Talk abstract:
Recent results in parallel numerical linear algebra
James W. Demmel, University of California, Berkeley
We discuss two recent developments in linear algebra algorithms
for high performance machines. The first development is SuperLU,
a new sparse Gaussian elimination routine for workstations and
SMPs. A variety of innovations, including some closely tied
to cache and superscalar architectures in modern workstations,
lets serial SuperLU attain up to 50% of peak machine performance,
depending on the sparse matrix. Parallel SuperLU attains up
to 6 or 7-fold speedup on an 8-CPU AlphaServer 8400, an 8-CPU
C90, or an 8-CPU SGI Power Challenge, again depending on the
sparse matrix. This is joint work with X. Li, J. Gilbert, S.
Eisenstat and J. Liu.
The second development is a parallel algorithm, due to B.
Parlett and I. Dhillon, for the symmetric tridiagonal eigenproblem.
While still under development, this algorithm, which is similar
to inverse iteration, promises to be highly accurate, guaranteed
O(n2) in cost, and embarrassingly parallel.
It could become the algorithm of choice on serial as well as
parallel machines.
Back to Workshop Schedule
1996-1997
Mathematics in High Performance Computing
|
