Some critical hpc improvements in numerical weather prediction workflows

• Autores: Raúl Moreno Galdón
• Directores de la Tesis: Francisco J. Tapiador Fuentes (dir. tes.) , Juan José Pardo Mateo (codir. tes.)
• Lectura: En la Universidad de Castilla-La Mancha ( España ) en 2019
• Idioma: español
• Tribunal Calificador de la Tesis: Maria del Carmen Carrion Espinosa (presid.) , Angelines Alberto Morillas (secret.) , Ismael Marín Carrión (voc.)
• Texto completo no disponible (Saber más ...)
• Resumen

  • Understanding the Earth System is of vital importance to many scientific fields and other civic issues.

    In particular, the understanding the atmosphere and its processes is very useful for forecasting weather, predicting climate change, or improving response to catastrophic weather events.

    Nowadays, the way to understand these processes is through the use of Numerical Weather Prediction (NWP) models that simulate the evolution of the atmosphere from initial conditions.

    These simulations can be very demanding from a computational point of view depending on the required resolution, which in many cases must be performed by a supercomputer.

    Therefore, these models implement High-Performance Computing (HPC) and parallel computing techniques that accelerate the computation through the use of platforms with multiple processors.

    Although NWP models use supercomputers, they need to be adjusted for optimal performance.

    This adjustment process can be very complex due to the different software and libraries involved in the process of configuring, compiling and executing HPC software.

    In this thesis some solutions are proposed to increase the performance of NWP models in HPC environments.

    An objective method based on an empirical score is developed to search for an optimal combination of compiler, communication libraries, hybrid computation and stability.

    Also, an algorithm has been developed for the optimal distribution of the processes in the computing nodes, together with a study on the balance between cache efficiency and communications overhead.

    A study of scalability and its effects on energy consumption has also been conducted, offering some ideas on how to balance energy and performance.

    The complexity of the setup process of NWP models prevents their use by a wider community, as only few users are able to deal with technical problems that might appear during the process.

    For this reason, this thesis presents a framework for automatically deploying and configuring HPC software.

    Through an interactive interface, the user can easily create and modify his experiments with a NWP model and run them on an HPC platform.

    In this way, users do not need to know how to operate a supercomputer to perform a simulation, making it usable for a wider community.

    The improvements proposed in this thesis aim to improve and facilitate the operation of NWP models, so that more results can be obtained from them and therefore our understanding and control of our planet.