mersenneforum.org  

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

Reply
 
Thread Tools
Old 2018-08-09, 16:38   #1574
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

114516 Posts
Default

Quote:
Originally Posted by axn View Post
Your argument is correct if we're verifying a prime. For verifying a composite, it is not. Because, now you don't have two results that match. Both could be wrong, and you wouldn't know it [Granted PRP w/ GEC is strong enough all on its own, but the principle still stands]
Agreed, up to a point. Matching 64-bit residues is an encouraging outcome. It seems highly reliable for detecting random errors. In the absence of some systematic error it seems consistent with a high probability of correct operation. It does not though ensure either were completely correct runs.

Two very separate ways of detecting composite seem to me to be a stronger scenario in a way. Two runs with different shifts and matching 64-bit residues could be both wrong if from the same software with some systematic error in its code . With more independent code performing separate algorithms, two independently developed programs agreeing a candidate is composite is more trustworthy in a sense. (They could be LL with Jacobi or PRP w/GEC in prime95/mprime, and PRP in OpenCl.) The LL without Jacobi checking was ~98% reliable; with Jacobi it should be ~99%. The PRP with GEC is higher yet, though not identically 1. Systematic errors embodied in the code of one are less likely to be present in the other, if their development was sufficiently independent, and the hardware they run on is sufficiently different.

A false negative of an actual prime is statistically easy to arrive at in a single test by error in the residue; a great many erroneous final residue values will do to indicate composite. A false positive is statistically hard; only one final residue value will do to indicate prime. A composite indication on an actual composite, that yields an erroneous residue that still indicates composite, may be technically wrong, but still indicates the correct conclusion. This is the most likely error case, as the product of the number of actual composites tested and the probability of wrong residues indicating composite. An error in computing a prime's final residue that still produces a primality indication seems highly improbable (unless it is a systematic error that produces a zero LL residue, which does happen).

There are systematic errors known in some of the primality testing software. While some raise the frequency of false positives, these attract attention and multiple rechecks. Others are known that produce patterns of residues, or a repeating single residue, that indicate composite. If the conditions were duplicated by different users on different hardware, with the same software, the erroneous residues are duplicated and will match. They're believed to be independent of shift. An example is CUDALucas for which some checks for known bad interim residues were implemented only last year in 2.06beta. Another example is clLucas for which some known bad interim residues go unchecked in the program still. Mprime/prime95 as I recall includes checks for certain bad interim residues in its LL code also. It's not even sufficient to run different software on different hardware in some cases. CUDALucas (LL on NIVIDIA gpus and clLucas (LL on AMD gpus) share bad residue 0x02. Some versions of CUDALucas and gpuOwL check for bad LL residue 0x00. Some bugs or user errors will produce a repeating 0x00 sequence (such as a reliably recurring error). Others will produce 0, -2 (ffff ffff ffff fffd), then 2 which repeats (such as a single occurrence of failure to copy memory, giving a zero residue, subsequently giving 0^2 -2, (-2)^2 -2=2, and looping on 2^2-2=2). Another type is a user specifying much too short an fft length for the exponent to be tested, perhaps accidentally by a typo, and the computations unavoidably going wrong as a result, because the program allows the user forcing the fft length below the functional minimum.

Judging from some threads here, these sorts of systematic errors are significant in the total error rate from some individual programs.

We can say that systematic errors should be found and remedied and running code updated. Or that checks for user error should be built in. Some have been known for years. Some users are still running vulnerable versions of code lacking checks for them despite patches being available for a long time. Some code has not been patched for known systematic errors. Some have been patched in linux but not Windows for years at a time. In the gpu computing realm, providing the patches is complicated by the combination of numerous OS and CUDA level variations, as well as the programming challenges. Getting it done is not easy or quick.

Last fiddled with by kriesel on 2018-08-09 at 16:40
kriesel is offline   Reply With Quote
Old 2018-08-14, 18:34   #1575
heliosh
 
Oct 2017
++41

7D16 Posts
Default

The GHz-Days received for manually submitted ECM work that was done using gmp-ecm is too high, especially for big B-sizes.
For a 10 CPU-days job* on an i5-4690 I got rewarded with 7384 GHz-Days or 738 GHz-days per day.
Compared to LL/PRP where I roughly got 36 GHz-days per day on the same machine.

