mersenneforum.org  

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

Reply
 
Thread Tools
Old 2010-06-30, 13:41   #1
davieddy
 
davieddy's Avatar
 
"Lucan"
Dec 2006
England

2×3×13×83 Posts
Default Is GIMPS grinding to a halt?

As if an increment in exponent of 4M/year wasn't
pessimistic enough wrt expectation of the next Mprime,
it seems the allocation wavefront is stuck below 51M ATM.

David
davieddy is offline   Reply With Quote
Old 2010-06-30, 13:57   #2
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

3·5·72·13 Posts
Default

Quote:
Originally Posted by davieddy View Post
it seems the allocation wavefront is stuck below 51M ATM.
Maybe the smart assignments are helping to fill in the gaps and the improved autoselection of "what makes sense" assignements is helping dole out more double checks.
Uncwilly is offline   Reply With Quote
Old 2010-06-30, 14:28   #3
petrw1
1976 Toyota Corona years forever!
 
petrw1's Avatar
 
"Wayne"
Nov 2006
Saskatchewan, Canada

110328 Posts
Default

Quote:
Originally Posted by davieddy View Post
As if an increment in exponent of 4M/year wasn't
pessimistic enough wrt expectation of the next Mprime,
it seems the allocation wavefront is stuck below 51M ATM.

David
I saw very similar events in V4 in 2008. It seemed to be stuck at 39M "for ever" because just when it cleaned up a wave of abandoned/expired exponents and was just starting to move forward the next wave would be released.
petrw1 is online now   Reply With Quote
Old 2010-06-30, 16:50   #4
davieddy
 
davieddy's Avatar
 
"Lucan"
Dec 2006
England

2×3×13×83 Posts
Default

Quote:
Originally Posted by petrw1 View Post
I saw very similar events in V4 in 2008. It seemed to be stuck at 39M "for ever" because just when it cleaned up a wave of abandoned/expired exponents and was just starting to move forward the next wave would be released.
That suggests a FFT boundary effect.

David

PS that must have been very early 2008 (like 2007:)
because the record 43M prime was found Aug 2008.
What fun that double find was!

Last fiddled with by davieddy on 2010-06-30 at 17:15
davieddy is offline   Reply With Quote
Old 2010-06-30, 18:12   #5
NBtarheel_33
 
NBtarheel_33's Avatar
 
"Nathan"
Jul 2008
Maryland, USA

111510 Posts
Default Several things are happening here

First is the sheer length of the newer tests. Granted, there are lots of shiny Core processors out there, but I would venture to guess that P4 boxen are the main poison of choice for GIMPS (especially in schools and universities - I bet most of curtisc's machines are P4s, for instance). Now P4's are fine crunchers, but on 50M exponents, they can take anywhere from 6-8 weeks. That's a yield of only 7 or 8 LL tests a year per core. This is going to get worse as time goes on. Betsy, a 3.06-GHz P4, says that she will take upwards of 4 years to crunch an LL on M1982xxxxx (I was flirting with the idea of making that one of this coming year's birthday projects - I was born in 1982 - but it would take four birthdays, not just one - maybe I'll try on a Quad!). Obviously, in the next few years, we are going to reach a point where P4s are going to have to be in the sole domain of the doublecheckers.

Next, how many cores are actually contributing regular results to GIMPS? I see something like 300,000 CPUs registered (in the "Today's Numbers" on the left hand side of the home page), but it's clear from the PrimeNet status report that only about 10-15% report (any kind of, whether TF, LL, P-1, or DC) results in an average month. Clearly, there are many of these cores which have either registered in error (say during a stress test), tried an assignment and given up after seeing the required time commitment, or are no longer even participating in GIMPS. In some ways, it might be worth flushing out any registered CPU not reporting at least a nominal amount of work - say 0.1 GHz-day or better - in a year's time. It would at least make the statistics a little more sensible.

Note that GIMPS has formidable computing power, but it's not necessarily large on a per-user basis. Consider that I'm almost in the lifetime top 100 (out of tens of thousands of registered users) for LL testing, and I only have 39 cores running. I'd bet that 90+ % of users are only running GIMPS on 1-4 cores, and possibly only part time due to business hours or electricity bill concerns.

