GPU-Accelerated Machine Learning Inference as a Service for Computing in Neutrino Experiments.

Area of Science:

• High-energy physics
• Computational science
• Particle physics

Background:

• Machine learning (ML) algorithms are crucial for event reconstruction in accelerator-based neutrino experiments.
• Increasing data volumes and computationally expensive ML models present significant computing challenges.
• Efficient processing of billions of neutrino events requires innovative computational solutions.

Purpose of the Study:

• To explore a novel computing model using heterogeneous computing with GPU coprocessors as a web service.
• To integrate GPU acceleration into the ProtoDUNE-SP reconstruction chain without disrupting existing workflows.
• To address the computational demands of large-scale neutrino experiments.

Main Methods:

• Developed Services for Optimized Network Inference on Coprocessors (SONIC) to provide elastic GPU resources.
• Integrated GPU acceleration into the ProtoDUNE-SP event reconstruction workflow.
• Focused acceleration on time-consuming tasks like track and particle shower hit identification.

Main Results:

• Achieved a 17x speedup for track and particle shower hit identification using GPU acceleration.
• Reduced overall processing time by a factor of 2.7 compared to CPU-only processing.
• Demonstrated a cost-effective solution with a ratio of 1 GPU per 68 CPU threads for the accelerated task.

Conclusions:

• Heterogeneous computing with GPU coprocessors offered as a web service (SONIC) is an effective solution for ML-intensive neutrino experiments.
• The SONIC framework significantly enhances computational efficiency and reduces processing time.
• This approach offers a scalable and cost-effective method for handling the growing data and computational needs of modern physics experiments.