GPU TF work and its impact on P-1

[ckdo]

Originally Posted by davieddy

I would guess that there are similar candidates in the 30-40M range which
are in the "available for DC" column, and they will get DCed before
the 40-50M range.
David

Nope. Any exponent which didn't have P-1 done would appear in the "available for P-1" column unless it was either double-checked already or currently assigned. Last year, I myself did P-1 on all exponents below 40M that weren't assigned at the time.

[ckdo]

Originally Posted by davieddy

Comparing the number of P-1 assignments for exponents < 10M with
the number of others, I think "perhaps" may be understating it

David

Considering the credit per assignment is like 1/10th for sub-10M assignments compared to mainstream assignments, it probably isn't.

[KingKurly]

Originally Posted by ckdo

Nope. Any exponent which didn't have P-1 done would appear in the "available for P-1" column unless it was either double-checked already or currently assigned. Last year, I myself did P-1 on all exponents below 40M that weren't assigned at the time.

I had typed a comment earlier that said something similar, but I guess I forgot to hit submit. I was at work at the time anyway, so it's highly likely I got distracted by that or something similarly important.

I am following in your footsteps, started with 40M and working my way up. I use a "tests saved" value of 1.1 instead of the recommended 1 (or the default 2, which is used for not-yet-LL'd exponents). Occasionally when I see a batch of not-yet-LL'd exponents below 50M, I'll take them and run P-1 with 2 tests saved. But recently, much of my work has been on "one LL'd" exponents in the 40-43M range.

I hope we see a day where I catch up to (and rejoin) the leading edge of P-1, but who knows... I probably need a lot more machines for that, and first I need this heat wave to end so I can turn all my cores back on!

[KingKurly]

Originally Posted by ckdo

Considering the credit per assignment is like 1/10th for sub-10M assignments compared to mainstream assignments, it probably isn't.

I do a fair amount of P-1 in the 1.9M range, which I do realize has no value in terms of finding new Mersenne primes. I'm sure that kills davieddy, but hey, I'm not here to please any one particular person. (Don't worry, we still love you, davieddy ) When I finish up that work, I'll bring those cores back into the fold of "leading edge" work.

Each assignment recently has been worth roughly 0.2029 GHz-days. I tend to use my slower machines for that work, but they still get results at a decent pace. And I've been finding plenty of factors.

[davieddy]

Originally Posted by KingKurly

I do a fair amount of P-1 in the 1.9M range, which I do realize has no value in terms of finding new Mersenne primes. I'm sure that kills davieddy,

I have been itching to point out that the project's acronym is GIMPS
and not GIMFS

My aim in this thread is to work out how to target the resources which
are trying to increase the probability of a prime per day most effectively.
Obviously the brute force "throw more resources at it" will help, but I am
more concerned optimizing the use resources we have.
Doing a LL without P-1 or adequate TF is obviously sub optimal.

I actually have more sympathy for those factoring DCed exponents:
at least they don't pretend to be helping GIMPS, unlike those factoring
well below the optimum bit level in the 100M+ range.
Our (quite modest) aim ATM is for GPUs to factor to 3 bit levels
higher than the limit for CPUs. That is 7 times as many GHz-days
as all the previous factoring if that had been done to the CPU limit.
If, as applies to virtually all exponents >61M, the factoring is below
the limit, all those tedious 20 digit factors which clog up the recent
cleared table are effectively worthless, as their credit shows.

Going back to the 40-53M range where 80% of LLs assignments come
from (and are returned to unfinished (rages)), I hope that the returns
can be P-1 ed (if necessary) and TFed an edtra 2 bits before being
assigned again. This might lead the new tester to treat his/her
carefully prepared exponent with the respect it deserves, and
actually complete the test!

David

[Christenson]

Umm, davieddy:
This is a strictly volunteer project, and, lots of those volunteers go off and do wild and distracting things as far as GIMPS proper is concerned. Go, look up GGNFS and such -- the truth is, mersenneforum (and P95) here support all reasonably effective computational number theory efforts.

Now, if you are telling me that 80% of 40-53M LL assignments go out and return unfinished, then I'd say that is because they are awfully large projects to take on for most individuals. I certainly wouldn't recommend that (or making a 1 meter telescope mirror) as a beginner project. I'd suggest a progression from small ECMs (might take 25 GHz days to get a factors, but assignments finish several times per day) to P-1s (takes a few days per assignment, might take 60 GHz-days to get a factor) to LL-Ds (might take 10-30 days per assignment) to full LLs (20-100 days per assignment), with the idea of giving early positive feedback. This could be implemented as the "whatever makes the most sense" path in P95.

We've certainly had some newcomers confused by the fact that feedback on the stats pages takes days or weeks, especially in this "instant" internet age.

In the mean time, I'm working on that P-1 cushion.....and every so often turning in a first-time LL result.

[davieddy]

Originally Posted by Christenson

Umm, davieddy:
This is a strictly volunteer project, and, lots of those volunteers go off and do wild and distracting things as far as GIMPS proper is concerned. Go, look up GGNFS and such -- the truth is, mersenneforum (and P95) here support all reasonably effective computational number theory efforts.

I was explaining in my post that I wasn't trying to be a recruiting sergeant.

But if optimum use of resources results in a speed up of the LL wave,
one could reasonably hope for a few more participants getting stuck
in where it hurts!

BTW You are making a splendid contribution to the project.

David

[cheesehead]

Originally Posted by davieddy

But if optimum use of resources results in a speed up of the LL wave,

What about optimum use of resources that results in a speed up of someone else's favorite wave or sub-wave?

You can advocate for yours without insinuating that it's the only legitimate favorite.

Why so insecure about LL? The only discoveries that bring fame and fortune are from LL tests, so they have an enduring advantage in attraction that none of the other assignments can match.

[davieddy]

Originally Posted by cheesehead

What about optimum use of resources that results in a speed up of someone else's favorite wave or sub-wave?

You can advocate for yours without insinuating that it's the only legitimate favorite.

Why so insecure about LL? The only discoveries that bring fame and fortune are from LL tests, so they have an enduring advantage in attraction that none of the other assignments can match.

Optimum use of resources currently engaged in finding a Mprime.

Optimum use of resources currently engaged in other waves as well.

Peace and goodwill to all men/women

David

[Christenson]

Do think upon all those LL assignments that return abandoned.....if we can do a little something for those folks, it will return many-fold.

[davieddy]

Originally Posted by Christenson

Do think upon all those LL assignments that return abandoned.....if we can do a little something for those folks, it will return many-fold.

Originally Posted by davieddy

Going back to the 40-53M range where 80% of LLs assignments come
from (and are returned to unfinished (rages)), I hope that the returns
can be P-1 ed (if necessary) and TFed to 72 bits before being
assigned again. This might lead the new tester to treat his/her
carefully prepared exponent with the respect it deserves, and
actually complete the test!

David

Originally Posted by davieddy

That would make a much more important consideration easier to recognize:
exponents <53M constitute about 80% of daily LL assignments. Yesterday
they didn't even get to 53M.
I have a hunch that George is on the case already, but returned/expired
exponents should be TFed to 72 bits (by GPU of course) before being
deemed "available for LL".

David

Do you ever actually read my posts?