ABSTRACT: Data centers (DCs) have been witnessing unprecedented growth in size, number, and complexity in recent years. They consist of tens of thousands of servers interconnected by fast network switches, hosting and enabling numerous applications with various traffic characteristics and requirements. As a result, DC networks have been presented with …
See more
ABSTRACT: Data centers (DCs) have been witnessing unprecedented growth in size, number, and complexity in recent years. They consist of tens of thousands of servers interconnected by fast network switches, hosting and enabling numerous applications with various traffic characteristics and requirements. As a result, DC networks have been presented with several unique challenges, pertaining to the scaling and allocation of network resources during the forwarding and moving of data across the different DC servers. Traffic routing in general and multicast routing, in particular, are important functions in DC networks, especially since modern cloud DCs tend to exhibit one-to-many communication traffic patterns. Unfortunately, recent multicast routing approaches that adopt IP multicast suffer from scalability and load balancing issues and do not scale well with the number of supported multicast groups when used for cloud DC networks. In this thesis, we propose a set of new, complementary schemes that overcome these challenges. More specifically, firstly, we study existing DC network topologies and propose Circulant Fat-Tree topology, an improvement over the traditional Fat-Tree topology with better properties to suit nowadays DC networks. Then, we review and classify recent studies that investigate and measure the traffic behavior of operational DC networks. We focus on the way they collect the traffic as well as on the key findings made in these studies. Secondly, we propose Bert, a source-initiated multicast routing scheme for DCS. Bert scales well with both the number and the size of multicast groups, and does so through clustering, by dividing the members of the multicast group into a set of clusters with each cluster employing its own forwarding rules. In essence, Bert yields much lesser multicast traffic overhead than state-of-the-art schemes. Thirdly, we propose Ernie, a scalable and load-balanced multicast source routing scheme. Ernie introduces a novel method for scaling out the number of supported multicast groups. In particular, it appropriately constructs and organizes multicast header information inside packets in a manner that allows core/root switches to only forward down the needed information. Ernie also introduces an effective multicast traffic load balancing technique across downstream links. Specifically, it prudently assigns multicast groups to core switches to ensure the evenness of load distribution across the downstream links.
See less
