Newsgroups: comp.parallel
From: karypis@in19.arc.umn.edu (George Karypis)
Subject: Version 2.0 of METIS: Graph Partitioning and Sparse Matrix Ordering Package
Organization: University of Minnesota
Date: 5 Sep 1995 14:43:39 GMT
Message-ID: <42hnmr$7ib@usenet.srv.cis.pitt.edu>

METIS: Unstructured Graph Partitioning and Sparse Matrix Ordering System

Release 2.0 of METIS is now available.

We would like to announce the release of version 2.0 of the METIS package 
for partitioning unstructred graphs (unstructured finite element meshes)
and for producing fill reducing orderings of sparse matrices.
METIS 2.0 contains a number of changes over METIS 1.0. The major changes 
are the following:

* A new faster k-way partitioning algorithm has been implemented
  that produces good partitions, and it is up to 5 times faster the 
  recursive bisection algorithm used in version 1.0.

* The algorithms have been extended to handle non-power of 2 partitions.  

* A stand-alone library is provided to interface with the partitioning 
  and ordering algorithms of METIS.

* A number of modifications to better handle weighted graphs.

* Speed improvements, particularly in the ordering code. METIS 2.0
  is about 10% to 15% faster than version 1.0.

* Better memory management. 


Overview of METIS
-----------------

Graph partitioning has extensive applications in many areas, including 
scientific computing, VLSI design, and task scheduling. The problem is 
to partition the vertices of a graph in k roughly equal parts, such that 
the number of edges connecting vertices in different parts is minimized.  
Graph partitioning is of particular importance in finite element 
computations on parallel computers, since a good partition significantly
reduces the amount of communication, increasing the performance.

Graph partitioning algorithms are also used to compute orderings for 
sparse matrices, in order to minimize the amount of fill and to increase 
the concurrency that can be exploited during parallel direct factorization.

METIS is a set of programs that implement various graph partitioning 
algorithms that are based on the multilevel paradigm. 
The advantages of METIS compared to other similar packages are the following:

- Provides high quality partitions!
    The partitions produced by METIS are consistently 10% to 50% better than 
    those produced by spectral partitioning algorithms, and 5% to 15% better 
    than those produced by Chaco multilevel.

- It is extremely fast!
    METIS is 20 to 160 times faster than multilevel spectral bisection, and 
    5 to 30 times faster than Chaco multilevel, for a wide variety of a graphs.
    Graphs with over 500,000 vertices can be partitioned in 256 parts, in 
    under a minute on scientific workstations. The run time of METIS is 
    comparable to (or even smaller than) the run time of geometric 
    partitioning algorithms that often produce much worse partitions.

- Provides low fill orderings!
    The orderings produced by METIS are significantly better than those 
    produced by multiple minimum degree, particularly for large finite 
    element graphs. Furthermore, unlike multiple minimum degree, the 
    elimination trees produced by METIS are suited for parallel direct 
    factorization.


Obtaining METIS
---------------

METIS is freely distributed. Information on how to get the source code
is available on WWW at
  URL: http://www.cs.umn.edu/~karypis/metis/metis.html

Alternatively, METIS can be obtained via anonymous ftp from
ftp.cs.umn.edu/dept/users/kumar/metis/metis-2.0.tar

Papers describing and analyzing the various algorithms implemented by 
METIS can be retrieved via WWW from:
    http://www.cs.umn.edu/users/kumar/papers/mlevel_serial.ps
    http://www.cs.umn.edu/users/kumar/papers/mlevel_kway.ps
    http://www.cs.umn.edu/users/kumar/papers/mlevel_analysis.ps
    http://www.cs.umn.edu/users/kumar/papers/mlevel_parallel.ps


METIS has been written by George Karypis and Vipin Kumar, at the 
Computer Science Department of the University of Minnesota.
If you have any questions or problems obtaining METIS, send
email to karypis@cs.umn.edu.




------------------------------------------------------------------
George Karypis                           email: karypis@cs.umn.edu 
University of Minnesota,       URL: http://www.cs.umn.edu/~karypis
------------------------------------------------------------------

