Newsgroups: comp.parallel.mpi
From: Mahesh Joshi <mjoshi@cs.umn.edu>
Subject: Announcing PSPASES and WSSMP
Organization: University of Minnesota
Date: Mon, 12 Jan 1998 16:00:14 -0600
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii; name="Announcement.news"
Content-Transfer-Encoding: 7bit
Content-Disposition: inline; filename="Announcement.news"
Message-ID: <34BA926E.9805346@cs.umn.edu>

  ________________________________________________________________________

                       Announcing PSPASES and WSSMP
  ________________________________________________________________________
  
  "PSPASES : A Scalable Parallel Direct Solver Library for Sparse Symmetric 
  Positive Definite Systems"

  Authors :

  Mahesh Joshi, George Karypis, Vipin Kumar (Department of Computer 
  Science, University of Minnesota, Mineapolis, MN) and Anshul Gupta (IBM 
  Thomas J. Watson Research Center, Yorktown Heights, NY).
  ________________________________________________________________________

  Abstract :

  We are glad to announce a beta release of PSPASES, a stand-alone MPI-based
  parallel library for solving linear systems of equations involving sparse 
  symmetric positive definite matrices. The library efficiently implements 
  the scalable parallel algorithms developed by authors, for each of the four 
  phases of direct solution method; viz. ordering, symbolic factorization, 
  numerical Cholesky factorization, and solution of triangular systems. 

  PSPASES is highly scalable, mainly because it uses a highly scalable 
  parallel multifrontal algorithm in the most expensive computational phase 
  of numerical factorization. All the other phases are also scalable by 
  themselves.

  In our testing, PSPASES solved a system of around 1 million equations in 
  just 62 seconds on 64 processors and in 38 seconds on 128 processors of 
  Cray T3E. This time included times required for all the four phases of 
  the solver. The highest performance clocked by PSPASES is 21.2 GFLOPS for
  the numerical factorization phase. This efficient and scalable behavior 
  is demonstrated while solving most of the systems appearing in practice. 

  PSPASES is implemented using standard MPI and BLAS, which makes it portable 
  to most of today's parallel computers and networks of workstations. We have 
  tested PSPASES extensively on IBM SP, network of IBM RS6000 workstations, 
  Cray T3E, SGI Origin 2000 and PowerChallenge architectures. 

  PSPASES uses ParMETIS and METIS as default libraries for computing fill-
  reducing ordering, but it also accepts user supplied ordering. Different 
  functional interfaces are provided for each of the phases of the solver 
  and a simple interface is also provided for easy use. The user can use 
  these interfaces to solve multiple systems with same nonzero structures; 
  to solve same system for multiple right hand sides; and to get different 
  statistical information such as the memory requirements of the solver 
  and the quality of the ordering.

  Visit the PSPASES web site at

                    http://www.cs.umn.edu/~mjoshi/pspases

  to obtain the software, the manual, and related technical papers. The 
  software can directly be obtained via an anonymous ftp to 
  "ftp.cs.umn.edu". Get the files "pspases-beta.tar.gz" and 
  "README.PSPASES" located in the directory "users/kumar/mahesh".

  Any comments, questions or bugs regarding PSPASES can be directed to 
  Mahesh Joshi (mjoshi@cs.umn.edu).

  ________________________________________________________________________

  "WSSMP: Watson Symmetric Sparse Matrix Package"

  Authors :

  by Anshul Gupta (IBM Thomas J. Watson Research Center, Yorktown Heights, 
  NY), Mahesh Joshi and Vipin Kumar (Department of Computer Science, 
  University of Minnesota, Mineapolis, MN).
  ________________________________________________________________________

  Abstract :

  A faster version of the solver with enhanced functionality is available 
  for IBM SP and RS6000 systems, as WSSMP.  The WSSMP package contains
  robust industrial strength code for serial and parallel solution of sparse
  symmetric positive definite and indefinite systems.  WSSMP has been
  clocked at up to 450 MFLOPS on an RS6000/397 workstation and up to 24 GFLOPS 
  on a 64-processor SP with model-397 nodes.  Documentation for WSSMP is
  available at ftp://ftp.cs.umn.edu/users/kumar/anshul/WSSMP-manual.ps.

  Any questions pertaining to WSSMP may be directed to Anshul Gupta
  (anshul@watson.ibm.com).

  ________________________________________________________________________