ppyNFS

paul0 · 2015-03-29, 21:06

Hi guys,

I've been asking lots of questions here to understand NFS, and people who answered really helped me understand it. So, I was able to eventually code NFS, it's a really good learning exercise.

I've uploaded the code at https://github.com/solidwrench/ppyNFS

Thanks,
paulo_

dleclair · 2015-03-30, 01:58

This is a very generous contribution. Thank you!

-Don

R.D. Silverman · 2015-03-30, 12:52

Quote:

Originally Posted by paul0

Hi guys,

I've been asking lots of questions here to understand NFS, and people who answered really helped me understand it. So, I was able to eventually code NFS, it's a really good learning exercise.

I've uploaded the code at https://github.com/solidwrench/ppyNFS

Thanks,
paulo_

Applause. Kudos. Congratulations.

It is always pleasing to see someone take the time to dig into a complicated
algorithm.

However, you will find that the numbers you can factor with your code are quite limited in size.

I did not dig into your polynomial root finder. May I ask what method you use?

Some suggestions:

(1) The most severe constraint for your code is the LA. You will find that Gaussian elimination sharply
restricts your factor base size. This, in turn, sharply restricts the numbers you can do.

(2) You need to implement a lattice siever. Use the more modern approach of Kleinjung et.al. rather than
Pollard's approach. [Note! I wish I could find the time to re-write my siever.] Note that a line-siever
will (with better LA, filter, sqrt code) allow you to perform factorizations up to (say) SNFS C180 or so.

(3) I did not look at your filtering/matrix preparation code. If you have not done so, you will need
to implement a clique based filter.

(4) Couveigne's sqrt algorithm is also rather limiting. It can't handle even-degree fields.

Ask if you need help.

paul0 · 2015-03-31, 10:48

Quote:

Originally Posted by R.D. Silverman

I did not dig into your polynomial root finder. May I ask what method you use?

I used the recursive "random splitting" algorithm as described in page 103 of Prime Numbers: A Computational Perspective. See functions getRootsModPFast() and getRootsModPSlow() in poly.py.

Quote:

Originally Posted by R.D. Silverman

(3) I did not look at your filtering/matrix preparation code. If you have not done so, you will need
to implement a clique based filter.

Though not currently uploaded, I have code that does singleton filtering.

Everything else you've mentioned is a future goal, but I'll need to invest time to study and implement them.

jasonp · 2015-04-02, 00:03

Nicely done.

The crappy thing about NFS is that in order to scale above small problems (70-80 digits) you have to implement all the features Bob lists. You can do line sieving with a single large prime per side, simple graph-based filtering, and keep the Gauss elimination, and that will get you up to 80-digit general numbers. But QS can factor numbers that size in a few minutes.

2015-03-29, 21:06	#1
paul0 Sep 2011 3·19 Posts	ppyNFS Hi guys, I've been asking lots of questions here to understand NFS, and people who answered really helped me understand it. So, I was able to eventually code NFS, it's a really good learning exercise. I've uploaded the code at https://github.com/solidwrench/ppyNFS Thanks, paulo_

2015-03-30, 01:58	#2
dleclair Mar 2003 7·11 Posts	This is a very generous contribution. Thank you! -Don

2015-04-02, 00:03	#5
jasonp Tribal Bullet Oct 2004 3,541 Posts	Nicely done. The crappy thing about NFS is that in order to scale above small problems (70-80 digits) you have to implement all the features Bob lists. You can do line sieving with a single large prime per side, simple graph-based filtering, and keep the Gauss elimination, and that will get you up to 80-digit general numbers. But QS can factor numbers that size in a few minutes.