Offline Primary Vertex Reconstruction for Heterogeneous Architectures

As the Large Hadron Collider (LHC) enters its future development phase, one of the key objectives is to increase its nominal luminosity, with the ultimate aim of achieving peak luminosities that would lead to up to 200 simultaneous proton collisions (pileup), presenting substantial hurdles for the CMS detector reconstruction process. The process of reconstructing the primary vertex (PV) in CMS involves two steps, namely vertex finding and fitting. Initially, the tracks originating from the same interaction vertex are clustered using the Deterministic Annealing algorithm. Subsequently, the Adaptive Vertex Fit is utilized to determine the most accurate estimate of the vertex position in either three or four dimensions. Due to the challenging conditions of High Luminosity LHC (HL-LHC), the PV reconstruction process can be extremely time-consuming, accounting for up to 6\% of the overall reconstruction time. To address this issue, this study explores the feasibility of rethinking and adapting the PV reconstruction algorithms for heterogeneous architectures, taking advantage of parallelization techniques to reduce processing time significantly. The results of this study demonstrate improvements in both computing and physics performance, particularly for HL-LHC conditions.