msieve153 GPU and GTX 10 Series problem

usermode · 2017-09-30, 06:30

hello, all!
I'm new in factoring and try to use tutorial from:
http://gilchrist.ca/jeff/factoring/n...ers_guide.html

I configure factmsieve.py, binaries and try.
On CPU msieve_153 all is good - all cores and threads are loaded on 100%.
But I have GTX 1060 and want to use also CUDA power with CPU.
I'm enable in script CUDA = True and while run example I have error in log and stoped process:

Quote:

Sat Sep 30 09:04:02 2017 -> factmsieve.py (v0.86)
Sat Sep 30 09:04:02 2017 -> This is client 1 of 1
Sat Sep 30 09:04:02 2017 -> Running on 4 Cores with 2 hyper-threads per Core
Sat Sep 30 09:04:02 2017 -> Working with NAME = example
Sat Sep 30 09:04:02 2017 -> Running polynomial selection ...
Sat Sep 30 09:04:02 2017 -> msieve -s example\example.dat -l example\example.log -i example\example.ini -nf example\example.fb -g 0 -v -np 1,4000 -t 4
Sat Sep 30 09:04:02 2017
Sat Sep 30 09:04:02 2017
Sat Sep 30 09:04:02 2017 Msieve v. 1.53 (SVN 1005)
Sat Sep 30 09:04:02 2017 random seeds: 4e7ed9c0 9c9f5908
Sat Sep 30 09:04:02 2017 factoring 2881039827457895971881627053137530734638790825166127496066674320241571446494762386620442953820735453 (100 digits)
Sat Sep 30 09:04:03 2017 searching for 15-digit factors
Sat Sep 30 09:04:03 2017 commencing number field sieve (100-digit input)
Sat Sep 30 09:04:03 2017 commencing number field sieve polynomial selection
Sat Sep 30 09:04:03 2017 polynomial degree: 4
Sat Sep 30 09:04:03 2017 max stage 1 norm: 1.58e+017
Sat Sep 30 09:04:03 2017 max stage 2 norm: 3.44e+015
Sat Sep 30 09:04:03 2017 min E-value: 8.85e-009
Sat Sep 30 09:04:03 2017 poly select deadline: 1317
Sat Sep 30 09:04:03 2017 time limit set to 0.37 CPU-hours
Sat Sep 30 09:04:03 2017 expecting poly E from 1.43e-008 to > 1.64e-008
Sat Sep 30 09:04:03 2017 searching leading coefficients from 1 to 4000
Sat Sep 30 09:04:03 2017 using GPU 0 (GeForce GTX 1060 6GB)
Sat Sep 30 09:04:03 2017 selected card has CUDA arch 6.1

any solution how to use last msieve gpu version with gtx 10XX?

wombatman · 2017-09-30, 14:01

Quote:

Originally Posted by usermode

hello, all!
I'm new in factoring and try to use tutorial from:
http://gilchrist.ca/jeff/factoring/n...ers_guide.html

I configure factmsieve.py, binaries and try.
On CPU msieve_153 all is good - all cores and threads are loaded on 100%.
But I have GTX 1060 and want to use also CUDA power with CPU.
I'm enable in script CUDA = True and while run example I have error in log and stoped process:

any solution how to use last msieve gpu version with gtx 10XX?

Code:

Sat Sep 30 09:04:03 2017 using GPU 0 (GeForce GTX 1060 6GB)
Sat Sep 30 09:04:03 2017 selected card has CUDA arch 6.1

You are! It should complete the first step very quickly.

usermode · 2017-09-30, 14:12

Quote:

Originally Posted by wombatman

Code:

Sat Sep 30 09:04:03 2017 using GPU 0 (GeForce GTX 1060 6GB)
Sat Sep 30 09:04:03 2017 selected card has CUDA arch 6.1

You are! It should complete the first step very quickly.

Sorry, full error log not written, I record it from console screen:
https://www.upload.ee/image/7509258/err_log.jpg

chris2be8 · 2017-09-30, 16:26

I use in on Linux, not windows, and SVN 988 not SVN 1005 so I'm not in the same setup. But msieve needs to have the PTX file and the sortlib either in the working directory or to be told where to find them. I had to make my script add a symlink to the PTX file in the working directory and pass msieve a parameter saying where the sortlib is to make it work. Which took me a little while to work out. But it works nicely now.

First look in the directory msieve is installed in, there should be a file called Readme.nfs with instructions on how to use msieve for NFS in it, including how to use a GPU. You should also have some files called something like stage1_core_sm20.ptx in there. If they are missing you need to reinstall msieve with GPU support enabled to generate them.

Then try running a GPU job in that directory. If that works you just need to get factmsieve.py to pass msieve the right parameters to make it work in other places. But I don't know python so can't help there.

Chris

usermode · 2017-09-30, 17:47

I try default settings, without .py script, with parameters:
msieve153_gpu -g 0 -v -np 1,4000 -t 4

and process failed with msieve gpu internal engine error:
CUDA_ERROR_FILE_NOT_FOUND

which file or argument is needed?

