Ferraris can't swim: A Case against Performance Portability (...for now)

While parallelism remains the main source of performance at extreme scales, architectural implementations and programming models change with each new hardware generation, often leading to costly application re-engineering. Most tools for performance portability require manual and costly application porting to yet another supercomputing architecture and/or programming model. We and others have developed such tools, as will be showcased in this talk. However, the main bottlenecks, which prevent performance portability at scales, aren't to be found in the lack of tools, but rather in other aspects of our HPC field. Mental barriers, physical barriers, and memory barriers will need to be broken down before a truly portable approach can be established. This talk will highlight some of the often overlooked aspects of performance portability. Whether the dream becomes reality depends on our visionary, aspirational, and young researchers, and knowing these hard and soft limits will help them in avoiding common pitfalls.