Fwd: Scotch v7.0 is out!
Hi Debian Science folk,
SCOTCH has made a new major release, and has also made a particular
announcement I'd like to draw your attention to.
To strengthen the long-term value of the software, they are creating a
Consortium to drive governance and development planning.
As part of that, they are requesting interested users to fill out a
feed-back form to help guide decisions. The form is found at
https://framaforms.org/feedback-on-the-scotch-software-1640960422
SCOTCH provides graph partitioning, which is crucial role for
implementing efficient large-scale parallelization. It's used by linear
algebra solvers such as PETSc and MUMPS, and therefore used by Finite
Element or Finite Volume solvers to solve differential equations or CFD.
It's a alternative package to ParMETIS which has a similar purpose (but
parmetis has a non-free licence).
Please fill out the feedback form if you've been using SCOTCH.
We might want to consider signing Debian up as a partner in their new
consortium.
I'm cc:ing Jonathan as DPL to let him know that might be interesting to
discuss further.
Drew
-------- Original Message --------
Subject: Scotch v7.0 is out!
Date: 2021-12-31 21:09
From: François PELLEGRINI <francois.pellegrini@u-bordeaux.fr>
To: François PELLEGRINI (UBx) <francois.pellegrini@u-bordeaux.fr>
We announce the release, as libre/free software, of version 7.0
(codename
"Sankara") of the SCOTCH and PT-SCOTCH software package and libraries
for
sequential and parallel graph partitioning, sequential and parallel
graph
clustering, sequential mesh/hypergraph partitioning and static mapping,
and
sequential and parallel sparse matrix block ordering.
THIS IS A MAJOR RELEASE, fruition of six years of development, which
brings
many innovative features.
CAUTION: While their interface is (as always) backward-compatible, and
their
overall quality is unchanged, SCOTCH and PT-SCOTCH v7.0 now use
dynamically
all threads available to their master (launch) thread. While this may
ease
the life of SCOTCH users, by bringing execution performance silently, it
may
raise issues for PT-SCOTCH users, because of thread allocation overlaps
with
MPI (duplicating threads for all MPI processes of a single hardware node
may
not yield the expected performance; all the contrary!). Hence,
PT-SCOTCH
users are invited to read carefully the user's manual to figure out
how to
launch MPI processes that will create non-overlapping pools of threads
for
every MPI node.
IMPORTANT: YOUR HELP IS NEEDED: in 2022, the SCOTCH project will turn
30.
In order to perpetuate software development around SCOTCH, while
preserving
its free availability to all under a free/libre software license, we
are on
the road to creating a consortium. The consortium will pool
development
goals and investments from its members. The investments will be used to
fund
engineering staff to carry out strategic R&D relevant to the
consortium
members.
We are currently identifying ambitious key parties interested in joining
the
consortium to shape the future of Scotch and drive the consortium. This
is a
great opportunity for people and organizations having needs in fast
and
efficient graph partitioning to support and ensure the sustainability of
the
Scotch software for its users all around the world.
Hence we make a call to our user community, especially on the industrial
and
research side: would you be so kind as to fill-in the questionnaire
below?
- The feed-back form is available here:
https://framaforms.org/feedback-on-the-scotch-software-1640960422
- The preparatory document for the Scotch consortium is available here:
https://framaforms.org/sites/default/files/forms/files/scotch_consortium_preparation_20211231.pdf
Important new features
======================
The most visible one is the complete refactoring of the thread
management
subsystem. Thread management is now fully dynamic. Thread creation
and
allocation is performed on the fly. Additionally, existing threads
can be
imported from the caller's environment, to avoid creating
unnecessary
threads and benefit from the mapping of existing threads.
Execution options, such as selecting either deterministic algorithms or
more
efficient yet non-deterministic versions, is now also dynamic
(while in
previous versions, it had to be prescribed at compile time).
A new class of objects is available in the API: execution contexts.
Context
objects allow users to encapsulate their thread pool, execution options
and
a private pseudo-random generator, so that concurrent SCOTCH tasks
(mapping
or partitioning, ordering, etc.) can be performed concurrently and
reprod-
ucibly on the same graph, or on different graphs at the same time.
The new thread model allowed us to design efficient multi-thread
versions of
many time-consuming algorithms, yielding much faster execution times
on
many-core machines. These algorithms concern the SCOTCH (formerly)
sequen-
tial library, as well as the PT-SCOTCH MPI-based library (which now
provides
hybrid parallel algorithms, mixing threads and MPI). Please make sure
your
MPI implementation supports this (else, it can be disabled at compile
time).
To ease migration from MeTiS, two versions of the libScotchMeTiS library
are
now provided: one for MeTiS v3, and the other for MeTiS v5. Hence, it is
no
longer necessary to compile Scotch twice to get the two versions.
Support for CMake has been added. It is still experimental, and it
will be
improved in the next releases. Feedback is welcome.
How to get SCOTCH
=================
SCOTCH can be freely downloaded, under the terms of the CeCILL-C
free
software license.
Since the Inria Gforge SVN repository has been decommissioned this fall,
the
new authoritative source for obtaining SCOTCH is the Inria Gitlab
repos-
itory, at this address:
https://gitlab.inria.fr/scotch/scotch
Happy new year to all! =^v^=
f.p.
Reply to: