Adjoining Hybrid Parallel Code
Michel Schanen, M. Foerster, Johannes Lotz, Klaus Leppkes, Uwe Naumann · Civil-comp proceedings · 2012
OpenMP and MPI are typically used in combination to implement distributed shared memory parallelization schemes for numerical simulations on hybrid systems.The performance of said numerical simulations is frequently linked to the efficient and accurate computation of derivatives.One method of acquiring this derivative information for a given code is algorithmic differentiation (AD).AD implementations are subdivided into either source transformation or overloading tools.We investigate a generic approach of applying AD tools on hybrid parallel codes.A dense matrixmatrix multiplication serves as a case study.