Modern storage systems have to face the surge of the amount of data to handle. At the current scale, it would be an illusion to believe that a single centralized storage device is able to store and retrieve all its users' data. While from the user's viewpoint the storage system remains a single interlocutor, its underlying architecture has become necessarily distributed. In others words, storage is no longer assigned to a centralized storage equipment, but is now distributed between multiple independent storage devices, connected via a network. Therefore, when designing networked storage systems, bandwidth should now be taken into account as a critical resource. In fact, the bandwidth of a system is intrinsically a limited resource which should be handled with care to avoid congestion. The focus of this thesis is to optimize the available bandwidth of distributed storage systems, lowering the impact of churn and failures. The objective is twofold, on the one hand the purpose is to increase the available bandwidth for data exchanges and on the other hand, to decrease the amount of bandwidth consumed by maintenance. We present three distinct contributions in this manuscript. The first contribution of this thesis presents an hybrid peer-to-peer architecture taking into account the low level topology of the network i.e., the presence of gateways between the system and the users. The second contribution proposes an adaptive and user-level timeout mechanism, based on a Bayesian approach. The third contribution describes a repair protocol especially designed for erasure-coded stored data. Finally, this thesis concludes on the possibility of employing interference alignment techniques in order to increase the efficiency of repair protocols especially designed for encoded data.
