Rapid Parallelisation of the Industrial Modelling Code PZFlex
Abstract
This paper presents the results of a recent collaborative
project between EPCC at the University of
Edinburgh and Scottish engineering consultants Weidlinger
Associates Ltd. The aim of this work was
to extend the solver capabilities, and in particular
the parallel solver capabilities, of an industrial modelling
code, PZFlex. Two key solvers were compared,
a preconditioned conjugate gradient (PCG) iterative
solver and a state-of-the-art direct solver (MFACT).
The PCG solver was parallelised using OpenMP, the
new standard for compiler directive based programming
of shared memory multiprocessor systems. Our
results show that OpenMP delivers parallel efficiencies
of up to 80% for the PCG solver. It also offers
good scalability, allowing the simple PCG algorithm
to beat the complex MFACT direct solver for typical
simulation runs of PZFlex.