Best type of work for my cpu

Hi, all.

I'm running Prime95 on an older computer of mine, which uses a 1st gen core i3. It runs at 2.3GHz on two cores (it's hyperthreaded but running four workers just generates more heat and each worker just runs half as fast).

Because it runs slower than the average computer, I set it up to do trial factoring only. In comparison, my 3rd gen i5 is about twice as fast clock per clock and runs at 4.6GHz and the LL tests it got are going to take 14 days. I don't really want to give my older computer a two month job.


This morning, I discovered GPU72 which has SOMEHOW flown completely under my radar. My desktop has a GTX 670 and I decided to run a single assignment on it as a test, to see how everything works and if I feel safe with the operating temperatures of my little baby. As it turns out, everything will run great and I am seriously considering running mfakct 24/7.

However, something caught my attention. The random assignment I got was for trial factoring in the 6.5 million area, from 71 to 74 bits. My old computer is currently running exponents of about 7 million from 70 to 71 bits and that is taking roughly a day per exponent, but mfakct did 71 to 72 in about sixteen minutes.


I understand that the architecture of a GPU is set up to run certain types of operations much, MUCH faster than a CPU, but this is ridiculous! I might as well not be running my 1st gen i3 at all if my GTX is going to accomplish a hundred times more work.

My concern here is that I'm wasting power, time and processor life by running trial factoring on my 1st gen i3. Should I set it up to do something else like P-1 factoring?

[Uncwilly]

Originally Posted by Unregistered

Should I set it up to do something else like P-1 factoring?

Maybe have one core do P-1 and the other doing double checks. Both are good solid choices for 'older' machines. If you take P-1, just be sure to give Prime95 as much memory as practical at night.

That machine has 8GB of memory, actually (I added some a while back) so I've got plenty there. The lower end double checks sound like a good idea too.

Basically I just want to find some kind of work that the GPU can't do or can't do way faster than a CPU. Double checks sound good. What about ECM factoring? I was never able to find much info on the different factoring methods...

Thanks

[Uncwilly]

Originally Posted by Unregistered

That machine has 8GB of memory, actually (I added some a while back) so I've got plenty there. The lower end double checks sound like a good idea too.

8GB is really good for P-1.

Originally Posted by Unregistered

What about ECM factoring? I was never able to find much info on the different factoring methods...

Currently that is only being used on lower numbers that have been double checked. One of the sub-projects is trying to find factors for all of the composite Mersenne numbers. This is where ECM is being used. ECM is related to P-1 (I don't understand it enough to talk too much.) But, if enough 'curves' are run with the right settings, it can be shown that there are no factors below a certain length. TF is the fastest (and most economical computer power-wise) way to find smaller factors, at some point P-1 and TF are about the same, there is an area where P-1 is better, then ECM becomes the choice. If after some reasonable level that doesn't find a factor, one of the sieves is employed. The sieving projects take a lot of fire-power, but will find the complete set of factors.
ECM is a low priority type of work.

I was actually wondering about that... I had once thought of asking if GIMPS is also trying to create a "profile" for each exponent.

I guess once you've done a double check, THEN found a factor and then found ALL factors you can be pretty sure a number is composite. Hah.


I've currently set the computer to use 2GB. Is anything more than that helpful, at all? Should I just go for the whole enchilada and give it 6GB?

[prgamma10]

Originally Posted by Unregistered

I guess once you've done a double check, THEN found a factor and then found ALL factors you can be pretty sure a number is composite. Hah.

Two Lucas-Lehmer Test (properly done) would be enough to be sure if a Mersenne number is prime or not.
Factoring is much much harder (factoring a 232-digit number takes years with many computers).

I've currently set the computer to use 2GB. Is anything more than that helpful, at all? Should I just go for the whole enchilada and give it 6GB?

2GB for each core doing P-1 (exponent 60-70M) would be enough.

[Mr. P-1]

Originally Posted by prgamma10

Factoring is much much harder (factoring a 232-digit number takes years with many computers).

That's only the case if the factors are large. If your number splits into P40 * P192 or thereabouts then you can expect to factor it in just a few hours.

