Strategic Double Clicking

[Mark Rose]

Originally Posted by endless mike

It looks like these all finished within the last few days while I was at work. I matched you on all of them.

Thank you!

Also thanks to ATH!

[Mark Rose]

Here is the list of unassigned Double Checks from this thread:

DoubleCheck=50454631,73,1
DoubleCheck=55299799,73,1
DoubleCheck=55435829,73,1
DoubleCheck=55516297,73,1
DoubleCheck=55551479,73,1
DoubleCheck=55555963,73,1
DoubleCheck=55623727,73,1
DoubleCheck=55635631,73,1
DoubleCheck=55875949,73,1
DoubleCheck=55984871,73,1
DoubleCheck=56020733,73,1
DoubleCheck=56020819,73,1
DoubleCheck=56114419,73,1
DoubleCheck=56131703,73,1
DoubleCheck=56385377,73,1
DoubleCheck=56443847,73,1
DoubleCheck=56458279,73,1
DoubleCheck=56702339,73,1
DoubleCheck=56715583,73,1
DoubleCheck=56754437,73,1
DoubleCheck=56801161,73,1
DoubleCheck=56826013,73,1
DoubleCheck=56958299,73,1
DoubleCheck=57014579,73,1
DoubleCheck=57143551,73,1
DoubleCheck=57484201,73,1
DoubleCheck=57652883,73,1
DoubleCheck=57664729,73,1
DoubleCheck=57670841,73,1
DoubleCheck=57792487,73,1
DoubleCheck=58374847,73,1
DoubleCheck=58773223,73,1
DoubleCheck=58974449,73,1
DoubleCheck=65021563,75,1
DoubleCheck=65022787,75,1
DoubleCheck=65022829,75,1
DoubleCheck=65022913,75,1
DoubleCheck=65070311,75,1
DoubleCheck=65070547,75,1
DoubleCheck=65070689,75,1
DoubleCheck=65070727,75,1
DoubleCheck=65095861,74,1
DoubleCheck=65434511,75,1
DoubleCheck=65434597,75,1
DoubleCheck=66437677,75,1
DoubleCheck=66437729,75,1
DoubleCheck=66437773,75,1
DoubleCheck=66437809,75,1
DoubleCheck=66455297,75,1
DoubleCheck=66823261,75,1
DoubleCheck=67943873,74,1
DoubleCheck=67943873,75,1
DoubleCheck=68002423,75,1
DoubleCheck=68565313,75,1
DoubleCheck=68685511,75,1
DoubleCheck=68841527,75,1
DoubleCheck=68912377,75,1
DoubleCheck=69021133,75,1
DoubleCheck=69332441,74,1
DoubleCheck=69414973,75,1
DoubleCheck=69614147,75,1
DoubleCheck=69636493,75,1
DoubleCheck=70351027,75,1
DoubleCheck=70402373,75,1
DoubleCheck=70402417,75,1
DoubleCheck=70663909,75,1
DoubleCheck=71227873,75,1
DoubleCheck=71274391,75,1
DoubleCheck=71861917,75,1
DoubleCheck=72029689,75,1
DoubleCheck=72054019,75,1
DoubleCheck=72116249,75,1
DoubleCheck=72162301,75,1
DoubleCheck=72237013,75,1
DoubleCheck=72486503,75,1
DoubleCheck=72486509,75,1
DoubleCheck=72638201,75,1
