mersenneforum.org  

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

Reply
 
Thread Tools
Old 2003-12-13, 17:36   #1
PrimeCruncher
 
PrimeCruncher's Avatar
 
Sep 2003
Borg HQ, Delta Quadrant

2×33×13 Posts
Default Time to change LL vs. TF weight?

From the PrimeNet FAQ:

Quote:
Primality test time is purposefully biased.
...
...about a 10:1 ratio of LL CPU time to factoring CPU time.
...
I needed a way to balance the zoom-ahead factoring so those machines don't run out of things to do, which is easy to do until stronger factoring programs are available. The best way to do that is not to reward the time spent on it as much as the necessary but slow LL tests.
Given that we are now noting a convergence of the first-time LL tests and TFs maybe it is time to rank them a little more evenly to encourage people to spend a little more time on TFs.
PrimeCruncher is offline   Reply With Quote
Old 2003-12-13, 18:00   #2
nfortino
 
nfortino's Avatar
 
Nov 2003

3×5×11 Posts
Default

Quote:
Originally posted by PrimeCruncher
Given that we are now noting a convergence of the first-time LL tests and TFs maybe it is time to rank them a little more evenly to encourage people to spend a little more time on TFs.
I would hold off on that for two reasons.
1) It is hard to tell what will happen after the run-up due to M40.
2) At 25.3M, there is an FFT size cut-off, at which point LL tests will take 20% longer than they do now. Considering the TF leading edge is at 24.1M, and the LL leading edge is at 22.0M, it is safe to assume TF will reach the cut-off before LL catches it. I do agree, however, that something will have to be done before TF reaches 28M, and starts factoring to 2^68.

Last fiddled with by nfortino on 2003-12-13 at 18:01
nfortino is offline   Reply With Quote
Old 2003-12-13, 21:24   #3
GP2
 
GP2's Avatar
 
Sep 2003

2×1,289 Posts
Default Re: Time to change LL vs. TF weight?

Quote:
Originally posted by PrimeCruncher
From the PrimeNet FAQ:



Given that we are now noting a convergence of the first-time LL tests and TFs maybe it is time to rank them a little more evenly to encourage people to spend a little more time on TFs.
Putting things in perspective, the convergence is still quite some time in the future:
http://opteron.mersenneforum.org/png/leading_edge.png
GP2 is offline   Reply With Quote
Old 2003-12-13, 23:45   #4
PrimeCruncher
 
PrimeCruncher's Avatar
 
Sep 2003
Borg HQ, Delta Quadrant

12768 Posts
Default Re: Re: Time to change LL vs. TF weight?

Quote:
Originally posted by GP2
Putting things in perspective, the convergence is still quite some time in the future:
http://opteron.mersenneforum.org/png/leading_edge.png
The problem is that that is leading-edge TF vs. leading-edge LL. You have to look at it from a perspective of trailing-edge TF vs. leading-edge LL. From that perspective, the convergence has already ocurred. Now I'm not sure exactly how the server is handling this but I see one of two possibilities:

a) Assign LLs concurrently with TFs, which defeats the entire purpose of doing trial factoring

or

b) Assign LLs after the TF completes, which will lead to long delays in LL testing

Either alternative is unacceptable in my opinion, which means we need to get more PCs to do TFs.
PrimeCruncher is offline   Reply With Quote
Old 2003-12-14, 01:25   #5
GP2
 
GP2's Avatar
 
Sep 2003

A1216 Posts
Default Re: Re: Re: Time to change LL vs. TF weight?

Quote:
Originally posted by PrimeCruncher
The problem is that that is leading-edge TF vs. leading-edge LL. You have to look at it from a perspective of trailing-edge TF vs. leading-edge LL. From that perspective, the convergence has already ocurred.
How do you define the "trailing edge"? It doesn't exist. The leading edge is hard and sharp, but below that there's just a long exponential decay tail. Technically, you could claim that trailing-edge LL and leading-edge DC have converged too.

I'm pretty sure the server just waits for the TF to complete or expire before assigning that exponent for LL testing. Sometimes that causes long delays, but long delays for a small fraction of exponents are a fact of life on this project -- there are more than 1500 exponents under 17M that have not yet been LL tested to completion (or factored). Usually the delay is due to a slow LL test... in a few cases it will be due to a slow TF.

The main thing is, computers doing LL testing have a plentiful supply of already-trial-factored exponents to work on. Only about 1.5% of the exponents at the current LL leading edge have not yet been trial factored. By contrast, you have to go as far back as 7M below the LL leading edge before the number of outstanding LL tests drops to such a low percentage.

Clearly, the main cause of delay for the overwhelming majority of long-delayed exponents is not slow trial factoring, but slow LL testing.
GP2 is offline   Reply With Quote
Old 2003-12-14, 01:44   #6
PrimeCruncher
 
PrimeCruncher's Avatar
 
Sep 2003
Borg HQ, Delta Quadrant

12768 Posts
Default Re: Re: Re: Re: Time to change LL vs. TF weight?

Quote:
Originally posted by GP2
How do you define the "trailing edge"? It doesn't exist.
If you look at the main page, it does exist. It starts in the 20.1M range. Usually there is a long trail but you'll note that almost every range from 20.8M through the leading edge have dozens and even hundreds of TF tests still out.

