332.2M - 333.9M (aka 100M digit range)

[aurashift]

Originally Posted by lycorn

May I at least try to lure you into giving a hand to this small sub-project (GMP-ECM on very small numbers, namely the (in)famous M1277)? The idea of using all that memory is most enticing...

http://www.mersenneforum.org/showthread.php?t=20092

*maybe*

Can I be lazy and just select that as a work type in mprime? Or do I have to install something?
I think I tried ECM on the RAMBeasts (500G-1T) but they didn't use much RAM. I could split half the work between P-1 and ECM since P-1 maxes out the 500GB's

[Uncwilly]

Originally Posted by aurashift

Would it be possible to unreserve ~60 abandoned exponents shortly before I get those ~30 blades I'm expecting in?

You want to unreserve work that you have? Or do you want to unreserve work that someone else has?

The easiest way to do the former is to remove them from your worktodo file and log in to PrimeNet and unreserve them there.

I have no power to do the latter.

[aurashift]

Originally Posted by Uncwilly

You want to unreserve work that you have? Or do you want to unreserve work that someone else has?

The easiest way to do the former is to remove them from your worktodo file and log in to PrimeNet and unreserve them there.

I have no power to do the latter.

The latter.
That way 9/10ths of the CPU power doesn't go to waste while trial factoring those numbers. (I mentioned in the other thread I don't have time to coordinate GPU trial factoring with all these servers.)

How come trial factoring isn't multi-threaded anyways?

[lycorn]

Originally Posted by aurashift

Can I be lazy and just select that as a work type in mprime? Or do I have to install something?

I think you´d be better off doing anything else, then.
Fair enough.

[TObject]

AsRock just released a new mini-ITX motherboard with LGA2011-3 socket and, what is very important, quad channel memory.

Thinking about stacking twelve of these in four rows, in front of a box-fan; and going for the big prize.

[Uncwilly]

Originally Posted by aurashift

The latter.
That way 9/10ths of the CPU power doesn't go to waste while trial factoring those numbers. (I mentioned in the other thread I don't have time to coordinate GPU trial factoring with all these servers.)

This link will show the exponents that are available and have a reasonable amount of TF : http://www.mersenne.org/report_facto...=1&tftobits=72

There are a bunch available that have been taken to 78, but no P-1 on them.

I am focusing mainly on the TF in this range, not the LL. With the firepower that you have, you should take those that are available and ready. Run them and then pick up those 78 and above P-1 them then do the LL on them.

[aurashift]

Originally Posted by Uncwilly

This link will show the exponents that are available and have a reasonable amount of TF : http://www.mersenne.org/report_facto...=1&tftobits=72

There are a bunch available that have been taken to 78, but no P-1 on them.

I am focusing mainly on the TF in this range, not the LL. With the firepower that you have, you should take those that are available and ready. Run them and then pick up those 78 and above P-1 them then do the LL on them.

Seems like P-1 occurs after 76 bits, isn't there something like a 3% chance it'll find a factor at any bit level?

So on one of my cores it takes ~10 days to take it from 73 to 76 bits. Not sure if I'm looking at this right but it'll take another ~8 to do P-1.

Maybe I should start P-1 now on those 500GB-1T RAMBeasties I've got laying around, and then start LL to lay my claim to them. It's a shame there isn't a P-1 100M exponent job type...unless I'm missing something. (I'm still learning)

[VBCurtis]

Originally Posted by aurashift

*maybe*

Can I be lazy and just select that as a work type in mprime? Or do I have to install something?
I think I tried ECM on the RAMBeasts (500G-1T) but they didn't use much RAM. I could split half the work between P-1 and ECM since P-1 maxes out the 500GB's

Uh.... a Terabyte?

The ECM routine in Prime95 uses a slower stage 2 that uses very little memory. The stage 2 in GMP-ECM is much much faster, and would use a fair amount of your monsters' memory. However, as Madpoo has discovered recently, the fast way to do ECM on M1277 is to use Prime95 for stage 1 (low memory use), save the residues from stage 1, and feed those residues to GMP-ECM stage 2. If you're willing to do this, stage 2 will use something like 34GB per thread (according to Madpoo, who recently tried this) with B1 = 2.9e9.

You could try a few huge curves, such as B1 = 7.6e9 (75 digit factor level) or B1 = 25e9 (80 digit factor level, near world-record territory http://maths-people.anu.edu.au/~brent/ftp/champs.txt). Those massive bounds would use yet more memory, and you can try altering B2 sizes (e.g. invoking -k 2 flag when launching GMP-ECM) for fun and to use even more memory.

Really, I'm just envious of your toys.

[VictordeHolland]

Originally Posted by aurashift

Seems like P-1 occurs after 76 bits, isn't there something like a 3% chance it'll find a factor at any bit level?

So on one of my cores it takes ~10 days to take it from 73 to 76 bits. Not sure if I'm looking at this right but it'll take another ~8 to do P-1.

Maybe I should start P-1 now on those 500GB-1T RAMBeasties I've got laying around, and then start LL to lay my claim to them. It's a shame there isn't a P-1 100M exponent job type...unless I'm missing something. (I'm still learning)

Wauw! Nice resources, I would like to have those .
It is indeed strange that you can't manually reserve 100M digit P-1 assignments. The Work Distribution Map (http://www.mersenne.org/primenet/) shows there are 186 available. It would be a shame to let those servers run TF on a single core for 10 days .

[aurashift]

I just had an idea. I'll give those five 40 core RAMBeasts 40 100M exponent jobs and let them chew on trial factoring and P-1 now, and then quit and unreserve those exponents before I get those 30 blades in. Should also leave some room for anonymous folks to take and abandon a bunch too.

[Uncwilly]

Originally Posted by aurashift

I'll give those five 40 core RAMBeasts 40 100M exponent jobs and let them chew on trial factoring and P-1 now

I noticed that you found a couple of factors.


BTW, to any reader of this thread, I found a factor for an exponent in this range a few days ago. The exponent had been assigned over 2 years ago to someone else. They had failed to do the 23 GHz-days worth of work to take it from 75 to 76 during that whole period. That is a 4900 GHz-days LL test. There have been several more like this.

If my slow climb taking the lowest exponents up to the next bit level over runs assignments that are over a year old, I don't feel guilty. For quite some time I had avoided those assigned. Having looked at the status of the first 1000 that are assigned to LL or P-1, greater than 3/4ths have not had a check-in in a year. Most of those have not checked in since the exponent was assigned.