An evaluation of GPUs for use in an upgraded ATLAS High Level Trigger

ATLAS is a general purpose particle physics experiment located on the LHC collider at CERN. The ATLAS Trigger system consists of two levels, the first level (L1) implemented in hardware and the High Level Trigger (HLT) implemented in software running on a computing cluster of commodity CPUs. The HLT reduces the trigger rate from the 100 kHz L1 accept rate to 1 kHz for recording, requiring an average per-event processing time of ∼300 ms for this task. The HLT selection is based on reconstructing tracks in the Inner Detector and Muon Spectrometer and clusters of energy deposited in the calorimeters (electromagnetic and hadronic). Performing this reconstruction within the available HLT computing cluster resources presents a significant challenge. Future HLT upgrades will result in higher detector occupancies and, consequently, will harden the reconstruction constraints. General purpose Graphics Processor Units (GPGPU) are being evaluated for possible future inclusion in an upgraded HLT computing cluster. We report on a demonstrator that has been developed consisting of GPGPU implementations of the calorimeters clustering and Inner Detector and Muon tracking algorithms integrated within the HLT software framework. We give a brief overview of the algorithm implementation and present preliminary measurements comparing the performance of the GPGPU algorithms with the current CPU versions.