Accelerating Machine Learning Inference with GPUs in ProtoDUNE Data Processing.

Area of Science:

• High Energy Physics
• Computational Physics
• Data Science

Background:

• Neutrino physics experiments generate massive datasets requiring efficient processing.
• Event reconstruction is a computationally intensive task crucial for data analysis.
• Scalable computing resources are essential for handling current and future experimental demands.

Purpose of the Study:

• To evaluate the performance of a cloud-based GPU-accelerated inference server for neutrino data batch jobs.
• To compare the processing speed of GPU-accelerated algorithms against CPU-based methods.
• To identify potential bottlenecks and challenges in large-scale data reprocessing.

Main Methods:

• Utilized detector data from the ProtoDUNE experiment.
• Employed standard DUNE grid job submission tools for data reprocessing.
• Ran thousands of concurrent grid jobs, processing data with both GPU and CPU versions of the algorithm.
• Compared processing times between GPU and CPU implementations.

Main Results:

• A 100-GPU cloud-based server effectively met processing demands.
• The GPU version of the event processing algorithm was twice as fast as the CPU version on the newest CPUs.
• High data transfer rates during GPU runs posed network bandwidth challenges.

Conclusions:

• Cloud-based GPU acceleration offers a viable solution for speeding up neutrino data reconstruction.
• Network infrastructure must be considered to avoid bottlenecks during large-scale GPU data processing.
• Further optimization and distributed computing strategies are needed for future experiments.