Ordonnancement périodique d'entrées/sorties pour super-ordinateurs

With the ever-growing need of data in HPC applications, the congestion at theI/O level becomes critical in super-computers. Architectural enhancement such asburst-buffers and pre-fetching are added to machines, but are not sufficient toprevent congestion. Recent online I/O scheduling strategies have been put inplace, but they add an additional congestion point and overheads in thecomputation of applications. In this work, we show how to take advantage of the periodic nature of HPCapplications in order to develop efficient periodic scheduling strategiesfor their I/O transfers. Our strategy computes once during the job scheduling phase a pattern where itdefines the I/O behavior for each application, after which the applications runindependently, transferring their I/O at the specified times. Our strategy limitsthe amount of I/O congestion at the I/O node level and can be easily integratedinto current job schedulers. We validate this model through extensive simulationsand experiments by comparing it to state-of-the-art online solutions, showing thatnot only our scheduler has the advantage of being de-centralized and thus overcoming theoverhead of online schedulers, but also that it performs better than thesesolutions, improving the application dilation up to 13% and the maximumsystem efficiency up to 18%.Dans cet article, nous nous intéressons à des techniques de gestion d'entrées-sorties dans les super-ordinateurs. La nouveauté de ce travail est la prise en compte de certaintes caractéristiques et arguments structurels sur les applications haute performance, leur périodicité, dans la conception de nos algorithmes.Nous nous comparons à des solutions récentes et montrons un gain en efficacité système atteignant 18% et en dilation atteignant 13%.Nous montrons comment facilement intégrer ces solutions sur des super-ordinateurs