[Mr. P-1]

Originally Posted by Unregistered

I was actually wondering about that... I had once thought of asking if GIMPS is also trying to create a "profile" for each exponent.

I guess once you've done a double check, THEN found a factor and then found ALL factors you can be pretty sure a number is composite. Hah.

Just one factor is sufficient. Finding more factors, preferably all of them is an end in itself.

I've currently set the computer to use 2GB. Is anything more than that helpful, at all? Should I just go for the whole enchilada and give it 6GB?

With 6 or 7GB it should run a few percent faster, or it might chose a higher E number resulting a slightly higher chance of finding a factor. It's worthwhile if the memory would otherwise lie unused, but nothing to worry about if you can't afford it.

Just make sure it doesn't thrash.

[Mr. P-1]

Originally Posted by Uncwilly

8GB is really good for P-1.
Currently that is only being used on lower numbers that have been double checked. One of the sub-projects is trying to find factors for all of the composite Mersenne numbers. This is where ECM is being used. ECM is related to P-1 (I don't understand it enough to talk too much.) But, if enough 'curves' are run with the right settings, it can be shown that there are no factors below a certain length.

Only in a probabilistic sense. With enough ECM the likelihood that a small factor remains undiscovered becomes vanishingly small. But it's not a proof that none exists.

TF is the fastest (and most economical computer power-wise) way to find smaller factors, at some point P-1 and TF are about the same, there is an area where P-1 is better,

P-1 will find all factors smaller than B1, but I'm pretty certain it is always slower than TF.

There are variations of the algorithm which will find all factors smaller than B2 and which, given sufficient memory, can reach a high B2 much faster than TF can. These variants however are much less likely to find factors greater than B2, so we don't use them. The P-1 we do use augments rather than replaces TF.

then ECM becomes the choice.

ECM is a slower and more memory-hungry algorithm, which can't (unlike P-1) make use of the 2kp+1 form of Mersenne factors. As a result it just isn't cost effective to use it at all, if the only reason you're searching for factors is to quickly eliminate primality test candidates. ECM is worthwhile if you're interested in the factors for their own sake, and are prepared to invest a lot of computer effort to find them.

If after some reasonable level that doesn't find a factor, one of the sieves is employed. The sieving projects take a lot of fire-power, but will find the complete set of factors.

ECM is highly sensitive to the size of the factors, but relatively insensitive to the size of the number to be factored. The sieve methods don't care at all about the size of the factors, but are extremely sensitive to the size of the number to be factored. Up to 100 digits or so, the fire-power required to use a sieve method is relatively modest. 150 digits is feasible on a single (modern) computer or a small cluster if you're willing to spend weeks or months on it. Any larger, and you need significant resources. 250 digits is about the limit of feasibility for even a well-resourced distributed computing effort.

[Uncwilly]

Originally Posted by Mr. P-1

Only in a probabilistic sense. With enough ECM the likelihood that a small factor remains undiscovered becomes vanishingly small. But it's not a proof that none exists.

Thanks for all of that clarification.

[cheesehead]

Originally Posted by Unregistered

< snip > I might as well not be running my 1st gen i3 at all if my GTX is going to accomplish a hundred times more work.

As I've pointed out elsewhere, 101 mph ("fast" + "slow" systems, or GPU + CPU) is faster than 100 mph ("fast" system alone, without contribution from "slow" system; or GPU alone without CPU).

GIMPS work on "slow" systems or components isn't wasted as long as it's not unnecessarily duplicated. Preventing unnecessary duplication is the purpose of the PrimeNet reservation system.

IMO the primary consideration in choosing a type of GIMPS work is what you'd be happiest accomplishing. GIMPS can use contributions from any speed of system, as long as it's coordinated so as to avoid unnecessary duplication.

My concern here is that I'm wasting power, time and processor life by running trial factoring on my 1st gen i3.

Re: power, time and processor life --
Compare the power, time and life effects of whatever you'd do instead of TF. If you're unhappy about doing TF after that comparison, then ...