| Resumo: | Algebraic reconstruction algorithms are iterative algorithms that are used in many area including medicine, seismology or meteorology. These algorithms are known to be highly computational intensive. This may be especially troublesome for real-time applications or when processed by conventional low-cost personnel computers. One of these real time applications is the reconstruction of water vapor images from Global Navigation Satellite System (GNSS) observations. The parallelization of algebraic reconstruction algorithms has the potential to diminish signi cantly the required resources permitting to obtain valid solutions in time to be used for nowcasting and forecasting weather models. The main objective of this dissertation was to present and analyse diverse shared memory libraries and techniques in CPU and GPU for algebraic reconstruction algorithms. It was concluded that the parallelization compensates over sequential implementations. Overall the GPU implementations were found to be only slightly faster than the CPU implementations, depending on the size of the problem being studied. A secondary objective was to develop a software to perform the GNSS water vapor reconstruction using the implemented parallel algorithms. This software has been developed with success and diverse tests were made namely with synthetic and real data, the preliminary results shown to be satisfactory. This dissertation was written in the Space & Earth Geodetic Analysis Laboratory (SEGAL) and was carried out in the framework of the Structure of Moist convection in high-resolution GNSS observations and models (SMOG) (PTDC/CTE-ATM/119922/2010) project funded by FCT.
|
|---|