Parallel Linear Algebra Package (PLAPACK) Release R3.1

Background Previously created software packages for parallel computers are inefficient because data is continuously passed back and forth between separate parallel machines. These applications have been unsuccessful in building parallel linear algebra libraries for distributed memory supercomputers.

Invention Description PLAPACK is a library infrastructure for the parallel implementation of linear algebra algorithms and applications on distributed memory supercomputers such as the Intel Paragon, IBM SP2, Cray T3D/T3E, SGI PowerChallenge, and Convex Exemplar. This infrastructure allows library developers, scientists, and engineers to exploit a natural approach to encoding so-called blocked algorithms, which achieve high performance by operating on submatrices and subvectors. This feature, as well as the use of an alternative, more application-centric approach to data distribution, sets PLAPACK apart from other parallel linear algebra libraries, allowing for strong performance and significantly less programming by the user.

Benefits

High-level abstraction obtained User-friendly Convenient interface for applications Reduced coding required to maintain libraries

Features

Use of abstract programming interface Use of message passing interface Data distribution that is vector-centric rather than matrix-centric

Market Potential/Applications PLAPACK is an infrastructure for building software libraries; thus, it is inherently designed to allow adding functionality. Such libraries are used for scientific applications in industry and academia.

UT Researcher Robert A. Van de Geijn, Ph.D., Computer Sciences, The University of Texas at Austin John A. Gunnels, Ph.D., T.J. Watson Research Center, IBM, Inc.

Type of Offer: Licensing



« More Computer Science Patents

---