Lamellar: Enabling Memory-Safe Distributed HPC in Rust

In this invited talk, a speaker from Sandia National Laboratories presents Lamellar, an HPC runtime for memory-safe distributed computing written in Rust. The talk opens with an introduction to Rust, framed around three pillars of reliability (memory and thread safety with compile-time guarantees), performance (no garbage collector, zero-cost abstractions), and productivity (the Cargo package manager and helpful compiler messages). The speaker notes that major companies attribute around 70 percent of vulnerabilities to memory-related issues that Rust addresses, and highlights emerging GPU support such as NVIDIA's cuTile RS, along with Rust's ownership and borrowing model. Lamellar is described as a distributed asynchronous tasking runtime that also exposes PGAS APIs. Its architecture builds on active messages (distributed async tasks or remote procedure calls), DARCs (distributed atomic reference counted objects that manage global data lifetimes), and high-level Lamellar arrays offering read-only, atomic, local-lock, global-lock, and unsafe variants with different safety guarantees. The runtime uses a Rust async executor thread pool and a network abstraction layer called Lamellae, supporting InfiniBand and emerging Slingshot backends via Libfabric and UCX. The speaker walks through code examples including a histogram application and explains the distinction between safe and unsafe Rust. Performance results cover bandwidth put/get curves, histogram, and single-element atomics, noting safe arrays incur active-message overhead while an upcoming version 0.8 improves single-element performance via NIC acceleration. Networking primitives were implemented in-house rather than using MPI.