifically advance our knowledge of good/bad machines... a guy can dream, right?

I completely agree. If I weren't dedicating most of my CPU to SoB, I'd be doing "strategic double checks" if the option existed.

I think any result from a machine with no double check, a result with no match, etc., could be lumped into one category of being strategic. When a result is returned, update a new "strategic" column in the exponents table appropriately. Then filter by that column when getting assignments based on a toggle setting like for smallest DC. And you could join-update the strategic column based on your manual analysis, making getting strategic DC automagic.

[UBR47K]

Requesting double checks on:

Doublecheck=N/A,37767209,74,1
Doublecheck=N/A,37713733,74,1

[0PolarBearsHere]

Originally Posted by UBR47K

Requesting double checks on:

Doublecheck=N/A,37767209,74,1
Doublecheck=N/A,37713733,74,1

edit: never mind, madpoo got them.

[kladner]

Originally Posted by kladner

I'll queue this one:
.....DoubleCheck=39205577,71,1

This one finally finished with a mismatch.

[Madpoo]

Originally Posted by Mark Rose

I completely agree. If I weren't dedicating most of my CPU to SoB, I'd be doing "strategic double checks" if the option existed.

I think any result from a machine with no double check, a result with no match, etc., could be lumped into one category of being strategic. When a result is returned, update a new "strategic" column in the exponents table appropriately. Then filter by that column when getting assignments based on a toggle setting like for smallest DC. And you could join-update the strategic column based on your manual analysis, making getting strategic DC automagic.

True enough. Right now my calculations of good/bad machines is being done ad-hoc using some temp tables. If this ever became "a thing" then these types of stats could/should be generated nightly or something and then queried as part of the process to figure out which exponents to assign...

There are some interesting options to explore along these lines, it just comes down to what's workable (and finding time to set some of this up if we go that direction).