Integration of Scheduler Knowledge into CiGri Control Loop
Résumé
Even with the progress done in the field of Cloud and High Performance Computing in hardware and programming structures, state of the art schedulers of these systems are still not able to take advantage of computing nodes at full capacity. A more economical and environmental friendly desires, require higher utilization of computing cluster's power. Harvesting idle resources in data-centers can be done by injecting best-effort jobs using CiGri a sclable software, in a way that minimizes unused resources while not perturbing higher priority jobs. Controlling this injection has been done using control-theory approaches, in a purely reactive fashion, independently of the scheduler and Resource and Jobs Management System (RJMS) in the system. This work explores a more pro-active approach, where information from the scheduler is used in a Feed-Forward Control loop, in order to achieve better performance. This approach can be done by conducting experimental campaigns to access newly identified scheduler's identified signals.
Origine : Fichiers produits par l'(les) auteur(s)