mersenneforum.org

mersenneforum.org (https://www.mersenneforum.org/index.php)
-   Information & Answers (https://www.mersenneforum.org/forumdisplay.php?f=38)
-   -   Worker Windows - Optimal settings (https://www.mersenneforum.org/showthread.php?t=13649)

Unregistered 2010-07-24 17:17

Worker Windows - Optimal settings
 
i had several computers running prime95 for a few years, but stopped a couple years ago due to heat related issues on one of them. none of my prior experience was with hyperthreading cpus.

but now i decided to restart, and installed the program on 3 computers; one is single core, one is quad core, one is quad core with hyperthreading (i7)

the quad core with hyperthreading; i set this for 4 worker windows and 2 cpus to use, but all i get is 50% cpu usage. the load seems to be balanced on all 8 threads.

does this make sense? or is something set incorrectly? i tried various options, searched the forums (found a lot of conflicting info) but am not sure the best settings here.

currently i have it set for factoring only so i can get some quick results, and if i have to reset something i dont lose much cpu time.

thanks

lycorn 2010-07-24 20:15

Welcome back!...

For multi core PCs, the most efficient arrangement in terms of overall throughput is one worker window per core; a different test wil hence run on each core.
Hyperthreading is not useful for Prime05 purposes. On your HT machine, just assign 1 CPU per each of the 4 worker windows.
May I just suggest one thing? Your quad core machines are, in terms of overall project advancement, "underused" for Trial factoring. They are modern machines, and very capable at LL test work. The project is currently overpowered as far as TF goes, but a little short on the LL side, particularly for double-checks. P-1 is also a type of work that has few people working at. You might wish to turn some of your fire power to these types of work...
Please don´t construe my words as "pushing" you. It´s just that you´ve been away for some time, and I just wanted to put you in the current picture. Do as you see fit... and have fun.

Rhyled 2010-07-24 23:45

Core i7 suggestions
 
For what it's worth, I run an overclocked (3.7 GHz) Core i7 PC. I'll share what I learned along the way.

Hyperthreading on the i7 for Prime95 is a bad idea. Not only does the total throughput drop by 5%, but the core temperatures climb 5-10C. I simply turned off hyperthreading in BIOS. Whether or not you wish to do so may depend on the non-Prime95 work you do.

I hated going for weeks without a single result as several LL tests chugged along in parallel. I changed one worker to P-1 tests, which gives a 4-5 GHz-day payoff every 31 hours or so, and staggered the remaining 3 workers on LL duty to give me one result every 7-10 days instead of 3 results all on the same day. This staggering makes absolutely no difference to the overall project progress, but I get more consistent feedback that I'm actually contributing.

If you do allocate a worker to P-1, make sure you assign it a reasonable amount of memory. I think the default is 8 MB, but 1000+MB significantly boosts the chance of finding a factor. Most Core i7 systems have a decent amount of memory to match their processor.

On my secondary PC - an older Athlon X2, I run double-checks. Those LL-D checks take about 1/4 the time new LL checks take (with much higher exponents). I wasn't willing to wait 2.5 months for a result to appear.

Thank you for returning to Prime95. Whatever allocation you decide to make, we appreciate your contributions.

S485122 2010-07-25 06:36

[QUOTE=Rhyled;222704]If you do allocate a worker to P-1, make sure you assign it a reasonable amount of memory. I think the default is 8 MB, but 1000+MB significantly boosts the chance of finding a factor. Most Core i7 systems have a decent amount of memory to match their processor.[/QUOTE]The minimum amount of memory to allocate to Prime95 to receive P-1 assignments is 300 MiB, see the assignment rules on the GIMPS site. (8 MiB is the default memory allocation, if you do not change anything after installing. With that memory allocation the program cannot execute the second stage of the P-1 factorisation.)

Indeed 1000 Mib or more gives better chances of finding a factor, because the program will use higher bounds in its search. A thing to keep in mind : the core that will do P-1 work will use that memory two thirds of the time since stage 2 takes about twice the time of stage 1.

Jacob

Unregistered 2010-07-30 21:49

thanks for all the input!

i actually allocated 1600mb, but have plenty more if it would benefit from it (running 16gb of ram, so there is plenty)

so its been running about a week, and with the 7 cores (i7, quad core amd, and an old pentium) i already have done about 50% GHz days of what i did the last time i was involved; which took about an entire YEAR! unbelievable what these new cpu's can do (with unlimited memory). i actually set them on small trial factoring just to make sure everything was working right (since LL took months before, i wanted some feedback i had the worker threads set right since i hadnt done this with hyperthreading before), and, well, i look at the results page and see it has done about 8 thousand factors! so time to switch to something more productive i think....:)


All times are UTC. The time now is 00:05.

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