Installation of GGNFS - Page 2

VBCurtis · 2017-01-19, 03:24

How large a factorization are you planning to do with your setup? Are you doing mostly SNFS or GNFS numbers?

The GPU support is for poly select only, the first stage of GNFS projects. It's massively faster than CPU select, but the overall difference in project length is tiny for inputs below 140 digits, acceptably small for 140-150 digits, and not so small for 150+ digits. If you're planning to tackle GNFS-150+, you should consider using msieve with GPU support; that means learning to find a poly separately from the python script.

I have a GPU handy, and still don't bother using it for inputs less than the high 130 digit range. The extra human time doesn't save enough silicon time for my interests below 138ish. A rough estimate is that GPU select can knock half a digit off the effective difficulty (by finding a polynomial that performs 5-7% better, in less wall-clock time, than a CPU search). You'll find threads in the msieve forum with guidance for specific invocations.

jasonp · 2017-01-19, 12:55

Actually Msieve runs GMP-ECM by default for 15-digit factors, but if you run with '-e' it will do more work and look for larger factors, up to the point where switching to QS is faster. For large inputs you can run ECM looking for up to 40-digit factors, and it tries to be smart about the number of curves and the stage 1 bound, but the ECM library itself uses only the default level of optimization so your actual ECM curves could probably run quicker. ECM with bigger bounds is not turned on by default because it can potentially take a lot of time, it's not multithreaded and if you're using Msieve through a script or to finish off a large NFS run then you don't even want ECM anyway.

The Msieve makefile expects ecm.h to be in a system header directory and expects libecm.a to be in a system library directory. Usually 'make install' for GMP-ECM will put them there, and by default they are dropped into /usr/local/include and /usr/local/bin respectively, but those need to be in the default paths for your compiler. Your error message says one of those things hasn't happened.

chris2be8 · 2017-01-19, 16:45

Quote:

Originally Posted by LegionMammal978

Also, just curious, the Python script describes Msieve GPU support, is that anything I should care about?

Only if you have a Nvidia GPU. If not forget it.

You can also use a Nvidia GPU to do ECM stage 1. Getting it all working was "interesting" but I've managed to script it so I just feed a number into the process and it gets ECMed and poly selected with GPU support, then passed to another system to factor without any more human intervention needed.

Chris

wombatman · 2017-01-19, 16:50

Quote:

Originally Posted by chris2be8

Only if you have a Nvidia GPU. If not forget it.

You can also use a Nvidia GPU to do ECM stage 1. Getting it all working was "interesting" but I've managed to script it so I just feed a number into the process and it gets ECMed and poly selected with GPU support, then passed to another system to factor without any more human intervention needed.

Chris

Any chance you'd be willing to share said script?

LegionMammal978 · 2017-01-19, 22:05

Quote:

Originally Posted by chris2be8

Only if you have a Nvidia GPU.

Good thing I got a GTX 1060 last year :P

chris2be8 · 2017-01-20, 16:49

Quote:

Originally Posted by wombatman

Any chance you'd be willing to share said script?

There are several parts, make_polys.sh is called at intervals which calls check_if_done.pl and mkpolys.pl to do the work. I've got one version of the shell script for each system, each is run from the directory it's in to process any .n or .poly files placed in that directory.

Rename them to remove .txt off the end of the names.

Be warned that mkpolys.pl is set to use 3 cores because the system with my GPU needs to use 3 cores to do ECM stage 2 as fast as the GPU can do stage 1. Adjust that to fit your system. And it calls several scripts in ecm.3.tgz to run the ECM in parallel on 3 cores. They will need re-writing to use a different number of cores.

Sorry it's a bit untidy. I never intended to share it.

Chris

wombatman · 2017-01-20, 19:49

No worries on the tidiness. You're very kind to share the scripts

2017-01-19, 03:24	#12
VBCurtis "Curtis" Feb 2005 Riverside, CA 1001011111101₂ Posts	How large a factorization are you planning to do with your setup? Are you doing mostly SNFS or GNFS numbers? The GPU support is for poly select only, the first stage of GNFS projects. It's massively faster than CPU select, but the overall difference in project length is tiny for inputs below 140 digits, acceptably small for 140-150 digits, and not so small for 150+ digits. If you're planning to tackle GNFS-150+, you should consider using msieve with GPU support; that means learning to find a poly separately from the python script. I have a GPU handy, and still don't bother using it for inputs less than the high 130 digit range. The extra human time doesn't save enough silicon time for my interests below 138ish. A rough estimate is that GPU select can knock half a digit off the effective difficulty (by finding a polynomial that performs 5-7% better, in less wall-clock time, than a CPU search). You'll find threads in the msieve forum with guidance for specific invocations. Last fiddled with by VBCurtis on 2017-01-19 at 03:26

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Installation and running of mkfactc	daxmick	GPU Computing	44	2018-03-28 02:04
GMP-ECM - installation - how to compile - Ubuntu 15	ThomRuley	GMP-ECM	2	2017-01-06 16:54
installation of OS trouble	wildrabbitt	Linux	5	2015-12-22 16:51
Don't know where installation folder went	jinydu	Information & Answers	10	2009-10-14 01:29
GMP-ECM 6.2-rc1 - installation under MinGW failed	Andi47	GMP-ECM	5	2008-04-30 17:46

2017-01-19, 12:55	#13
jasonp Tribal Bullet Oct 2004 3,541 Posts	Actually Msieve runs GMP-ECM by default for 15-digit factors, but if you run with '-e' it will do more work and look for larger factors, up to the point where switching to QS is faster. For large inputs you can run ECM looking for up to 40-digit factors, and it tries to be smart about the number of curves and the stage 1 bound, but the ECM library itself uses only the default level of optimization so your actual ECM curves could probably run quicker. ECM with bigger bounds is not turned on by default because it can potentially take a lot of time, it's not multithreaded and if you're using Msieve through a script or to finish off a large NFS run then you don't even want ECM anyway. The Msieve makefile expects ecm.h to be in a system header directory and expects libecm.a to be in a system library directory. Usually 'make install' for GMP-ECM will put them there, and by default they are dropped into /usr/local/include and /usr/local/bin respectively, but those need to be in the default paths for your compiler. Your error message says one of those things hasn't happened.

2017-01-20, 19:49	#18
wombatman I moo ablest echo power! May 2013 6E9₁₆ Posts	No worries on the tidiness. You're very kind to share the scripts