How to connect the dots very fast? High-Performance Heterogeneous Particle Track Reconstruction for the ATLAS Phase-II High Level Trigger

The ATLAS experiment is a key project in high-energy particle physics exploring particle collisions at unprecedented energies to recreate early-universe conditions and probe phenomena that occurat extreme scales. The upcoming high luminosity upgrade of the LHC will significantly increase the collision rate and energy; the higher data volume and collision complexity necessitate a major upgrade of the ATLAS detector to consist of a more granular tracking system. Due to much higher data rate and processing complexity, it is crucial to optimise the collision selection system (trigger system) and its most computationally expensive component - the track reconstruction algorithms. This work explores advanced optimisation methods, including graphics card acceleration and machine learning, to enhance computational efficiency and effectively manage the increased data throughput and complexity of the upgraded detector.

The first considered approach optimises the track reconstruction algorithm used in the ATLAS trigger system. By employing the track seeding on the graphic card accelerator and adjusting the track seed selection criteria, the final performance was improved by 95%, achieving an average processing time per event of 1.16 s. The performance was evaluated on different graphics cards, considering their limitations, with NVIDIA RTX 5000 Ada achieving the best results due to its exceptionally high number of processing cores.

The second part of this work focuses on the application of machine learning techniques to particle track reconstruction. A novel Interaction Graph Neural Network (IGNN) demonstrates competitive reconstruction accuracy; however, it is known to be resource-consuming. To address these computational challenges, two optimisation strategies are proposed, aimed at reducing both memory consumption and inference time without compromising model performance.

The instantaneous memory footprint of the model was reduced by partial processing (substepping). Memory consumption can be decreased by approximately 30% without an increase in processing time. Further memory reductions are achievable by adjusting the size of partitions, enabling the deployment of the IGNN on memory-constrained GPUs and allowing parallel processing, depending on the available hardware resources.

The second discussed compression technique is structured pruning of IGNN, where by removing the least important groups of parameters, the model size is reduced. A selection of pruning techniques applied to Graph Neural Networks (GNN) was analysed to determine the most effective methodology for GNN compression. The final pruning configuration achieves up to 20% improvement in computational performance without compromising model accuracy. Furthermore, per-layer sensitivity was analysed and incorporated in the pruning strategy to guide layer-wise pruning aggressiveness, enabling further model size reduction by 20% while maintaining reconstruction accuracy. The pruning strategy was evaluated on the standard GNN benchmark models, demonstrating satisfying performance gains. The performance of IGNN was evaluated on different graphics cards, considering their limitations, with NVIDIA RTX A100 achieving the best results due to its highly efficient memory throughput.