ABSTRACT
Hans P. Zima Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA Institute of Computational Science, University of Vienna, Austria
Efficient management of locality is a key requirement for today’s high performance computing (HPC) systems, most of which have a physically distributed memory. The standard programming paradigm for these systems has been based for more than a decade on the extension of sequential programming languages with message-passing libraries, in a processor-centric model for programming and execution. It is commonly understood that this approach leads to complex and error-prone programs, due to the way in which algorithms and communication are inextricably interwoven.
