Strategic Double Clicking

[rudi_m]

Originally Posted by Mark Rose

I still think this should be managed internally. A new work type. A new column for Madpoo to mark, and a new index for finding any SDC efficiently. Apply the DC cat 0/1/2 rules (no cat 3/4). Return normal DC if no SDC available. Prevent DC from returning assignments marked SDC.

Or maybe simply mark the candidates as suspect, so that the normal LL subscribers would do this work.

After one year strategic double checking, spending 55000 GHz Days for 577 exponents, I found about 170 mismatches. In other words I did basically 170 first time checks using 324 GHz Days per exponent. That's only about 30% less throughput than doing current larger first time checks. So people, who want's to find primes should not have any reason to complain about geting SDC sometimes.

As a side effect we would slightly increase the overall DC throughput to keep the gap smaller.

[storm5510]

Originally Posted by Mark Rose

I still think this should be managed internally. A new work type. A new column for Madpoo to mark, and a new index for finding any SDC efficiently. Apply the DC cat 0/1/2 rules (no cat 3/4). Return normal DC if no SDC available. Prevent DC from returning assignments marked SDC...

I see references to categories and a new work type. Does this only apply to Prime95? If one desires to control the data this closely then manual reservations may not be desirable. This would leave CUDALucas users out-in-the-cold, would it not?

[Mark Rose]

Originally Posted by rudi_m

Or maybe simply mark the candidates as suspect, so that the normal LL subscribers would do this work.

After one year strategic double checking, spending 55000 GHz Days for 577 exponents, I found about 170 mismatches. In other words I did basically 170 first time checks using 324 GHz Days per exponent. That's only about 30% less throughput than doing current larger first time checks. So people, who want's to find primes should not have any reason to complain about geting SDC sometimes.

As a side effect we would slightly increase the overall DC throughput to keep the gap smaller.

That works for the low exponents, but for high exponents there is the risk of being assigned to a slow machine and not getting a check for years.

I ran a bunch of SDC of high exponents and not every one was a mismatch, so depending on the exponent range there can be a lot less "first time" throughput.

[Mark Rose]

Originally Posted by storm5510

I see references to categories and a new work type. Does this only apply to Prime95? If one desires to control the data this closely then manual reservations may not be desirable. This would leave CUDALucas users out-in-the-cold, would it not?

It's not so much controlling the data, but automating the process. An available SDC exponent report could always be published, too. That shouldn't be hard to implement either, and it would save Madpoo from copying and pasting results in this thread.

I got tired of fiddling with work queues for fifteen CPUs. So I just run normal DC now.

Automating the process would also allow future functionality such as looking at CPU heuristics and automatically SDC'ing work from flaky machines. Or periodically queuing the lowest exponents from machines without a DC. Or prioritizing TC work. Etc.

[rudi_m]

Originally Posted by Mark Rose

That works for the low exponents, but for high exponents there is the risk of being assigned to a slow machine and not getting a check for years.

I ran a bunch of SDC of high exponents and not every one was a mismatch, so depending on the exponent range there can be a lot less "first time" throughput.

I see the point, but we have the "suspect" category already now, and there is already the risk that it's "just a double check". Assuming now 70-80% of the "suspects" are *really* bad, I see no big problem if Madpoo's semi-automatic selections would decrease this probability to 30-40%.

I remember some years ago I've had a problematic machine and I emailed George to set all it's completed LL''s to "suspect". He did it and all these LL's got double checked quickly. About 25% were bad, all fine.

Anyways to solve your point, maybe all suspect's could be assigned to cat-1 if that's possible to implement.

[Madpoo]

Originally Posted by rudi_m

I see the point, but we have the "suspect" category already now, and there is already the risk that it's "just a double check". Assuming now 70-80% of the "suspects" are *really* bad, I see no big problem if Madpoo's semi-automatic selections would decrease this probability to 30-40%.

I remember some years ago I've had a problematic machine and I emailed George to set all it's completed LL''s to "suspect". He did it and all these LL's got double checked quickly. About 25% were bad, all fine.

