Performance Optimization of the ATLAS Detector Simulation

In the thesis at hand the current performance of the ATLAS detector simulation, part of the Athena framework, is analyzed and possible optimizations are examined. For this purpose the event based sampling profiler VTune Amplifier by Intel is utilized. As the most important metric to measure improvements, the total execution time of the simulation of $t\bar{t}$ events is also considered. All efforts are focused on structural changes, which do not influence the simulation output and can be attributed to CPU specific issues, especially front end stalls and vectorization. The most promising change is the activation of profile guided optimization for Geant4, which is a critical external dependency of the simulation. Profile guided optimization gives an average improvement of $8.9\%$ and $10.0\%$ for the two considered cases at the cost of one additional compilation (instrumented binaries) and execution (training to obtain profiling data) at build time.