[QUOTE=KyleAskine;279442]The efficiency of GPUs on TF makes it so GHz/days is not a good comparison metric when going across work classes. I promise the project is saving work.[/QUOTE]

Indeed.

GPUs have completely changed the metrics. The more advanced ones produce something like 100 times the output of a CPU for trial factoring. It's a bit like a Battleship circling an Ironclad...

But this is the metric which PrimeNet uses.

James and I are simply calculating what PrimeNet calculates.

[QUOTE=chalsall;279446]
James and I are simply calculating what PrimeNet calculates.[/QUOTE]
Well, I am not really thinking in numbers, I have no idea how they are computed, and I am not really interested to find out today. I am thinking in time.

Say like that, I am TF-ing for GPU272 since 5-6 days and I took 7 exponents from 67-69 to 72 and a couple others to intermediary steps. Meantime, I did not find any factor. In all this time I used say 30% of one of my GPU cards, that could run a LL with CudaLucas in less then 24 hours, for a 25-30M exponent. Now, that would be full power, but in this particular case, with 30%, say it takes 3 days to clear one exponent. In 5-6 days I could clear 2. I cleared none, and the LL still has to be done for all exponents I touched.

Now, endless discussion could go on into the fact that this exponents had plenty of P-1 done on them already, therefore decreasing the chance to find a factor, and that is why I found none, but I could argue that exactly here is the point where the time is NOT saved. For an exponent with no P-1 and no LL, there should be reasonable to TF it to higher and higher limits, because in case when a factor is found, we save plenty of P-1 and 2 (two!) LL tests.

But for lower exponents with one LL and a lot of P-1 done????

To be profitable, this has to produce a factor at every 4-5-6 exponents taken from 67 to 72. That is because one need about 4-5 hours for one expo in the range of 25M (if the exponent is lower, the time is longer, there is about 11 minutes for the first bit, and the time doubles at every bit) on a very fast GPU, that would take 24 hours. If you find a factor, then one LL is saved and you still have to run another 3-4-5 LLs for the other exponents. But if you do not find a factor in these 5 trials, then better you could run CudaLucas and CLEAR one exponent "for sure" in 24 hours. Saving time. Because if you do not find a factor, then you still have to run CL for ALL 6 expos, and the time was just wasted.

And here it come the probability of finding factors for expos that were heavily P-1-ed, which is much lower then 1 in 6. Much lower then one in 10. Much lower then... well.. it is lower... :D

And by doing that, we take "the bread from the mouth" from many old computers involved in the project which are so old that they can do only DC. Reducing DC assignments for lower exponents, on long term, will put such computers out of business.

That is why I can not stop wondering and being a bit grumpy about it...

[QUOTE=LaurV;279449]Say like that, I am TF-ing for GPU272 since 5-6 days and I took 7 exponents from 67-69 to 72 and a couple others to intermediary steps. Meantime, I did not find any factor. In all this time I used say 30% of one of my GPU cards, that could run a LL with CudaLucas in less then 24 hours, for a 25-30M exponent. Now, that would be full power, but in this particular case, with 30%, say it takes 3 days to clear one exponent. In 5-6 days I could clear 2. I cleared none, and the LL still has to be done for all exponents I touched. [/quote]
1. Since first time LL exponents should be taken to 72 and DC exponents should only be taken to 68 bits, I am guessing you're comparing apples and oranges (or did you take DC exponents to 72?! In which case, don't).
2. Statistics -- Specifically, low sample size. ~ 10 exponents is nowhere near a large enough sample size to make any conclusions. [Nonetheless, you were not really expected to find a factor with so few exponents done]

[QUOTE=axn;279451] (or did you take DC exponents to 72?! In which case, don't).[/QUOTE]
:blush: Well... in fact that is what I did... and where my credit was coming from... As the project is called "GPU to [COLOR=Red][B]72[/B][/COLOR]", and I did not find any explanation about where we should stop, I took all my assignments to 72, as you can see on my "view completed" page... Grrrr.... And this is where all the dilemma started... So, I worked in vain... common for me... it happens all the time... :smile:

So, let's start from the beginning:

How far should I take an exponent (TF bit level), depending on its size and former work done, LL, P-1, etc)??? Any tabular data or graphic? Like for dummies?

Edit: and what should I do with about 140 DC expo already reserved, waiting to be factored to... 72, in my worktodo? (I think let them finish is better, I am lazy to edit the file and this will complicate the things, as I have to return them to the project and reserve them again with a lower bound??)

Edit edit: for (2), I know the "statistics" part, that wasn't the point of my post. You can take first 3-4 people from the top of the status report on the project's page, who returned thousands of exponents, and tell me about it. Their "statistics" are even worse then mine, if I put in the GPU time spent.

72 for first time LL's, 68 for DC. Keep in mind that it also takes twice as long to do a DC exponent at a certain bit depth as the 45M first timers.

