Credit for trial factoring

[LaurV]

well.. we never experienced otherwise here, which is normal and correct according with the time we invested in factoring, in long term averaging.

(sorry for the formatting, I am a bit in hurry)

Manual testing    41482733    F    2014-02-14 05:22    0.0    343020020610099313049    0.3125
Manual testing    41483399    NF    2014-02-14 05:22    0.0    no factor for M41483399 from 2^68 to 2^69 [mfakto 0.13-Win cl_barrett15_69_gs_2]    1.4411
Manual testing    35525923    F    2014-02-14 01:16    0.0    1616268328913346819329    3.0502
Manual testing    35525807    NF    2014-02-14 01:16    0.0    no factor for M35525807 from 2^70 to 2^71 [mfaktc 0.20 barrett76_mul32_gs]    6.7311
Manual testing    35525141    NF    2014-02-13 23:13    0.0    no factor for M35525141 from 2^70 to 2^71 [mfaktc 0.20 barrett76_mul32_gs]    6.7312
Manual testing    41471389    F    2014-02-13 22:17    0.0    343995315434689936607    0.3185
Manual testing    41472661    NF    2014-02-13 22:17    0.0    no factor for M41472661 from 2^68 to 2^69 [mfakto 0.13-Win cl_barrett15_69_gs_2]    1.4415
Manual testing    35524277    NF    2014-02-13 19:05    0.0    no factor for M35524277 from 2^70 to 2^71 [mfaktc 0.20 barrett76_mul32_gs]    6.7314
Manual testing    41466101    F    2014-02-13 18:14    0.0    522256853656664233487    1.1870
Manual testing    41466527    NF    2014-02-13 18:14    0.0    no factor for M41466527 from 2^68 to 2^69 [mfakto 0.13-Win cl_barrett15_69_gs_2]    1.4417
Manual testing    35523797    NF    2014-02-13 17:00    0.0    no factor for M35523797 from 2^70 to 2^71 [mfaktc 0.20 barrett76_mul32_gs]    6.7315
Manual testing    41463671    F    2014-02-13 16:13    0.0    490020082106298012337    1.0545
Manual testing    35523443    F    2014-02-13 14:57    0.0    2104958803532585306983    5.6160
Manual testing    35523473    NF    2014-02-13 14:57    0.0    no factor for M35523473 from 2^70 to 2^71 [mfaktc 0.20 barrett76_mul32_gs]    6.7315
Manual testing    35523259    F    2014-02-13 13:55    0.0    1964303073347245406057    4.9444
Manual testing    35523143    NF    2014-02-13 13:55    0.0    no factor for M35523143 from 2^70 to 2^71 [mfaktc 0.20 barrett76_mul32_gs]    6.7316
Manual testing    35516167    F    2014-02-12 23:26    0.0    2049830258650648675937    5.3594
Manual testing    35516353    NF    2014-02-12 23:26    0.0    no factor for M35516353 from 2^70 to 2^71 [mfaktc 0.20 barrett76_mul32_gs]    6.7329
Manual testing    41431069    F    2014-02-12 15:25    0.0    377275909099099752481    0.5111
Manual testing    41431237    NF    2014-02-12 15:25    0.0    no factor for M41431237 from 2^68 to 2^69 [mfakto 0.13-Win cl_barrett15_69_gs_2]    1.4429

[bayanne]

I have now had a look at the new stats available on mersenne.ca thus:
http://www.mersenne.ca/exponent/68047649 which gives a credit of 3.941 Ghz/day rather than the 0.6650 Ghz/day I was awarded

Still confused ...

[ramgeis]

Shouldn't the credit for finding a factor actually be almost 0.0?
At the moment submitting a factor only includes this as the information.
What is missing is the information that (maybe) additional effort has been spent.
The server implicitly assumes that also everything below has been checked and gives credit for that but IMO this should be reported as well - and that's the submission the user gets the credit for, like:

no factor for M987654321 from 2^77 to 2^77.7777777 (credit given)
M987654321 has a factor: 259086082177383985712319 (=2^77.7777777, no credit given)

[bayanne]

Originally Posted by ramgeis

Shouldn't the credit for finding a factor actually be almost 0.0?
At the moment submitting a factor only includes this as the information.
What is missing is the information that (maybe) additional effort has been spent.
The server implicitly assumes that also everything below has been checked and gives credit for that but IMO this should be reported as well - and that's the submission the user gets the credit for, like:

no factor for M987654321 from 2^77 to 2^77.7777777 (credit given)
M987654321 has a factor: 259086082177383985712319 (=2^77.7777777, no credit given)

However this does not take into account if there is another factor for the same exponent. Is it 'ethical' to not factor each exponent to the end of the bit depth ...

