Revision 0560307f1161b3b7f573f8e399ed6e7c777c3a97 authored by Emmanuel Thomé on 25 February 2019, 15:54:57 UTC, committed by Emmanuel Thomé on 25 February 2019, 15:54:57 UTC
NEWS
Main changes between cado-nfs-2.3.0 and cado-nfs-3.0.0:
* C++11 is now a requirement.
* The sieving region is no longer limited to I=16.
* Finer-grain timings in the lattice sieving code can be obtained with -T
Main changes between cado-nfs-2.2.0 and cado-nfs-2.3.0:
* When tasks.threads is not given, it is set to the number of logical cores
(including hyperthreading if any) instead of the number of physical cores.
* The estimated time of arrival (ETA) is now printed during the linear
algebra phase too.
* Speedup in the linear algebra phase (mksol) when m > n
* DLP support: got rid of the Magma dependency
* Fixed compiler warnings with new versions of compilers.
* Several bug fixes.
* The output files have been shortened. Previously, we had filenames of
the form cxxx.TASKNAME.PROGRAMNAME.foo, where TASKNAME was actually an
implementation detail. In cado-nfs-2.3.0, TASKNAME goes away.
* In parallel to the change above, the parameter hierarchy no longer
includes the TASKNAME, which is an implementation detail. Previously,
the finest grain to which one could specify parameters was
tasks.PATH_PREFIX.TASKNAME.PROGRAMNAME.foo where PATH_PREFIX is set in
scripts/cadofactor/cadotask.py (among polyselect, sieve, filter,
linalg, reconstructlog, descent -- some toplevel relevant parameters
have no PATH_PREFIX). Now the TASKNAME goes away. This
means that as before, setting tasks.sieve.threads works for setting a
parameter which is obeyed by las (and actually makefb too), but changes
happen when one specifies parameters on a finer grain. While
tasks.sieve.sieving.las.threads or
tasks.sieve.factorbase.makefb.threads used to work, now these become
simply tasks.sieve.las.threads or tasks.sieve.makefb.threads.
* The --verboseparam option to cado-nfs.py now also prints an
automatically generated, and by default comprehensive list of
parameters that cado-nfs recognizes (among other things
which this option still does as it always has). Note that the lists for
factoring and DLP differ slightly.
* The database which stores the computation state can now also be backed
by a mysql server. This is optional, and requires the
python3-mysql.connector package (on Debian linux -- actual package name
depends on your OS distribution). To activate this feature, pass the
database URI as an extra parameter to cado-nfs.py, e.g.
"database=db:mysql://USER:PASS@host:port/foobar", where port, PASS, and
USER are optional. The same syntax may be used to have an sqlite3
database as usual, but stored in a different location (which can be
handy to work around some filesystem deficiencies). For that, use
"database=db:sqlite3:///tmp/foo.db", where obviously "/tmp/foo.db" must
be adapted to your preferred database path (you do need the triple /).
Main changes between cado-nfs-2.1.1 and cado-nfs-2.2.0:
* The toplevel scripts (factor.sh, cadofactor.py, wuclient2.py) have
changed. Now there are only two main scripts to be used by most users:
- cado-nfs.py
This combines the roles of the previous factor.sh and cadofactor.py.
The most basic use is ./cado-nfs.py <N>, and it will factor N on the
local host. It can also become a server in a multi-host
computation. See README for more details.
- cado-nfs-client.py
This is mostly a renaming of wuclient2.py. This starts a client that
can contribute to a running computation, possibly on another
machine.
By default, factorizations running on the localhost will use all
available cores.
* Main algorithmic improvements:
- polynomial selection: algorithm of Bai, Bouvier, Kruppa, Zimmermann
(Better polynomials for GNFS, 2015) has been implemented.
- relation collection: multi-level buckets for better handling of
medium-sized primes; cofactorization strategies à la Kleinjung. Both
features are experimental, and not activated by default.
- linear algebra: support for non-binary base field (for dlp),
including the removal of the overhead due to SMs.
* Main speedups:
- the creation of the factor base is now multithread.
- linear algebra speedup taking into account machine topology thanks to
HWLOC (http://www.open-mpi.org/projects/hwloc/).
