Christof Voemel's PhD thesis defense

   PhD Thesis defense of
                             Christof Voemel
                                  on
   "Contributions to research in high performance scientific computing for
                              sparse matrices"

Date and location : 10.00 am.  March 20, 2003. CERFACS, Toulouse France.

Members of the Jury:
     P. Amestoy, M. Arioli, I.S. Duff, E. Ng, A. Pothen, J. Roman
         (E. Ng and J. Roman acting also as referees)

Abstract :
In this thesis, we present our research on methods in high performance
scientific computing for sparse matrices. Specifically, we are
concerned with the development of a new algorithm for incremental norm
estimation, the reference model implementation of the standard
for Basic Linear Algebra Subprograms for sparse matrices (Sparse
BLAS), and the design of a new task scheduler for MUMPS,
an asynchronous distributed memory multifrontal direct solver for
sparse linear systems.

Our new incremental norm estimator has the advantage of being applicable
to both dense and sparse systems, in contrast to previously existing
incremental schemes. Applications include for example monitoring of a QR,
Cholesky or LU factorization. If the explicit inverse of a triangular
factor is available our scheme can be used to estimate the condition
number of a matrix.

Our Fortran 95 reference model is currently the only existing implementation
of the interfaces in the standard for the Sparse BLAS.
We identify many of the complicated issues regarding the representation
and the handling of sparse data structures that
underlie the high-level description of sparse vector and matrix
operations in the standard but are avoided there in order to
leave enough freedom for vendors to provide efficient implementations.

With our new task scheduler for MUMPS, we address concerns about
lack of scalability and performance on large numbers of processors
that arose in a comparison with the SuperLU parallel solver.
In the new approach, we determine, during
the analysis of the matrix, candidate processes for the tasks
that will be dynamically scheduled during the subsequent factorization.
This approach significantly improves the scalability of the solver
in terms of time and memory needed, as we show by comparison with the
previous version.

-----------------------
Note that the thesis will later appear as a CERFACS report on the web site:
http://www.cerfacs.fr/algor/reports/index.html