I have a few questions about getting my GPU working for GIMPS

[jwaltos]

I apologize in advance if this not the appropriate thread.
I recently rescued a Dell 690, Tesla D870 (still looking for two appropriate PCI graphics cards for the unit) and a Tesla C1060. The Dell has 64 GB Ram and 2 Xeon 5355's and has a Quadro FX 5600. I'm running it with Mageia 5 (alpha 2) using the most recent Cado-nfs and gmp-ecm. I have a custom build, ASUS P9X79, 64 GB and a 6 core oc'd 3960 with a GTX 580 running Mageia/Win7 Pro. I have two other boxes holding 32 GB
and 16 GB. I intend on obtaining a Cuda enabled graphics card, highest cc and lowest cost, that will allow me to write/test gpu programs within
one of the latter boxes compiling the working versions so they will work on the cc 1.0 and cc 2.0 gpu's.

One of the questions I have is, can a polynomial file be created with the appropriate parameters that matches an integer perfectly. To restate,
working in reverse, given a polynomial, what numeric values can it represent, as exactly as possible, relative to the parameters provided.

I can use many CAS systems but prefer Maple and FORM; I have used Knuth's Stanford Graphbase over 20 years ago on a 486 dual booting Win95/RHL 4.1x - I have his AoCP and have read many of his books/papers. I prefer to delve into state-of-the-art theory by studying the most current papers and learning from those professionals (many exist within this forum) who have hands-on experience with the concepts/implementations and limitations thereof.

The distribution of prime numbers and general integer factorization are two things I have been studying as an amateur - certain forum
members have been a great help while others have been complete assholes. So to all concerned, having received invaluable assistance as well as unbridled scorn for dumb questions, I've learned from both.

Again, if anyone has used the deprecated hardware I intend to revitalize and can provide a pointer on the best graphics cards to use I would
sincerely appreciate it - I don't know if I can run those boxes and Christmas lights at the same time later this year.;)

[kladner]

I don't know if I can run those boxes and Christmas lights at the same time later this year.;)

I have only looked at a couple of pieces of your power equation. First, the twin Xeons. This shows that you might end up well north of 200 W just for the CPUs running hard.

Nvidia rates the Tesla C1060 at 187.8 W, the Quadro at 171 W. Figure out a few more components, including PSU efficiency estimate, and you could get a reasonably good idea what the whole thing will draw.

EDIT: If you live in a place where it is going to get really cold, remember that computers make dandy space heaters, so you might be consuming that energy with a dumb device that only puts out heat, instead of one which does calculations as a byproduct. :)

[Red Raven]

Oh, and Batalov, I know the chain was already explained, that's how I know the order, but I'm looking for more details. IE, right now I think GPU 72 automatically takes a block of assignments from primenet and doles it out to GPU 72 users, but is that true? How does MISFIT decide what a GPU will be most effective at? Does MISFIT report back to GPU 72 with the results, which GPU 72 uses to remove the corresponding exponents from the block it reserved? Does GPU 72 the report the results to primenet, or does MISFIT send them to both GPU 72 and primenet? Or does it just send them to primenet and GPU72 automatically removes exponents from its list once it assigns them? In that case, primenet would make the exponents available again if their results were never reported, and I'm guessing GPU72's algorithm for choosing exponents would them up from primenet again and restart the cycle. Basically, I'm just trying to figure out how the path of exponents and results between primenet and mfakto works with GPU72 and MISFIT involved. I'm also trying to figure out what happens to exponents that become expired with this setup.

[Red Raven]

I just read the recent replies, and I honestly don't understand why there's so much conflict. I'll try to clear things up a bit.

1) I'm here for the science. More knowledge, even if it's use isn't totally understood yet, is always good.
2)Competition makes the science a bit more interesting, so I embrace it. I know it doesn't actually matter, but it makes me a bit more invested in the project, which I'm fine with.
3) I know the competition is essentially pointless. Successful TF tests happen randomly, so the faster GPU setup will have a better chance of winning. That's why I consider it to be a friendly competition, and don't take it seriously.
4) The competition is pretty much like waiting for the grass to grow, but people have competed over dumber things with less purpose.
5) I am interested in learning more about how the project works at a mathematical level, but at the moment I'm also a full time freshman college student trying to push through the gen ed classes so I can study electrical engineering. I don't mind not having the spare time to learn how Prime95's math works though, because the university gives me as much electricity as I want to run Prime95 with my tuition fee. That means I'll be adding my laptop's down time and possibly an old Dell with a P4 I left at home to the project.
6) Everyone has their own reasons for participating in the project. No particular reason is bad, unless you plan to use the data from the project to hurt someone, somehow. It's probably unlikely that anyone is, so there's no reason to attack people for their reason in joining the project.

Now, I'm going to go download Prime95 to my laptop so I can have it crunch so numbers while I for my Intro to Music Enjoyment test tomorrow. Clearly, I need to be told how to enjoy music despite my 17+ GB library because it doesn't include music from the Baroque or Classical eras. It's an essential skill for making circuits that I can guarantee won't kill people.

[chalsall]

@RedRaven et al...

I'm just out of one meeting, heading into another. But to try to answer your questions quickly...

1. Large blocks of candidates are imported into GPU72 manually from time-to-time if they have been sub-optimally TF'ed. These will appear as being assigned to "GPU Factoring" as TF.

