The FEniCS Project on AWS Graviton3

We show initial performance results executing the FEniCS Project finite element software on Amazon Web Services (AWS) c7g instances with Graviton3 processors. Graviton3 processors are based on the ARM64 instruction set and provide Scalable Vector Extensions (SVE) for single instruction, multiple data (SIMD) operations. The c7g instances include a fast Elastic Fabric Adaptor (EFA) interconnect for low-latency high-bandwidth Message Passing Interface (MPI) based parallel communication. Comparing clang 15 and GCC 12 series compilers for compiling a high-order elasticity finite element kernel our results show that GCC emitted more vectorised loops with variable width SVE instructions than clang. The runtime performance of the GCC compiled kernel was 20% faster than the clang compiled kernel. We also tested multi-node weak scalability of a Poisson solver on the EFA interconnect up to 512 MPI processes. We find that overall performance and weak scalability of the AWS provisioned cluster is similar to a dedicated AMD EPYC x86-64 HPC installed at the University of Luxembourg.