Strategic Double Clicking

[Madpoo]

Originally Posted by frmky

I just submitted the first two results of the eight I reserved. Both matched the previous runs.

http://www.mersenne.org/report_expon...4654643&full=1
http://www.mersenne.org/report_expon...4682537&full=1

Cool. I didn't know anyone was going to start double-checking any, but that's cool.

The list I put up for extra TF work was using looser rules for figuring out if work was done by a bad machine or not. Once all the extra TF'ing is done, I'll generate a list of work done where it's far more likely the first residue will be bad.

I don't know what it is, but it's kind of fun and satisfying to do a test and realize your result (which is hopefully the correct one) is different from whomever ran it earlier. I had 6-7 more today.

[chalsall]

Just for those "playing", I've made a quick report showing the Strategic Double Check TF'ing status. This isn't linked anywhere from the site as this is temporary and of special interest only.

Separately, Mark found a factor a little while ago, and I noticed that as soon as it was factored the status changed to show the full residue marked as "Verified (Factored)".

Since this whole exercise is to find bad machines, perhaps showing the full residue isn't a good idea, since there's a chance the residue isn't in fact correct.

Thoughts?

[Mark Rose]

Originally Posted by chalsall

Just for those "playing", I've made a quick report showing the Strategic Double Check TF'ing status. This isn't linked anywhere from the site as this is temporary and of special interest only.

Awesome! I was wondering if the 70-bit 50-52M was assigned or not. I'll scoop those.

Edit: Or not. What Makes Sense seems to ignore the exponent range. No big deal :)

Separately, Mark found a factor a little while ago, and I noticed that as soon as it was factored the status changed to show the full residue marked as "Verified (Factored)".

Since this whole exercise is to find bad machines, perhaps showing the full residue isn't a good idea, since there's a chance the residue isn't in fact correct.

Thoughts?

I agree.

I also found a factor for M47549387, which is in the list.

[chalsall]

Originally Posted by Mark Rose

Edit: Or not. What Makes Sense seems to ignore the exponent range. No big deal :)

Ah... Thanks for pointing that out. A legacy conditional. Please try again.

[Madpoo]

Originally Posted by chalsall

Just for those "playing", I've made a quick report showing the Strategic Double Check TF'ing status. This isn't linked anywhere from the site as this is temporary and of special interest only.

Separately, Mark found a factor a little while ago, and I noticed that as soon as it was factored the status changed to show the full residue marked as "Verified (Factored)".

Since this whole exercise is to find bad machines, perhaps showing the full residue isn't a good idea, since there's a chance the residue isn't in fact correct.

Thoughts?

"Factored" status on LL tests is tricky... it's kind of a no-go status in terms of determining computer quality. We may never know in those cases.

Sometimes people go back and do extra factoring on an exponent that was already verified. I guess I could look for any cases where there was a reject residue that should be included in the "known bad" column.

For the factors found when it was still unverified, yeah, just no way of knowing. I could at least look at their error code and see if it was originally marked "suspect".

For now I'm not even bothering to include the "factor found later" runs in my tallies.

EDIT: By the way, I'm going to employ a little bit of a cheat when counting how many bad results a computer has done.

In the cases where I've done a double (or triple+) check and there's still no match, I'm going to claim technical superiority and say my result is correct even though it's unverified. After all, I'm running systems with ECC at stock speeds in climate controlled datacenters. I have a mere 3 results that were bad, and in all 3 I'm inclined to chalk them up to some kind of Prime95 issue. If you look at them you'll see why (the first few bits of the residue are zero, but the rest match). Plus, they were all part of my manic "triple check everything below 2M" thing:
M8291
M12281
M801883

I think that gives me solid footing to claim I'm right, they're wrong, etc. I wonder if I could expand that to do the same for other mismatches, and if one is from a very solidly reliable computer, assume that's right and the other is bad, for purposes of this project.

[Mark Rose]

Originally Posted by chalsall

Ah... Thanks for pointing that out. A legacy conditional. Please try again.

Didn't work. I asked for 96 assignments to 72, 50M to 53M, WMS, and it gave me back a bunch of 41M assignments. I tried the same with to 71, and got a bunch of 43M assignments.