Anyways to solve your point, maybe all suspect's could be assigned to cat-1 if that's possible to implement.

Suspect results hover around 50% bad (last time I bothered to analyze the data...maybe a year back).

I'm not totally sure what the "hit" rate of my lists are... if they at least matched the 50% mismatch rate that you get from suspect, it might be worth it, but it tends to vary quite a bit from "yeah, this result is almost 100% going to be bad" to "maybe 10% chance it's bad".

As I've learned from helping implement some of the recent Primenet changes to handle PRP (and George/James did most of the heavy lifting anyway), I realize it's more involved to just add a new work-type... lots of code changes ripple throughout a bunch of stuff.

[rudi_m]

Originally Posted by Madpoo

Suspect results hover around 50% bad (last time I bothered to analyze the data...maybe a year back).

I'm not totally sure what the "hit" rate of my lists are... if they at least matched the 50% mismatch rate that you get from suspect, it might be worth it, but it tends to vary quite a bit from "yeah, this result is almost 100% going to be bad" to "maybe 10% chance it's bad".

My last year hit rate for about 450 - 570 exponents (can't say exactly) was between 30-38%. BTW some of the larger exponents listed here were already marked suspect anyways. My stats have counted them as LL instead of DC.

As I've learned from helping implement some of the recent Primenet changes to handle PRP (and George/James did most of the heavy lifting anyway), I realize it's more involved to just add a new work-type... lots of code changes ripple throughout a bunch of stuff.

Could you manually move an exponent to a certain "cat"? Or is this only possible with continuous ranges?

Maybe just set SDC's suspect and "move" to "LL/cat-x".
Also move triple checks to "DC/cat-x"

[Madpoo]

Here's a new list... I limited this to exponents below 50M just to mix it up a bit.

