high performance computing on graphics processing units: hgpu.org

Increasing heterogeneity among HPC platforms requires applications to be frequently ported and tuned, adding burden to developers. Fast evolution of hardware mandates adaptation of algorithms and data structures to get higher performance, while application complexity constantly grows accordingly. Ensuring portability while preserving high performance at large scale along with minimal changes to an already existing application is an actual challenge. Separation of concerns to decouple performance from semantics in simulation codes are typically required. We describe a specialized programming framework for N-Body simulations that provides such separation. It allows one to develop computation kernels in the form of sequential-looking functions, while self-generating multi-level parallelism. exaNBody possesses both an application layer with its own input data format, a way to define specific computation kernels and a separate runtime system that can address both CPUs and GPUs. The framework enables performance portability for N-Body simulations, bringing both flexibility and a set of handy tools. Performance results and speedups up to 32k cores with two distinct applications based on exaNBody are discussed.