mersenneforum.org  

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

Reply
 
Thread Tools
Old 2020-02-08, 01:35   #12
retina
Undefined
 
retina's Avatar
 
"The unspeakable one"
Jun 2006
My evil lair

124748 Posts
Default

Quote:
Originally Posted by Uncwilly View Post
PRP can't prove a prime.
Or more specifically, it does prove a composite. And only needs to defer to LL when it finds a suspected prime.

But, yes, the error check is the main reason to use PRP. The user can have higher confidence in the result, and can have fewer runs corrupted by random bit flips, thus producing more accurate outputs.

We can expect the error rate to be significantly reduced when PRPs become the dominant test.
retina is offline   Reply With Quote
Old 2020-02-08, 02:44   #13
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

24·239 Posts
Default

Quote:
Originally Posted by storm5510 View Post
A bit off-topic: Something I have been thinking about. There is PRP and then there is LL. If I understand it correctly, PRP is probabilistic and LL is deterministic. It seems to take the same amount of time to run a PRP as it does a LL. I do not understand so much emphasis being placed on PRP when the same test will need to be repeated by a LL test. Anyone?
Ideally, LL gets executed perfectly and gives a definitive answer or composite.
Meanwhile, PRP gets executed perfectly and gives a definitive answer composite, or a probably prime indication that has a slight chance of being wrong. I think less than 1ppm chance. The chance of a PRP test's probably prime indication being a false positive is so low that nobody minds the need for an LL confirmation. There are more than enough volunteers.

In the real world, LL executes wrong and gives a wrong residue 2% of the time. Half of that 2% can be caught with the Jacobi test in mprime/prime95, but not in other apps that have not implemented the Jacobi check. (CUDALucas, cllucas, very early gpuowl, etc).

Meanwhile, PRP with the excellent Gerbicz error check is thought to be less than 1ppm, maybe a LOT less than 1ppm in error in the real world, providing there's been no transcription error in manual submission or intentional user forging going on.

A practical demonstration of the power of of PRP with GEC over LL with Jacobi is


Run times are essentially the same, from the programmer's point of view, per George Woltman, in the absence of error. There are differences in efficiency of implementation in some cases; recent commits of gpuowl's PRP are faster on most gpus than CUDALucas on the same gpus.

So, in the real world, with fallible hardware and fallible users: PRP is as fast or faster, and considerably more reliable. Some people even claim it does not need to be double checked. (I'm not one of them.)
kriesel is offline   Reply With Quote
Old 2020-02-08, 03:35   #14
ewmayer
2ω=0
 
ewmayer's Avatar
 
Sep 2002
Rep├║blica de California

2·7·797 Posts
Default

Quote:
Originally Posted by kriesel View Post
So, in the real world, with fallible hardware and fallible users: PRP is as fast or faster, and considerably more reliable. Some people even claim it does not need to be double checked. (I'm not one of them.)
I would phrase that latter bit as "may not require a DC, assuming an error-free implementation which avoids all possible forms of 'silent' corruption of the GEC residue itself, and assuming the test result has not been forged by a party with ill intent."

In practice, this means a DC will still always be required - but hopefully far fewer 3Cs, 4Cs, etc.
ewmayer is offline   Reply With Quote
Old 2020-02-08, 04:26   #15
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

24·239 Posts
Default

Quote:
Originally Posted by ewmayer View Post
I would phrase that latter bit as "may not require a DC, assuming an error-free implementation which avoids all possible forms of 'silent' corruption of the GEC residue itself, and assuming the test result has not been forged by a party with ill intent."

In practice, this means a DC will still always be required - but hopefully far fewer 3Cs, 4Cs, etc.
Another possibility is the PRP is manual and there is some sort of accidental transfer of data error in the manual reporting process. GPU result lines don't have ECC built in.


My sentence fragment in the previous post didn't get finished before the edit timeout. Should read:
A practical demonstration of the power of of PRP with GEC over LL with Jacobi is the history of a flaky old AMD box I call albatross. A PRP with GEC run in prime95 became gradually less reliable, to the point that GEC was catching so many errors it could no longer make progress. The assignment was moved and completed. Then I cleaned the heat sink fuzz off, so if anything it should have run cooler and become more reliable, and I gave it another chance. It had 9 Jacobi errors in an LL DC run it finished, and the final res64 was wrong. I cleaned it some more and it's trying a DC again, and if it fails to match again after this compressed-air cleaning, it's getting recycled.
kriesel is offline   Reply With Quote
Old 2020-02-08, 04:46   #16
Fan Ming
 
Oct 2019

6810 Posts
Default

I remembered that LL assignment can be treated and reported as PRP progress if written as the format of PRP work in worktodo.txt using just the same AID in Prime95. Not sure about manual testing.
I remembered that the trailing is 0. But probably it's not important, I think gpuowl can recognize it anyway.

Last fiddled with by Fan Ming on 2020-02-08 at 04:51
Fan Ming is offline   Reply With Quote
Old 2020-02-08, 07:55   #17
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

22×1,709 Posts
Default

Quote:
Originally Posted by ewmayer View Post
I have several first-time LL assignments queued up on my old Haswell quad, which I would like to move from the CPU-side of things to the new Radeon 7 GPU...
One reason not to do this is if there was a previous suspect LL test run. These are handed out again as first-time tests and sometimes your LL test matches for a completed DC. The other reason is if your LL assignment was a recently recycled LL assignment. Sometimes a recycled assignment reports a result -- again your LL test might then complete a DC.

You can go to the exponent status web page, https://www.mersenne.org/report_exponent/ for each of your exponents and see if these "edge cases" applies to your assignments.
Prime95 is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
New user for manual results (mfaktc), any word on account verification time frame so I can submit? Realizm Information & Answers 15 2019-08-17 00:14
Long time P95 user now experiencing errors with new version. Sugam Software 6 2016-03-24 10:06
Old User Unregistered Information & Answers 1 2012-10-18 23:31
The user CP has gone :( retina Forum Feedback 5 2006-12-05 16:47
help for a Mac user drakkar67 Software 3 2003-02-11 10:55

All times are UTC. The time now is 06:21.

Mon Jun 1 06:21:15 UTC 2020 up 68 days, 3:54, 1 user, load averages: 1.38, 1.40, 1.41

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.