Accelerated Computing and its Implications for Computational Astrophysics

In the paradigm of accelerated computing, CPUs, GPUs, and NICs have to be used together to solve computational problems. The density of compute power on a modern server node means that developers must think carefully about how to program their applications so that they do not hit strong scaling limits too early. Additionally, the complexity and diversity of the hardware available on HPC clusters means that there are a number of approachers developers can take to write code that works on today's and tomorrow's clusters. In this presentation I will discuss what some developers have been doing in computational astrophysics and related fields to achieve performance portability (or something approximating it), and what might be different in the future to make it easier. I will specifically discuss what needs to change about code originally written for CPUs when it is adapted for today's heterogeneous systems. Although these changes can be time-consuming and challenge some favored software paradigms (and, in some cases, even some favored programming languages), I will argue that the benefits of this adaptation are clearly worth it.