Preparing Students to Master Hybrid and Co-Processing Methods for High Performance Computing

Abstract Students learning methods of parallel programming have a unique opportunity to develop a distinguished career in scalable scientific computing by learning scale-up, scale-out, and co-processing computer architecture. Teaching a course that covers traditional parallel programming methods for scalable high-performance computing has included scale-up shared memory methods such as OpenMP, scale-out distributed memory methods such as MPI (Message Passing Interface) and more recently co-processing methods such as CUDA (Compute Unified Device Architecture). Learning these three major methods is challenging for senior year undergraduate or first year graduate students, but now, quantum computing, and specifically quantum co-processing has emerged as another challenge for future high-performance computing software developers. It is accepted that quantum computing will show advantages in specific areas such as cryptanalysis, optimization, and quantum simulation, but will not soon replace traditional digital logic high performance computing anytime soon, if ever. Instead, much like a GP-GPU (General Purpose Graphics Processing Unit) acts as a co-processor for specific parallel work off-load, so will quantum computers, providing a QPU (Quantum Processing Unit). In this paper the methods for helping students deal with the triple challenge of learning parallel programming, writing correct code for any scale, and verifying scalability are presented along with new methods to incorporate quantum computing as another hybrid option. The paper provides details of how the triple challenge can be extended to include the fourth challenge of advanced co-processing methods. Techniques used are problem-based learning for mastery and contract-based projects where students demonstrate their achievement of key learning objectives.