Strategic Double Clicking

[Madpoo]

Originally Posted by sdbardwick

We match.

Ah, the famous "Phil Frakes" account was the bad one there. When I first started this project out on my own, looking for bad systems, that was the very first account I identified as having an unusually high amount of bad stuff.

Since then I've broken it down more so instead of just by the user's account, I look at the CPU and now also by year. By those breakdowns his account isn't as bad as I originally thought, but of the 211 total records (account/cpu/year) for him, only 27 of those had more bad than good. Total, he has 721 good, 173 bad, and another 261 unknown.

Not the worst track record out there in terms of the hit/miss ratio.

So, not to single him out or anything, but it did trigger my interest in the idea and helped me refine the analysis.

[LaurV]

Originally Posted by Madpoo

Good... I'm already doing triple-checks on 3 more of your self-verified work.
You may have noticed that I picked up a bunch of your older results and did a double-check for you, free of charge. LOL

You were very welcome to do that for "self-doublechecked" part of my old work, especially for v4 work (where I was contributed with over 50 computers, but gave up when EFF money were awarded - some of those computers did produce bad results, and DC/TC made sense).

But it makes no sense for you to continue to TC my own, new self-DC work, especially if the first check and the DC came in the same time day:hour:minute and at least one of them has a valid UID, and they have different shifts*. Which is almost everything above 50M. You just waste your time, you can use the resources to do a more useful job. Enough people knows me here already to say I won't "cheat", I am a "credit whore" but I don't go over some borders

Moreover, I have no reason to "hide" a prime.

I was doing LL tests in parallel, in different hardware, with different shifts, and the tests were combined in such a way to take about the same time, so I can compare the residues at every check point, like 1M or 100k iterations, or so, depending on the exponent, if they are different, then BOTH tests will resume from a previous checkpoint. If they match, previous checkpoints are deleted (always only the last two are kept, this in case the last one had errors when the file was written, but never happened). This is as simple as comparing file names, because cudaLucas saves the residue in the file name (I had long arguing with people compiling/working on cudaLucas, to get this feature inside, you may not remember, you are kind of "newer" here ).

And this saves me a lot of time, and for the project too, because when a mismatch is found, no time is wasted - contrarily to the classical work path when a test with a mistake in the beginning will still continue to the end, and be reported, to find the mismatch, for both LL and DC, a lot of wasted time.

When the tests end with a match, I report both. Because I have done both. I used two times the amount of hardware, even if the time was spent only once. Why should I not get credit for both? This is (partially) the reason we pushed to implement "shifting" in cudaLucas. The reports come in the same time.

I also proposed (in fact, sustained, it was not my proposition, but other's forum member) in the past to implement a feature on the server side to be able to mark an exponent for TC, with a low priority. If that would be implemented, you could use it to mark all my self-DC work, instead of wasting your time to TC it.

I can understand that my hardware can produce crap and some of my tests are bad, so other people will DC them and find them bad, but under no circumstance I would accept the fact that a self-DC which I did and matched could be found bad. You TC them if you like, is your time and resources.

And I will continue to do that, in spite of what you or others say, because I feel better to know that I didn't miss a prime , and I didn't report a wrong result by mistake.

----------
* all these conditions have some importance to avoid the case when someone can report fake results using my name, but I don't think anybody has any reason, and would waste his/her time to do such a stupid thing anyhow.

[chalsall]

Originally Posted by chalsall

The lowest six of my runs from this batch have now completed; three matched, three didn't.

OK, all fourteen from this batch have now completed. Five matched, nine didn't. Looks like you're onto something here Aaron.

[Madpoo]

Originally Posted by chalsall

OK, all fourteen from this batch have now completed. Five matched, nine didn't. Looks like you're onto something here Aaron.

Woot!

Here are some more. These are 3:1 bad/good and for this query, I loosened up what I consider an "awesome" system to include anything with 15 good and zero bad. Most are from a single new system that has 10 solo-checked stuff... If someone grabs all of these, I might suggest trying that smallest one first just as a trial to get an idea of how the rest might go. Still, that same system has 7 suspect/7 mismatches which means all of those were probably bad too and just awaiting a triple-check.

In fact, when the # of suspect is close to the # of mismatches, that means someone double-checked their suspect results and got a different result, so more often than not (like 90%+ of the time) those will be bad...you can mentally include those in the bad column if it helps. LOL

exponent	Bad	Good	Unk	Sus	Solo	Mis	worktodo
43101913	3	0	4	0	4	0	DoubleCheck=43101913,72,1
43102177	3	0	4	0	4	0	DoubleCheck=43102177,72,1
43985531	4	1	2	2	2	2	DoubleCheck=43985531,72,1
44802613	3	1	10	7	10	7	DoubleCheck=44802613,72,1
44954269	3	1	10	7	10	7	DoubleCheck=44954269,72,1
45678811	3	1	10	7	10	7	DoubleCheck=45678811,72,1
46013971	3	1	10	7	10	7	DoubleCheck=46013971,72,1
46296641	4	1	2	2	2	2	DoubleCheck=46296641,72,1
47973161	3	1	10	7	10	7	DoubleCheck=47973161,72,1
48686327	3	1	10	7	10	7	DoubleCheck=48686327,72,1
48761711	3	1	10	7	10	7	DoubleCheck=48761711,72,1
49334413	3	1	10	7	10	7	DoubleCheck=49334413,72,1
49482271	3	1	10	7	10	7	DoubleCheck=49482271,72,1

[chalsall]

Originally Posted by Madpoo

Here are some more.

Mine.

[Madpoo]

Originally Posted by LaurV

You were very welcome to do that for "self-doublechecked" part of my old work, especially for v4 work (where I was contributed with over 50 computers, but gave up when EFF money were awarded - some of those computers did produce bad results, and DC/TC made sense).

I'm just OCD about it I guess. Nothing personal. I know you wouldn't cheat, but on general principle I think it "looks cool" if different a different person does the double-check.

And yeah, you do have the occasional bad result. In fact, if you want, you can do a triple-check of this one I just turned in a result for:

M45196597

[LaurV]

Originally Posted by chalsall

Looks like you're onto something here Aaron.

Agree, my second one is also a mismatch, up to now, two from two. Interesting they were done by the same user before. They are both still unassigned.

Originally Posted by Madpoo

And yeah, you do have the occasional bad result. In fact, if you want, you can do a triple-check of this one I just turned in a result for:
M45196597

Sure boss! Queued! Thanks! (it seems that was not a GPU test, which computer does it come from?) scrap that, it was a GPU test, I see from James' site. I have to go home to look into the logs to see which card sent it (in few hours. Lunch break here).

[endless mike]

Originally Posted by endless mike

Maybe a few more

I'm halfway through the 24 triple checks that I took earlier in the thread; no quadruple checks needed yet. It does mean that Madpoo should hopefully have a few more machines for his list that are no longer perfect.

[henryzz]

@Madpoo It might make sense if you import the data into R and do some machine learning to find candidates. Logistic regression and random forests would be fairly decent for this.

[firejuggler]

a long overdue doublecheck on M36433591

Verified 2007-09-13 Yu Peng Chen 4103C5EBC1F27D48
Verified 2015-11-02 firejuggler 4103C5EBC1F27D48

[LaurV]

Originally Posted by Madpoo

M45196597

And done, matching yours.
Note that the period is mid-of-Jan to mid-of-March 2012, when I was testing cudaLucas (switching from powers of 2 to non powers of two), as we already discussed. So this is not a hardware failure, for the peace of my heart
You may find more like that, from the same period.

Next time however, when you find a mistake in my tests, you should find a prime! Otherwise we both are wasting the time...