exponent	Bad	Good	Unk	worktodo
41749109	1	2	1	DoubleCheck=41749109,72,1
42015691	2	1	1	DoubleCheck=42015691,72,1
42205903	1	3	3	DoubleCheck=42205903,72,1
42233071	1	1	1	DoubleCheck=42233071,72,1
42269677	1	2	1	DoubleCheck=42269677,72,1
42292573	1	3	1	DoubleCheck=42292573,72,1
42293327	1	3	2	DoubleCheck=42293327,72,1
42328829	1	2	1	DoubleCheck=42328829,72,1
42354199	7	24	7	DoubleCheck=42354199,72,1
42416909	1	2	1	DoubleCheck=42416909,72,1
42447133	1	3	1	DoubleCheck=42447133,72,1
42597067	1	2	1	DoubleCheck=42597067,72,1
42740827	1	3	3	DoubleCheck=42740827,72,1
42933509	1	3	1	DoubleCheck=42933509,72,1
43085579	10	36	8	DoubleCheck=43085579,72,1
43137767	1	3	2	DoubleCheck=43137767,72,1
43274137	7	24	7	DoubleCheck=43274137,72,1
43292021	2	2	1	DoubleCheck=43292021,72,1
43302949	1	3	1	DoubleCheck=43302949,72,1
43315511	1	1	3	DoubleCheck=43315511,72,1
43358989	1	2	1	DoubleCheck=43358989,72,1
43373563	2	7	1	DoubleCheck=43373563,72,1
43475947	1	2	1	DoubleCheck=43475947,72,1
43523143	1	2	1	DoubleCheck=43523143,72,1
43546729	2	6	1	DoubleCheck=43546729,72,1
43559209	2	6	1	DoubleCheck=43559209,72,1
43574287	1	2	1	DoubleCheck=43574287,72,1
43593229	1	2	1	DoubleCheck=43593229,72,1
44000899	2	2	3	DoubleCheck=44000899,72,1
44030093	1	3	2	DoubleCheck=44030093,72,1
44071549	1	2	1	DoubleCheck=44071549,72,1
44155721	2	2	3	DoubleCheck=44155721,72,1
44185403	1	3	3	DoubleCheck=44185403,72,1
44271839	1	2	1	DoubleCheck=44271839,72,1
44354813	1	2	1	DoubleCheck=44354813,72,1
44367319	8	18	9	DoubleCheck=44367319,72,1
44450647	7	24	7	DoubleCheck=44450647,72,1
44522461	1	2	1	DoubleCheck=44522461,72,1
44587723	10	36	8	DoubleCheck=44587723,72,1
44602699	2	3	3	DoubleCheck=44602699,72,1
44664989	1	3	1	DoubleCheck=44664989,72,1
44693777	8	18	9	DoubleCheck=44693777,72,1
44716459	1	2	2	DoubleCheck=44716459,74,1
44729521	1	2	1	DoubleCheck=44729521,72,1
44736221	10	35	6	DoubleCheck=44736221,72,1
44752111	1	3	4	DoubleCheck=44752111,72,1
44777701	9	17	21	DoubleCheck=44777701,72,1
44838923	2	6	2	DoubleCheck=44838923,72,1
44842349	9	17	21	DoubleCheck=44842349,72,1
44849569	1	3	1	DoubleCheck=44849569,72,1
44858269	1	3	4	DoubleCheck=44858269,72,1
44890003	1	3	4	DoubleCheck=44890003,72,1
44938771	10	35	6	DoubleCheck=44938771,72,1
44976683	1	2	8	DoubleCheck=44976683,72,1
45054767	1	2	2	DoubleCheck=45054767,74,1
45056507	1	2	4	DoubleCheck=45056507,72,1
45085349	8	18	9	DoubleCheck=45085349,72,1
45089663	2	4	3	DoubleCheck=45089663,72,1
45128969	2	4	3	DoubleCheck=45128969,72,1
45166157	1	3	4	DoubleCheck=45166157,72,1
45230543	1	3	4	DoubleCheck=45230543,72,1
45308309	2	4	3	DoubleCheck=45308309,72,1
45309533	1	3	2	DoubleCheck=45309533,72,1
45322567	4	14	5	DoubleCheck=45322567,72,1
45323713	4	14	5	DoubleCheck=45323713,72,1
45329093	1	3	1	DoubleCheck=45329093,72,1
45389653	1	2	2	DoubleCheck=45389653,72,1
45422627	8	18	9	DoubleCheck=45422627,72,1
45495647	12	18	1	DoubleCheck=45495647,72,1
45501793	1	1	2	DoubleCheck=45501793,72,1
45599839	2	3	2	DoubleCheck=45599839,72,1
45623021	4	11	34	DoubleCheck=45623021,72,1
45675743	1	2	2	DoubleCheck=45675743,72,1
45702901	1	3	2	DoubleCheck=45702901,72,1
45719057	2	7	1	DoubleCheck=45719057,72,1
45732961	4	11	34	DoubleCheck=45732961,72,1
45770513	1	1	2	DoubleCheck=45770513,72,1
45770519	1	1	2	DoubleCheck=45770519,72,1
45782861	9	17	21	DoubleCheck=45782861,72,1
45943889	1	1	5	DoubleCheck=45943889,72,1
45954151	10	35	6	DoubleCheck=45954151,72,1
45958819	2	6	1	DoubleCheck=45958819,72,1
46020787	1	2	2	DoubleCheck=46020787,73,1
49110287	1	2	4	DoubleCheck=49110287,73,1
49137329	1	2	4	DoubleCheck=49137329,73,1
49176067	1	2	11	DoubleCheck=49176067,73,1
49298971	1	2	5	DoubleCheck=49298971,73,1
49334843	1	2	1	DoubleCheck=49334843,73,1
49346747	1	3	2	DoubleCheck=49346747,73,1
49395569	1	2	6	DoubleCheck=49395569,73,1
49512989	1	2	2	DoubleCheck=49512989,73,1

[Prime95]

To mix things up I took from the middle: 43.6M to 44.0M

[Madpoo]

Originally Posted by Prime95

To mix things up I took from the middle: 43.6M to 44.0M

Cool, I edited the list and removed those to be helpful to the next volunteer.

[rudi_m]

I tool all below 45M.