mfaktc: a CUDA program for Mersenne prefactoring - Page 135

TheJudger · 2011-12-29, 12:17

Hi,

Quote:

Originally Posted by BigBrother

I did. I disabled it and now it seems to run without performance dips. I'll try to check it out further tonight.

just for curiosity, can you run 'mfaktc --sleeptest' on your system and post the result? Thank you!

Oliver

Dubslow · 2011-12-29, 15:02

Quote:

Originally Posted by kladner

I do have AllowSleep=1. Most of the time the GPU runs at 95-98%. I do see variations in throughput, but lots of times it has to do with other things going on in the system which is taking over more CPU time and starving mfaktc. For example, anything involving Adobe apps almost always eats a chunk of performance. This shows up as big dips in GPU utilization. Norton and its like can certainly do the same. Are you sure what else is going on in your machine?

For me, I just sat and watched the computer that whole time, so only something in the background could have done it. If the process list in the Task Manager is anything to go by, the only thing I could think of would be Steam doing something, but like I said, nothing appeared on screen. The only thing changing were the GPU-Z window and the mfaktc window.

BigBrother · 2011-12-29, 20:24

Quote:

Originally Posted by TheJudger

Hi,

just for curiosity, can you run 'mfaktc --sleeptest' on your system and post the result? Thank you!

Oliver

Code:

my_usleep(  1000): t_min =    492us, t_max =   1083us, t_avg =    999us
my_usleep(  2000): t_min =   1668us, t_max =   2057us, t_avg =   1999us
my_usleep(  4000): t_min =   3631us, t_max =   4070us, t_avg =   3998us
my_usleep(  8000): t_min =   7700us, t_max =   8112us, t_avg =   7997us
my_usleep( 16000): t_min =  15727us, t_max =  16116us, t_avg =  15996us
my_usleep( 32000): t_min =  31617us, t_max =  32033us, t_avg =  31989us
my_usleep( 64000): t_min =  63661us, t_max =  64070us, t_avg =  63985us
my_usleep(128000): t_min = 127679us, t_max = 128055us, t_avg = 127955us

TheJudger · 2011-12-29, 23:40

Quote:

Originally Posted by BigBrother

Code:

my_usleep(  1000): t_min =    492us, t_max =   1083us, t_avg =    999us
my_usleep(  2000): t_min =   1668us, t_max =   2057us, t_avg =   1999us
my_usleep(  4000): t_min =   3631us, t_max =   4070us, t_avg =   3998us
my_usleep(  8000): t_min =   7700us, t_max =   8112us, t_avg =   7997us
my_usleep( 16000): t_min =  15727us, t_max =  16116us, t_avg =  15996us
my_usleep( 32000): t_min =  31617us, t_max =  32033us, t_avg =  31989us
my_usleep( 64000): t_min =  63661us, t_max =  64070us, t_avg =  63985us
my_usleep(128000): t_min = 127679us, t_max = 128055us, t_avg = 127955us

close to perfect results, this can't be the root issue. I've noticed that some systems have a not so good accuracy, e.g. my own Windows box has a bad sleep accuracy:

Code:

my_usleep(  1000): t_min =   4699us, t_max =  15736us, t_avg =  15612us
my_usleep(  2000): t_min =  15552us, t_max =  15649us, t_avg =  15623us
my_usleep(  4000): t_min =  15564us, t_max =  15645us, t_avg =  15623us
my_usleep(  8000): t_min =  15571us, t_max =  15648us, t_avg =  15622us
my_usleep( 16000): t_min =  31197us, t_max =  31257us, t_avg =  31247us
my_usleep( 32000): t_min =  46819us, t_max =  46879us, t_avg =  46872us
my_usleep( 64000): t_min =  78075us, t_max =  78129us, t_avg =  78120us
my_usleep(128000): t_min = 140576us, t_max = 140627us, t_avg = 140618us

This box has three OSes installed: Windows XP SP3, Windows 7 professional 64bit, openSUSE 11.1 64bit. Both Windows installations have a bad accuracy while under Linux I get close to perfect sleep results. So I guess this is a software issue, not an hardware issues. I've observed multiples of 15.6ms asweel as 7.8ms and 3.9ms for different Windows installations on different machines.

Oliver

nucleon · 2011-12-30, 02:32

I'm having weird an unusual slowdowns as well with version 0.18.

Seems to be across all my machines.

Only affects when multiple instances run on same video card.

As an example, when I start both instances the first reported times per class are 17.796s & 11.830s, after a couple of classes it increases to 20.059s & 13.568s where it stays. M48121669 71-73, and M47551541 72-73. Initial run has CPU wait <5% and increases to over 20%.

Machine is 2600k@4.5GHz with GTX580.

The initial times seem more in line with version 0.17.

I tried mucking around with sieve primes, grid size, cpu streams, num streams and no luck.

Any ideas?

-- Craig

nucleon · 2011-12-30, 03:00

Aaah scratch that.

I put a faster fan in front of the case dropped the GPU further a few degrees and it stays constant.

I thought I ruled at GPU thermal throttling as the GPU% didn't drop. Oh well.

