mersenneforum.org  

Go Back   mersenneforum.org > Great Internet Mersenne Prime Search > Hardware > GPU Computing

Reply
 
Thread Tools
Old 2015-08-23, 04:59   #1
dragonbud20
 
dragonbud20's Avatar
 
Mar 2014

34 Posts
Default a question about GPU to 72 assignments.

I was wondering if anyone could tell me what LMH Bit-first and depth-first are. I've been curious for a while about those options.
dragonbud20 is offline   Reply With Quote
Old 2015-08-23, 05:29   #2
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

3×2,917 Posts
Default

Quote:
Originally Posted by dragonbud20 View Post
I was wondering if anyone could tell me what LMH Bit-first and depth-first are. I've been curious for a while about those options.
They are two different ways of ultimately doing the same list of tasks.

Bit-first or bit-wise takes a group of exponents and runs the trial factoring on them all from the starting bit (we will use 65 for our example) to the next bit level (66). Then it will go through the same group of exponents (minus any that had a factor found) and take them to the next bit level (67). This is repeated until the ultimate goal bit level is achieved (we will use 72). So there were 7 passes through the exponents made.

Depth first takes the list of exponents and takes each one in order from the starting bit level (65) to the ultimate goal level (72) in sequence.

There are reasons to like both. There are reasons to use one or the other in certain situations.
Uncwilly is offline   Reply With Quote
Old 2015-08-23, 07:36   #3
dragonbud20
 
dragonbud20's Avatar
 
Mar 2014

34 Posts
Default

Ahh I see; huh that lines up with sample assignments now that I think about it. what kinds of situations would make one or the other preferable?
dragonbud20 is offline   Reply With Quote
Old 2015-08-23, 14:30   #4
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
Jun 2011
Thailand

22×7×317 Posts
Default

When the front wave of LL is burning our backs, we will struggle to give more "well TF-ed" expos to LL-ers, so deep first will be preferred. To make more credit, deep first is preferred too, for some Nvidia cards (due to the fact that mfaktc does not have a penalty for higher levels, and at higher levels you handle less exponents, the overload being minimal). Breadth first is preferred in all other situations, well.. arguably, this may be a subjective taste... or not. For example, you will find more factors doing lower bitlevels first, which are also faster to do (getting even more factors, i.e. clearing more exponents per time unit). This is subjective, because you do more 'overload', handling more exponents (including printing on screen, done by the same card), and you may get lower credit, and the higher bitlevels will still need to be done for the expos you have no factors. The choice becomes objective if you have an AMD card (because mfacto has a penalty for higher bits) - your card is more efficient in doing lower bitlevels. At the end, it depends on your hardware, and on the fact that you are TF-ing close to the LL front or not. The rest is subjective, a matter of taste. Everybody should do the work he likes, but all the work has to be done at the end. But don't worry, there is enough work for everyone!

Last fiddled with by LaurV on 2015-08-23 at 14:34
LaurV is offline   Reply With Quote
Old 2015-08-23, 14:36   #5
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

3·2,917 Posts
Default

What he said
</edit>

A few quick examples of where one type would make more sense than the other.

If you are trying to keep other machines fed with exponents for LL testing (and have just more than enough power to do so). It often makes sense to do the depth-first. Take each one to the goal bit depth first.

If you have one machine that is very efficient up to a certain bit depth and another that is equally fast everywhere. It then might make sense to use the first machine up to one bit depth, then the other to the deeper bits. This would call for bit-first.

If you don't have enough power to feed all of the machines exponents for LL, bitwise is generally called for. Getting all of the numbers up to x-bits removes about 1/x of the remaining exponents. As the bit level goes up, this number gets smaller. So, to eliminate the best fraction before the wave over takes you, doing it bit-wise will factor out the most numbers.

GPU72 balances this with the assignment type "Let GPU72 decide."

Last fiddled with by Uncwilly on 2015-08-23 at 14:38
Uncwilly is offline   Reply With Quote
Old 2015-08-26, 08:01   #6
dragonbud20
 
dragonbud20's Avatar
 
Mar 2014

34 Posts
Default

Right now I run this rig: http://pcpartpicker.com/user/dragonbud20/saved/72Trxr for all the work I do. Looking through what you guys have said I could either keep doing auto assign for both my p95 and gpu72 or I could factor and then feed my own LL machine or I guess find someone doing a specific project that requires whatever parameters. does this all sound about right?

Last fiddled with by dragonbud20 on 2015-08-26 at 08:01
dragonbud20 is offline   Reply With Quote
Old 2015-08-26, 12:10   #7
manfred4
 
manfred4's Avatar
 
Mar 2014
Germany

23·3·5 Posts
Default

Yes, that sounds fine. If you just take the automatic assignments from primenet and GPU72 then you will the most needed work that you can do - so for the project that is good!
If you wanted to you also could join small projects on this forum that do the the same work but just in a different order, but you don't need to to do the 'right' thing.
Have fun!
manfred4 is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Noob question on assignments stalag Information & Answers 9 2018-02-02 13:52
Question about GPU72 and expiring DC assignments Mark Rose GPU to 72 2 2014-09-04 21:45
Quick Question about assignments Dubslow PrimeNet 291 2011-11-17 11:50
assignments mpeisl11 Information & Answers 4 2009-04-28 19:22
Question about assignments and v4 days stealthaxe PrimeNet 2 2008-12-17 03:48

All times are UTC. The time now is 05:25.

Sat Oct 31 05:25:34 UTC 2020 up 51 days, 2:36, 2 users, load averages: 2.14, 2.07, 1.93

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.