Performance Portable science on the world’s largest supercomputers

Tom Deakin

This project will look at how to get a scientific simulation application running efficiently across the diverse range of supercomputing processors. Supercomputers connect together many thousands of processors in order to reach enormous levels of performance. However, the landscape is diverse, with no one solution leading. As such, the community has to be able to efficiency run their applications across a range of advanced and highly parallel CPUs and GPUs from many different vendors, including Intel, AMD, NVIDIA and Fujistu/Arm, among others. Achieving a similar level of high performance across different processors is known as Performance Portability, but this is a significant challenge.

A PhD in this area will look at how to measure performance portability and work with a range of partners in achieving performance portability in different scientific domains. Avenues of research include how best to write applications for a multi-accelerator nodes at scale, the effectiveness of the latest parallel programming models towards performance portability, how to optimise accelerator-to-accelerator communication, how the performance bounds might change depending on scale/architecture, etc.