SVM Support in the Vienna Fortran Compilation System

Vienna Fortran, a machine-independent language extension to Fortran which allows the user to write programs for distributed-memory systems using global addresses, provides the forall loop construct for specifying irregular computations that do not cause inter-iteration dependences. Compilers for distributed-memory systems cannot generate efficient code for such loops using compile-time techniques only. This paper presents the runtime message passing implementation and the SVM implementation. We compare both approaches and present first performance results for a realistic loop. Keywords: distributed-memory systems, shared virtual memory systems, compile-time optimization, runtime analysis, Fortran language extensions 1 Introduction Massively parallel computers (MPP) offer an immense peak performance. Current architectures consist of hundreds of nodes with physically distributed memory and are either pure distributed memory systems (Paragon), hardware-supported shared memory systems (K..