new list
 

This is an updated list of fun things to double-check early.

It's a mix of:
Tests done by systems with > 3:1 bad/good
or
Tests done by systems with zero good, >= 2 bad, and at least one unverified mismatch at some point

That second clause will grab some that only have a pair of bad ones, but a strong suspicion that more of them are also bad. I've had pretty good luck doing some of those on my own (and I've even been doing some with a single bad result so far with about 50/50 mine was different).

If anyone is ever interested, there are ~270 exponents in a big list of more speculative work. Systems with zero good, zero bad, but 2+ mismatches. I've only had my result be different in about a third of those, but that's still higher than the typical, random rate. My goal with those is to take the smallest exponent from each machine and test it... either I add to it's good or bad count, but either way we'll have more info than before. So if anyone felt like helping with that, and a smaller chance that your result will be different, let me know. The cool part is sometimes you'll find a machine that really is bad and has a lot of "solo" checks, so you can follow it through and get a bunch of other stuff out of it. I did that with one machine that had a dozen other checks which were all bad (they were all above 60M though so I didn't share it here). That was fun.

[CODE]exponent Bad Good Unk Sus Solo Mis worktodo
36109043 2 0 7 0 6 1 DoubleCheck=36109043,71,1
36475279 2 0 7 0 6 1 DoubleCheck=36475279,71,1
37900663 6 1 8 2 4 6 DoubleCheck=37900663,71,1
41321141 15 5 5 0 5 0 DoubleCheck=41321141,72,1
41448403 4 0 14 1 7 8 DoubleCheck=41448403,72,1
41757647 2 0 3 6 3 6 DoubleCheck=41757647,72,1
41794073 4 0 10 2 7 5 DoubleCheck=41794073,72,1
41856721 3 0 4 1 3 2 DoubleCheck=41856721,72,1
41877499 15 5 5 0 5 0 DoubleCheck=41877499,72,1
43371121 2 0 5 1 5 1 DoubleCheck=43371121,72,1
43502383 2 0 5 1 5 1 DoubleCheck=43502383,72,1
43721149 2 0 5 1 5 1 DoubleCheck=43721149,72,1
43749311 2 0 5 0 4 1 DoubleCheck=43749311,72,1
45601877 2 0 2 1 2 1 DoubleCheck=45601877,72,1
45710017 2 0 2 1 2 1 DoubleCheck=45710017,72,1
45870623 2 0 5 1 5 1 DoubleCheck=45870623,72,1
46102687 2 0 5 1 5 1 DoubleCheck=46102687,72,1
46567757 2 0 5 1 5 1 DoubleCheck=46567757,72,1
47921893 2 0 5 0 4 1 DoubleCheck=47921893,72,1
[/CODE]

I'll try 36109043

I took 36475279.

I took all below 46M

[QUOTE=Madpoo;411500]Hmmm... I'm a bit more skeptical. I just looked at those 5 exponents along with the good/bad counts of the other systems involved.

For the most part, the other, non-AirSquirrels cpu has an okay track record.[/QUOTE]

You were correct.

AirSquirrels, beat the heck out of that new machine. Both of my triple checks matched the original assignee, but not you.

I'll take M46102687, M46567757, M47921893

[QUOTE=Madpoo;410934]Here's an updated list. First part is exponents from systems with at least 2x as many bad as good and already have 2 mismatches under their belt. The pair of exponents at the bottom were from systems with > 3x as many bad as good and one mismatch so far:
[CODE]exponent Bad Good Unk Sus Solo Mis worktodo
35551543 8 4 4 1 3 2 DoubleCheck=35551543,71,1
36459239 8 4 4 1 3 2 DoubleCheck=36459239,71,1
37720493 8 4 4 1 3 2 DoubleCheck=37720493,71,1
41065553 18 6 2 2 2 2 DoubleCheck=41065553,72,1
41069953 2 1 8 1 7 2 DoubleCheck=41069953,72,1
41079221 12 5 11 7 5 13 DoubleCheck=41079221,72,1
41079751 3 0 3 0 1 2 DoubleCheck=41079751,72,1
41106281 4 2 6 2 4 4 DoubleCheck=41106281,72,1
----------------
36166439 3 0 2 0 1 1 DoubleCheck=36166439,71,1
37846999 3 1 3 0 2 1 DoubleCheck=37846999,71,1[/CODE][/QUOTE]

Chris confirmed my residue for 41106281, so this set of 10 exponents is now finished. 4 out of the 10 original residues were wrong.

I have 3.5 hours to go on 41031493. I am a little leery of my own result, though perhaps without real cause. The higher than usual number of resets makes me nervous. :ermm:

I matched David J. Stucki on[INDENT]M( 41031493 )C, 0x7545027971b19a58, offset = 12185, n = 2304K, CUDALucas v2.05.1[/INDENT]

[QUOTE=kladner;411662]I have 3.5 hours to go on 41031493. I am a little leery of my own result, though perhaps without real cause. The higher than usual number of resets makes me nervous. :ermm:[/QUOTE]

I've had several runs where the FFT must have been right on the border and it chose the lower one. Resulted in lots of re-doing using the safer method, but in the end it matched anyway. Even when the error wasn't repeatable, it still passed. The error code will reflect all of that info, and from what I've seen, it's *usually* non-critical.

I did an analysis of error codes at one point, checking which ones are more likely to wind up bad. Those round-off errors didn't seem any more likely to be bad than a totally clean run.

If your result is flaky enough that it got marked as suspect, then you do have a much higher chance that it's bad. I haven't analyzed that exactly since I'm not totally sure which error codes will trigger a "suspect" flag, but that's on my to-do list.

There is an "undoc" option that may avoid that, and making your test a bit faster (each iteration slower, but avoids repeated iterations) and safer, by increasing the FFT on "borders". You may look to "undoc.txt" for "NearFFTLimitPct=value", and respective "SoftCrossovers=n" and "SoftCrossoverAdjust=n".