*) M2591, 1187 curves, B1=26e7, B2=3079973376496
heliosh is offline   Reply With Quote
Old 2018-08-21, 06:08   #1576
LaurV
Romulan Interpreter
 
LaurV's Avatar
 
Jun 2011
Thailand

25×3×7×13 Posts
Default

I was so busy last weekend that I forgot the most important part... Happy birthday Server!... hehe... Five years by now?
(Edit: backup server, whatever... )


Last fiddled with by LaurV on 2018-08-21 at 06:16
LaurV is offline   Reply With Quote
Old 2018-08-21, 10:54   #1577
ET_
Banned
 
ET_'s Avatar
 
"Luigi"
Aug 2002
Team Italia

476610 Posts
Default

Quote:
Originally Posted by LaurV View Post
I was so busy last weekend that I forgot the most important part... Happy birthday Server!... hehe... Five years by now?
(Edit: backup server, whatever... )

ET_ is offline   Reply With Quote
Old 2018-08-24, 22:27   #1578
James Heinrich
 
James Heinrich's Avatar
 
"James Heinrich"
May 2004
ex-Northern Ontario

2×1,511 Posts
Default

Quote:
500 Can't connect to v5.mersenne.org:80 (Bad hostname)
Was this hostname taken down intentionally, or is something broken?
James Heinrich is offline   Reply With Quote
Old 2018-08-25, 02:42   #1579
Chuck
 
Chuck's Avatar
 
May 2011
Orange Park, FL

5×173 Posts
Default Resolved

I got this message for five hours; it has since been resolved.
Chuck is offline   Reply With Quote
Old 2018-08-29, 03:24   #1580
Madpoo
Serpentine Vermin Jar
 
Madpoo's Avatar
 
Jul 2014

2×1,637 Posts
Default

Quote:
Originally Posted by Chuck View Post
I got this message for five hours; it has since been resolved.
Hmm... no idea. I was on vacation last week, no internet. Heavenly.

I just glanced at the web logs for that day for v5.mersenne.org and initially it seems like it had regular traffic patterns all day. I don't know, maybe it was a DNS resolution issue in some areas, based on the "bad hostname" error?

I don't see any errors on the server, so I'm going with "internet gremlins" as the answer this time, pending any other discoveries.
Madpoo is offline   Reply With Quote
Old 2018-10-19, 19:05   #1581
tServo
 
tServo's Avatar
 
"Marv"
May 2009
near the Tannhäuser Gate

11×47 Posts
Default Huh?

This morning, there were different statistics presented on 'reports->top producers-Trial factoring' and the numbers looked horribly wrong. A customized report looked off also.
It looks like the inital page ( 1 yesr summary ) is better now but it still dropped 12,000 GHzDays from my total.
I didn't check the other work types.
Database problems?
tServo is offline   Reply With Quote
Old 2018-10-19, 20:33   #1582
ramgeis
 
ramgeis's Avatar
 
Apr 2013

22×29 Posts
Default

Quote:
Originally Posted by tServo View Post
This morning, there were different statistics presented on 'reports->top producers-Trial factoring' and the numbers looked horribly wrong. A customized report looked off also.
It looks like the inital page ( 1 yesr summary ) is better now but it still dropped 12,000 GHzDays from my total.
I didn't check the other work types.
Database problems?
Looking at your lifetime stats I assume you experience the same phenomenon you already observed recently on others: https://www.mersenneforum.org/showthread.php?t=23579
Let me guess, you joined the project roughly a year ago?
ramgeis is offline   Reply With Quote
Old 2018-10-20, 01:00   #1583
tServo
 
tServo's Avatar
 
"Marv"
May 2009
near the Tannhäuser Gate

10000001012 Posts
Default

Quote:
Originally Posted by ramgeis View Post
Looking at your lifetime stats I assume you experience the same phenomenon you already observed recently on others: https://www.mersenneforum.org/showthread.php?t=23579
Let me guess, you joined the project roughly a year ago?
Yes, I remember that thread.
No, I joined in Spring around March or April.
I started by doing P-1 work for a month or so until I switched to TF.
Of course, looking at those numbers is fun but trivial. However, I wonder if
some posted results got lost.
Since I save my results, I shall investigate further in a day or 2.
tServo is offline   Reply With Quote
Old 2018-10-21, 21:46   #1584
ramgeis
 
ramgeis's Avatar
 
Apr 2013

22×29 Posts
Default

Quote:
Originally Posted by tServo View Post
Yes, I remember that thread.
No, I joined in Spring around March or April.
I started by doing P-1 work for a month or so until I switched to TF.
Of course, looking at those numbers is fun but trivial. However, I wonder if
some posted results got lost.
Since I save my results, I shall investigate further in a day or 2.
ok, as the difference between your lifetime stats and the one for the last year is the mentioned 12,000 GHzDays I guessed that there is the missing connection but my assumption was wrong.
ramgeis is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Official "Faits erronés dans de belles-lettres" thread ewmayer Lounge 39 2015-05-19 01:08
Official "all-Greek-to-me Fiction Literature and Cinema" Thread ewmayer Science & Technology 41 2014-04-16 11:54
Official "Lasciate ogne speranza" whinge-thread cheesehead Soap Box 56 2013-06-29 01:42
Official "Ernst is a deceiving bully and George is a meanie" thread cheesehead Soap Box 61 2013-06-11 04:30
Official "String copy Statement Considered Harmful" thread Dubslow Programming 19 2012-05-31 17:49

All times are UTC. The time now is 05:19.

Wed Sep 23 05:19:15 UTC 2020 up 13 days, 2:30, 0 users, load averages: 2.27, 2.01, 1.91

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.