ECM users - version 30.9/30.10 (see post#168)

[axn]

Originally Posted by congsz

I pick M2222219, which has completed to t25. I will work on it until it reached at least t35.

It is probably better to do t30 on 10 exponents rather than t35 on a single one. You're more likely to find a factor.

Originally Posted by congsz

If I want to reach t35 on the exponent, should I pick B1=250000, complete 640 curves, and then pick B1=1000000 and complete 1566 curves? Or just choose B1=1000000 from the start of my work?

If you're determined to do t35, going straight to 1M would probably be better. However, either approach would take roughly the same time ... because the server "prorates" t30 curves to the equivalent t35 curves. For example, if you do 640 curves at 250K, that will count as 160 curves at 1M, so you'll only have to do another (1566-160) curves at 1M to complete t35.

[congsz]

Originally Posted by axn

It is probably better to do t30 on 10 exponents rather than t35 on a single one. You're more likely to find a factor.

Yes, this is completely correct. But I prefer to stick to my choice, because I don't actually care too much about finding a factor or not in this ECM try. Denial140 is working in this range (2.22M-2.23M) now and he will make the major factoring effort here.

Originally Posted by axn

If you're determined to do t35, going straight to 1M would probably be better. However, either approach would take roughly the same time ... because the server "prorates" t30 curves to the equivalent t35 curves. For example, if you do 640 curves at 250K, that will count as 160 curves at 1M, so you'll only have to do another (1566-160) curves at 1M to complete t35.

Thanks for your explanation. :D

[seatsea]

Just got running ECM with 30.9 on my PC with a 1700X 8c/16t, and 16GB of RAM


Given the new stage 2 method memory use, if I run multiple workers, very frequently one or several workers will find itself waiting "Exceeded limit on number of workers that can use lots of memory" which is almost always only one at a time, even if there may be enough ram for 2 workers to do this. I have already set resources limits to the highest I can for RAM.


I thus assume it's better to run fewer workers with more threads to avoid having them wait even if multithreading is a bit less efficient. Is this a correct assumption? And/or am I overlooking a way to allow more tasks to use a lot of memory?

[Denial140]

If you are running curves on many exponents, simply having the exponents in the worktodo file will make it run stage 1 on one of the later exponents if the memory is not available. A similar effect for many curves on one exponent can be achieved by putting the exponent several times in the worktodo file.

The other thing that can be done is to set MaxHighMemWorkers=2 (or higher) in local.txt; the default setting is 1.

[seatsea]

Thanks!
I've seen it look ahead for stage 1 work, but despite this it's still not uncommon for it to run out of available low memory work, given that not all stage 2 use the full 14GB of available ram this should help.

Given this I've set 4 workers with 2 threads.

[lycorn]

Still, you should try toying with the MaxHighMemWorkers=n parameter. Set n to 3, and depending on the B1 and B2 sizes (more precisely, the relative durations of Stage 1 and Stage 2), you may almost never see the message about exceeding the number of workers using lots of memory.
What exponent ranges are you working on?

[seatsea]

I did set MaxHighMemWorkers=2 and stopped getting that message at all, I think because I had been running 8 single thread workers before, I had accumulated a backlog of stage 2.
As for exponents, I've set it to run ECM on mersenne cofactors

[SuikaPredator]

After observing 2 workers running on my machine for several days, I came up with the following question:

Does it make sense to let one worker to run incompleted ECM stage 2 when all the other workers are running stage 1?

To be more precise, assume that ECM jobs indexed a,b,c are assigned to worker #1, ECM jobs d,e,f are assigned to worker #2, MaxHighMemWorkers=1. If stage 1 of job a,d are completed at (about) the same time, then worker #1 will continue to run job a stage 2, and worker #2 will switch to job e(or f) stage 1. Worker #2 will switch back to job d stage 2 immediately after job a stage 2 completed, and will continue to do the next curve (stage 1) of job d. If after each curve of jobs a,d the two workers collide on stage 2, then worker #2 will finish job e,f stage 1 while waiting and have nothing to do.

My proposal is that instead of continuing job d stage 1 while worker #1 is doing stage 1 works, worker #2 should probably finish job e stage 2 so that it can have more choices of low memory tasks in the future.

Since time(stage2):time(stage1) is about 0.3, three ECM jobs will be enough to make full use of the "idle time" in the long run by this strategy.

See illustration below for better understanding (?)



I don't know if I missed something. Please don't hesitate to correct me if I considered something wrong.

[Prime95]

30.9 build 2

If you aren't doing ECM, do not download! PRE-BETA software!

What's new over build 1? Should be a little more stable and a little faster. Fixed nasty bug that sometimes caused subsequent curves to incorrectly calculate curve starting point.

Caveats? Same as build 1:
Save files during ECM stage 2 are still broken (probably a crash)
Stage 2 time estimates and optimal B2 bounds could be off
Accurate estimates of stage 2 memory consumed may be off
Further stage 2 multithreading improvements are needed.
Stage 2 is pretty verbose, there's lots of code cleanup in my future.

Windows 64-bit: https://mersenne.org/ftp_root/gimps/p95v309b2.win64.zip
Linux 64-bit: https://mersenne.org/ftp_root/gimps/...linux64.tar.gz

[axn]

A cosmetic bug in b2
nQx complete. Time: 5.552 sec.
PolyR built. Time: 1664077385.142 sec.
Poly compress. Time: 0.510 sec.

[axn]

Originally Posted by axn

PolyR built. Time: 1664077385.142 sec.

Apparently this is the unix time