mersenneforum.org  

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

Reply
 
Thread Tools
Old 2011-10-25, 02:58   #738
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
"name field"
Jun 2011
Thailand

24·613 Posts
Default

The credit in fact should be based not only on how difficult an assignment is, but also on how "needed" it is. I need this more than that, so I would pay more money for it, even if its value is the same, or lower. This is the key of any "marketing" system. So, George or other guys there at the helm of the boat, could give any credit they like, if they consider that some work is more necessary then the other, to stimulate the people to go for that "most needed" assignments. I started to do P-1 since this thread started**, because I was reading some arguments written by some guys who seem they know what they are talking about, in this very current thread. And I feel ok doing it, in spite of the fact that I never did it for credit, I still feel "important", as I am doing some "useful" work

**edit: correction: This thread started long ago, so in fact I am doing P-1 since I read this thread, in june, when I joined the forum and read all the interesting topics from the beginning. The argument that P-1 could be more needed than other work types still sounds reasonable for me now, especially after GPU's took over with trial factoring and cudaLL tests.

Last fiddled with by LaurV on 2011-10-25 at 03:07
LaurV is offline   Reply With Quote
Old 2011-10-25, 03:46   #739
Christenson
 
Christenson's Avatar
 
Dec 2010
Monticello

5×359 Posts
Default

Quote:
Originally Posted by Dubslow View Post
Didn't P95 say that was due to a lack of FFT size information? Or is this something else I missed?
Indeed P95 did say this was due to the server having to guess the FFT size.

I put this in the same class as getting odd amounts of credit for mfaktc factors found...not a very big problem.

Indeed, my CPUs are doing majority P-1...and P-1 is making some headway due to you guys help. At the moment, credit seems to be effort-based, with a big correction due because TF is so much faster on CPUs. The advantage of an effort-based system is that it is relatively easier to figure out the minimum effort to get the maximum effect, that is to prove the most mersenne exponents are not in fact associated with mersenne primes in the smallest amount of time.
Christenson is offline   Reply With Quote
Old 2011-10-25, 03:50   #740
Christenson
 
Christenson's Avatar
 
Dec 2010
Monticello

5·359 Posts
Default

Quote:
Originally Posted by kladner View Post
To whom it may concern (Mr. P-1,etc.):

I am currently running 3 1090T cores on P-1. 2 cores are feeding mfaktc on a GTX 460. (But that belongs in another thread.) 1 core is cleaning up the LL assignments which I moved from the other cores. In Win7-64 I have 8GB RAM, so I can handle giving fully adequate amounts to 4 cores once the LL's finish.

I have been taking whatever PrimeNet gives me in the P-1 department, without changing the depth.

I would gladly accept assignments from other sources, if it would be helpful.
At this point, PM either ckdo (27M TF assignments) or Mr P-1 (48-50M TF assignments) to get the best assignments. Mr P-1, I *assume* (with all the risks thereunto) that what the server is handing out for P-1 assignments is good, though I did note a 60M P-1 assignment in one of my queues this weekend. Or should I be running some of those non P-1'ed expired assignments you've been looking at?
Christenson is offline   Reply With Quote
Old 2011-10-25, 10:22   #741
nucleon
 
nucleon's Avatar
 
Mar 2003
Melbourne

5×103 Posts
Default

Quote:
Originally Posted by LaurV View Post
The credit in fact should be based not only on how difficult an assignment is, but also on how "needed" it is.
Yes! I was thinking something similar. A weighted Ghz-days (WGD) metric.

So the rankings would be based on a WGD. Then the project people can adjust metrics and people change the workload. You'd have to keep them similar as you don't want _all_ the nodes to do the highest rank.

Weighting examples:
TF: 0.9
LL:1.0
LLD:1.2
P-1: 1.5

Yes I understand it may not be feasible due to the extra work it would create for the server.

-- Craig
nucleon is offline   Reply With Quote
Old 2011-10-25, 13:03   #742
kladner
 
kladner's Avatar
 
"Kieren"
Jul 2011
In My Own Galaxy!

2·3·1,693 Posts
Default

Quote:
Originally Posted by Christenson View Post
At this point, PM either ckdo (27M TF assignments) or Mr P-1 (48-50M TF assignments) to get the best assignments. Mr P-1, I *assume* (with all the risks thereunto) that what the server is handing out for P-1 assignments is good, though I did note a 60M P-1 assignment in one of my queues this weekend. Or should I be running some of those non P-1'ed expired assignments you've been looking at?
Thanks, Christenson.
kladner is offline   Reply With Quote
Old 2011-10-25, 15:25   #743
Mr. P-1
 
Mr. P-1's Avatar
 
Jun 2003

100100100012 Posts
Default

Quote:
Originally Posted by Christenson View Post
At this point, PM either ckdo (27M TF assignments) or Mr P-1 (48-50M TF assignments) to get the best assignments.
My assignments range from 45M-52M, with a very small number higher, but still within the Test wavefront.

Also, with all due respect to ckdo, I question whether the assignments he supplies are "the best", whether your criteria is value to the project, or factors found. The value of a TF bit level of a DC exponent is the same as that four bits higher on an LL exponent twice the size. Most of my assignments are from 69-70. Unless ckdo's are 65-66 or lower, his are not as valuable as mine. Also many of mine have not had P-1 done, so there are more factors to find.

