| ATLAS Slides | |
| Report number | ATL-SOFT-SLIDE-2024-499 |
| Title | Improving Computational Performance of ATLAS GNN Track Reconstruction Pipeline |
| Author(s) |
Lazar, Alina (Youngstown State University (US)) ; Vallier, Alexis (Centre National de la Recherche Scientifique (FR)) ; Collard, Christophe (Centre National de la Recherche Scientifique (FR)) ; Murnane, Daniel Thomas (University of Copenhagen (DK)) ; Torres, Heberth (Centre National de la Recherche Scientifique (FR)) ; Burzynski, Jackson Carl (Simon Fraser University (CA)) ; Stark, Jan (Centre National de la Recherche Scientifique (FR)) ; Burleson, Jared Dynes (Univ. Illinois at Urbana Champaign (US)) ; Chan, Jay (Lawrence Berkeley National Lab. (US)) ; Condren, Levi Harris Jaxon (University of California Irvine (US)) ; Neubauer, Mark (Univ. Illinois at Urbana Champaign (US)) ; Pham, Minh Tuan (University of Wisconsin Madison (US)) ; Calafiura, Paolo (Lawrence Berkeley National Lab. (US)) ; Liu, Ryan (Lawrence Berkeley National Lab. (US)) ; Caillou, Sylvain (Centre National de la Recherche Scientifique (FR)) ; Ju, Xiangyang (Lawrence Berkeley National Lab. (US)) |
| Corporate author(s) | The ATLAS collaboration |
| Submitted to | 27th International Conference on Computing in High Energy & Nuclear Physics, Kraków, Pl, 19 - 25 Oct 2024 |
| Submitted by | [email protected] on 28 Oct 2024 |
| Subject category | Particle Physics - Experiment |
| Accelerator/Facility, Experiment | CERN LHC ; ATLAS |
| Free keywords | Track Reconstruction ; Graph Neural Networks |
| Abstract | Track reconstruction is an essential element of modern and future collider experiments, including the ATLAS detector. The HL-LHC upgrade of the ATLAS detector brings an unprecedented tracking reconstruction challenge, both in terms of the large number of silicon hit cluster readouts and the throughput required for budget-constrained track reconstruction. Traditional track reconstruction techniques often contain steps that scale combinatorically, which could be ameliorated with deep learning approaches. The GNN4ITk project has been shown to apply geometric deep learning algorithms for tracking to a similar level of physics performance with traditional techniques while scaling sub-quadratically. In this contribution, we compare the computational performance of a variety of pipeline configurations and machine learning inference methods. These include heuristic-and-ML-based graph segmentation techniques, GPU-based module map graph construction, and studies of high throughput graph convolutional kernels. In this contribution, we present benchmarks of latency, throughput, memory usage, and power consumption of each pipeline configuration. |
Record created 2024-10-28, last modified 2024-10-30
PDF