Accelerating the ATLAS Trigger system with Graphical Processing Units
PT-CERN Call 2021/2
Patricia Conde Muino
Instituto Superior Técnico (Universidade de Lisboa)
Laboratório de Instrumentação e Física Experimental de Partículas
The LHC is the highest energy particle accelerator ever built. The gigantic ATLAS experiment records proton and ion collisions produced by the LHC to study the most fundamental matter particles and the forces between them. A major upgrade, expected for the years 2025-27, will increase the LHC collision rate up to a factor 7 with respect to the nominal values, to allow acquiring a huge amount of data and pushing the limits of our understanding of Nature. The online event selection system (trigger) is a crucial part of the experiment. It analyses in real time, the 40 MHz event rate, selecting only the potentially interesting collisions for later analysis. After the LHC upgrade, the estimated increase in collision rate, and consequently event size, will lead to much longer event reconstruction times, that are not matched by the slower expected growth in computing power at fixed cost. This implies a change in paradigm, increasing parallelism and/or using hardware accelerators, such as GPUs or FPGAs. The study of hardware accelerators is also interesting for the offline ATLAS reconstruction software, given the availability of High Performance Computers. The LIP Portuguese team was responsible for the calorimeter reconstruction algorithms of the first ATLAS Trigger GPU prototype. The results obtained showed the potential gain but also the limitations of the architecture and implementation done. The objective of this PhD thesis is to contribute to the design, implementation and optimisation of trigger and offline reconstruction algorithms using GPUs as hardware accelerators. In particular, the project focuses on the ATLAS calorimeter clustering algorithm TopoCluster and its GPU counterpart, the Topo-Automaton Clustering.