Parallel computing in bound constrained quadratic programming (Q2718097)

The purpose of this paper is to develop an efficient parallel mathematical software for solving large-scale convex constrained quadratic programming problems on high performance computers. The authors compare the parallelism of active set and interior point strategies, more precisely, of the projected gradient and potential reduction algorithm, respectively. They outline the key computational kernels arising in the implementation of the algorithms under consideration and elaborate parallelizations for the computational algebra kernels. Thereafter some computational results of the implementation of the parallel algorithms on a distributed memory computer are presented. On the basis of the computational results they establish that on sequential computers active set strategies can be competetive with interior point algorithms but on multiprocessors the potential reduction algorithm is the most attractive.