Parallel mapping of program graphs into parallel computers by self-organization algorithm
The optimization problem of mapping program graphs to parallel distributed memory computers is considered. An algorithm which is based on the self-organizing neural networks is proposed. We tried to apply the ability of Kohonen neural networks to compute a neighbourhood preserving mapping to complicated topologies of program and processor graphs. The goal of the algorithm is to produce an allocation with the minimal communication cost and the computational load balance of processors.
Unable to display preview. Download preview PDF.