HPX beats OpenMP by 26% in tiled Cholesky decomposition benchmark

A new research paper by Alexander Strack et al. benchmarks tiled Cholesky decomposition—a canonical irregular kernel—comparing four parallelization strategies across two shared-memory runtimes: OpenMP (GCC and LLVM) and the HPX asynchronous many-task (AMT) runtime. The variants span classical fork-join, collapsed fork-join (exposing more inner-loop parallelism), synchronous tasking, and asynchronous tasking with explicit data dependencies. The study uses their new Cholesky-Bench tool on a dual-socket 128-core AMD Zen 2 node, testing multiple tile and problem sizes to isolate the true performance impact of runtime design.

Key results show that HPX outperforms OpenMP at the optimal tile size by 15–30% across all variants. Specifically, asynchronous HPX tasks are up to 26% faster than equivalent OpenMP tasks, and HPX exhibits roughly 3.8x smaller task creation and scheduling overhead. Removing redundant synchronization barriers—an optimization possible with async runtimes—yields an additional improvement of 7% for OpenMP and 14% for HPX. The collapsed fork-join variants close most of the gap to synchronous tasking, but asynchronous tasking still wins. A GCC vs. LLVM comparison also reveals compiler-specific differences in fork-join scheduling and task-creation costs, underscoring that runtime and compiler choices matter for real-world HPC performance.