Next: Dynamic Load Balancing Algorithms
Up: The Static Load Balancing
Previous: Parallel partitioning algorithms
In choosing a partitioning algorithm, it is necessary to balance the
benefit of the reduction in communication cost, against that of the
cost of generating the partitioning. If the parallel computer
has slow communication, or the application is
fine-grained where there
is no large chunk of computation to ``shield'' the cost of
communication, it is very important to have a partition
of good quality. On the other hand, if the parallel computer has
very fast communication, or the application is coarse grained
where there is a lot of computation in between communication,
a big reduction of the communication cost will only
result in a fractional reduction of the overall cost,
and in such a case a simple and cheap partitioning algorithm,
such as the greedy algorithm, is quite adequate.
Some of the most recent parallel partitioning codes [49,83]
give very good quality partitions. They are scalable in memory
and yet are comparable in
cost to the very simple algorithms such as the greedy algorithm.
A number of packages are publicly available.
Many of them are constantly being updated.
Further details should be checked from the contact addresses listed.
- CHACO
(http://www.cs.sandia.gov/CRF/chac.html)
perhaps the first publicly available package. The latest
version CHACO 2.0 uses multilevel combined with a number of
available methods such as K-L refinement, spectral bisection.
Available by request.
- JOSTLE
(http://www.gre.ac.uk/jostle/)
multilevel graph partitioning using -way refinement.
Both the sequential and parallel version are
available free for academic use as an executable,
after signing an agreement.
- METIS
(http://www.cs.umn.edu/~karypis/metis/metis.html)
multilevel graph partitioning using -way refinement.
Both the sequential and the parallel
version are available free
for down-load.
- PARTY
(http://www.uni-paderborn.de/fachbereich/AG/monien/RESEARCH/
PART/party.html)
sits on top
of existing packages, offers
greedy, multilevel spectral bisection, coordinate bisection,
inertia bisection and other algorithms.
Version 1.1 is available after signing a license.
- PMRSB
( stb@renaissance.cray.com)
parallel multilevel spectral bisection. Available only
on CRAY platforms.
- S-HARP [77]
(http://www.cs.njit.edu/sohn/sharp/)
multilevel graph partitioning using multiple
eigenvectors and (spectral) inertia bisection on the coarsest level.
- SCOTCH
(http://www.labri.u-bordeaux.fr/Equipe/ALiENor/membre/
pelegrin/scotch/)
multilevel implementation of a number of algorithms including K-L.
The code is free for down-load for academic use.
- TOP/DOMDEC [21]
( charbel@boulder.colorado.edu)
Greedy algorithm, recursive coordinate and graph bisection,
recursive inertia bisection,
multilevel spectral bisection algorithm
with tabu search, simulated annealing and stochastic evolution
as smoothers.
- WGPP [31,32]
multilevel graph partitioning. Similar to METIS with some new features.
Next: Dynamic Load Balancing Algorithms
Up: The Static Load Balancing
Previous: Parallel partitioning algorithms
2000-03-21