mersenneforum.org  

Go Back   mersenneforum.org > New To GIMPS? Start Here! > Information & Answers

Reply
 
Thread Tools
Old 2018-03-22, 02:32   #1
carpetpool
 
carpetpool's Avatar
 
"Sam"
Nov 2016

32610 Posts
Post LLR times are not precise

I left LLR running on my computer for some time, I went to the 'results' file to check out the log and I found this (the first 12 times):

... Time: 294.154 sec.
... Time: 286.107 sec.
... Time: 300.044 sec.
... Time: 285.638 sec.
... Time: 294.911 sec.
... Time: 300.298 sec.
... Time: 304.192 sec.
... Time: 301.783 sec.
... Time: 342.110 sec. <--- Woah, where'd this leap come in ??? I'm confused.
... Time: 338.485 sec.
... Time: 350.073 sec.
... Time: 367.194 sec.
... Time: 363.413 sec.
... Time: 380.094 sec.
... Time: 391.251 sec.
... Time: 400.150 sec.
... Time: 381.600 sec.
... Time: 397.539 sec.
... Time: 410.327 sec.
... Time: 415.241 sec.
... Time: 409.323 sec.
... Time: 419.111 sec.
... Time: 413.120 sec.
... Time: 425.692 sec.
... Time: 425.570 sec.
... Time: 437.457 sec.
... Time: 431.243 sec. <--- How'd we get from 294 seconds to 431 seconds that quick ???

I do NOT know what happened to the times I pointed in red, especially given that

1) There were no other programs running throughout the whole process
2) The priority level was at 1.
3) My screen saver is turned off or muted.
4) I did not change the priority at all, I do not know if it would make it run faster.

LLR is, really faster than PFGW, however, this problem never occurs with PFGW unless I start background programs that steal CPU space, only then, the times would go up, otherwise, they would remain precise. Is there any way that LLR can be just as precise as PFGW if I were not to disturb the process at all? Seeing the sudden time change makes me think twice about using LLR, and also switching back to PFGW.

Thanks for help.
carpetpool is offline   Reply With Quote
Old 2018-03-22, 02:49   #2
wombatman
I moo ablest echo power!
 
wombatman's Avatar
 
May 2013

6D116 Posts
Default

What OS? What numbers? Single-threaded or multi-threaded? CPU?

Last fiddled with by wombatman on 2018-03-22 at 02:50
wombatman is offline   Reply With Quote
Old 2018-03-22, 03:57   #3
carpetpool
 
carpetpool's Avatar
 
"Sam"
Nov 2016

5068 Posts
Post

Quote:
Originally Posted by wombatman View Post
What OS? What numbers? Single-threaded or multi-threaded? CPU?
As far as I'm aware of it is single thread computation, Windows 10 OS, Intel(R) Core(TM) i5-3230M Processor @ 2594.13 MHz.

Attached NewPgen file I was working on (line 27).
Attached Files
File Type: txt sieve-file.txt (183.7 KB, 179 views)
carpetpool is offline   Reply With Quote
Old 2018-03-22, 04:14   #4
axn
 
axn's Avatar
 
Jun 2003

2×5×7×71 Posts
Default

Quote:
Originally Posted by carpetpool View Post
As far as I'm aware of it is single thread computation, Windows 10 OS, Intel(R) Core(TM) i5-3230M Processor @ 2594.13 MHz.

Attached NewPgen file I was working on (line 27).
Did the problem happen only on this number? How often is the time being output (iterations between output?). Is this repeatable -- i.e. if you were to re-run the same number, do you get the same timing distribution? What test is LLR doing (PRP, SPRP, some other test?)

EDIT:- What base is LLR using, if it is a PRP test? At what iteration the output started slowing down?

Last fiddled with by axn on 2018-03-22 at 04:15
axn is offline   Reply With Quote
Old 2018-03-22, 04:42   #5
paulunderwood
 
paulunderwood's Avatar
 
Sep 2002
Database er0rr

E5D16 Posts
Default

There are number of factors: Your OS uses resources at random, heating up of CPU and consequential throttling, more iterations per number and FFT size changes

Last fiddled with by paulunderwood on 2018-03-22 at 04:43
paulunderwood is offline   Reply With Quote
Old 2018-03-22, 05:37   #6
carpetpool
 
carpetpool's Avatar
 
"Sam"
Nov 2016

2×163 Posts
Post

Quote:
Originally Posted by axn View Post
Did the problem happen only on this number? How often is the time being output (iterations between output?). Is this repeatable -- i.e. if you were to re-run the same number, do you get the same timing distribution? What test is LLR doing (PRP, SPRP, some other test?)

