Méthodes d'apprentissage automatique appliquées à l'analyse des signaux d'utilisations des grands calculateurs

The aim of this thesis is to determine what statistical methods can currently be used to improve the understanding of the use of a computing center. We decompose the computer into three parts : hardware, software and users in order to identify three relevant research directions. We propose a model allowing the prediction of the power consumption of a computer before it is placed in the queue, so that the software thatmanages this queue can control the computer’s consumption. We also seek to visualise more easily the data relating to events in the computing center which can be textual or a number of occurrences. Finally, we propose to group and slice in relevent parts time series from sensors installed on the CEA’s computers. These methods are therefore very useful for computer scientists and can be original for statisticians.