The ATLAS experiment will undergo a series of upgrades in association with the High-Luminosity LHC program. Given the new high-luminosity conditions and the predicted increase in event rates at the ATLAS High-Level Trigger by a factor of 10, additional computational load will be placed on the trigger farm. One possibility to accommodate this is the use of hardware accelerators, such as GPUs, for the cost and energy efficiency they offer. Among the algorithms being assessed for GPU acceleration is Topological Clustering, the main and most computationally demanding stage of calorimeter reconstruction. A more GPU-friendly variant of the algorithm, dubbed Topo- Automaton Clustering, has been implemented, reaching the significant milestone of 100% agreement with the CPU algorithm and maximum speed-ups in excess of a factor of 10. A significant bottleneck remains in conversion between the data representation used within the GPU and the equivalent CPU data structures, which can consume up to two thirds of the total execution time of the algorithm. The development, optimization and integration of Topo-Automaton Clustering with the ATLAS trigger will be described, including the latest benchmarks and ongoing efforts to develop a framework for general description of GPU-friendly data structures to mitigate the current bottleneck.
Nuno dos Santos Fernandes (Wed,) studied this question.