Compiling Scientific Programs for Parallel Computers

                      John Mellor-Crummey
                  Department of Computer Science
                        Rice University
                          Houston, TX

Over the last several decades, computational modeling and simulation
has become a cornerstone of scientific inquiry.  To reduce the
execution time for large-scale calculations, scientists have turned
to parallel computers ranging from tightly-coupled multiprocessors to
loosely-coupled networks of workstations or personal computers.  The
key stumbling block to broader use of parallel computation in
engineering and science is the difficulty of developing programs for
parallel systems.

This talk will focus on parallelizing compiler technology designed to
simplify the task of developing scientific applications for parallel
computers. I will describe analysis and code generation challenges we
have encountered for parallelizing scientific applications and how we
address these challenges using a novel data-parallel compilation
framework based on the manipulation of sets of integer tuples.  This
approach enables us to combine aggressive program optimization with
sophisticated computation partitioning strategies. Using this
approach, the dHPF compiler, a parallelizing compiler for High
Performance Fortran under development at Rice University, has begun to
generate parallel code that achieves near hand-coded performance for
tightly-coupled computational fluid dynamics simulations. The talk
will conclude with an assessment of progress to date and the
challenges that remain.

This is joint work with Daniel Chavarria-Miranda. 

-----------------------------------------------------------------------

Dr. John Mellor-Crummey earned a BSE degree magna cum laude in
electrical engineering and computer science from Princeton University
in 1984, and MS (1986) and Ph.D. (1989) degrees in computer science
from the University of Rochester.  In 1989, he joined the Department
of Computer Science and the Center for Research on Parallel
Computation (CRPC) at Rice University, where he currently holds the
rank of Senior Faculty Fellow.  Dr. Mellor-Crummey was a member of the
technical steering committee of the CRPC (a National Science
Foundation Science and Technology Center), and is currently a member
of the Executive Committee for the Los Alamos Computer Science
Institute--a joint venture between Los Alamos National Laboratory and
Rice University.  He is best known for multi-processor synchronization
algorithms he developed with Michael Scott (University of Rochester).
In recent years, his research has focused on compilers, run-time
libraries, and programming environments for parallel processing. He is
the leader of the dHPF compiler project at Rice University--a
long-term research effort that has focused on developing technology
for compilers and tools to support construction of efficient
data-parallel scientific applications.

Dr. Mellor-Crummey is a member of the Association for Computing
Machinery, IEEE Computer Society, Tau Beta Pi, and Phi Beta Kappa.