A Parallel Hierarchical Approach for Automatic Differentiation
We evaluate in parallel first-order derivatives given a sequential computer program of the function to be differentiated. Our parallel implementation of an automatic differentiation (AD) algorithm is based on a hierarchical approach. The parallel method is developed by considering as a parallel computational model a shared-memory paradigm. The performance of the derivative codes is evaluated by considering a SGI Origin 2000 and by using the OPENMP standard library. In our computational experiments, we have considered the Flow in a Driven Cavity function belonging to the MINPACK-2 test problem collection. The computational results show the performance gain of the parallel approach over both the sequential one and the stripmining technique.
Unable to display preview. Download preview PDF.