Held November 13 and 14 at the University of Tennessee and partially funded by the CRPC, the Basic Linear Algebra Subprograms (BLAS) Technical Workshop attracted 52 researchers from around the country to discuss the possibility of extending currently accepted BLAS standards to provide greater coverage of sparse matrices and additional facilities for parallel computing. Organized by CRPC researcher Jack Dongarra of the University of Tennessee and Oak Ridge National Laboratory, Iain Duff of Rutherford Appleton Laboratory, and Mike Heroux of Cray Research, Inc., the workshop was divided into a birds of a feather segment and eight other sessions that included a total of 28 talks by experts in the field of linear algebra. A discussion period followed each session.s

Talks focused primarily on aspects of developing a set of Parallel BLAS and related interfaces for linear algebra, including the Sparse BLAS, Sparse Parallel BLAS, Basic Linear Algebra Communication Subprograms (BLACS), and extensions to the BLAS. Participants discussed the future development of a set of standards for basic matrix data structures, both dense and sparse, as well as calling sequences for a set of low-level computational kernels for the parallel and sequential settings. These new standards would complement and supplement the existing ones for sparse and parallel computation. One of the major goals of this work is to enable both public domain and commercial linear algebra libraries to interoperate efficiently.

Attendees concluded that there was sufficient interest to form a group to investigate establishing a standard for the Parallel BLAS. They also concluded that researchers will need lower-level routines to ensure high performance on small problems and stressed the importance of interoperability between programming languages. They will look into extensions and alternative entry points that allow different levels of error checking and profiling to the BLAS, and will give further consideration to the facilities and functionality offered by the BLAS in existing software packages.

The next BLAS workshop will be scheduled in spring 1996. For additional information about the November 1995 workshop, including many of the presentations, see .

Table of Contents