mersenneforum.org  

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

Reply
 
Thread Tools
Old 2018-10-22, 14:28   #309
kriesel
 
kriesel's Avatar
 
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest

2×5×7×73 Posts
Default

Quote:
Originally Posted by James Heinrich View Post
edit: also note two things about the bounds on mersenne.ca:
1) the P-1 bounds in the GPU72 line are wrong, they are not adjusted as described above, I need to figure out how to do so.
2) the P-1 bounds in general are close to magic numbers, they're currently plucked off an Excel graph I did in 2008 based on observed bounds for exponents 1M-500M, so while they might be reasonably close to reality, they'll never match exactly.
Thanks for the explanation. I had thought that the red font used for negative numbers in the difference row meant that they were considered somehow insufficient.
Similar questions had arisen in regard to CUDAPm1, which would be a little easier to manage worktodo entry adjustment on, since it is all manually assigned and manually reported, rather than sometimes being assigned and completed quicker than I would have time to intervene as on primenet-connected prime95.

I've seen some manual P-1 work done in CUDAPm1 be marked as expired when the result was reported, and that seemed to occur when one or both the bounds reported were lower than the primenet values at mersenne.ca.
A recent example is https://www.mersenne.org/report_expo...0000797&full=1 https://www.mersenne.ca/exponent/200000797
A particularly interesting case is https://www.mersenne.ca/exponent/200000551, where I pushed the manual P-1 to runs saved = 3, and tf and p-1 bounds all exceed primenet values, yet the computed probability fell a bit short.
kriesel is online now   Reply With Quote
Old 2018-10-22, 17:33   #310
James Heinrich
 
James Heinrich's Avatar
 
"James Heinrich"
May 2004
ex-Northern Ontario

22·839 Posts
Default

Quote:
Originally Posted by James Heinrich View Post
1) the P-1 bounds in the GPU72 line are wrong, they are not adjusted as described above, I need to figure out how to do so.
I transcribed the Prime95 P-1 bound estimator to PHP so I should now be able to get the same numbers are Prime95 (I'm going on the assumption that plenty of RAM is available, as in sufficient to do the whole P-1 in a single pass, for simplicity, so bounds may be slightly higher than what you may get in real life). However the code is quite ponderous due to the (large) number of iterations it goes through to find the "optimal" solution... 66k iterations at 276 and 116k iterations at 271, which has a non-trivial runtime (several seconds) such that I wouldn't want to use it on a live webpage. Now you know why it takes forever to get an estimated completion date if you have many P-1 assignments in Prime95. I'll have to see if I can optimize it somewhat while remaining relatively accurate...

The numbers I got for M89200591:
Code:
TF=71
[b1] => 950000
[b2] => 26600000
[success_rate] => 0.059250926679056

TF=76
[b1] => 725000
[b2] => 14318750
[success_rate] => 0.031280274205779
James Heinrich is offline   Reply With Quote
Old 2018-10-22, 23:23   #311
irowiki
 
Sep 2018

3×23 Posts
Default

So I just realized two of my "higher end" machines (Haswell or so) are 32 bit windows 7 (don't laugh too much, I have some old stuff). Is that going to impact LL/PRP testing or should I redo them into 64 bit?

I saw a thread from 12 years ago that said it wouldn't affect it so I thought I'd ask now.

Last fiddled with by irowiki on 2018-10-22 at 23:25
irowiki is offline   Reply With Quote
Old 2018-10-23, 00:04   #312
James Heinrich
 
James Heinrich's Avatar
 
"James Heinrich"
May 2004
ex-Northern Ontario

D1C16 Posts
Default

Quote:
Originally Posted by James Heinrich View Post
I'll have to see if I can optimize it somewhat while remaining relatively accurate...
Indeed. By picking some more sensible starting, ending and step points in the loops I cut the iterations from 65000 to 1500 and a much more sane runtime while still getting more-or-less the same results.

Looking back at https://www.mersenne.ca/exponent/89200591 the table at the top should now show better values for P-1 on the GPU72 line, and more sensibly show that despite the lower P-1 bounds the overall factor probability was higher due to more TF.
Attached Thumbnails
Click image for larger version

Name:	M89200591.png
Views:	59
Size:	14.6 KB
ID:	19161  
James Heinrich is offline   Reply With Quote
Old 2018-10-23, 14:35   #313
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

164758 Posts
Default

Quote:
Originally Posted by irowiki View Post
are 32 bit windows 7 ....Is that going to impact LL/PRP testing or should I redo them into 64 bit?
Redo. 10-20% gain
Prime95 is offline   Reply With Quote
Old 2018-11-13, 16:35   #314
irowiki
 
Sep 2018

3×23 Posts
Default

Quote:
Originally Posted by Prime95 View Post
Redo. 10-20% gain
Okay, for science I figured I'd post my results here. I redid the machine, and flipped it from Win7 32 bit to Win10 64 bit.