Then you have the weird way in which assignments were handed out, I'd say roughly from 2007 to early 2009. Before the improvements that came with v5, you'd see stuck exponents, exponents held up for literally years with no discernible progress, &c. George had to manually kick PrimeNet every couple of days it seemed to release really low exponents - I can remember waiting for the 6am GMT low exponent rush, and collecting scores of sub-10Md exponents whenever the wavefront was in the high-30Ms. This is now all done automatically on v5, thank Heavens, which is why, for instance, the other day I got two LLs - one right after the other - but one was a 50M and one was a 40M. The upshot of all this was that the wavefront got artificially far along - we zoomed from the mid- and high-30Ms into the 40Ms, and then because of v5 testing, we skipped overnight from 43Ms to 46Ms. There is a post (I believe of uncwilly's) in late 2008 about breaching the 50M barrier - we had no business being that high then, but many of the mid-40M exponents were locked up in the v5 transition, and we were waiting to make sure the v4-bridged-users were able to finish anything lower before reassigning them.

So, overall, I'd say we're still making progress at roughly 4M a year (the primes found in 2008 were in the 43M area, which should put us at 51M by Aug. - Sept. 2010, which is where we're looking to be). I think what we're seeing is not a slowdown, but simply a return to the normal order in which numbers should have been handed out in the first place. With George's improvements and optimizations in the coming Prime95 v26, we might see things move a little faster yet.

The next issue that IMHO needs to be addressed is double-checking. I started keeping an eye on the milestone page starting on May 21, and as of today, given the pace of returned results from May 21 - today, this is your wait to prove that various Mxx really are Mxx:

M40 - 20996011 - whenever someone just poaches the last darn number (probably October, given that is when the current assignment would expire)

M41 - 24036583 - October 10, 2010 (this could happen before M40, although things will probably slow down on the last handful of exponents, as always)

M42 - 25964951 - August 31, 2011

Hope you had fun with those...because it's going to be a while now:

M43 - 30402457 - February 16, 2014

M44 - 32582657 - May 2, 2015

M45 - 37156667 - October 8, 2017

M46 - 42643801 - August 17, 2020

M47 - 43112609 - November 17, 2020

It almost wouldn't hurt to put a moratorium on first-time LL testing (especially on 100Md numbers, for Heaven's sake) for six months, just to bring these double-check dates a little closer to Earth. I wonder if the folks interested in a proof that Mxx is really Mxx outnumber the folks interested only in the glory of finding a new prime.
NBtarheel_33 is offline   Reply With Quote
Old 2010-06-30, 18:25   #6
Mini-Geek
Account Deleted
 
Mini-Geek's Avatar
 
"Tim Sorbera"
Aug 2006
San Antonio, TX USA

10000101010112 Posts
Default

Quote:
Originally Posted by NBtarheel_33 View Post
M41 - ...this could happen before M40...
Um, no. No, it couldn't. It could happen at the same time as M40* being proven as M40 (how likely? in my mostly-uninformed opinion, fairly likely unless there are more stragglers in that fairly large space), or it could be proven that there are no primes between M40* and M41*, but it (that M41* is M41) can't happen before it (that M40* is M40).
* meaning that the numbering assumes no as-yet-unknown primes change it

Last fiddled with by Mini-Geek on 2010-06-30 at 18:34
Mini-Geek is offline   Reply With Quote
Old 2010-06-30, 18:28   #7
mdettweiler
A Sunny Moo
 
mdettweiler's Avatar
 
Aug 2007
USA (GMT-5)

3·2,083 Posts
Default

Quote:
Originally Posted by NBtarheel_33 View Post
It almost wouldn't hurt to put a moratorium on first-time LL testing (especially on 100Md numbers, for Heaven's sake) for six months, just to bring these double-check dates a little closer to Earth. I wonder if the folks interested in a proof that Mxx is really Mxx outnumber the folks interested only in the glory of finding a new prime.
Interesting idea. I remember the Seventeen or Bust project (for those not familiar, it aims to prove the Sierpinski conjecture and in fact uses its own distinct PrimeNet v5 system at this point to handle assignments via Prime95) did something very similar a while back--their doublecheck threshold was getting quite far behind their firstpass leading edge, so they switched over to doublechecks as the primary assignment given out. This was back when they were on their "old" assignment system (proprietary, not PrimeNet) so while firstpass tests were still available, one had to actually edit the client's configuration file to talk to the server on a different port (rather than just flipping a GUI switch as in Prime95). Lo and behold, before long they actually found a previously-missed prime in their doublechecking! (Since then the doublechecks have caught up somewhat and I think they're back on firstpass tests.)