I'm giving myself the sub-sub-sub project of taking the sdc exponents above 50M to 72, then to 73.

[chalsall]

Originally Posted by Mark Rose

Didn't work. I asked for 96 assignments to 72, 50M to 53M, WMS, and it gave me back a bunch of 41M assignments. I tried the same with to 71, and got a bunch of 43M assignments.

Weird. Yes, I saw (in the logs) that the code is changing your 50M to be 40M.

Could you please give it one more go? I've removed every case where the "Low" is modified to be 40M.

[Mark Rose]

Originally Posted by chalsall

Weird. Yes, I saw (in the logs) that the code is changing your 50M to be 40M.

Could you please give it one more go? I've removed every case where the "Low" is modified to be 40M.

Same result. Oh well, it's not like there's a huge amount of TF'ing to be done. I'll just do it in the order GPU72 currently wants to give it :)

[chalsall]

Originally Posted by Mark Rose

Same result. Oh well, it's not like there's a huge amount of TF'ing to be done. I'll just do it in the order GPU72 currently wants to give it :)

Grrrr... I hate Newton's Method of software debugging....

There was yet another legacy conditional, which this time brought your low down to 35M. This has been removed.

Sorry about this; the next time you try it /should/ work as requested.

[Mark Rose]

Originally Posted by chalsall

Grrrr... I hate Newton's Method of software debugging....

There was yet another legacy conditional, which this time brought your low down to 35M. This has been removed.

Sorry about this; the next time you try it /should/ work as requested.

Qapla'!

Thanks for the bug fixes :)

[Madpoo]

Okay, let's do some strategic double checking.

To start out here I've generated a list of exponents that are pretty likely to wind up with one of our double checks NOT matching the first one. The criteria here is:
Bad > 6 and Bad > Good*6 and Unknown > 0

Essentially meaning they have at least one unknown (so there'd actually be something to check), and they either have 6 times as many bad as good, or if good happens to be zero they have 6+ bad ones.

The list includes the exponent, and then for each exponent there's a worktodo entry plus some stats on the computer that did the first time check: # of bad, # of good, # of unknown, and # of suspect

I have a hunch that in these cases, if they have a bunch of proven bad results, the suspect results are probably also bad but they're waiting on triple checks to find out for sure. I don't factor them into the calculation, but if you're eyeballing some likely suspects then feel free to use that # as a guide as well.

These are all currently unassigned, so I'd say you can try what George mentioned and simply add it to your worktodo and I guess the server will create an assignment for it when it phones home.

I'm starting out small with these 16 just to make sure the process works (plus there's a good # assigned out to GPU72 for carrying them forward the few extra bits).

Oh, and before picking any of these up, look at the exponent report page and see if it's already assigned to someone.
The URL for that is simply:
www.mersenne.org/M<exponent>
and look in the assignment section for any active ones.

Be sure to report results back here for feedback... like did your residue match, or did you get a different (and hopefully correct) one? That'll guide us going forward.

exponent	Bad	Good	Unk	Sus	worktodo
35064059	10	1	5	0	DoubleCheck=35064059,71,1
36497473	14	2	3	2	DoubleCheck=36497473,71,1
36517909	19	3	4	0	DoubleCheck=36517909,71,1
36572957	19	3	4	0	DoubleCheck=36572957,71,1
36702241	9	1	1	10	DoubleCheck=36702241,71,1
36779077	7	1	4	1	DoubleCheck=36779077,71,1
37053197	14	2	3	2	DoubleCheck=37053197,71,1
37303081	9	1	4	1	DoubleCheck=37303081,71,1
37521863	16	1	4	1	DoubleCheck=37521863,71,1
40564793	9	1	1	1	DoubleCheck=40564793,72,1
41216807	13	2	4	8	DoubleCheck=41216807,72,1
41350963	9	1	4	1	DoubleCheck=41350963,71,1
41636321	13	1	3	6	DoubleCheck=41636321,72,1
42159511	7	1	4	1	DoubleCheck=42159511,71,1
43207327	7	1	3	4	DoubleCheck=43207327,72,1
49520501	15	2	30	14	DoubleCheck=49520501,72,1