(Win7x64)

swellman · 2017-10-01, 01:02

I'm not sure what's tripping you up - might be an older version of msieve code not built for the GTX 1060 card? There are several versions of msieve out there.

Try the following steps just to see if msieve will run and engage your GPU.

http://www.mersenneforum.org/showpos...0&postcount=12

wombatman · 2017-10-01, 05:15

Quote:

Originally Posted by usermode

I try default settings, without .py script, with parameters:
msieve153_gpu -g 0 -v -np 1,4000 -t 4

and process failed with msieve gpu internal engine error:
CUDA_ERROR_FILE_NOT_FOUND

which file or argument is needed?

(Win7x64)

What PTX files are in the folder with msieve, and what files beginning with "sort_engine" are there as well?

usermode · 2017-10-01, 08:23

swellman
I'm try, but have same error: CUDA_ERROR_FILE_NOT_FOUND, maybe some in my input was incorrect.

Quote:

Originally Posted by wombatman

What PTX files are in the folder with msieve, and what files beginning with "sort_engine" are there as well?

I have default included files:
stage1_core_sm20.ptx
stage1_core_sm30.ptx
stage1_core_sm35.ptx

and dir "cub" with sort_engine.dll inside.

I can set, f.e. gpu_mem_mb=5500, but I don't understand what x argument is need to sortlib=x. If it passed to file, like, sortlib=sort.dat, then it not worked with same CUDA error.
In doc sayed what:

Quote:

By default, the library will attempt to load the sort engine that is
appropriate for your GPU, but this can be overriden with additional
arguments:

sortlib=X use GPU sorting library in file X
gpu_mem_mb=X use at most X megabytes of GPU memory

I think what msieve load defaults as well for used GPU. But I can't to use gpu version of msieve by default.

jasonp · 2017-10-01, 12:03

The memory use per GPU has an upper limit of 300MB so more than that will have no effect. The release of v1.53 needed a fix to compile the sort library for compute capability 5.0 devices; it's in the latest SVN

usermode · 2017-10-01, 12:20

Quote:

Originally Posted by jasonp

The memory use per GPU has an upper limit of 300MB so more than that will have no effect. The release of v1.53 needed a fix to compile the sort library for compute capability 5.0 devices; it's in the latest SVN

It's hard for me, I'm not work with VC compiler, and I don't understand how to fix it.
Any chances for me who can fix latest gpu msieve for the latest cuda devices?

Thanks to all, I will know.

wombatman · 2017-10-02, 03:32

Try dropping both of these files into the same folder as your msieve executable. It might work for you as I think they're good for CC5.2 cards.

2017-09-30, 17:47	#5
usermode Sep 2017 3² Posts	I try default settings, without .py script, with parameters: msieve153_gpu -g 0 -v -np 1,4000 -t 4 and process failed with msieve gpu internal engine error: CUDA_ERROR_FILE_NOT_FOUND which file or argument is needed? (Win7x64) Last fiddled with by usermode on 2017-09-30 at 17:54

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Best 4XX series GPU	siegert81	GPU Computing	47	2011-10-14 00:49
PR 4 # 33 -- The last puzzle from this series	Wacky	Puzzles	31	2006-09-14 16:17
An interesting series	Citrix	Math	0	2005-11-02 05:33
Another Series	Gary Edstrom	Puzzles	7	2003-07-03 08:32
Series	Rosenfeld	Puzzles	2	2003-07-01 17:41

2017-09-30, 16:26	#4
chris2be8 Sep 2009 4036₈ Posts	I use in on Linux, not windows, and SVN 988 not SVN 1005 so I'm not in the same setup. But msieve needs to have the PTX file and the sortlib either in the working directory or to be told where to find them. I had to make my script add a symlink to the PTX file in the working directory and pass msieve a parameter saying where the sortlib is to make it work. Which took me a little while to work out. But it works nicely now. First look in the directory msieve is installed in, there should be a file called Readme.nfs with instructions on how to use msieve for NFS in it, including how to use a GPU. You should also have some files called something like stage1_core_sm20.ptx in there. If they are missing you need to reinstall msieve with GPU support enabled to generate them. Then try running a GPU job in that directory. If that works you just need to get factmsieve.py to pass msieve the right parameters to make it work in other places. But I don't know python so can't help there. Chris

2017-10-01, 01:02	#6
swellman Jun 2012 2²×13×59 Posts	I'm not sure what's tripping you up - might be an older version of msieve code not built for the GTX 1060 card? There are several versions of msieve out there. Try the following steps just to see if msieve will run and engage your GPU. http://www.mersenneforum.org/showpos...0&postcount=12

2017-10-01, 12:03	#9
jasonp Tribal Bullet Oct 2004 3541₁₀ Posts	The memory use per GPU has an upper limit of 300MB so more than that will have no effect. The release of v1.53 needed a fix to compile the sort library for compute capability 5.0 devices; it's in the latest SVN