Adaptive control of HPC clusters for server overload avoidance
Résumé
This thesis focuses on the maximal usage of an High-Performance Computing (HPC) Cluster with particular attention to the overloading of the storage, namely the file server. HPCs have become important in the scientific domain because of their computational power which allows calculation that otherwise would be impossible with a single computer. On the other hand, this powered structure has a considerable cost, and unfortunately also a short life cycle. As a consequence, it becomes important to use as best as possible all the resources that this new structure gives. In this context, Autonomic Computing and the collaboration between control theory and computer science were born.
This Mater Thesis continues a broader project where a PI controller was designed for an HPC Cluster. The aim of the controller is the minimization of idle resources inside the cluster. However, the controller is built on a model in a nominal configuration which leads to heterogeneous results for different working conditions. Moreover, conditions far from the nominal one bring the file server close to the overloading which is a challenging condition for the system.
The idea of this project is to design an identification algorithm in order to end up with an Adaptive PI that changes its own parameters to achieve specification with various working conditions. We design three different algorithms that are going to be tested in a simulation environment (SIMULINK) and then only one will be carried to the real set-up. In the end, there is a comparison between the PI and the Adaptive PI. Here it is shown how the Adaptive PI has a more homogeneous result than the case of the simple PI and how the overloading of the file server is avoided. However, the drawback is the settling time which becomes larger than in the case of classic PI.
Origine : Fichiers produits par l'(les) auteur(s)
