Fault-Tolerant Network Topologies for Datacenters

Abstract

Data centers are an integral part of cloud computing infrastructure to support various cloud-based services such as web search, email, social networking, distributed file systems and scientific computing. Data centers provide huge computational power and storage, reliability, availability, and cost-effective solutions needed by the cloud applications. A data center network (DCN) topology connects thousands of servers within the datacenter and to the external world. The topology is vulnerable to failures due to the presence of huge number of servers, switches and links. Several data center network topologies have been proposed and implemented; however, most of them lack the ability to recover from failures. One of the biggest challenges in DCN is to provide a graceful degradation in performance in the event of a link or server failure. Fault-tolerance in a DCN topology can be provided by adding extra hardware (switches, links) or by provisioning of multiple redundant routing paths among servers. This thesis proposes two new fault-tolerant DCN topologies derived from the standard 𝐷𝑐𝑒𝑙𝑙 topology. The proposed topologies, 𝐷𝑐𝑒𝑙𝑙−𝑆𝑡𝑎𝑟 and 𝐷𝑐𝑒𝑙𝑙−𝑅𝑖𝑛𝑔, are both cost-effective and scalable. In addition, the proposed topologies enhance the overall performance (throughput and latency) of 𝐷𝑐𝑒𝑙𝑙 topology, and offer graceful performance degradation in the case of a link or server failure. Furthermore, we propose a new mechanism to select the optimal path between the hosts in the topology using Genetic Algorithm (GA). Performance evaluation of the proposed topologies and techniques is done through a simulation study using realistic intra-datacenter traffic models, and the results are compared with the standard 𝐷𝑐𝑒𝑙𝑙 topology. The comparison is done in terms of various metrics such as throughput, latency, diameter, and average shortest path length. The simulation results show that the proposed topologies outperform the standard 𝐷𝑐𝑒𝑙𝑙 topology due to the availability of multiple alternate shortest paths between any pair of servers, resulting in an improvement of about 5% in throughput even for a small-size network. GA algorithm for the path selection is applied to the two proposed topologies, and it is found that there is a further improvement of about 2% in the throughput of the topologies.