mersenneforum.org  

Go Back   mersenneforum.org > New To GIMPS? Start Here! > Information & Answers

Reply
 
Thread Tools
Old 2020-03-17, 06:30   #1
phillipsjk
 
Nov 2019

111102 Posts
Default Best use of one fortnite of CPU time?

I moved my assignments from my (old) dual Xeon machine to my quad AMD Bulldozer. They are now expected to complete [within] about a month (at an estimated power cost of $70 (600W, 30 days, ~$0.16/kWh) instead of 18 months (estimated power cost of $200 or more).



One of my (underclocked to 1.4Ghz) 16 core CPUs with 32GB of Quad channel RAM is expected to be idle for about 2 weeks while the other CPUs complete their assignments. I feel I have enough RAM to do trial factoring, but the consistent advice on this forum seems to be let the GPUs do trial factoring.


The thread by the new person was helpful. While I can complete two LL double-check assignments with this CPU in two weeks (I have one CPU that was going to run out of (DD) work in 3 days doing 4 of those tests), it was pointed out factoring may be more efficient long-term.



It was also mentioned that you people desperately want PRP reliability data. Presumably PRP double checking will take longer due to a larger exponent in use.


Assuming I decide to do factoring against advice, which type of factoring is most helpful?


I appear to have (2 - Trial factoring), (4 - P-1 factoring), and (5 - ECM for first factors of Mersenne numbers) to choose from. I really want any work I choose to complete in two weeks: because the machine takes 300W to just idle. I would save only about $4 (75W, 2 weeks, ~$0.16/kWh) leaving the one CPU idle (presumably the missing 100W is cooling). The system is running from an 8GB USB stick, in case that is a consideration.

Last fiddled with by phillipsjk on 2020-03-17 at 06:31 Reason: in->within
phillipsjk is offline   Reply With Quote
Old 2020-03-17, 07:02   #2
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
Jun 2011
Thailand

23·29·37 Posts
Default

Don't do TF on CPU !
It will be a completely waste of resources.

With 32G of QC RAM, that machine is ok to do P-1. Give it few gigs for stage 2 and here you go. You will finish many (like in 20 to 50, depending of how high B1 and B2) assignments in 2 weeks, one P-1 can take from few hours up (to how long you like).

P-1 and LL/PRP do the same type of calculus (i.e. floating point, FFT multiplication) and it is well suited in both CPU and/or GPU.

TF is different, it involves manipulating lots of small integers, moving them from one side of your computer to the other side and back, and it is better suited for GPU (which was designed to move a lot of pixels on screen from here to there, or from green to blue). Haha, I like how this turned up. Anyhow, your CPU will be much slower doing that. Like in 10 to 200 times slower. Times, not percent. But you can try anyhow, for a short time, to see how it goes. Then, once convinced, move to more serious things

Your rig, your money.

Last fiddled with by LaurV on 2020-03-17 at 07:07
LaurV is offline   Reply With Quote
Old 2020-03-17, 14:44   #3
phillipsjk
 
Nov 2019

368 Posts
Default

OK, I did not realize there was such a distinction between Trial Factoring and P-1 factoring.


I suppose the problem with ECM is that it is probabilistic, and only helps the project if I actually find new factors.
phillipsjk is offline   Reply With Quote
Old 2020-03-17, 17:34   #4
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

2×13×157 Posts
Default

Quote:
Originally Posted by phillipsjk View Post
OK, I did not realize there was such a distinction between Trial Factoring and P-1 factoring.
PRP, LL, and P-1 factoring use long fft length double precision floating point computation to do multi-megabit math modulo the Mersenne number. ("Irrational base discrete weighted transform")

TF does sort of grammar-school style 32-bit few-word powering modulo the candidate factor being tried, approx 72 to 86 bits depending on assignment and needed depth that is a function of exponent. Plus lots of sieving and other optimizations. GPU TF uses the hundreds or thousands of parallel processor units on a gpu to do many batches of candidate factors in parallel.

The ratio between fft DP GhzD/day (primality and P-1) and integer GhzD/day (TF) on a cpu is generally 0.7 to 1.4 depending on processor type. The ratio on a gpu is typically TF/fft = 11.0 to 42.0

TF has the smallest memory footprint, primality testing larger, P-1 factoring stage 2 largest of all.

Set prime95 or mprime's allowed memory to something substantial (8G, 16G, 24G, not the mere default 8M) and set day and night values the same so it does not restart the P-1 factoring at allowed memory changes daily.

For more background, see https://mersenneforum.org/showthread.php?t=24607
kriesel is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Time to End davar55 Lounge 4 2013-02-23 02:40
P95 est time to go seems off bcp19 Software 1 2012-08-03 22:39
How much time Unregistered Information & Answers 4 2008-12-20 21:00
We'll be done in no time!!!! petrw1 Hardware 6 2006-11-30 08:49
P3 TF time PrimeCruncher Software 30 2003-12-21 05:26

All times are UTC. The time now is 06:08.

Mon Jul 13 06:08:35 UTC 2020 up 110 days, 3:41, 0 users, load averages: 2.08, 2.03, 1.99

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2020, 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.