Performance Evaluation of Unstructured Mesh Physics on Advanced Architectures

Unstructured mesh physics codes tend to exhibit different performance characteristics than other types of codes such as structured mesh or particle codes, due to their heavy use of indirection arrays and their irregular memory access patterns. For this reason unstructured mesh mini-apps are needed, alongside other types of mini-apps, to evaluate new architectures and hardware features. This paper uses one such mini-app, PENNANT, to investigate performance trends on architectures such as the Intel Xeon Phi, IBM BlueGene/Q, and NVIDIA K40 GPU. We present basic results comparing the performance of these platforms to each other and to traditional multicore CPUs. We also study the usefulness for unstructured codes of various hardware features such as hardware threading, advanced vector instructions, and fast atomic operations.