This thesis evaluates the High Performance Fortran (HPF) language as a candidate
for implementing ocean general circulation modeling (OGCM) software on
parallel architecture computer systems. The thesis reviews major HPF language
features and discusses general algorithmic issues common to broad classes of
OGCM codes. We discuss the use of intrinsic functions, data distribution
directives and explicitly parallel constructs to optimize performance by
minimizing communication requirements in a portable manner. We show that HPF
provides convenient language structures for implementing several widely used
OGCM algorithms, including finite difference solvers that use regular grids. We
focus on regular grids, since these can be efficiently represented by existing HPF
definitions.
We parallelize the algorithm by partitioning arrays among processors of IBM SP2
system. We use the DISTRIBUTE directive to specify how to distribute a
template among the processors, and the arrays are then aligned to the template. We
also use the FORALL parallel construct to take the advantage of parallel
computer architecture. We use the INDEPENDENT directive with FORALL
construct to provide additional assurance for parallelism.
We have demonstrated the basic HPF concepts through the some examples. We
have carried out many test runs for both FORTAN77 and HPF code on IBM SP2
platform using various numbers of processors. We obtain a nearly linear scaling
performance line for the HPF code, while the speed up data of the HPF code also
shows excellent improvement over the FORTAN77 code.