Optimization and validation of Geant4 detector simulation software for the ATLAS experiment at the LHC

Abstract: Simulations of large detectors such as the ATLAS detector at the LHC (Large Hadron Collider) are compute-intensive projects, and are according to current internal estimates expected to be consuming 40 % of the total ATLAS CPU resources in 2020, creating a need for software optimization strategies. We investigate build configurations of the Geant4 simulation software to find if a) a performance improvement can be made by switching build configuration and b) if such build configuration switches produce the same physics results. We performed measurements of CPU time per event and energy deposited per event using both the ATLAS detector geometry and the CMS detector geometry on two types of hardware. We find potential performance improvements using static linking of the Geant4 libraries of up to ∼19% using unsafe build options and up to ∼10 % using safe build options. We observe three sources of differences in the average energy deposition per event of order ∼0.1 %. These sources include the expected unsafe optimization methods but also differences between hardware platforms and from changing the compiler that is used to build the simulation.