On the acceleration of stencil operations in the data-parallel solution of PDEs

Abstract

We propose some non-standard, yet straightforward, and highly efficacious alternative modes of data assignment which induce a significant reduction in communication volume and hence in execution time for stencil operations, i.e. local iterative updates, implemented within a data-parallel programming environment. Performance results obtained in the solution of two three-dimensional elliptic partial differential equations (PDEs) using iterative methods entailing such updates indicate that substantial performance increases can be realized using these alternative data assignment schemes.<>