Discussion related to cado-nfs

[Emmanuel Thomé <Emmanuel.Thome@inria.fr>]

On Thu, Sep 10, 2015 at 11:40:39AM -0300, Douglas B. Staple wrote:
> Dear CADO-NFS team:
> 
> Earlier this year I helped Karl Dilcher and John Cosgrave factor several
> integers between 151 and 166 digits using CADO-NFS, version 2.1.1.  You can
> see the complete factorizations here:
> http://www.mersenneforum.org/showthread.php?t=17226&page=13
> John and Karl wrote a paper on generalized Fermat numbers where they make
> use of these factorizations.  I'll be sure to pass the reference on when
> it's published.

Thanks for the link !

> I can make one usability suggestion, if you're interested.  It's is somewhat
> awkward the way the number of threads and the number of clients have to be
> specified independently.  I ran the program with OpenMP, but no OpenMPI
> support, so I configured CADO to use a 1x16 split for linear algebra, but a
> 8x2 split for sieving.  As a user, I'm less concerned with the way 16
> threads are split between processes than I am with the fact that 16 threads
> per node should be used throughout.  It'd be a lot easier to just specify
> "-t 16" rather than "-t 2 -s 8 tasks.linalg.bwc.threads=16".

Last week I tweaked factor.sh, so the behaviour changed a bit.

You'd want to use "-t 2 -s auto" now.

I'm not satisfied at all with the way it works, though -- it's pretty
contrived.

Chances are that factor.sh will disappear in the next release, by the
way, once we've arranged so that the python script does the few remaining
bits that only factor.sh does.

> The above comment is most relevant for numbers below 160 digits, where it's
> still reasonable to run the job on a compute single node using your
> "factor.sh" convenience script.  In the end, I used a pair of scripts of my
> own concoction (attached) to drive your cadofactor.py and wuclient2.py,
> which you may find interesting.
> 
> It was quite a pleasure using your tool.

Thanks !

E.