Fall 1997, August 26, Education 216, 3 pm.

Strategies, code generation and results for the automatic parallelisation
of mesh based codes onto distributed memory architectures

Steve Johnson


Computational speed has become one of the most limiting factors in the
exploitation of computational mechanics simulation. An obvious route to
relaxing this restriction is the use of parallel computation to employ a
large number of processors to complete the task in a fraction of the
original serial time. This use of parallel machines is inhibited by the
effort required to port codes onto parallel computer architectures and
acheive the promised speed improvements. A number of strategies for the
effective parallelisation for a set of code classes have successfully been
applied. The challenge that remains is to ease the process of converting
serial code to a parallel form without sacrificing performance. 

In this talk, strategies for parallelisation will be discussed in
conjunction with the automation of their application to real world codes.
The implications of some of the commonly used techniques employed in
computational mechanics software will also be addressed. A demonstration
of the parallelisation of a simple CFD code will also be included. Results
will be presented for a number of codes to demonstrate that effective
parallelisation of real world codes can be achieved with minimal user
effort and without the need for manual code re-authoring.  The basic
paradigm is source-to-source transformation, with explicit insertion of
message passing, followed by conventional compilation.  A detailed
interprocedural dependency analysis is a necessary precursor.