Quote:

...long delays for a small fraction of exponents are a fact of life on this project...

Only about 1.5% of the exponents at the current LL leading edge have not yet been trial factored.
Small fraction being the key word there. If LL continues to advance on TF that small fraction is going to get a lot bigger.

Quote:

Clearly, the main cause of delay for the overwhelming majority of long-delayed exponents is not slow trial factoring, but slow LL testing.
I'm not talking about long-delayed exponents. We all know LL tests are slow while TFs are generally fairly fast. I completely agree with that point but I don't see how it is relevent.

In your post you say that you think the server waits for the TF to complete or expire. If we're waiting for TFs to complete or expire we can't be doing LL tests which means a shortage and maybe a nonexistance of first-time LLs. If we take LLs from beyond the leading edge of TF then we defeat the purpose of TF.

I'm not trying to be an alarmist here, but we have to face reality. At this rate LL tests will overtake TFs and then we will have a large problem to deal with. The solution will not be an easy one and it will take time to implement and allow the results to filter through, which is why we should implement it soon to prevent that from happening.
PrimeCruncher is offline   Reply With Quote
Old 2003-12-14, 01:50   #7
outlnder
 
outlnder's Avatar
 
Aug 2002

2×3×53 Posts
Default

I tend to agree that it is time to give TFing more value.

I don't really care about whether the LL tests are catching up or not, if my computer is doing valid work on this project, why shouldn't I get the same credit someone doing an LL or DC test does??????
outlnder is offline   Reply With Quote
Old 2003-12-14, 02:10   #8
GP2
 
GP2's Avatar
 
Sep 2003

2·1,289 Posts
Default

Attached is a graph drawn from a recent summary.txt.

I don't agree that 20.1M constitutes the trailing edge of LL testing. The graph doesn't show it. There's a small dip there, but that turns out to be due to a large number of manually-assigned exponents in that particular 0.1M range: half that range (20.104M to 20.15M) is reserved for manual assignment (but not assigned to anyone).

At the current leading edge of LL testing, we're waiting for only about 1.5% of the exponents to complete TF -- about 30 out of 2000 -- which leaves the other 98.5% available for LL testing. That's not a shortage.

There won't be a shortage of LL tests unless the leading edge of LL testing catches up to the leading edge of TF. But at current rates of convergence (http://opteron.mersenneforum.org/png/leading_edge.png) that won't happen for many years.
Attached Thumbnails
Click image for larger version

Name:	summary_histogram.png
Views:	350
Size:	4.9 KB
ID:	85  

Last fiddled with by GP2 on 2003-12-14 at 02:14
GP2 is offline   Reply With Quote
Old 2003-12-14, 02:16   #9
GP2
 
GP2's Avatar
 
Sep 2003

2×1,289 Posts
Default

For what it's worth, we're already seeing an increase in trial-factoring P90 CPU-years per day, probably due to post-M40 new accounts:
http://opteron.mersenneforum.org/png/TFspeed.png

By contrast, we haven't yet seen a boost in LL testing rate... in fact, it's actually dipped recently (to below 700 CPU-years / day).
GP2 is offline   Reply With Quote
Old 2003-12-14, 02:24   #10
GP2
 
GP2's Avatar
 
Sep 2003

2·1,289 Posts
Default

George,

Perhaps the manual range 20104000 to 20150000 (listed as '***** not assigned to anyone!!! ***') should be in fact be released to PrimeNet.

Especially considering the relative proximity to M40... if by any tiny chance an "island theory" twin of M40 is hiding in there, it would be nice to flush it out.
GP2 is offline   Reply With Quote
Old 2003-12-14, 02:32   #11
nfortino
 
nfortino's Avatar
 
Nov 2003

3·5·11 Posts
Default

Quote:
Originally posted by GP2

I don't agree that 20.1M constitutes the trailing edge of LL testing.
The trailing edge of TF is 20.1M, not that of LL. I must point out that any fix will only affect the leading edge, not the trailing edge. Also, I believe the only reason we have this problem is because there were two occasions, Aug 18 and Sep 20, in which the leading edge had a huge setback, presumably due to manual reservation ranges being released. Had the exponents farther out not been trial factored before the smaller ones, they would be finished by now. Finally, as it stands, it would take over 100 days for LL testing to reach the current leading edge of TF at a rate of 1000 P90 years per day (way above current.) A single TF should never take more than 100 days, even if it expires once, and is then reassigned. I still don't think we have enough of a problem to open the can of worms involved in changing the stats.
nfortino is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
New low weight project Citrix Riesel Prime Search 54 2014-01-08 23:55
Low weight k's kar_bon Riesel Prime Data Collecting (k*2^n-1) 18 2010-05-14 08:49
Low Weight Subsequences masser Sierpinski/Riesel Base 5 17 2007-02-14 02:04
Is it time to change the CPU year measurement? E_tron Lounge 7 2004-06-29 10:17
Low Weight 15k Citrix 15k Search 20 2004-06-20 21:00

All times are UTC. The time now is 13:52.

Mon Jun 1 13:52:23 UTC 2020 up 68 days, 11:25, 2 users, load averages: 1.55, 1.70, 1.73

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.