Development of a computing farm with Cloud Computing on GPU

All the work in this thesis has been developed in the context of the Cherenkov Telescope Array (CTA), which is going to be the major next-generation observatory for ground-based very-high-energy gamma-ray astronomy. The plan for this work is to use GPUs and Cloud Computing in order to speed up the computing demanding tasks, developing and optimizing data analysis pipelines. The thesis consists on two main parts: the first one is dedicated to the estimation of the future performances of CTA towards the observation of violent phenomena such as those generating Gamma Ray Bursts and Gravitational Waves, with a initial work done for the creation of the models for the First CTA Data Challenge. The second part of the thesis is related to the development of the pipelines for the reconstruction of the low-level data coming from the Monte Carlo simulations using the software library called ctapipe. In chapter 1 I go into the details of the CTA project, the telescopes and the performances of the array, together with the methods used to derive them from Monte Carlo simulations. The science goals of CTA and the Key Science Projects (KSPs) will be covered in chapter 2, with a focus on Gamma Ray Bursts and the follow-up of Gravitational Waves events. The work done for the First CTA Data Challenge (DC-1) is presented in chapter 3. More than 500 extragalactic sources have been modelled combining information from different catalogues in order to create a population of AGNs. This Challenge has been important both to involve more people in the analysis of CTA data and to compute the observation time needed by the different KSP. The simulations for the gravitational waves and gamma-ray bursts Consortium papers have been created with the ctools-pipe pipeline (presented in chapter 4), implemented around the libraries ctools and gammalib. The pipeline is composed of two main parts: the task to be executed (background simulation, model creation and detection) and in which computing centre. The second part of the thesis is focused on the development and optimization of the analysis pipelines to be used for the event reconstruction of simulated raw data and for the visualization of the events in a 3D space. This analyses have been performed using ctapipe, a framework for prototyping the low-level data processing algorithms for CTA. The structure of the library is presented in chapter 5 together a focus on the reconstruction methods that are implemented in ctapipe, including the so called ImPACT. This method uses a template of images created from the Monte Carlo simulations and a seed from the standard reconstruction method to fit between the templates to find a better estimation of the shower parameters. The time profiling and the strategies adopted to optimize the ImPACT pipeline are presented in chapter 6. The implementation of the a pipeline for the analysis of the Large Size Telescope observing in monoscopic mode and its GPU implementation with PyTorch is also presented. ctapipe has also been used and developed to estimate the performances of CTA when observing using the 'divergent pointing' mode, in which the pointing directions are slightly different with respect to the parallel pointing mode, so that the final hyper field-of-view of all the telescopes is larger with respect to the parallel pointing mode. The angular and energy resolutions and also the sensitivity are worse in this scenario, but having a wider hyper field-of-view can be good for other topics, such are searching for transient sources. The modifications to the reconstruction code introduced in ctapipe and some angular resolution plots for the simulated point source gammas are presented in chapter 7. The results presented in this thesis are a demonstration of the usage of advanced software techniques in very high energy astrophysics. (author)