Track reconstruction as a service for collider physics

Yuan-Tang Chou^1*, Haoran Zhao¹, Yao Yao⁴, Xiangyang Ju³, Yongbin Feng², William Patrick McCormack⁵, Miaoyuan Liu⁴, Jan-Frederik Schulte⁴, Kevin Pedro², Nhan Tran², Javier Duarte⁶, Philip Harris⁵, Shih-Chieh Hsu¹

¹Physics, University of Washington, Seattle, Washington, USA
²Fermi National Accelerator Lab, Illinois, USA
³Lawrence Berkeley National Lab, Berkeley, California, USA
⁴Physics, Purdue University, West Lafayette, IN, USA
⁵Physics, Massachusetts Institute of Technology, Cambridge, MA, USA
⁶Physics, University of California San Diego, San Diego, CA, USA

* Presenter:Yuan-Tang Chou, email:yuan-tang.chou@cern.ch

Tracking algorithms are vital in event reconstruction in Large Hadron Collider (LHC) experiments. However, tracking algorithms are also the most time-consuming component in the particle reconstruction chain. To reduce processing time, existing tracking algorithms have been adapted for massively parallel coprocessors, such as GPUs and FPGA. Nevertheless, challenges remain in fully utilizing the coprocessors' computational capacities. The talk will focus on the inference-as-a-service approach for tracking algorithms in LHC experiments and also discuss its application in astrophysics experiments. Two distinct benchmark tracking algorithms are used: Patatrack, a rule-based algorithm, and Exa.trkX, a Machine Learning algorithm. These implementations enhance GPU utilization and can process requests from multiple CPU cores concurrently without slowing each request down. Data transfer latency is observed to be minimal and insignificant compared to running on local coprocessors. This approach greatly improves the computational efficiency of charged particle tracking, solving the computing challenges anticipated in the High-Luminosity LHC era.

Keywords: LHC, Tracking, GPU, Machine Learning