Network virtualization is an extensively used approach to allow multiple tenants with different network architectures and services to coexist on a shared data center infrastructure. Core to its realization is the mapping (or embedding) of virtual networks onto the underlying substrate infrastructure. Existing approaches are not suitable for cloud environments as they lack its most fundamental requirement: elasticity. To address this issue, we introduce two new network primitives -expand and contract- which allow virtual networks to scale up and down. Mapping and scaling virtual networks over time, however, introduces fragmentation in the substrate network. This is akin to fragmentation in a file system where files are not laid out in contiguous physical blocks of the storage device. This problem impacts network performance and reliability for tenants and their applications. Instead of further improving embedding algorithms to tackle this problem, in this work, we present a yet unexplored approach: leveraging network migration techniques to defragment the network. We introduce network defragmentation as a new management primitive and propose algorithms to materialize it. We show through extensive simulations that our techniques significantly improve network performance while maintaining high utilization of the infrastructure, thus increasing provider revenue. On average, using defragmentation leads to 20% reduction in path length and utilization and cuts the number of very long paths (longer than half of the network diameter) between 52% and 62%. Moreover, it doubles the number of servers utilized by 50% or less as a result of consolidation.

Top