2. Some low candidates which aren't imported in step #1 above are sometimes automatically "recaptured" if we have the firepower available, and have been sub-optimally TF'ed. These will appear as being assigned to "For Research", my personal account which is "trusted", as either LL or P-1.

3. Candidates which are reserved via the spider described in step #2 are done first.

4. GPU72 never submits results itself; it is always by the worker's machine. GPU72 simply observes the status of Primenet to determine what work has been done, and who to credit.

5. Yes, this is all rather convoluted. Consider it "Evolution in action"...

I hope that makes sense....

Originally Posted by Red Raven

Oh, and Batalov, I know the chain was already explained, that's how I know the order, but I'm looking for more details. IE, right now I think GPU 72 automatically takes a block of assignments from primenet and doles it out to GPU 72 users, but is that true? How does MISFIT decide what a GPU will be most effective at? Does MISFIT report back to GPU 72 with the results, which GPU 72 uses to remove the corresponding exponents from the block it reserved? Does GPU 72 the report the results to primenet, or does MISFIT send them to both GPU 72 and primenet? Or does it just send them to primenet and GPU72 automatically removes exponents from its list once it assigns them? In that case, primenet would make the exponents available again if their results were never reported, and I'm guessing GPU72's algorithm for choosing exponents would them up from primenet again and restart the cycle. Basically, I'm just trying to figure out how the path of exponents and results between primenet and mfakto works with GPU72 and MISFIT involved. I'm also trying to figure out what happens to exponents that become expired with this setup.

[Mark Rose]

Originally Posted by Red Raven

2)Competition makes the science a bit more interesting, so I embrace it. I know it doesn't actually matter, but it makes me a bit more invested in the project, which I'm fine with.

Likewise. I've learned some math from participating, but participating in the mammoth project is the real fun for me. Some people only care about the math. No reason we can't all participate.

3) I know the competition is essentially pointless. Successful TF tests happen randomly, so the faster GPU setup will have a better chance of winning. That's why I consider it to be a friendly competition, and don't take it seriously.

Random, but predictable. There's something like a 1% chance of finding a factor each TF level. Don't ask me to explain the math because I don't know it ;)

5) I am interested in learning more about how the project works at a mathematical level, but at the moment I'm also a full time freshman college student trying to push through the gen ed classes so I can study electrical engineering. I don't mind not having the spare time to learn how Prime95's math works though, because the university gives me as much electricity as I want to run Prime95 with my tuition fee. That means I'll be adding my laptop's down time and possibly an old Dell with a P4 I left at home to the project.

Borg the labs when no one else is using the machines. I've done that, but not for GIMPS.

[legendarymudkip]

Originally Posted by Mark Rose

Random, but predictable. There's something like a 1% chance of finding a factor each TF level. Don't ask me to explain the math because I don't know it ;)

As far as I know it's to do with distribution of primes and factors.

Someone correct me if I'm wrong, but it's mainly because of how each prime as you go up has less chance of being a factor and also because of the distribution of primes - every factor tested with the powering algorithm has less chance of being prime because of this. Every factor tested tested with the powering algorithm has less chance of being a factor the larger the factor. The best way I can explain this is like this:
If testing a number for divisibility by 2 (obviously general rather than for Mersennes) then it can either be 0 or 1 (mod 2) so it has a 1/2 chance of being a factor. However, if testing a number for divisibility by 1279, it can be anywhere from 0 to 1278 (mod 1279), so there is only a 1/1279 chance of it being factor because n + 1 (mod k) = (n mod k) +1 (mod k), basically because they're in sequence and there's nothing random about 2 adjacent integers mod k.

[R.D. Silverman]

Originally Posted by legendarymudkip

As far as I know it's to do with distribution of primes and factors.

.

Mertens' Theorem.

[Mark Rose]

Originally Posted by legendarymudkip

As far as I know it's to do with distribution of primes and factors.

Someone correct me if I'm wrong, but it's mainly because of how each prime as you go up has less chance of being a factor and also because of the distribution of primes - every factor tested with the powering algorithm has less chance of being prime because of this. Every factor tested tested with the powering algorithm has less chance of being a factor the larger the factor. The best way I can explain this is like this:
If testing a number for divisibility by 2 (obviously general rather than for Mersennes) then it can either be 0 or 1 (mod 2) so it has a 1/2 chance of being a factor. However, if testing a number for divisibility by 1279, it can be anywhere from 0 to 1278 (mod 1279), so there is only a 1/1279 chance of it being factor because n + 1 (mod k) = (n mod k) +1 (mod k), basically because they're in sequence and there's nothing random about 2 adjacent integers mod k.

Ahh, that makes sense. Thanks!

[Gordon]

Originally Posted by LaurV

Referrring to Batalov - (and you didn't detect the irony, someone who asks if there is a database with all known primes is hopeless, let them invest their time and resources in TF :P)

That's what happens when someone sets them self up as arbiter of what is good and proper and they don't fully understand the language they are spouting off in.

[Gordon]

Originally Posted by Batalov

Ah, but that's easy.
But for the intellectually lazy - here's the link. Click the current primes?

I like the Eisenstein-Mersenne prime and a couple of Gaussian-Mersenne primes best, personally. But more than anything, I like variety. Not one of the people who order the same Chicken Piccata every time they go to an eatery. Or those who ate Chicken Piccata once in their life and then tell about it for the rest of their life.

I've looked at your link, still don't see any Mersenne primes on there...