DoubleCheck=72639043,75,1
DoubleCheck=72639407,75,1
DoubleCheck=72832189,75,1
DoubleCheck=73000217,75,1
DoubleCheck=73000237,75,1
DoubleCheck=73000241,75,1
DoubleCheck=73000267,75,1
DoubleCheck=73000273,75,1
DoubleCheck=73000289,75,1
DoubleCheck=73000471,75,1
DoubleCheck=73000549,75,1
DoubleCheck=73000561,75,1
DoubleCheck=73000583,75,1
DoubleCheck=73000607,75,1
DoubleCheck=73000619,75,1
DoubleCheck=73000639,75,1
DoubleCheck=73000687,75,1
DoubleCheck=73000883,75,1
DoubleCheck=73000913,75,1
DoubleCheck=73000927,75,1
DoubleCheck=73001119,75,1
DoubleCheck=73003673,75,1
DoubleCheck=73116257,75,1
DoubleCheck=73116607,75,1
DoubleCheck=73135091,75,1
DoubleCheck=73138057,75,1
DoubleCheck=73139081,75,1
DoubleCheck=73143079,75,1
DoubleCheck=73143311,75,1
DoubleCheck=73146001,75,1
DoubleCheck=73146041,75,1
DoubleCheck=73146043,75,1
DoubleCheck=73146079,75,1
DoubleCheck=73146841,75,1
DoubleCheck=73146901,75,1
DoubleCheck=73147051,75,1
DoubleCheck=73148011,75,1
DoubleCheck=73333049,75,1
DoubleCheck=73385537,75,1
DoubleCheck=73385681,75,1
DoubleCheck=73446011,75,1
DoubleCheck=73469003,75,1
DoubleCheck=73469057,75,1
DoubleCheck=73477013,75,1
DoubleCheck=73481381,75,1
DoubleCheck=73491073,75,1
DoubleCheck=73499077,75,1
DoubleCheck=73514201,75,1
DoubleCheck=73566947,75,1
DoubleCheck=73647019,75,1
DoubleCheck=73648027,75,1
DoubleCheck=73649033,75,1
DoubleCheck=73649089,75,1
DoubleCheck=73659077,75,1
DoubleCheck=73692011,75,1
DoubleCheck=73692041,75,1
DoubleCheck=73692083,75,1
DoubleCheck=73694801,75,1
DoubleCheck=73697053,75,1
DoubleCheck=73697069,75,1
DoubleCheck=73698103,75,1
DoubleCheck=73755637,75,1
DoubleCheck=73888403,75,1
DoubleCheck=74064449,75,1
DoubleCheck=74271289,75,1
DoubleCheck=74396849,75,1
DoubleCheck=74417531,75,1
DoubleCheck=74683481,75,1
DoubleCheck=74953217,75,1
DoubleCheck=75015701,75,1
DoubleCheck=75015803,75,1
DoubleCheck=75089249,75,1
DoubleCheck=75161209,75,1
DoubleCheck=75256669,75,1
DoubleCheck=75322987,75,1
DoubleCheck=75477893,75,1
DoubleCheck=75694397,75,1
DoubleCheck=75818773,75,1
DoubleCheck=75875957,75,1
DoubleCheck=76006829,75,1
DoubleCheck=76077787,75,1
DoubleCheck=76134293,75,1
DoubleCheck=76141001,75,1
DoubleCheck=76176511,75,1
DoubleCheck=76193773,75,1
DoubleCheck=77187083,75,1
DoubleCheck=77187181,75,1
DoubleCheck=77888123,75,1
DoubleCheck=79395941,75,1

