A Project-Based HPC Course for Single-Box Computers

Throughout three iterations and six years we have developed a project-based course in HPC for single-box computers tailored to science students in general. The course is based on strong premises: showing that assembly is what actually runs on machines, dividing parallelism in three dimensions (ILP, DLP, TLP), and using them incrementally in a single numerical simulation throughout the course working in interdisciplinary pairs (CS, non-CS). The final goal is to explore how to use all the available transistors in a die. Assembly proved a great tool to show how bare-metal works, an alternative-semantics approach to programs, and a tool to demystify compiler technology. Parallelism is tackled gradually with a clear division into instruction, data, and thread parallelism. GPUs, through CUDA in particular, are used as a radically different approach to the three dimensions of parallelism. Each dimension is explored in a gradual manner, starting from a sequential toy-yet-interesting numerical simulation. After using each form of parallelism and submitting a short report, the experiences are put together in group discussion unveiling the strengths and weaknesses of each form of parallelism for each class of numerical simulation. Although there is a high variance in the students' background, CS and non-CS students pair well in project development, generating understanding and value of the disciplines. The experience proved successful, with former students producing parallel accelerated code of their own in their disciplines.