GTC 2020: Running Multi-Messenger Astrophysics with IceCube Across All Available GPUs in the Cloud

GTC 2020 S22206
Presenters: Igor Sfiligoi ,UC San Diego - San Diego Supercomputer Center; Benedikt Riedel,University of Wisconsin-Madison and Wisconsin IceCube Particle Astrophysics Center
Abstract
We’ll report on a computational experiment that marshaled all globally available for-sale NVIDIA GPUs across AWS, Azure, and GCP. The net result was a peak of about 51,000 GPUs of eight different kinds, with an aggregate peak of about 380 PFLOPS fp32.
The experiment used simulations for the IceCube Neutrino Observatory, an array of some 5,000 optical sensors buried deep within a cubic kilometer of ice at the South Pole. The sensors detect the signatures of shock waves created by particles from neutrino interactions passing through the ancient ice sheets. Simulation is needed to properly account for natural ice imperfections, and photon propagation codes are a natural fit for GPGPU computing. We’ll provide both a summary overview and technical details of the infrastructure needed to create a supercomputer-like environment across multiple cloud providers, as well as an overview of the science behind IceCube and how GPU compute helps in advancing the scientific goals.

Watch this session
Join in the conversation below.