And three suspect results that will be assigned for a new first-time LL:

DoubleCheck=71862187,75,1
DoubleCheck=71862229,75,1
DoubleCheck=73154539,75,1

[rudi_m]

Originally Posted by Mark Rose

Here is the list of unassigned Double Checks from this thread:

DoubleCheck=50454631,73,1

I guess this one is correct. I've checked many exponents from "Jan van Niekerk", posted firstly here
http://mersenneforum.org/showthread....631#post413651

[Mark Rose]

Originally Posted by rudi_m

I guess this one is correct. I've checked many exponents from "Jan van Niekerk", posted firstly here
http://mersenneforum.org/showthread....631#post413651

I was wondering why it was no longer reserved by you :)

[endless mike]

From the triple check listed Madpoo posted I took

50000429	DoubleCheck=50000429,73,1
50028227	DoubleCheck=50028227,73,1
50030707	DoubleCheck=50030707,73,1
50059861	DoubleCheck=50059861,73,1
50066651	DoubleCheck=50066651,73,1
50073451	DoubleCheck=50073451,73,1
50214673	DoubleCheck=50214673,73,1
50219087	DoubleCheck=50219087,73,1
50290931	DoubleCheck=50290931,73,1
50292971	DoubleCheck=50292971,73,1
50308267	DoubleCheck=50308267,73,1
50316529	DoubleCheck=50316529,73,1
50317901	DoubleCheck=50317901,73,1
50368427	DoubleCheck=50368427,73,1
50436173	DoubleCheck=50436173,73,1
50456453	DoubleCheck=50456453,73,1
50457371	DoubleCheck=50457371,73,1
50478761	DoubleCheck=50478761,73,1
50506513	DoubleCheck=50506513,73,1
50509889	DoubleCheck=50509889,73,1
50516021	DoubleCheck=50516021,73,1
50539081	DoubleCheck=50539081,73,1
50552189	DoubleCheck=50552189,73,1
50557649	DoubleCheck=50557649,73,1
50581847	DoubleCheck=50581847,73,1
50592107	DoubleCheck=50592107,73,1
50647801	DoubleCheck=50647801,73,1
50658007	DoubleCheck=50658007,73,1
50663017	DoubleCheck=50663017,73,1
50668547	DoubleCheck=50668547,73,1
50672927	DoubleCheck=50672927,73,1
50734591	DoubleCheck=50734591,73,1
50743633	DoubleCheck=50743633,73,1
50744017	DoubleCheck=50744017,73,1
50745449	DoubleCheck=50745449,73,1
50793713	DoubleCheck=50793713,73,1
50797427	DoubleCheck=50797427,73,1
50828207	DoubleCheck=50828207,73,1
50843557	DoubleCheck=50843557,73,1
50867627	DoubleCheck=50867627,73,1
50883047	DoubleCheck=50883047,73,1
50884123	DoubleCheck=50884123,73,1
50902693	DoubleCheck=50902693,73,1
50921669	DoubleCheck=50921669,73,1
51032617	DoubleCheck=51032617,73,1
51035741	DoubleCheck=51035741,73,1
51044563	DoubleCheck=51044563,73,1
51057593	DoubleCheck=51057593,73,1
51064217	DoubleCheck=51064217,73,1
51081739	DoubleCheck=51081739,73,1
51172777	DoubleCheck=51172777,73,1
51186683	DoubleCheck=51186683,73,1

[GP2]

Originally Posted by rudi_m

I guess this one is correct. I've checked many exponents from "Jan van Niekerk", posted firstly here
http://mersenneforum.org/showthread....631#post413651

When datamining for juicy double-check targets, you look at some user+computer names and it seems like a promising vein because there is a pretty high ratio of bad results or mismatches. So you start double-checking all the other exponents returned by that user+computer, only to find them all turning out good, one after another...

Often the explanation is simple. LL results that have an associated error code are more likely to be bad (to be more precise, the ones with asymmetric error codes), so those got double-checked right away, while the results without an error code only get checked considerably later. So the statistics for a given user+computer can get front-loaded with bad results, making it look like a very promising vein to mine... and then you start double-checking the normal non-error-code results and they're all good....

Of course, for some other user+computer names, it turns out that they mostly churned out errors regardless of whether the error code got set or not... and then you've found the motherlode.

[Madpoo]

Originally Posted by GP2

When datamining for juicy double-check targets, you look at some user+computer names and it seems like a promising vein because there is a pretty high ratio of bad results or mismatches. So you start double-checking all the other exponents returned by that user+computer, only to find them all turning out good, one after another...

Often the explanation is simple. LL results that have an associated error code are more likely to be bad (to be more precise, the ones with asymmetric error codes), so those got double-checked right away, while the results without an error code only get checked considerably later. So the statistics for a given user+computer can get front-loaded with bad results, making it look like a very promising vein to mine... and then you start double-checking the normal non-error-code results and they're all good....

