Overview¶
High Energy Physics (HEP) is entering an exciting era for potential scientific discovery. There is now overwhelming evidence that the Standard Model (SM) of particle physics is incomplete. A targeted program, as recommended by the Particle Physics Project Prioritization Panel (P5), has been designed to reveal the nature and origin of the physics Beyond Standard Model (BSM). Two of the flagship projects are the upcoming high luminosity upgrade of the High Luminosity Large Hadron Collider (HL-LHC) and its four main detectors and the Deep Underground Neutrino Experiment (DUNE) at the Sanford Underground Research Facility (SURF) and Fermi National Accelerator Laboratory (Fermilab). Only by comparing these detector results to detailed Monte Carlo (MC) simulations can new physics be discovered. The quantity of simulated MC data must be many times that of the experimental data to reduce the influence of statistical effects and to study the detector response over a very large phase space of new phenomena. Additionally, the increased complexity, granularity, and readout rate of the detectors require the most accurate, and thus most compute intensive, physics models available. However, projections of the computing capacity available in the coming decade fall far short of the estimated capacity needed to fully analyze the data from the HL-LHC. The contribution to this estimate from MC full detector simulation is based on the performance of the current state-of-the-art and LHC baseline MC application Geant4, a threaded CPU-only code whose performance has stagnated with the deceleration of clock rates and core counts in conventional processors.
General-purpose accelerators offer far higher performance per watt than Central Processing Units (CPUs). Graphics Processing Units (GPUs) are the most common such devices and have become commodity hardware at the U.S. Department of Energy (DOE) Leadership Computing Facilities (LCFs) and other institutional-scale computing clusters. However, adapting scientific codes to run effectively on GPU hardware is nontrivial and results both from core algorithmic properties of the physics and from implementation choices over the history of an existing scientific code. The high sensitivity of GPUs to memory access patterns, thread divergence, and device occupancy makes effective adaptation of MC physics algorithms especially challenging.
Our objective is to advance and mature the new GPU-optimized code Celeritas 1 to run full-fidelity MC simulations of LHC detectors. The primary goal of the Celeritas project is to maximize utilization of HEP computing facilities and the DOE LCFs to extract the ever-so-subtle signs of new physics. It aims to reduce the computational demand of the HL-LHC to meet the available supply, using the advanced architectures that will form the backbone of high performance computing (HPC) over the next decade. Enabling HEP science at the HL-LHC will require MC detector simulation that executes the latest and best physics models and achieves high performance on accelerated hardware.
- 1
This documentation is generated from Celeritas *unknown release*.