EDIT:- What base is LLR using, if it is a PRP test? At what iteration the output started slowing down?
Directly inputting a NewPgen file into LLR, I do not know how to do PRP test only if it is k*b^n+-1 with n fixed. But as you asked, here is what LLR is doing for each of the numbers (I already tested this number):

Code:
Base factorized as: 2*7*23
Base prime factor(s) taken: 23
Starting N+1 prime test of 5*322^50004-1
Using AVX FFT length 32K, Pass1=128, Pass2=256 a=3
5*322^50004-1 is not prime.  RES64: 88DA4B86F5AD8BFA.  OLD64: 9A8EE294E108A3EB  Time : 79.472 sec.
Edit: I probably picked an unlucky k range. Still searching for primes of this form.

Last fiddled with by carpetpool on 2018-03-22 at 05:39
carpetpool is offline   Reply With Quote
Old 2018-03-22, 13:58   #7
wombatman
I moo ablest echo power!
 
wombatman's Avatar
 
May 2013

6D116 Posts
Default

With fixed n and small changes in k, your time to test should stay more or less constant for each of the numbers. I think paulunderwood is right--on my Windows 10 system, I see a slowdown when things like the file indexing service run. It might be worthwhile to open task manager and check it every now and then to see if any background processes are using CPU and/or reading/writing to disk.
wombatman is offline   Reply With Quote
Old 2018-03-22, 15:16   #8
pinhodecarlos
 
pinhodecarlos's Avatar
 
"Carlos Pinho"
Oct 2011
Milton Keynes, UK

22·5·13·19 Posts
Default

I’m on windows 7 and I have almost windows services shutdown since it’s a dedicated machine, I even have my antivirus shutdown.

Also be aware that Firefox and chrome now have like 6 instance each running if you use that machine for internet. Also Check Skype or similar softwares.

Last fiddled with by pinhodecarlos on 2018-03-22 at 15:18
pinhodecarlos is offline   Reply With Quote
Old 2018-03-22, 21:45   #9
gd_barnes
 
gd_barnes's Avatar
 
May 2007
Kansas; USA

22×2,593 Posts
Default

In looking at your sieve file, I'm fairly sure that the following is the problem because I have run fixed-n tests starting from a small k before:

The FFT length is much smaller for very small k such as k=5 resulting in much faster testing times. That is why people like to do prime searches with fixed-k to a very high search depth when k is very small. The FFT then jumps up fairly quickly as the k's get higher when starting from such small k. It is exponential. So a jump from, say, k=1M to k=1.1M is not likely to affect the FFT length. But a jump from k=5 to k=1000 may be quite large. You may even jump two FFT lengths in that range, which I believe is what you have experienced.

Long story short; check your results file or the screen for the FFT length of each of the tests. If you see it increase 1, 2, or 3 times within your small range of tests that is why your testing times are jumping up quickly.

PFGW should do the same thing for these tests. Both programs use similar libraries. Try doing this same testing with PFGW and see if you get the same jump in test times.

Last fiddled with by gd_barnes on 2018-03-22 at 21:59
gd_barnes is offline   Reply With Quote
Old 2018-03-22, 23:29   #10
gd_barnes
 
gd_barnes's Avatar
 
May 2007
Kansas; USA

22×2,593 Posts
Default

I ran the beginning of your file with PFGW 3.7.7 on a known stable 10-year-old AMD machine running Windows 7. Here were the results:

5*322^50004-1 is composite: RES64: [88DA4B86F5AD8BFA] (168.9425s+0.0002s)
17*322^50004-1 is composite: RES64: [BE608236D02295A2] (247.6629s+0.0009s) <-----
42*322^50004-1 is composite: RES64: [821C121F327A3F43] (249.9622s+0.0007s)
48*322^50004-1 is composite: RES64: [074CFFD98FD30366] (247.3925s+0.0008s)
54*322^50004-1 is composite: RES64: [A4793D287AA3D5B5] (253.2692s+0.0008s)
60*322^50004-1 is composite: RES64: [DB90CA5AA25AC40A] (244.6137s+0.0008s)
93*322^50004-1 is composite: RES64: [A36E803FE718D83D] (242.7812s+0.0006s)
98*322^50004-1 is composite: RES64: [20D0E79ED53CA431] (243.8956s+0.0007s)
140*322^50004-1 is composite: RES64: [AA295EA766714FCD] (254.8522s+0.0002s)
.
.
.
204*322^50004-1 is composite: RES64: [691DBAE1EAF0F1BC] (251.9508s+0.0002s)
237*322^50004-1 is composite: RES64: [AEC14A44290288FB] (250.0411s+0.0008s)
242*322^50004-1 is composite: RES64: [9D3213CBEBE79CD4] (290.5276s+0.0001s) <-----
257*322^50004-1 is composite: RES64: [2FB58BEF0C171EB4] (297.2559s+0.0007s)
.
.
.
315*322^50004-1 is composite: RES64: [09B9412C510B7CB9] (299.6575s+0.0001s)
.
.
.
500*322^50004-1 is composite: RES64: [002EE3A57557F1F3] (301.6779s+0.0001s)


Notice the FFT length increases between k=5 & 17 -and- k=237 & 242 causing an increase in test times.

I believe this is what you are experiencing. You will see much less of this as the k gets much larger.
gd_barnes is offline   Reply With Quote
Old 2018-03-23, 02:59   #11
carpetpool
 
carpetpool's Avatar
 
"Sam"
Nov 2016

32610 Posts
Post

Quote:
Originally Posted by gd_barnes View Post
I ran the beginning of your file with PFGW 3.7.7 on a known stable 10-year-old AMD machine running Windows 7. Here were the results:

5*322^50004-1 is composite: RES64: [88DA4B86F5AD8BFA] (168.9425s+0.0002s)
17*322^50004-1 is composite: RES64: [BE608236D02295A2] (247.6629s+0.0009s) <-----
42*322^50004-1 is composite: RES64: [821C121F327A3F43] (249.9622s+0.0007s)
48*322^50004-1 is composite: RES64: [074CFFD98FD30366] (247.3925s+0.0008s)
54*322^50004-1 is composite: RES64: [A4793D287AA3D5B5] (253.2692s+0.0008s)
60*322^50004-1 is composite: RES64: [DB90CA5AA25AC40A] (244.6137s+0.0008s)
93*322^50004-1 is composite: RES64: [A36E803FE718D83D] (242.7812s+0.0006s)
98*322^50004-1 is composite: RES64: [20D0E79ED53CA431] (243.8956s+0.0007s)
140*322^50004-1 is composite: RES64: [AA295EA766714FCD] (254.8522s+0.0002s)
.
.
.
204*322^50004-1 is composite: RES64: [691DBAE1EAF0F1BC] (251.9508s+0.0002s)
237*322^50004-1 is composite: RES64: [AEC14A44290288FB] (250.0411s+0.0008s)
242*322^50004-1 is composite: RES64: [9D3213CBEBE79CD4] (290.5276s+0.0001s) <-----
257*322^50004-1 is composite: RES64: [2FB58BEF0C171EB4] (297.2559s+0.0007s)
.
.
.
315*322^50004-1 is composite: RES64: [09B9412C510B7CB9] (299.6575s+0.0001s)
.
.
.
500*322^50004-1 is composite: RES64: [002EE3A57557F1F3] (301.6779s+0.0001s)


Notice the FFT length increases between k=5 & 17 -and- k=237 & 242 causing an increase in test times.

I believe this is what you are experiencing. You will see much less of this as the k gets much larger.
Ok, I understand what is going on (if you have N = k*b^n+1, an easy way to compute a^(N-1) modulo N is to compute g = a^k modulo N, and then repeatedly raise g to the bth power modulo N, n times. Obviously the smaller k the easier the first part will be and the smaller b the more easier it seems.) Just curious, did you find any primes of this form?

Last fiddled with by carpetpool on 2018-03-23 at 03:09
carpetpool is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
TF times Unregistered Information & Answers 7 2012-04-14 11:45
What are your per-iteration times? LiquidNitrogen Hardware 22 2011-07-12 23:15
Different bases = different times? roger Information & Answers 1 2007-04-25 14:35
LLRNET Times masser Sierpinski/Riesel Base 5 6 2007-03-05 18:17
40 Times the speed of PC's lpmurray Hardware 2 2007-02-17 19:53

All times are UTC. The time now is 23:50.

Tue May 18 23:50:53 UTC 2021 up 40 days, 18:31, 0 users, load averages: 1.74, 1.58, 1.51

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.