Graph Neural Network-based Track finding as a Service with ACTS

Recent progress in track finding for the High-Luminosity Large Hadron Collider (HL-LHC) has demonstrated the effectiveness of Graph Neural Network (GNN)-based algorithms. While these algorithms offer high efficiency and reasonable resolutions, their computational demands on CPUs hinder real-time processing, requiring accelerators like GPUs. However, the large size of the involved graphs poses a challenge for facilities lacking high-end GPUs.

To address this, we propose deploying the GNN-based track-finding algorithm as a service in the cloud or high-performance computing centers such as the NERSC Perlmutter system with over 7000 A100 GPUs. We have implemented a tracking-as-a-service prototype within A Common Tracking Software (ACTS), a toolkit for charged particle track reconstruction.

This approach is algorithm-agnostic, allowing the incorporation of various algorithms as new backends through interactions with the client interface in ACTS. In this contribution, we showcase the versatility of the as-a-service approach by implementing the GNN-based track-finding workflow using the Nvidia Triton Inference Server within ACTS. We assess track-finding throughput and GPU utilization, exploring the scalability of the inference server across the NERSC Perlmutter supercomputer and cloud resources.