Discussion related to cado-nfs

[Paul Leyland <paul.leyland@gmail.com>]

On Sat, 2013-04-06 at 14:05 +0200, Zimmermann Paul wrote:
> Paul,
> 
> > From: Paul Leyland <paul.leyland@gmail.com>
> > Reply-To: paul.leyland@gmail.com
> > 
> > On Fri, 2013-04-05 at 09:36 +0200, Zimmermann Paul wrote:
> > > Paul,
> > > 
> > > I hope the issue is fixed on your side now.
> > 
> > Very probably.  purge.log now ends with
> > 
> >    new_nrows=19342364 new_ncols=19185740 (156624) at 400.68
> > excess < 0.10 * #primes. See -required_excess argument.
> > 
> > The previous version had an excess of 1749.  #primes is ~53M so another
> > two or three iterations should be enough.
> 
> any news of your first factorization with CADO-NFS?

Yes, though due to my incompetence the news isn't good.

I tried to add another machine as a siever but didn't really understand
the format of the mach_desc file and, in particular, the scope of the
bindir declaration.  When cadofactor.pl restarted it tried to de-dup and
purge the relations found so far but used the wrong path.  Instead of
dying with an error "program not found" or similar, it deleted all the
relation files.  After several restores from backups it still wasn't
working properly; as the backups were almost a day old I'd lost a lot of
work anyway.  So I gave up and started again with the polynomial finder.
Right now two machines are sieving and have found 24M relations.

Incidentally the old (1.1) polynomial file had
# Murphy's E(Bf=10000000,Bg=5000000,area=1.00e+16)=5.69e-12
and the new (git) file has
# MurphyE(Bf=10000000,Bg=5000000,area=1.00e+16)=6.23e-12

which is markedly better, so not as much computation has been lost as
might otherwise been thought.

I'm not too disappointed.  A steep learning curve was expected from the
start.  Using previous NFS implementations, such as the CWI suite and
msieve / ggnfs, has been at least as difficult.

> > > Anyway we will change our code so that when an error occurs in purge,
> > > the factorization is aborted, to avoid it runs forever as you 
> > > experienced.
> > 
> > Thanks.  When I've finished this factorization and perhaps another one,
> > I'll put together a list of suggestions based on my experience with this
> > code and with other NFS implementations.   I may even be able to
> > implement one or two of them!
> 
> you are most welcome!

Another addition to the list should be apparent from the tale above.

> 
> Paul


ATB,
        Paul (the other one)