How to use prime95 for stage 1 & GMP-ECM for stage 2

[philmoore]

Originally Posted by R.D. Silverman

However, it is clear that rather than run a SINGLE ECM curve to high limits (which is what P-1 to high limits accomplishes)
it is much better to run MANY ECM curves with lower limits.

This is what I thought, too, but Ernst and I did some calculations a few years back that indicated it was generally worthwhile to spend on the order of 5-6% or so of your total ECM effort on Mersenne or Fermat numbers to run P-1 to high limits. I was surprised that the percentage was that high, but the fact that some factors of P-1 were already known boosted the success probability considerably. I would guess that for generic numbers this percentage would be lower.

[Madpoo]

Originally Posted by R.D. Silverman

A better question:

Why would you want to? If people would ever bother to READ my ECM paper they might actually learn something.

Running P-1 is equivalent to running a single ECM curve. If it is known a priori that P-1 is divisible by (say) q, then one
simply reduces the size of the factor you are looking for by log(q) and then chooses the ECM parameters appropriately
for running just one curve.

However, it is clear that rather than run a SINGLE ECM curve to high limits (which is what P-1 to high limits accomplishes)
it is much better to run MANY ECM curves with lower limits.

This @&*!&*#%^ fascination with running P-1 to high limits is simply WRONG-HEADED. [unless of course, ECM is not
available]

A colorful reply as always.

I only asked because I was lurking on the forum where they were talking about the 332M-333M range (or whatever...I'm sure that's not exactly right). The concept of factoring some grandfathered LL assignments in that range came up, and I'd noted that quite a few of those assignments hadn't had any P-1 done at all. They were going to do some extra TF work on them up to 79 bits or something and I just wondered about the feasibility of doing the P-1 work on some of my systems with lots of RAM.

I have zero built-in idea of how long it takes to run P-1 work at all in that range, for any B1/B2 values, so I thought I'd first see if GMP-ECM would be advantageous.

[bur]

I am currently running stage 1 ECM on M22543 with B1=3M with 9 workers.


I noticed they all write their residues to results.txt. I then checked GMP ECM's loop option and found that it specifically is incompatible with -resume.


So how do I run several instances of GMP-ECM parallel for stage 2? Or isn't this possible? If not, can I salvage the stage 1 residues for an mprime stage 2 run? Doing 4000 curves on one core would take a bit too long... :)

[kruoli]

You would need to divide the stage 1 residues into n seperate files, then start n processes of GMP-ECM, each with their own file. Please be beware that this might use more memory than you have! In this case, you would need to use the -maxmem switch.

[bur]

Your reply sounds like this is not usually done, I can imagine due to the additional manual work. I'm wondering if splitting work over prime95/gmp-ecm is worth it for small numbers of curves and small B1.

Is there a table or similar for timings difference of prime95 vs gmp-ecm?

[kruoli]

In this case, I can only comment on how frequent I am doing it like this, but when I have a lot of curves that should go to GMP-ECM, I will definitely do it the way I described for the reasons you mentioned. If it is more than one or two days of work on a single core, I would deem it beneficial. For less than that, I would not care to spend manual work on it. But this is my personal preference.

Since we now got 30.9 as a (pre?) beta or alpha, such a table would change drastically. For 30.8 and older, I had better stage 2 even with exponents as large as 200k. In this case, one should experiment, but as George said, up to 50k, GMP-ECM is definitely the better choice.

[bur]

Ok, thanks, that leaves the question of how to report the results. In the early posts of this thread the solution was "mail it to George" who already then was understandably reluctant.


Can I just tailor a report.json.txt entry?

[kruoli]

Originally Posted by bur

In the early posts of this thread the solution was "mail it to George" who already then was understandably reluctant.

This would suprise me since mailing everything to George is the default way to report GMP-ECM results.

Edit: Your JSON would have a wrong/missing checksum.

[kriesel]

For reporting gmp-ecm ECM results manually, see http://www.mersenneforum.org/showthread.php?t=23143 posts 2, 5, 8
post 2 (Prime95):
"You have to reformat the results to look like it came from prime95. The manual results web page will then complain that the checksum isn't right. However, I can add you to the trusted users list and have that web page accept your results anyway."