Data center networks are large-scaled interconnects which demand high scalability and fault tolerance. Production systems highly depend on regular structures (e.g., the FAT tree) to organize servers in direct networks (or switches in indirect networks). While the recent proposals are scalable to accommodate an increasing number of servers added to the system, the regular topologies are not flexible as servers (or switches) cannot be interconnected freely and installation errors may incur due to network operators, thus increasing their provisioning and maintenance costs. In this paper, we present a randomized topology for data center network. Our proposed network is not only scalable and fault-tolerant, but supports flexible provisioning. Servers (or switches) can link to one another in some probability. Our proposed networks exhibit nice properties, including the exploitation of geographical hierarchy and the low diameter of the networks. We assess our proposal in rigorous theoretical analysis.
All Science Journal Classification (ASJC) codes
- Computer Networks and Communications