mersenneforum.org  

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

Reply
 
Thread Tools
Old 2012-11-30, 23:15   #1
swl551
 
swl551's Avatar
 
Aug 2012
New Hampshire

14508 Posts
Default Does trial factoring to 75+ make sense

I have factored dozens of my assignment all the way to 75 and both MFAKTO and MFAKTC handled it fine. GIMPS gave me full credits. However my question is more on the statistical side. Does it provide any value to factor that high. I know GPU72 defaults to 73 but why isn't the range higher. Why is the "system/primenet/gpu72" as whole constantly handing off the next range. In other words why do I get an assignment from primenet of 69,70 just to see 70,71 to handed off to someone else and so on. Why am I not assigned 69,75 or 69,80 or whatever the limit is of the T.F. engines right from the start?
swl551 is offline   Reply With Quote
Old 2012-11-30, 23:58   #2
kracker
ἀβουλία
 
kracker's Avatar
 
"Mr. Meeseeks"
Jan 2012
California, USA

32·241 Posts
Default

Because it just isn't worth it. I'm sure you know as you go up, every bit level up takes approx. double from before, it's just worth it to do a damn LL test instead of having the "chance" to find a factor
kracker is offline   Reply With Quote
Old 2012-12-01, 00:08   #3
swl551
 
swl551's Avatar
 
Aug 2012
New Hampshire

23×101 Posts
Default

Quote:
Originally Posted by kracker View Post
Because it just isn't worth it. I'm sure you know as you go up, every bit level up takes approx. double from before, it's just worth it to do a damn LL test instead of having the "chance" to find a factor
if you factor an exponent 69 to 70 and then I factor that same exponent from 70 to 73 how did that benefit you more than me? Each bit level doubles the work and primenet gives the correct increased credits. (factoring 69->75 is approx 121 ghz days credits)

So I not sure I see your logic other than the lower ranges could find a factor faster, but finding a factor does not get you some huge ghzdays bonus.
swl551 is offline   Reply With Quote
Old 2012-12-01, 00:22   #4
kracker
ἀβουλία
 
kracker's Avatar
 
"Mr. Meeseeks"
Jan 2012
California, USA

41718 Posts
Default

If you're doing it just for GHz-days, yeah of course. But if you're looking for *primes* with the best "efficiency", it's not worth going over 72-3 (gpu72)

Last fiddled with by kracker on 2012-12-01 at 00:23
kracker is offline   Reply With Quote
Old 2012-12-01, 00:37   #5
swl551
 
swl551's Avatar
 
Aug 2012
New Hampshire

80810 Posts
Default

I'm really looking for a scientific or statistical answer. At what point does TF factoring become a waste because additional effort is better served by X.
And what is X (LL testing?)
swl551 is offline   Reply With Quote
Old 2012-12-01, 00:55   #6
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

249C16 Posts
Default

Quote:
Originally Posted by swl551 View Post
I'm really looking for a scientific or statistical answer. At what point does TF factoring become a waste because additional effort is better served by X.
And what is X (LL testing?)
Look up your GPU here:
http://www.mersenne.ca/cudalucas.php?model=26
It will tell you how far you should go before LL would be a better use of the resources.
Uncwilly is online now   Reply With Quote
Old 2012-12-01, 01:11   #7
swl551
 
swl551's Avatar
 
Aug 2012
New Hampshire

23×101 Posts
Default

Quote:
Originally Posted by Uncwilly View Post
Look up your GPU here:
http://www.mersenne.ca/cudalucas.php?model=26
It will tell you how far you should go before LL would be a better use of the resources.
BINGO!

Thanks that is exactly what I was trying to understand.
swl551 is offline   Reply With Quote
Old 2012-12-01, 16:21   #8
bcp19
 
bcp19's Avatar
 
Oct 2011

7·97 Posts
Default

Quote:
Originally Posted by swl551 View Post
BINGO!

Thanks that is exactly what I was trying to understand.
Best way to answer your question is that when the bit levels were first set up, you looked at the length of time it would take to do an LL test. Let's call this X. At the point in time it takes longer than 2X for the CPU to find a TF factor on an exponent (For our purposes we'll say 1 in 100 tests [1%] have a factor), it is now more efficient to perform the LL and DC tests time wise. So if X = 30, 2X = 60 .01*2X = .6 or 14.4 hours is the TF breakeven point.

