Aachen 2019 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 59: Grid-Computing und Software

T 59.3: Vortrag

Mittwoch, 27. März 2019, 16:30–16:45, S11

Speeding Up Reconstruction of Low Energy Neutrino Events in IceCube with GPUs — •Maicon Hieronymus, Sebastian Böser, Bertil Schmidt, and Elisa Lohfink for the IceCube collaboration — Johannes Gutenberg-Universität, Mainz, Deutschland

The IceCube Neutrino Observatory can reconstruct neutrinos with just a few GeV of energy, even though only few hits per event are detected. The multi-dimensional parameter space where each event has an energy deposit at the vertex, potentially an energy and direction of the emerging muon track as well as the usual position and time parameters makes this a non-trivial problem. In addition, the light diffusion process can not be calculated ab-initio but has to be taken from simulation stored in so called photosplines as look-up tables. A maximum likelihood approach is used for the reconstruction. The vertex and the direction are optimized using MultiNest, a multi modal nested sampling algorithm that handles degenerated likelihoods. Track length and energy are determined separately. The likelihood is constructed from hit probabilities stored in spline tables. The overall process takes on average 10 mins per event on a CPU. The lookups in the spline tables are the bottleneck of the reconstruction. In this talk I present approaches using a CUDA-enabled GPU to significantly speed up the reconstruction. The focus lies on balanced workload, coalesced memory access and overlapping computation and communication, which might result in a speed-up of 15-20 over the CPU implementation for O(10⁴) likelihood evaluations.