[LaurV]

Originally Posted by ramgeis

no factor for M987654321 from 2^77 to 2^77.7777777 (credit given)
M987654321 has a factor: 259086082177383985712319 (=2^77.7777777, no credit given)

This is exactly what I explained in my initial post, and exactly what the server is doing, because it assumes you did TF with p95, which tests the factor candidates in order. The problem is that there are programs which do not test the candidates in order: mfaktc, mfakto, blah blah. Imagine you want to find the factors of x=473*z, with z being a big prime, but have a program which tests first the factor candidates which are 1 (mod 3), then the ones which are 2 (mod 3). For whatever reasons, say that the program is faster doing that. So, after filtering, you check if 7, 13, 19, 31, etc is a factor, and you will find first the factor 43, and only (much) later the factor 11. In this case, if you stop after you find 43, and you report it, you will get more credit than you would get by reporting the 11, because the server assume you took them "in order". He doesn't know how much time you spent, or what the program does. He is either not aware that a smaller factor was missed (11) because he is in other modularity class, not yet tested. That is why you always have to report the fact that the bitlevel was "partially tested", smaller factors may exist if the test was done with mfaktX.

And taking the time into account would not be right for the people having faster hardware, or overclocking (i.e. investing more money in electricity!).

The only thing you can take into account is the size of the factor. In time, this "averages" to the right credit. There is no harm done. Sometime you are lucky, finding a big factor fast, because it was in the first class, and get big credit. Sometime is viceversa, the small factor lies in the last tested class. So what?

[axn]

Originally Posted by LaurV

because it assumes you did TF with p95, which tests the factor candidates in order.

Nuh uh. It does not. It too uses factor classes.

[TheMawn]

Frankly I couldn't care less about how much credit I get for finding a factor. I'll say it again: I am as much of a GHz-Days whore as LaurV but I fully understand that their value is completely meaningless.

If I TF'ed and found 2^66.1 as a factor, why am I getting credited more or less than if the factor was 2^73.9? I cleared the exponent just the same.

You might argue that finding the factor was more work so I should be awarded more credit, which is a perfectly legitimate claim. But what of the fact that I saved two LL tests (or one if we're DC-TF'ing)? Should I be credited for the work I saved? But then DC-TF to ⁷² should be worth less than LL-TF to 2⁷² if a factor is found?

It's all kind of subjective and confusing.

[Bdot]

Originally Posted by ramgeis

Shouldn't the credit for finding a factor actually be almost 0.0?

Seems the good times are over, and this has been "fixed"

got assignment: exp=65986033 bit_min=72 bit_max=73 (14.50 GHz-days)
Starting trial factoring M65986033 from 2^72 to 2^73 (14.50GHz-days)
Using GPU kernel "cl_barrett15_73"
No checkpoint file "M65986033.ckp" found.
W4 done|    ETA |     GHz |time/class|    #FCs | avg. rate | SieveP. |CPU idle
 24.0% |  5h35m |   47.35 |  27.555s |   1.64G |  59.36M/s |   30951 |   2.06%
M65986033 has a factor: 4822533015970008558073

found 1 factor for M65986033 from 2^72 to 2^73 (partially tested) [mfakto 0.13-Win cl_barrett15_73_2]
tf(): total time spent:  1h 46m 16.954s (196.40 GHz-days / day)

20140222_102728 INFO:  M65986033 submitted; 0.4389 GHz Days credit.

But that is OK, I've been "overpaid" for a long time - I just need to change mfakto's GHz-days calculation ...

Just one question: was the mfaktc/o detection also fixed, meaning, can I remove submitting my "dummy no factor mfakto" along with factors without the server changing "F" to "F-PM1"?

[LaurV]

Originally Posted by Bdot

Just one question: was the mfaktc/o detection also fixed, meaning, can I remove submitting my "dummy no factor mfakto" along with factors without the server changing "F" to "F-PM1"?

Not fixed. Continue to submit a "no factor" line in front, when you only report factors.

[Bdot]

Thanks, I'll keep it.

Did you see a post anywhere that details the new calculation for "F" results? Or, do you happen to know?

[LaurV]

Originally Posted by Bdot

Did you see a post anywhere that details the new calculation for "F" results? Or, do you happen to know?

If you read my former posts, you will see that for me it always was like it is now. I still can't believe that you could get, for finding a factor, a higher credit than for a "no factor" result, because, in ANY case, no mater what program you used, you spent LESS time to find that factor.

The "logical" procedure to compute the right credit, would be:
A) assuming the factor was found with p95 or some program which tests the factors consecutively, or splits them in classes, but checks all classes in parallel, then the credit is c*(f-s)/(F-s), where c is the credit you get for "no factor", f is the factor you found, F is the highers factor candidate in the interval, s is the starting candidate (the smallest candidate). Or something like that.

B) assuming the candidates are split in classes which are done serial (i.e. one by one, as mfaktx is doing) then if you found the factor in class x, you should get the credit for completing (x-1) from 960 classes, plus the part of the x-class credit till you found the factor.

Only one method should be used, to avoid cheating, or the max credit between the two can be given (!), in this case there is no tempting to cheat (like finding a factor with one program and reporting it with another program, to get bigger credit), and the users would be happy they got a small bonus for finding a factor. As opposite of the case when no factor is found - no bonus is given. In time, this normalizes anyhow to the right credit (plus a small bonus) for long runners.

Edit: remark that in all cases, finding a factor (except additional bonus which accumulates over time) has to (and it would!) bring you less credit than for a "no factor" result. You just spent less time finding it, regardless the program you used. Of course, a "bonus" can be given because you found a factor, like some constant "c" added to the calculus above. This would be a different story.

However, I believe George/James can enlighten you about the formulas. They (the formula, not the guys) are also somewhere in the source codes, or on the web page, or wiki, I remember I saw them in the past.