high performance computing on graphics processing units: hgpu.org

Scheduling applications tasks across heterogeneous clusters is a growing problem, particularly when new upgraded components are added to a parallel computing system that may have originally been homogeneous. We describe how automatic and just-in-time source code generation techniques can be used to make the best parallel decomposition for whatever resource is available in a heterogeneous system consisting of graphical processing unit accelerators and multi-cored conventional CPUs. We show how a high level domain specific language approach to our set of target simulation applications can be used to cater for a variety of different GPU and CPU models and scheduling circumstances. We present some performance and resource utilisation data illustrating the scheduling issue for heterogeneous systems in computational science. We discuss the future outlook for this code generation approach in software engineering.