PGAS-based Parallel Branch-and-Bound Algorithms for Ultra-Scale Supercomputers

in Talks

Speaker

Guillaume Helbecque

When

April 24th, 2026 at 3pm

Where

Alan Turing room

Title

PGAS-based Parallel Branch-and-Bound Algorithms for Ultra-Scale Supercomputers

Abstract

Branch-and-Bound algorithms are widely used for the exact solution of combinatorial optimization problems, but their parallelization remains challenging due to the dynamic generation of large and highly irregular search trees. In the exascale era, supercomputers are composed of heterogeneous nodes combining multi-core CPUs and GPU accelerators, leading to complex multi-level parallelism (intra-node, GPU, and inter-node). In this context, this talk presents the PGAS (Partitioned Global Address Space) approach as an alternative to MPI+X, offering a unified and high-level abstraction of parallelism. PGAS languages, such as Chapel, also enhance code portability across heterogeneous architectures by abstracting communication details and reducing the need for hardware-specific adaptations.

Keywords: Combinatorial Optimization, Parallel Computing, Branch-and-Bound, GPU computing, PGAS, Chapel