You were right to consider time, instead of GHz-days. Divide the GHz-Days/Factor by 100, and that's the 'GPU-Days' (give or take, my card being mid range and capable of ~100GHz-Days/Day). The GPU-Days per factor (around 3.5, according to the most recent data) is a [I]lot[/I] less than days per LL test. With CUDALucas thrown into the mix it's not as shocking a gap, but for my 100GHzDay/Day graphics card, it takes about ~15 days to LL test a similar sized exponent (not DC) so that's still 5 times as many factors as LL tests.

[QUOTE=LaurV;279452]:blush: Well... in fact that is what I did... and where my credit was coming from... As the project is called "GPU to [COLOR=Red][B]72[/B][/COLOR]", and I did not find any explanation about where we should stop, I took all my assignments to 72, as you can see on my "view completed" page... Grrrr.... And this is where all the dilemma started... So, I worked in vain... common for me... it happens all the time... :smile:[/quote]

Well, the rule of thumb is "4 bits less for DC than for LL". Since we're taking first time tests to 72 (why the project was called "GPU to 72"), it follows that we should take DC's to 68. But [URL="http://www.mersenne.info/trial_factored_tabular_delta_7/1/0/"]here[/URL], I see that most people are taking it to 69. Which is probably not advisable. [Alternatively, if it _is_ worth taking DC to 69, then it is worth taking first time LL to 7[B]3[/B]]

[QUOTE=LaurV;279452]Edit: and what should I do with about 140 DC expo already reserved, waiting to be factored to... 72, in my worktodo? (I think let them finish is better, I am lazy to edit the file and this will complicate the things, as I have to return them to the project and reserve them again with a lower bound??)[/QUOTE]

That would be a colossal waste of resources. Can you coordinate with chalsall to see how best to handle this? As you said, you're probably better off DC-ing them outright, rather than taking them to 72 bits.

I think, then, the next question is why not go further then 72; there are (à mon avis) two answers. The first is breadth; get the most exponents done, rather than highest bit-level; the other answer is that the above analysis ignores things like 'mfaktc takes a CPU core' (--> CUDALucas + plus the extra core means <15 days per test) and 'my CPU can do 4 tests at once' instead of one instance of mfaktc.

Oops, double posted. This is in response to my previous post, not having noticed axn's.

Ok. Now it started to make more sense. Thanks a lot axn and Dubslow.

So, going to the table of the overall progress, it means the project found a LL-front factor in about 352 GHzd, that is 3.5 to 4 days, which indeed is saving a lot of time, assuming my card can do one LL in 5-6 days (about 130 hours) at this LL-front-size, and we saved 2 LLs and some P-1 too. So, we do the 10-12 day job in 4 days. This really makes sense!!

Also, no discussion about P-1 factors, this is quite profitable, I never had any bit of doubt, and the proof is in the table (average, that is one factor at each half day, which is more than wonderful!).

But please allow me to still doubt about factoring at DC-front. That 220 GHzd per factor is [B]much over double of the time[/B] DC will take. It is two days and about a half for a factor, when a DC-LL would take 24 to 30 hours. Of course, this come from the fact that a lot of P-1 was already done there, which filtered out part of the exponents with factors in the targeted range.

[QUOTE=axn;279454]
That would be a colossal waste of resources. [/QUOTE]

Ok. Trusting in your judgement. I will unreserve them all when I arrive home after the working day ends (in about 5 hours), except for the expos which are currently under process, and re-reserve other, with more reasonable limits. Don't need to involve chalsall in this, it can be done from the assignment page. It will require some work from my side for wortodo files (cut, paste, blabla) as I spreaded the assignments for all mfaktc copies.

[QUOTE=LaurV;279458]Ok. Trusting in your judgement. I will unreserve them all when I arrive home after the working day ends (in about 5 hours), except for the expos which are currently under process, and re-reserve other, with more reasonable limits. Don't need to involve chalsall in this, it can be done from the assignment page. It will require some work from my side for wortodo files (cut, paste, blabla) as I spreaded the assignments for all mfaktc copies.[/QUOTE]

The reason I wanted you to talk to chalsall was that may be you don't need to do any release/assignment at all.

I am not sure, but if you just edit your worktodo and change the "72" to "68" (or "69" if some of them are already at "68"), and then report your work, the tool would still accept it. After that you can just release them.

But chalsall would know the best way.

Anyways... up to you. If you're comfortable just releasing/reassigning, go for it. :smile:

[QUOTE=Dubslow;279453]72 for first time LL's, 68 for DC. Keep in mind that it also takes twice as long to do a DC exponent at a certain bit depth as the 45M first timers.[/QUOTE]

Unverified exponents not at 68 already are rare, to put it mildly, below 32M. Since extending TF depth on expos above 32M is not exactly of any urgency, taking everything below to 69 (and no further) is the route to follow, for now.