Of course, for some other user+computer names, it turns out that they mostly churned out errors regardless of whether the error code got set or not... and then you've found the motherlode.

Sometimes that is indeed the case, although I've noticed that for some really bad systems, the bad stuff they check in tend to be "error free" as far as they know.

More often what happens is that sometimes I don't have enough info on a system to break it down into results by year or month (they may only turn in a few per year) so instead I wind up looking at their total results over their lifespan.

In those cases, there may in fact be only a small period of time over a few scattered months or a year where it spat out bad results, but up until then it was doing okay.

I find that I have the strongest predictability when a machine spits out a lot of results each month. That simply provides more data to go on and it's kind of simple to see how it trends good/bad over time.

For the low output systems, the full picture only reveals itself as more double-checks are done, and by the time we know if they were trending good or bad, we may have finished all of their double-checks.

By way of example, consider a computer that ran for 2 years and turned in 10 first time checks. Maybe one of them was found bad and the rest were unknown. Speculatively, I may have picked the lowest unchecked exponent of theirs and did another test, finding it bad as well. Now I have 2 of them and I start to think "aha!" and distribute their other work to the group. But then we start getting more and more good results turning up.

Once they're all done, I could look back and see that the 2 bad results were actually the last two it turned in over that 2 year span, and the other 8 were done earlier and the machine must have been stable. But that 2 year span could in fact span 3 calendar years, and no one month in there had more than a single result, so there really is no trend to work from besides a hunch.

That's actually far more common that you might imagine... lots of folks just run (or ran) Prime95 for fun on a low end machine...it wasn't fast, but it ran all the time and turned in a decent amount of work before going off to silicon heaven. We only find out years down the road that it had some bad times...

That's why I'm hopeful the new assignment rules will help out. Assigning periodic double-checks by default to make sure systems are on the straight and narrow should help us find out about these things further in advance.

Now that AirSquirrels and I have the end of the triple-checking backlog in sight, and the low hanging fruit of the strategic double-checking has been spotted, I may turn my attention to doing those validation runs again of systems that have never had a single result double-checked. I'm sure most of them will turn out okay (should be 95%, if the statistics hold true) but I may also find new bad machines in the process.

My methodology on that is pretty simple... just find machines without any double-checks being done on their work and pick their lowest exponent to DC. Oh, and only for systems with more than X amount of unverified work. Doesn't make much sense to go after systems with only 1 or 2 results because by the time you've spotted a trend, it's all over. LOL

I guess I could pick their *last* or *first* result rather than the lowest exponent... I just do the smallest one because it'll test quicker. Their last result tends to be larger and I want to cover more ground.

[rudi_m]

Originally Posted by Madpoo

I just do the smallest one because it'll test quicker. Their last result tends to be larger and I want to cover more ground.

That's understandable, but larger exponents may have higher probability of having bad results. Maybe we should go for something better like "validating at least 100 GHz-Days of each machines past calculations".

Also we could try to check exponents first which were checked during the hot times of the year Is there any statistics about the bad-LL distribution other the months?

[S485122]

Originally Posted by rudi_m

...
Also we could try to check exponents first which were checked during the hot times of the year Is there any statistics about the bad-LL distribution other the months?

You would have to know where the machines are to be able to do that.

Jacob

[bgbeuning]

Originally Posted by Mark Rose

I have a few low exponents that need a triple, if anyone is interested:

DoubleCheck=40171801,72,1
DoubleCheck=48405233,72,1
DoubleCheck=48603307,72,1

These are done and matched your results.
I should mention the machine doing these checks had
its first "C - Bad" result a couple of days ago.

[Mark Rose]

Originally Posted by bgbeuning

These are done and matched your results.
I should mention the machine doing these checks had
its first "C - Bad" result a couple of days ago.

Thanks!