Topology-aware Gossip Dissemination for Large-scale Datacenters

Gossip-based protocols are very robust and are able to distribute the load uniformly among all processes. Furthermore, gossip-protocols circumvent the oscillatory phenomena that are known to occur with other forms of reliable multicast. As a result, they are excellent candidates to support the dissemination of information in large-scale datacenters. However, in this context, topology oblivious approaches may easily saturate the switches in the highest level of the datacenter network fabric. This thesis presents a novel gossip protocol for datacenters, named Bounded Gossip, that provides an adequate load distribution among the different layers of the switching fabric of the datacenter, avoiding being a source of network bottlenecks. Bounded Gossip embodies techniques from previous protocols, such as Hierarchical Gossip and CLON, and combines them with a topology-aware membership maintenance scheme and a topology aware rate-based flow control scheme. The benefits from our solution are illustrated by an experimental evaluation that compares the performance of Bounded Gossip with that of other competing protocols, in terms of imposed load in the routing topology, overall cost of communication, and dissemination latency. Palavras Chave Keywords