high performance computing on graphics processing units: hgpu.org

The demise of frequency scaling, which is the easiest way to improve computing performance, in addition to the growing gap between CPU and memory speeds and the increase in arithmetic intensity in current problems, has given rise to a new range of devices created to improve performance. Heterogeneous Computing (HC), and many-cores are examples of this new range of devices. However, the complexity of these new hardware architectures is not easily hidden from the programmer. In this thesis, I propose a set of tools that seek to exploit (through source-to-source (S2S) compilers) the capabilities and peculiarities of parallel computing and HC to speed up and increase the energy efficiency of originally sequential source code. The proposed modular programs are implemented as a set of tools that help port sequential source code to OpenMP, MPI, and HMPP, demonstrating how the input code can effectively automatically be translated. Through a real-life example, I show how the proposed dependency analysis tool trivializes the task of parallelizing sequential code, breaking the first performance barrier. The OMP2MPI experiments generate code that is more than 60x faster than its sequential version and also faster than its original OpenMP code. The OMP2HMPP experiments obtain an average speedup of 31x and average increase in energy efficiency of 5.86x. Both tools were tested with OpenMP, obtaining successful results that demonstrate the feasibility of using this set of tools for exploring HC.