Quote:
Mr P-1, I *assume* (with all the risks thereunto) that what the server is handing out for P-1 assignments is good, though I did note a 60M P-1 assignment in one of my queues this weekend. Or should I be running some of those non P-1'ed expired assignments you've been looking at?
My own assumptions are that there is enough GPU capacity working on Primenet-assigned TFs too keep that wavefront ahead of the LLs, even if some are diverted to TFing at the LL wavefront, but there is insufficient CPU capacity to keep the P-1 wavefront ahead of the LLs.

If those assumptions are correct, then diverting P-1 effort to the LL test range will be like digging holes further up the road, in order to fill them in where you are now.

Despite this, I still think there is some benefit to doing P-1s at the LL test range. First, your results are put to use immediately rather than months or years down the line. Second, the P-1 bounds calculated are more likely to be based upon the actual amount of TF that will likely ever be done on the exponent, thus are more likely to be optimal. Third, we cannot tell the future. Perhaps the amount of P-1 capacity will increase so that by the time we reach the holes further up the road, they will have been filled in.

I currently have a small number of P-1 ripe exponents in the 45M range and above. As more and more TF workers report back, I will soon have a large number of them. There will be very little additional work involved in farming these out to whoever wants them.

Last fiddled with by Mr. P-1 on 2011-10-25 at 15:34
Mr. P-1 is offline   Reply With Quote
Old 2011-10-25, 15:43   #744
davieddy
 
davieddy's Avatar
 
"Lucan"
Dec 2006
England

194A16 Posts
Default

Quote:
Originally Posted by nucleon View Post
Yes! I was thinking something similar. A weighted Ghz-days (WGD) metric.

So the rankings would be based on a WGD. Then the project people can adjust metrics and people change the workload. You'd have to keep them similar as you don't want _all_ the nodes to do the highest rank.

Weighting examples:
TF: 0.9
LL:1.0
LLD:1.2
P-1: 1.5

Yes I understand it may not be feasible due to the extra work it would create for the server.



-- Craig
This is the daftest post I've ever read.

It is tricky enough to come up with a sensible measure of "work done",
without a subjective "value-added" component.

This is what is important:

GPUs TF as far as poss before 1st time LLs are assigned to CPUS.
The CPU does P-1 if necessary, and completes its test.

DCs will look after themselves.

End of story.

David

Last fiddled with by davieddy on 2011-10-25 at 15:58
davieddy is offline   Reply With Quote
Old 2011-10-25, 15:57   #745
James Heinrich
 
James Heinrich's Avatar
 
"James Heinrich"
May 2004
ex-Northern Ontario

67128 Posts
Default

Quote:
Originally Posted by davieddy View Post
This is the daftest post I've ever read.
Congrats nucleon, that's quite the achievement! (You even beat out post #711 on the previous page of this thread )

Back to topic, I understand where Craig was going with his idea, but it's probably best to achieve the desired effect with more tweaking to the "Whatever makes sense" PrimeNet worktype. If L-Ls were only handed out to machines that can't reasonably run P-1 (due to lack of RAM) and anyone cable of running it was assigned P-1, then the assignment landscape would change very quickly.
James Heinrich is online now   Reply With Quote
Old 2011-10-25, 16:22   #746
Dubslow
Basketry That Evening!
 
Dubslow's Avatar
 
"Bunslow the Bold"
Jun 2011
40<A<43 -89<O<-88

3·29·83 Posts
Default

Quote:
Originally Posted by James Heinrich View Post
Congrats nucleon, that's quite the achievement! (You even beat out post #711 on the previous page of this thread )
Ehh...
I swear that's not a usual occurrence.
Usually.

I did have fun making that post though, for what it's worth.

Last fiddled with by Dubslow on 2011-10-25 at 16:23
Dubslow is offline   Reply With Quote
Old 2011-10-25, 22:44   #747
davieddy
 
davieddy's Avatar
 
"Lucan"
Dec 2006
England

2·3·13·83 Posts
Default Cable physicist

Quote:
Originally Posted by James Heinrich View Post
If L-Ls were only handed out to machines that can't reasonably run P-1 (due to lack of RAM) and anyone cable of running it was assigned P-1, then the assignment landscape would change very quickly.
I allowed 400MB of my Gig of RAM for stage two, and noted here the effect
it had on my browsing. But it only took a day or so, so I don't really see
what the fuss is about.

David
davieddy is offline   Reply With Quote
Old 2011-10-25, 23:40   #748
fivemack
(loop (#_fork))
 
fivemack's Avatar
 
Feb 2006
Cambridge, England

22×32×179 Posts
Default

Quote:
Originally Posted by Mr. P-1 View Post
The relevant metric here, surely, is not "factors per GHz-Days" but "expected GHz-Days saved by running this P-1".
But optimising one is the same as optimising the other; you save GHz-days only by finding a factor, and the number of GHz-days you save (for the two LL tests) is not a function of the effort required by the P-1.

(a 54M takes about 180 GHz-days for two LL tests, so P-1 testing that costs less than 180 GHz-days per expected factor is worth doing; but parameters that cost 65 GHz-days per expected factor should definitely be used ahead of parameters that cost 90 GHz-days per expected factor)
fivemack is offline   Reply With Quote
Reply

Thread Tools


All times are UTC. The time now is 00:41.


Thu Dec 2 00:41:02 UTC 2021 up 131 days, 19:10, 1 user, load averages: 0.56, 0.62, 0.79

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.