mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > PrimeNet

Reply
 
Thread Tools
Old 2018-10-10, 16:14   #1
irowiki
 
Sep 2018

3·23 Posts
Default ECM work vs LL work

I notice slow machines get ECM work at first before getting a double check, and if I transfer that ECM work to a fast machine, the fast machine chews right through it.

Is there a benefit to the community/gimps to have a few fast machines chewing out ECM work or should I leave those on LL tests?
irowiki is offline   Reply With Quote
Old 2018-10-10, 16:54   #2
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

47·107 Posts
Default

Quote:
Originally Posted by irowiki View Post
I notice slow machines get ECM work at first before getting a double check, and if I transfer that ECM work to a fast machine, the fast machine chews right through it.

Is there a benefit to the community/gimps to have a few fast machines chewing out ECM work or should I leave those on LL tests?
ECM factoring is not relevant to the current GIMPS wavefronts of activities related to finding new Mersenne primes: TF, P-1 factoring, LL, PRP, p>80M.

I think the conventional wisdom is PRP & P-1 are the best uses of cpus, and TF the best use of gpus, for advancing the wavefronts and finding the next Mersenne prime.
P-1 assignments complete quicker than PRP or LL (primality tests) on the same hardware. Consequently P-1 is more suitable for slower cpus than primality testing.
But run what makes you smile the most, and keeps you from getting bored and quitting.
Running a mixed workload is fine, and prime95 makes it easy to do so automatically on multicore machines.
kriesel is online now   Reply With Quote
Old 2018-10-10, 17:55   #3
irowiki
 
Sep 2018

3·23 Posts
Default

Wow, that's a lot to digest.

So you're saying doing a P-1 is "quicker" and you do more of them but would give the same credit as doing a longer LL, in the end?

Either way the following (in the blue brackets) furthers the cause?



https://media.discordapp.net/attachm...64/unknown.png
irowiki is offline   Reply With Quote
Old 2018-10-10, 18:33   #4
Mark Rose
 
Mark Rose's Avatar
 
"/X\(‘-‘)/X\"
Jan 2013

3·977 Posts
Default

Quote:
Originally Posted by irowiki View Post
Wow, that's a lot to digest.

So you're saying doing a P-1 is "quicker" and you do more of them but would give the same credit as doing a longer LL, in the end?

Either way the following (in the blue brackets) furthers the cause?



https://media.discordapp.net/attachm...64/unknown.png
P-1 will give you approximately the same credit per CPU time as doing LL or PRP. P-1 assignments take less time. P-1 assignment do take a fair amount of RAM though, so if you have less than 1 GB, I wouldn't do them.

If the machines were given ECM automatically, it's probably the best work for them. ECM won't help find new primes, but will help factor numbers we know not to be prime. Some people are interested in that as a sub-project. The LL double check is to check if the hardware is reliable. By default, each machine will do a single LL double check per year. This helps us find unreliable machines so their work can be double checked earlier: there's a reasonable chance a prime has been missed.

Unreliable hardware should only run PRP as it can reliably detect errors and retry, whereas the other jobs cannot.
Mark Rose is offline   Reply With Quote
Old 2018-10-10, 19:21   #5
irowiki
 
Sep 2018

3·23 Posts
Default

Well I've had a few machines where Primenet detects them as P4 100mhz equiv to start, assigns them ECM work, and then after a while of ECM work, it realizes that "oh, this is actually a 5 ghz P4 equiv, here, have a doublecheck"

The only machine I had only get ECM work only was an Atom Netbook that I brought on as a test, and it took it days to do one assignment so I retired it.

I started in august, every machine I've brought online (almost 50) has done a few doublechecks then moved on to a LL check.

Oddly, when I bumped everything up to "get 10 days of work" they got confused and heaped on more doublechecks.


Edit: so P-1 Factoring helps weed out things so the LL tests are better, and I should get machines with a ton of ram and make them do P-1 tests?

Last fiddled with by irowiki on 2018-10-10 at 19:25
irowiki is offline   Reply With Quote
Old 2018-10-10, 22:18   #6
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

47×107 Posts
Default

Quote:
Originally Posted by irowiki View Post
Well I've had a few machines where Primenet detects them as P4 100mhz equiv to start, assigns them ECM work, and then after a while of ECM work, it realizes that "oh, this is actually a 5 ghz P4 equiv, here, have a doublecheck"

The only machine I had only get ECM work only was an Atom Netbook that I brought on as a test, and it took it days to do one assignment so I retired it.

I started in august, every machine I've brought online (almost 50) has done a few doublechecks then moved on to a LL check.

Oddly, when I bumped everything up to "get 10 days of work" they got confused and heaped on more doublechecks.


Edit: so P-1 Factoring helps weed out things so the LL tests are better, and I should get machines with a ton of ram and make them do P-1 tests?
For the basics see https://www.mersenne.org/various/works.php and https://www.mersenne.org/various/math.php
P-1 factoring bounds are tested for each run for the values which give optimal time savings given the probabilities of finding a factor or not, given the estimated time costs of P-1 factoring versus performing a number of LL or PRP tests on the exponent. P-1 done extensively may save the project time. But the system time you allocate to P-1 factoring won't find a Mersenne prime, although it may find some impressively large factors.
P-1 and PRP are performed with very similar calculations (3 raised to a power mod the Mp using DP fft transforms), and LL is close too, so the credit per cpu hour expended is close. If P-1 takes 1/40 the time of a primality test and has a 3% chance of a factor, you'll get 1000 P-1 factored in the time it takes to do 25 primality tests, and find about 30 factors, eliminating the need for ~30-60+ primality tests. (First LL, LL DC, and the occasional third test when residues don't match, for 60+. Or 30 PRPs.) So the project would be 5 primality tests ahead in that hypothetical case.
P-1 speed is helped by more RAM available to it, in stage 2. But slow machines can run it with 1GB available, at the current wavefront.
Double checks are welcome both because they test a system's reliability and because they help cut down on the growing double-check backlog.
kriesel is online now   Reply With Quote
Old 2018-10-11, 21:17   #7
irowiki
 
Sep 2018

3·23 Posts
Default

Quote:
Originally Posted by kriesel View Post
If P-1 takes 1/40 the time of a primality test and has a 3% chance of a factor, you'll get 1000 P-1 factored in the time it takes to do 25 primality tests, and find about 30 factors, eliminating the need for ~30-60+ primality tests. (First LL, LL DC, and the occasional third test when residues don't match, for 60+. Or 30 PRPs.) So the project would be 5 primality tests ahead in that hypothetical case.
Thank you, I had not wanted to put prime on my newish laptop because it's not on all the time, so it would get hung up doing LL tests, but if it can crank out P-1 factoring at a decent rate when it is on, that gives it something to do.
irowiki is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
The best work for my CPU MacMagnus Information & Answers 57 2013-11-22 16:27
How to calculate work/effort for PRP work? James Heinrich PrimeNet 0 2011-06-28 19:29
No Work Pilgrim Information & Answers 1 2008-01-31 18:53
Work to do for old CPU Riza Lone Mersenne Hunters 7 2006-03-15 22:57
It seems to work, but why ? T.Rex Math 15 2005-10-15 10:38

All times are UTC. The time now is 17:12.

Sun Apr 18 17:12:41 UTC 2021 up 10 days, 11:53, 1 user, load averages: 1.90, 1.89, 1.81

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.