Heterogeneous CPU-GPU software framework for DAG's in high performance computing

Recent advancements in High Performance Computing (HPC) infrastructure with traditional computing systems augmented with accelerators like graphic processing units (GPUs) and coprocessors like Intel Xeon Phi have successfully enabled predictive simulations specifically Computational Fluid Dynamics (CFD) with more accuracy and speed. One of the most significant challenges in high-performance computing is to provide a software framework that can scale efficiently and minimize rewriting code to support diverse hardware configurations. Algorithms and framework support have been developed to deal with complexities and provide abstractions for a task to be compatible with various hardware targets. Software is written in C++ and represented as a Directed Acyclic Graph (DAG) with nodes that implement actual mathematical calculations. This thesis will present an improved approach for scheduling and execution of computational tasks within a heterogeneous CPU-GPU computing system insulting application developers with the inherent complexity in parallelism. The details will be presented within a context to facilitate the solution of partial differential equations on large clusters using graph theory.