- the square root step is now multithread.
- default parameters files were optimized for sieving with OPAL/nomad.
* DLP support: now fully functional over prime fields, including the
descent step. Still requires Magma for a small number-theoretic
computation at the beginning (dependency will be removed sooner or later).
See README.dlp for further details.
* In local.sh, the customization of the build_tree using the $up_path
variable used to be phrased as: build_tree="${up_path}some/sub/path".
Now it should be build_tree="${up_path}/some/sub/path", as $up_path no
longer embodies a trailing /
* The polynomial file now supports another format. For giving a
polynomial, a line like
poly0=17,42,34,55
can be written, to set the polynomial 17+42*x+34*x^2+55*x^3 on side 0.
The old format with X, c, Y,... is still supported, with c and X for
side 1 and Y for side 0.
* References to rational or algebraic sides have been reduced (at least
internally). This means that several parameters have now aliases
corresponding to sizes. For instance, the parameter -lpb1 for las is an
alias of -lpba. The common convention is that the rational side (if
any) is on side 0, but everything will work if a polynomial file is imported
with the rational side on side 1.
* Various changes in the parameter list for the linear algebra task,
in particular, the linalg.bwc.mn=... parameter is replaced by 2
separate parameters for m and n. Also the rhs=... parameter allows
non-homogeneous system solving (useful for handling SMs with no
overhead in DLP mode).
* An estimated time of arrival (ETA) is now printed during the polynomial
selection and sieving steps. It corresponds to the estimate finish time
of each of those steps (not the full factorization).
Main changes between cado-nfs-2.1 and cado-nfs-2.1.1:
* Added test-case in wuclient2.py for buggy Python MIME encoder to prevent
data corruption during file upload under future Python versions
* Bug fixes to multi-threading in linear algebra code
Main changes between cado-nfs-2.0 and cado-nfs-2.1:
* Polynomial selection now runs in two separate phases, size optimization
and root optimization, with improved parameters
* Unit tests added, please run "make check" after building
* Various small speed-ups
* Various bug-fixes
* The wuclient2.py script no longer requires an external download tool
(wget or curl) to download securely via SSL under Python 2
* better support of DL in GF(p); still not fully automatic, nor
independent of third-party software. See README.dlp.
Main changes between cado-nfs-1.1 and cado-nfs-2.0:
* the default parameters in the params subdirectory have been improved
* major speedups in the polynomial selection, sieving and filtering steps
* the filtering step produces a smaller matrix (about 5% less rows),
see http://hal.inria.fr/hal-00734654 for more details.
* the replay step is faster and uses less memory
* fixed bug in the lattice siever (las) on 32-bit computers
* allow compilation on MinGW32. Performance is poor, however. MinGW64 is
currently not supported
* new cadofactor.py script written in Python, allows larger-scale distribution
with client/server model
* files used/generated by cado-nfs-1.1 are not entirely compatible with
cado-nfs-2.0. Relations files are compatible, but for pretty much
everything else, it is preferrable to start afresh. Likewise,
command-line options of several binaries have changed.
* implementations of NFS-DL and NFS-FFS (experimental, requires some
background knowledge to use)
Main changes between cado-nfs-1.0 and cado-nfs-1.1:
* the polynomial selection uses the new algorithm presented by Kleinjung at
the 2008 CADO workshop (supports multithreading)
* the polynomial selection uses a faster root sieve (rootsieve5)
* the lattice siever (las) supports a larger sieving region (I=16). This is
useful for integers greater than 200 digits
* minor bug fixes, small speedup in sieving (las) and filtering (purge)
* reorganization of the binaries in the linear algebra step and speedup
when using many machines with MPI
* new experimental scripts to execute the sieve on a cluster.
These scripts rely on the OAR job scheduler being used, and exploit its
"besteffort" mode. Some even more experimental scripts for linear
algebra are in the development version.
* addition of new parameter files and a few new parameters (cf params.c91)
* the linear algebra step now makes use of dynamically loaded shared
libraries. This might result in configuration issues, which we hope the
scripts get it right. See the README file.
Computing file changes ...