A Workload Partitioning Strategy for PDEs by a Generalized Neural Network

Abstract

We consider the partitioning of a workload defined over a discrete geometrical data Sb"Ucture in a way that balances it across multiple processors while minimizing the communication/synchronization among them.We fannulate this problem in the contex.t of the numerical solution of partial differential equations in distributed multiprocessor hardware environments and we explore a neural network approach for determining its solution.Speci lically we are developing four neural network models for the corresponding geometric graph partitioning problem.examine the optimality of the obtained solution and argue about their suitability in solving these types of problems.