Speaker
Description
In recent years, numerous Machine Learning–based algorithms have been developed within particle physics experiments to accelerate the reconstruction of complex detector objects, notably at CERN in the context of the HL-LHC and, for example, within the Belle II experiment. A significant fraction of these approaches relies on Deep Geometric Learning, and in particular on Graph Neural Networks (GNNs). These algorithms are typically integrated into end-to-end pipelines combining graph construction from detector data, GNN model inference, and dedicated post-processing steps for the final reconstruction of physics objects.
We present Module Map Graph (MMG), a generic high-performance library optimized for hybrid CPU and GPU architectures, providing a unified implementation of all stages of such reconstruction workflows. MMG delivers specialized algorithms implemented as dedicated kernels, with a strong focus on memory efficiency through fixed per-event pre-allocation strategies and the use of stride-based data structures. MMG also leverages state-of-the-art inference frameworks (e.g., NVIDIA TensorRT) for the GNN step. The library exploits large-scale parallelism and supports fully asynchronous execution through CPU multithreading and CUDA streams, enabling excellent scalability on heterogeneous architectures. MMG follows modern software quality standards and provides Python interfaces as well as integration with experiment-independent track reconstruction toolkit ACTS [1].
We present the performance of MMG measured on benchmarks deployed on a realistic, production-like hybrid architecture setup representative of HL-LHC computing environments, and report a standardized set of metrics including latency, peak memory usage, energy consumption, and scalability. While these developments are motivated by the reconstruction of data from LHC experiments, MMG can straightforwardly be extended to other applications of GNNs in high-energy physics.
[1] The Acts project: track reconstruction software for HL-LHC and beyond, url: https://doi.org/10.1051/epjconf/202024510003