-- Craig

BigBrother · 2011-12-30, 16:16

Quote:

Originally Posted by TheJudger

close to perfect results, this can't be the root issue. I've noticed that some systems have a not so good accuracy, e.g. my own Windows box has a bad sleep accuracy:

Code:

my_usleep(  1000): t_min =   4699us, t_max =  15736us, t_avg =  15612us
my_usleep(  2000): t_min =  15552us, t_max =  15649us, t_avg =  15623us
my_usleep(  4000): t_min =  15564us, t_max =  15645us, t_avg =  15623us
my_usleep(  8000): t_min =  15571us, t_max =  15648us, t_avg =  15622us
my_usleep( 16000): t_min =  31197us, t_max =  31257us, t_avg =  31247us
my_usleep( 32000): t_min =  46819us, t_max =  46879us, t_avg =  46872us
my_usleep( 64000): t_min =  78075us, t_max =  78129us, t_avg =  78120us
my_usleep(128000): t_min = 140576us, t_max = 140627us, t_avg = 140618us

This box has three OSes installed: Windows XP SP3, Windows 7 professional 64bit, openSUSE 11.1 64bit. Both Windows installations have a bad accuracy while under Linux I get close to perfect sleep results. So I guess this is a software issue, not an hardware issues. I've observed multiples of 15.6ms asweel as 7.8ms and 3.9ms for different Windows installations on different machines.

Oliver

Here are two other results:

Code:

my_usleep(  1000): t_min =    927us, t_max =  15734us, t_avg =  14110us
my_usleep(  2000): t_min =   1748us, t_max =  15776us, t_avg =  13341us
my_usleep(  4000): t_min =  15308us, t_max =  15722us, t_avg =  15597us
my_usleep(  8000): t_min =  10336us, t_max =  20743us, t_avg =  15595us
my_usleep( 16000): t_min =  15911us, t_max =  31302us, t_avg =  25924us
my_usleep( 32000): t_min =  31827us, t_max =  46911us, t_avg =  43794us
my_usleep( 64000): t_min =  77618us, t_max =  78120us, t_avg =  77972us
my_usleep(128000): t_min = 140214us, t_max = 140414us, t_avg = 140367us

my_usleep(  1000): t_min =    861us, t_max =  15852us, t_avg =  11617us
my_usleep(  2000): t_min =   1889us, t_max =  16597us, t_avg =   8094us
my_usleep(  4000): t_min =  15170us, t_max =  15762us, t_avg =  15597us
my_usleep(  8000): t_min =   7921us, t_max =  18629us, t_avg =  13621us
my_usleep( 16000): t_min =  15933us, t_max =  31917us, t_avg =  28255us
my_usleep( 32000): t_min =  31753us, t_max =  46860us, t_avg =  41721us
my_usleep( 64000): t_min =  77831us, t_max =  78041us, t_avg =  77993us
my_usleep(128000): t_min = 127993us, t_max = 140137us, t_avg = 129739us

By the way, AllowSleep=0 didn't fix my problem.

Dubslow · 2011-12-31, 00:51

What was the old format for a found factor? I want to manually edit them until PrimeNet is updated.

James Heinrich · 2011-12-31, 00:55

Quote:

Originally Posted by Dubslow

What was the old format for a found factor? I want to manually edit them until PrimeNet is updated.

Makes no difference. Primenet doesn't look at that part of the result line in detail.

But it was like this:

Quote:

M148989163 has a factor: 92743592786721447289
found 1 factor(s) for M148989163 from 2^64 to 2^68 (partially tested) [mfaktc 0.13-Win 71bit_mul24]

Dubslow · 2011-12-31, 00:59

Thanks. I just wanted to make sure it wasn't recorded as P-1.

James Heinrich · 2011-12-31, 01:03

Quote:

Originally Posted by Dubslow

Thanks. I just wanted to make sure it wasn't recorded as P-1.

That's a server-side issue that is as-yet unaddressed. Doesn't matter what format you submit for manual results, Primenet will (currently) ignore the differences and process it however it sees fit. Which is usually right, but not always.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
mfakto: an OpenCL program for Mersenne prefactoring	Bdot	GPU Computing	1676	2021-06-30 21:23
The P-1 factoring CUDA program	firejuggler	GPU Computing	753	2020-12-12 18:07
gr-mfaktc: a CUDA program for generalized repunits prefactoring	MrRepunit	GPU Computing	32	2020-11-11 19:56
mfaktc 0.21 - CUDA runtime wrong	keisentraut	Software	2	2020-08-18 07:03
World's second-dumbest CUDA program	fivemack	Programming	112	2015-02-12 22:51

2011-12-30, 02:32	#1479
nucleon Mar 2003 Melbourne 203₁₆ Posts	I'm having weird an unusual slowdowns as well with version 0.18. Seems to be across all my machines. Only affects when multiple instances run on same video card. As an example, when I start both instances the first reported times per class are 17.796s & 11.830s, after a couple of classes it increases to 20.059s & 13.568s where it stays. M48121669 71-73, and M47551541 72-73. Initial run has CPU wait <5% and increases to over 20%. Machine is 2600k@4.5GHz with GTX580. The initial times seem more in line with version 0.17. I tried mucking around with sieve primes, grid size, cpu streams, num streams and no luck. Any ideas? -- Craig

2011-12-30, 03:00	#1480
nucleon Mar 2003 Melbourne 203₁₆ Posts	Aaah scratch that. I put a faster fan in front of the case dropped the GPU further a few degrees and it stays constant. I thought I ruled at GPU thermal throttling as the GPU% didn't drop. Oh well. -- Craig

2011-12-31, 00:51	#1482
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3×29×83 Posts	What was the old format for a found factor? I want to manually edit them until PrimeNet is updated.

2011-12-31, 00:59	#1484
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 7221₁₀ Posts	Thanks. I just wanted to make sure it wasn't recorded as P-1.