The ND Framework - Towards Grid-Based HPC Numerical Discretisation Schemes in Rust

This invited talk by Timo Betcke of University College London, developed jointly with Matthew Scroggs, presents the nd framework, an effort to build grid-based numerical discretization software for HPC in Rust, funded through UK ExCALIBUR exascale grants. The motivation came from limits of a Python/OpenCL boundary integral equation code, which could not scale to large virus simulations, prompting a rewrite in Rust rather than C++. Betcke surveys the maturing Rust numerical ecosystem, including nalgebra, ndarray, the Burn machine learning framework, and the rsmpi MPI bindings, highlighting procedural macros that manipulate the abstract syntax tree at compile time to generate MPI types and stack-allocated arrays. He then describes the group's libraries: rlst, a linear algebra toolbox with dense, sparse, and distributed operators built on a trait-based operator algebra; ndelement, a finite element definition and tabulation library mirroring the DefElement reference; and ndmesh, a distributed unstructured grid library. Benchmark runs on the UK's Archer2 system evaluated boundary integral operators on meshes up to 134 million elements, and the talk emphasizes that deploying Rust on HPC systems is essentially no harder than C or C++. Future work includes function space and full FEM/BEM discretization libraries and evaluation of CubeCL for GPU compute, since the next UK national system will be GPU-based. Betcke concludes that Rust is not easier to learn than C++ but yields near-Python productivity through compile-time safety and freedom from threading bugs.