Optimization of high-throughput real-time processes in physics reconstruction

The current thesis has been developed in collaboration between Universidad de Sevilla and the European Organization for Nuclear Research, CERN. The LHCb detector is one of four big detectors placed alongside the Large Hadron Collider, LHC. In LHCb, particles are collided at high energies in order to understand the difference between matter and antimatter. Due to the massive quantity of data generated by the detector, it is necessary to filter data in real-time. The filtering, also known as High Level Trigger, processes a throughput of 40~Tb/s of data and performs a selection of approximately 1\,000:1. The throughput is thus reduced to roughly 40~Gb/s of data output, which is then stored for posterior analysis. The High Level Trigger process is subdivided into two stages: High Level Trigger 1 (HLT1) and High Level Trigger 2 (HLT2). HLT1 occurs in real-time, and yields a reduction of data of approximately 30:1. HLT1 consists in a series of software processes that reconstruct particle collisions. The HLT1 reconstruction only analyzes the trajectories of particles produced at the collision, solving a problem known as track reconstruction, that determines whether the collision data is kept or discarded. In contrast, HLT2 is a finer process, which requires more time to execute and reconstructs all subdetectors composing LHCb. Towards 2020, the LHCb detector and all the components composing the data acquisition system will be upgraded. As part of the data acquisition system, the servers that process HLT1 and HLT2 will also be upgraded. In addition, the LHC accelerator will also be updated, increasing the data generated in every bunch crossing by roughly 5~times. Due to the accelerator and detector upgrades, the amount of data that the HLT will require to process is expected to increase by 40~times. The foreseen scalability of the software through 2020 underestimated the required resources to face the increase in data throughput. As a consequence, studies of all algorithms composing HLT1 and HLT2 and code modernizations were carried out, in order to obtain a better performance and increase the processing capability of the foreseen hardware resources in the upgrade. In this thesis, several algorithms of the LHCb recontruction are explored. The track reconstruction problem is analyzed in depth, and new algorithms are proposed. Since the analyzed problems are massively parallel, these algorithms are implemented in specialized languages for modern graphics cards (GPUs), due to their inherently parallel architecture. From this work stem two algorithm designs. Furthermore, four additional decoding algorithms and a clustering algorithms have been designed and implemented, which are also part of HLT1. Apart from that, a parallel Kalman filter algorithm has been designed and implemented, which can be used in both HLT stages. The developed algorithms satisfy the requirements of the LHCb collaboration for the LHCb upgrade. In order to execute the algorithms efficiently on GPUs, a software framework specialized for GPUs is developed, which allows executing GPU reconstruction sequences in parallel. Combining the developed algorithms with the framework, an execution sequence is completed as the foundations of a GPU HLT1. During the research carried out in this thesis, the aforementioned developments and a small group of collaborators co-coordinated by the author lead to the completion of a full GPU HLT1 sequence. The performance obtained on GPUs allows executing a reconstruction sequence in real-time, under LHCb upgrade conditions. The developed GPU HLT1 constitutes the first GPU high level trigger ever developed for an LHC experiment. Finally, various possible realizations of the GPU HLT1 to integrate in a production GPU-equipped data acquisition system are detailed.