It is an i3-3220 with 4 GB (2x2).

It's doing about 9% better. Score!
irowiki is offline   Reply With Quote
Old 2018-11-20, 18:12   #315
Cruelty
 
Cruelty's Avatar
 
May 2005

23·7·29 Posts
Default

Using latest version 29.4.b8 (Win64) on Skylake X platforms, I have noticed that in main thread after launching Prime95 there is:
Code:
Optimizing for CPU architecture: Corei3/i5/i7, L2 cache size: 256 KB, L3 cache size: 14080 KB
Whereas L2 cache should be reported as 1MB
In undoc.txt there is a section concerning L2 cache size, but it's said to relate to Pentium 4 architecture:
Code:
CpuL2CacheSize=128 or 256 or 512
CpuL2CacheLineSize=32 or 64 or 128
CpuL2SetAssociative=4 or 8
After setting it to:
Code:
CpuL2CacheSize=1024
CpuL2CacheLineSize=64
CpuL2SetAssociative=16
Program reports it properly:
Code:
Optimizing for CPU architecture: Corei3/i5/i7, L2 cache size: 1 MB, L3 cache size: 14080 KB
Question: should detection of L2 cache parameters affect Skylake X performance? After setting it to proper values I haven't noticed improvement in iteration times, at least for FMA3 FFT length of 864K.

Another minor issue I have noticed is the representation of CPU information - the line below L1 cache information is unreadable (see attachment)
Attached Thumbnails
Click image for larger version

Name:	cpu.png
Views:	58
Size:	8.8 KB
ID:	19299  
Cruelty is offline   Reply With Quote
Old 2018-11-20, 18:59   #316
kladner
 
kladner's Avatar
 
"Kieren"
Jul 2011
In My Own Galaxy!

2·3·1,693 Posts
Default

Sorry. This is about the latest version. No thread, yet.
I switched to 295B4 in the past two days. I just noticed that Manual Communication and Unreserve Exponent menu items are greyed out. I tried commenting the GPU72 proxy out of prime.txt, with no change.

Did I do something wrong? I don't think I've ever seen this, before. This persisted after program and system restarts.

EDIT: Found the cause. In prime.txt I found 'UsePrimenet=0'. I guess this was a misstep in transferring the prime.txt information to the new version. Changed to '=1' and all is well.
Maybe not. On contacting the server, I was told that the assignment key for a 60% done DC (482955xx) 'belongs to another user.' However, this did not remove the assignment from P95. What I see is that GPU72 is still aware of this assignment, but Primenet does not update progress. However, Primenet Exponent Status still lists me as the owner.

Last fiddled with by kladner on 2018-11-20 at 20:23
kladner is offline   Reply With Quote
Old 2018-11-20, 19:56   #317
Chuck
 
Chuck's Avatar
 
May 2011
Orange Park, FL

2×443 Posts
Default

Quote:
Originally Posted by Cruelty View Post
Another minor issue I have noticed is the representation of CPU information - the line below L1 cache information is unreadable (see attachment)
Mine also cuts off the last line; is it because the AVX512F wraps to a new line?
Chuck is offline   Reply With Quote
Old 2018-11-20, 20:11   #318
Prime95
P90 years forever!
 
Prime95's Avatar
 
Aug 2002
Yeehaw, FL

748510 Posts
Default

I will widen the dialog box.

The L2 cache mis-detection is a known bug, I do plan to fix it. The bug does not impact performance.
Prime95 is offline   Reply With Quote
Old 2018-11-20, 23:52   #319
irowiki
 
Sep 2018

3×23 Posts
Default

Quote:
Originally Posted by kladner View Post
Maybe not. On contacting the server, I was told that the assignment key for a 60% done DC (482955xx) 'belongs to another user.' However, this did not remove the assignment from P95. What I see is that GPU72 is still aware of this assignment, but Primenet does not update progress. However, Primenet Exponent Status still lists me as the owner.
Are you still listed with your ID in File -> Primenet? I've had some of my clients randomly switch to anon and I had to switch it back.
irowiki is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Prime95 version 27.3 Prime95 Software 148 2012-03-18 19:24
Prime95 version 26.3 Prime95 Software 76 2010-12-11 00:11
Prime95 version 25.5 Prime95 PrimeNet 369 2008-02-26 05:21
Prime95 version 25.4 Prime95 PrimeNet 143 2007-09-24 21:01
When the next prime95 version ? pacionet Software 74 2006-12-07 20:30

All times are UTC. The time now is 03:05.

Mon May 10 03:05:18 UTC 2021 up 31 days, 21:46, 0 users, load averages: 1.61, 2.10, 2.06

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.