New work types for GPU72?

[kladner]

Whatever informed opinion comes up with is fine with me. I'm sure I will find something useful to do.

[chalsall]

Originally Posted by ET_

Anyway, I'd like to advance the DC wave, so I'm going to use LL-DC more often.

Yeah... That's what I've personally done as well (other than one very sad little GPU which is doing some TFing).

Originally Posted by ET_

That's why I'd be more interested in keeping the DC TF ahead of limits, and eventually rising one bit the DC-TF limits, if some TF work should be chosen.

OK, but keep in mind that thanks to James' analysis we know that it doesn't make "economic" sense to TF in the DC range past 69 bits below 32M. And we're currently more than 500 days ahead of the DC wave front.

[KingKurly]

I do a fair amount of P-1 on once-LL'd exponents, usually from curtisc. Started in the 39M and now it appears we are in the 45M area. (I also do P-1 assignments from GPU272; the exact percentages vary.)

I could be persuaded to change (I saw an argument above that we should use 2.6 LL tests saved) but I would argue that we should use 1.1 tests saved. I believe I read (but could be 'misremembering') that the average exponent needs 2.1 LL tests due to potential errors and such. One has already been performed. That leaves 1.1 LL left.

It is pretty easy to blow through these assignments (time-wise), and while it's perhaps not as comprehensive a search as 2.6 LL, I don't think a case can be made for doing once-LL'd candidates to better bounds than no-LL'd candidates (which use 2, by default).

I would not be opposed to grabbing the work from GPU272 instead, since it's largely a manual process anyway, using James' p1small.php as assistance. About the only downside I see is that one would have to look in two places to determine where the tail of 'non P-1 DC candidates' is. And... meh, it's tough to get that upset about that.

(To avoid any confusion, I do not modify the 'tests saved' on P-1 assignments that I get from GPU272 today; I leave it at 2. But on "once LL'd" candidates, I set to 1.1.)

[Dubslow]

Actually, everyone should know, for both Test= and Pfactor=, Prime95 automatically incorporates a 2*error_rate fudge factor for any tests_saved value in the workfile line. Even if you put tests saved = 1 or tests saved = 3, it still still add in the extra 2*err fudge factor. As of 27.7, the error rate is estimated at 1.8%, so if you put in saved = 2, the bounds are actually calculated against 2.036, not 2.000. (If you put in 2.6, bounds will be calculated with saved==2.636.)

(For the interested, the relevant line is 6109 of ecm.c, inside guess_pminus1_bounds(). [The error rate is defined in commonc.h.])



PS Presumably James would help chalsall get this extra P-1 code working so that we'd only need to look in one place.

[Uncwilly]

Originally Posted by chalsall

Since we're only a few weeks away from taking everything below 60M to at least 72 bits, I thought now might be a good time to talk about the possibility of adding some new work types coordinated by the system.

Thoughts? Any other work types anyone would like to see?

*cough* 100million digit range *cough*
factoring from 74 bits to 79 or 80.

[kracker]

Hmm, I don't have much firepower compared to others, but maybe I should do some TF DC...

[chalsall]

Originally Posted by kracker

Hmm, I don't have much firepower compared to others, but maybe I should do some TF DC...

You're better than me -- I only have (part time access to) a measly Quatro FX 1800...

But every TF run is valuable; please do whatever work type you enjoy the most.

[chalsall]

Originally Posted by diamonddave

I would like it very much if the exp that need P-1 were TF to 71. It would mean allot less manual work for me.

OK, I have started the work for this. The first step was to bring in some of the DC candidates without P-1, and made them available on the DCTF request form.

For those who are already doing DCTF work, please consider selecting the new option "No P-1 done". This will set your pledge to 71, and give you candidates above 40M. Those who aren't doing DCTF work at the moment, please don't stop LLTFing for this.

Note that I have some back-end code changes to make -- I never expected the DC and LL ranges to cross. Thus, for example, the "Expected Completion" and "Returned Level" reports are slightly broken in the 46M range.

Additionally, I want to consult with James as to how best to create the P-1 assignment lines before exposing the DC P-1 assignment form.

[kracker]

Originally Posted by chalsall

You're better than me -- I only have (part time access to) a measly Quatro FX 1800...

But every TF run is valuable; please do whatever work type you enjoy the most.

Ok, I see.

Right now, I'm getting about 25 GHz Days per day, with this 7770 I'll be getting around est. ~75-85

[Xyzzy]

Just a question:

If we start factoring real big numbers isn't this going to greatly affect our "work saved" metric?

That is the metric that is the most important, right?

[LaurV]

Originally Posted by Xyzzy

Just a question:

If we start factoring real big numbers isn't this going to greatly affect our "work saved" metric?

Clearly! With this output (20-50 factors per day, for a 332M exponent, which "saves" tremendous amount of LL) we would be far ahead of you in the top

(maybe the right solution should be to register a separate "metric", or to register none if the work unit takes less then few hours - the equivalent of taking a 50M expo to 71, 72 ??).