Heterogeneous clusters with nodes containing one or more accelerators, such as GPUs, have become common. While MPI provides a mechanism and management of interaddress space communication, and OpenCL provides a way to manage computation and communication within a process with access to heterogeneous computational resources, programmers are forced to write hybrid programs that manage the interaction of both of these systems. This paper describes an array programming interface that provides users with automatic or manual distributions of data and work. Using the distribution and information about what data is used and defined by kernels, communication among processes and among devices in a process is performed automatically. The interface provides a unified programming model to the user, thus simplifying program development.
Parallel Programming Model, Distributed Shared Memory, Heterogeneous Systems, MPI, OpenCL
Date of this Version