PDSEC 2015 Keynote - Domain-specific approaches in scientific computing

Summary form only given. Parallel programming with low-level interfaces has been the most viable choice in scientific computing for a long time. In such models, different parallelisms require different parallel programming interfaces, e.g., message passing for parallelism across nodes, threading for intranode parallelism, and vector processing for SIMD and GPUs. Often applications are confronted with these multiple interfaces to fully exploit the current and future large-scale machines. We present our work toward higher-level programming models, allowing for a single program to run on different parallel platforms without much human intervention, and at the same time to achieve close to hand-tuned performance.