Reconstruction of Low Energy Neutrino Events with GPUs at IceCube

Abstract

IceCube is a cubic kilometer neutrino observatory located at the South Pole that produces massive amounts of data by measuring individual Cherenkov photons from neutrino interaction events in the energy range from few GeV to several PeV. The actual reconstruction of neutrino events in the GeV range is computationally challenging due to the scarcity of data produced by single events. This can lead to run times of several weeks for the state-of-the-art reconstruction method – Pegleg – on CPUs for typical workloads of many ten-thousand events. We propose a GPU version of Pegleg that probes the likelihood space with several hypotheses in parallel while adapting the amount of parallel sampled hypotheses dynamically in order to reduce computation time significantly. Our results show an average speedup of 14 (with a maximum of over 200) for 5262 reconstructed neutrino events of different flavors on a Titan V GPU compared to the multithreaded CPU version, which enables quicker and broader analysis of IceCube events.