While this may or may not be the way to go for GIMPS, my point is that something similar has been tried before and went over without a huge amount of complaint from the membership. Granted, their primes didn't have a cash prize, so there's less disincentive to do doublechecks, so the situation might be a tad different here (where even a non-100M digit prime gets a comparatively small prize).
mdettweiler is offline   Reply With Quote
Old 2010-06-30, 18:37   #8
NBtarheel_33
 
NBtarheel_33's Avatar
 
"Nathan"
Jul 2008
Maryland, USA

5·223 Posts
Talking

Quote:
Originally Posted by Mini-Geek View Post
Um, no. No, it couldn't. It could happen at the same time as M40* being proven as M40 (how likely? in my mostly-uninformed opinion, fairly likely unless there are more stragglers in that fairly large space), or it could be proven that there are no primes between M40* and M41*, but it (that M41* is M41) can't happen before it (that M40* is M40).
* meaning that the numbering assumes no as-yet-unknown primes change it
LOL, duhhhhh!!! That is what happens when you blindly make conclusions from an Excel spreadsheet without considering what the data actually represent - wait, that's how Wall Street does so well, right?

Although, if Chuck Norris joined GIMPS....he could prove M41 is M41 without M40 being M40!
NBtarheel_33 is offline   Reply With Quote
Old 2010-06-30, 20:54   #9
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

3×5×72×13 Posts
Default

Quote:
Originally Posted by NBtarheel_33 View Post
It almost wouldn't hurt to put a moratorium on first-time LL testing (especially on 100Md numbers, for Heaven's sake) for six months, just to bring these double-check dates a little closer to Earth. I wonder if the folks interested in a proof that Mxx is really Mxx outnumber the folks interested only in the glory of finding a new prime.
Rather than that: have all new machines that want to do 100M numbers prove themselves first. Have them do a minimum number of double checks to verifiy that they are running okay. Maybe at least 4 DC's with no serious errors and at least 3 of them matching the first timer. Make it 4 or 5 for a single core and 3 or 4 per core devoted to 100m. For any mis-matched results, have it do the same quantity again.
Uncwilly is offline   Reply With Quote
Old 2010-06-30, 20:55   #10
Uncwilly
6809 > 6502
 
Uncwilly's Avatar
 
"""""""""""""""""""
Aug 2003
101×103 Posts

255316 Posts
Default

Quote:
Originally Posted by davieddy View Post
it seems the allocation wavefront is stuck below 51M ATM.
I think that they may have also run into the area where they are having to do their own P-1. This would slow things down.
Uncwilly is offline   Reply With Quote
Old 2010-06-30, 22:56   #11
lycorn
 
lycorn's Avatar
 
Sep 2002
Oeiras, Portugal

2·36 Posts
Default

Quote:
Originally Posted by Uncwilly View Post
Rather than that: have all new machines that want to do 100M numbers prove themselves first. Have them do a minimum number of double checks to verifiy that they are running okay. Maybe at least 4 DC's with no serious errors and at least 3 of them matching the first timer. Make it 4 or 5 for a single core and 3 or 4 per core devoted to 100m. For any mis-matched results, have it do the same quantity again.
I second that suggestion, and also restate what I have proposed in the last paragraph of this post: http://www.mersenneforum.org/showpos...1&postcount=52
lycorn is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
128-bit OS'es and GIMPS? ixfd64 Software 22 2011-10-31 22:23
GIMPS on PS3 flouran Hardware 202 2010-04-30 09:06
GIMPS uses only 1 cpu Unregistered Information & Answers 7 2009-01-10 20:01
My Prime95 always Halt when FFT size is 16K overpro Hardware 1 2004-10-21 08:07
Why do you run GIMPS ? Prime Monster Lounge 12 2003-11-25 19:04

All times are UTC. The time now is 17:23.

Sat May 8 17:23:06 UTC 2021 up 30 days, 12:03, 0 users, load averages: 3.13, 3.26, 3.06

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.