Newer milestone thread - Page 333

S485122 · 2022-11-18, 15:42

Quote:

Originally Posted by retina

Math is hard !

I could try to invent excuses, but I can only say :
Oops !

slandrum · 2022-11-26, 20:22

Every exponent through 10⁹/9 has been tested at least once.

SethTro · 2023-03-04, 01:23

All exponents below 63,000,000 have been double checked!

I'm excited that with the move to PRP we've slowly (but surely) reducing the backlog of DC and that it should head to zero around 2030!

Rubiksmath · 2023-03-04, 01:49

I think one of the more remarkable things about this is that Ryan Propper is single-handedly producing more than 1/3 of the entire DC effort throughput. They complete the equivalent of a 0.7M range or so every time the entire rest of the DC effort clears a 1M range.

S485122 · 2023-03-04, 10:52

Quote:

Originally Posted by SethTro

All exponents below 63,000,000 have been double checked!
...

A pity that the last trailing exponent has been done by a poacher : M62961323

(Of course the user who had the non expired assignments should have cancelled them when stopping or reducing his participation to GIMPS.)

Andrew Usher · 2023-03-04, 13:43

Eliminating the DC backlog, in practical terms, means that two different milestones are no longer needed - instead, just 'tested and verified'. There will always be some DCs needed but when only a few, those below the wavefront can be targeted rapidly. I would like to see also a PRP-cofactor milestone: though new factors are continually discovered, meaning there will never be completeness, saying that all exponents below X (unless prime or fully factored) have had a PRP-cofactor test at some time would still be meaningful.

Ryan Propper is indeed remarkable - and he uses LL. So his results might be used to give more modern (better) estimates of LL reliability than the ancient one that Kriesel repeats - his rate of mismatches is an upper limit on the LL error rate in the region (as residues will not match if either was wrong). The server will not give accurate statistics here because those cleared by LL-TC are not distinguished, but comparing with his rate in the last few days shows it to be no more than 0.2% - much less than the roughly 0.7% that the antiquated data would predict here.

So LL really is better than PRP for double-checks (the threshold should be about 0.5% mismatches), if your machine is reliable, and it will probably remain so (given decreasing hardware error rates over time for the same length of computation) up to the current FTC wavefront, above which there are too few LL-DC candidates to really matter. I'm sure he (Propper) has figured this out - at his volume, it would be hard not to.

Mark Rose · 2023-03-05, 07:40

As of a few minutes ago, all current DCTF for LL results has been completed, up to 1000M.

Uncwilly · 2023-03-05, 08:07

Cool. Time to hit the DC-LL's hard. Might be fun for someone to start working on the DC's starting at 110M and working backward. Then when the wave front gets up there, zoooom!

kriesel · 2023-03-05, 08:09

See Remi Lucet's LLDC activity downward from M74207281 (Mp#49*); ~165,783 DC to go. And there are "only" ~53,771 DC to do between Mp49* and Mp50* (M77232917).

Andrew Usher · 2023-03-05, 14:03

He's doing a couple of thousand a year, and there are not that many thousands ... I crudely estimate the total DC work remaining at 200 million GHz-days, of which about 10% is above 100M. Ultimately it doesn't matter in what order it's done.

S485122 · 2023-04-15, 08:03

We seem to have missed the following :

2023-04-11 All exponents below 112 million tested at least once.

2023-03-05, 07:40	#3659
Mark Rose "/X\(‘-‘)/X\" Jan 2013 https://pedan.tech/ 2⁴×199 Posts	As of a few minutes ago, all current DCTF for LL results has been completed, up to 1000M. Last fiddled with by Mark Rose on 2023-03-05 at 08:39

2023-03-05, 08:09	#3661
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 3²×11×79 Posts	See Remi Lucet's LLDC activity downward from M74207281 (Mp#49); ~165,783 DC to go. And there are "only" ~53,771 DC to do between Mp49 and Mp50* (M77232917). Last fiddled with by kriesel on 2023-03-05 at 08:27

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Newer X64 build needed	Googulator	Msieve	75	2022-06-13 14:22
Performance of cuda-ecm on newer hardware?	fivemack	GMP-ECM	14	2015-02-12 20:10
Cause this don't belong in the milestone thread	bcp19	Data	30	2012-09-08 15:09
Newer msieves are slow on Core i7	mklasson	Msieve	9	2009-02-18 12:58
Use of large memory pages possible with newer linux kernels	Dresdenboy	Software	3	2003-12-08 14:47

2022-11-26, 20:22	#3654
slandrum Jan 2021 California 1060₈ Posts	Every exponent through 10⁹/9 has been tested at least once.

2023-03-04, 01:23	#3655
SethTro "Seth" Apr 2019 762₈ Posts	All exponents below 63,000,000 have been double checked! I'm excited that with the move to PRP we've slowly (but surely) reducing the backlog of DC and that it should head to zero around 2030!

2023-03-04, 01:49	#3656
Rubiksmath Sep 2022 125₁₀ Posts	I think one of the more remarkable things about this is that Ryan Propper is single-handedly producing more than 1/3 of the entire DC effort throughput. They complete the equivalent of a 0.7M range or so every time the entire rest of the DC effort clears a 1M range.

2023-03-04, 13:43	#3658
Andrew Usher Dec 2022 3×13² Posts	Eliminating the DC backlog, in practical terms, means that two different milestones are no longer needed - instead, just 'tested and verified'. There will always be some DCs needed but when only a few, those below the wavefront can be targeted rapidly. I would like to see also a PRP-cofactor milestone: though new factors are continually discovered, meaning there will never be completeness, saying that all exponents below X (unless prime or fully factored) have had a PRP-cofactor test at some time would still be meaningful. Ryan Propper is indeed remarkable - and he uses LL. So his results might be used to give more modern (better) estimates of LL reliability than the ancient one that Kriesel repeats - his rate of mismatches is an upper limit on the LL error rate in the region (as residues will not match if either was wrong). The server will not give accurate statistics here because those cleared by LL-TC are not distinguished, but comparing with his rate in the last few days shows it to be no more than 0.2% - much less than the roughly 0.7% that the antiquated data would predict here. So LL really is better than PRP for double-checks (the threshold should be about 0.5% mismatches), if your machine is reliable, and it will probably remain so (given decreasing hardware error rates over time for the same length of computation) up to the current FTC wavefront, above which there are too few LL-DC candidates to really matter. I'm sure he (Propper) has figured this out - at his volume, it would be hard not to.

2023-03-05, 08:07	#3660
Uncwilly 6809 > 6502 """"""""""""""""""" Aug 2003 101×103 Posts 3×7×17×31 Posts	Cool. Time to hit the DC-LL's hard. Might be fun for someone to start working on the DC's starting at 110M and working backward. Then when the wave front gets up there, zoooom!

2023-03-05, 14:03	#3662
Andrew Usher Dec 2022 3·13² Posts	He's doing a couple of thousand a year, and there are not that many thousands ... I crudely estimate the total DC work remaining at 200 million GHz-days, of which about 10% is above 100M. Ultimately it doesn't matter in what order it's done.

2023-04-15, 08:03	#3663
S485122 "Jacob" Sep 2006 Brussels, Belgium 3642₈ Posts	We seem to have missed the following : 2023-04-11 All exponents below 112 million tested at least once.