Enter the GPU. The GPU can TF much faster than a CPU, but uses part or all of the CPU in addition to the GPU. Since Nvidia GPUs have the capability of performing LL tests, you now have a more complicated formula to determine breakeven points. It still takes 2X for the CPU to run the LL and DC tests, but if the GPU was running the LL it could finish in Y. I no longer have access to the testing I had done, but let's say it takes 30 days for the CPU to run the LL and 6 days for the GPU to run it. Since the GPU barely impacts the CPU when running LL tests, we will just use the given times. So in the 60 days given, the GPU could crank out 10 tests to the CPU's 2. 60/12 = 5 days average. We do this cause we lose both the GPU and the CPU when we are TFing. 5 days * .01 = 1.2 hours. We estimate the GPU to be ~100 faster than the CPU on TF, so the 14.4 hours would become ~8 min 42 sec. The next bit levels would then be 17m 24s, 34m 48s and 69m 36s. Since we are looking at 1.2 hours, you can see the LL would become the faster option again.
bcp19 is offline   Reply With Quote
Old 2012-12-02, 04:11   #9
swl551
 
swl551's Avatar
 
Aug 2012
New Hampshire

23·101 Posts
Default

Nice write up.

Thanks.


I'm experimenting with CudaLucas as I write this.
swl551 is offline   Reply With Quote
Old 2012-12-03, 02:06   #10
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
Jun 2011
Thailand

52×7×53 Posts
Default

Quote:
Originally Posted by bcp19 View Post
...<snip>...
Good post!

@OP: for more detailed explanation, you can read the gpu-to-72 status thread. We gave a lot of explanations and arguments why the limits (69, 70, and 72, 73 bits) were chosen for the ranges they were chosen. But most probably you already got the idea, from bcp's post. You have to keep in mind always that the goal is to clear the exponents.

For example: if you have a gtx580 board and you want to work at DC front (25M-30M exponents), then you need about 14-17 hours to clean one exponent. If you TF to 69 bits, you may need half hour or a bit less for one assignment, and if you find a factor, you "clear" the exponent, therefore "saving" the time for one (single) LL test. But you will find a factor in 40 or 50 "trials", therefore you will need 20-25 hours to clear an exponent (take this numbers just as an example only, they are not exactly the real figures)

In this light, doing DC-TF to 69 bits is "at the limit", on the 25-29M range of exponents, and this was my argument in that thread, you can do it if you want, but you can be better and faster doing just DC-LL. Of course, you will get much more credit doing TF, but you may help the project less - the goal is to clear as many as possible exponents in as short as possible time. Doing TF to 70 (or more) bits in this case will be a waste of time and resources.

Last fiddled with by LaurV on 2012-12-03 at 02:09
LaurV is offline   Reply With Quote
Old 2012-12-03, 21:14   #11
aketilander
 
aketilander's Avatar
 
"Åke Tilander"
Apr 2011
Sandviken, Sweden

2×283 Posts
Smile "time"

"time"

as mentioned above is a good way of measuring things since the production of a videocard is very, very different depending on if it is doing TF or LL.

Example GTX 580
TF: 287.5 GHz-days/day
LL: 35.4 GHz-days/day (at best)

Besides this it is needed to take into account the error-rate (disadvantage for LL since each workunit is larger) and that there are not so many users who use their videocards for either TF or LL, most users only use them for TF I think.
aketilander is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
GPU Trial Factoring FAQ garo GPU Computing 100 2019-04-22 10:58
trial factoring and P-1 jocelynl Math 8 2006-02-01 14:12
What is Randomness and does relative randomness make sense? sakreha Data 1 2005-02-04 23:29
Trial factoring Citrix Software 7 2004-02-26 03:24
About trial factoring gbvalor Math 4 2003-05-22 02:04

All times are UTC. The time now is 14:15.

Fri Mar 5 14:15:00 UTC 2021 up 92 days, 10:26, 0 users, load averages: 2.39, 2.60, 2.68

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.