Lightning Talk: MPIWasm: Executing WebAssembly on HPC Systems – Mohak Chadha, Technical University of Munich
This talk will describe MPIWasm in detail, present experimental results, and motivate the development of other Wasm runtimes for executing parallel HPC applications.
MPIWasm is a Wasm runtime that enables the execution of applications that utilize the Message Passing Interface (MPI) standard on HPC systems. It builds on Wasmer and supports the execution of MPI-based HPC application modules on both x86_64 and aarch64 processor architectures. To facilitate its adoption and suitability in HPC environments, it supports the high-performance execution of HPC application modules and has low overhead for MPI calls through zero-copy memory operations. The former is achieved by leveraging the LLVM compiler for translating Wasm instructions to native machine code AoT, while the latter is achieved by transparently translating between the host and the Wasm module’s linear memory address space. In addition, MPIWasm supports high-performance networking interconnects such as Infiniband that are used by MPI libraries for inter-rank communication by linking against the host MPI library at runtime and providing a translation layer between the Wasm modules and the host library. Our experiments on a production HPC system demonstrate that MPIWasm delivers competitive native application performance.
