Doughnutie: An efficient and low‐latency cloud data center network architecture

Abstract

Data center network provides the supporting infrastructure for cloud computing. Torus interconnection network is widely employed in the cloud data center networks. However, its long network diameter and average path length (APL) are prohibitive factors for supporting the delay-sensitive applications. With the aim of meeting latency, network capacity, fault-tolerance, scalability, and energy-efficiency requirements of cloud data center networks, we propose a new data center network architecture, named Doughnutie. Doughnutie is built by adding a number of flyway connections to the base Torus network to ameliorate its performance. Doughnutie does not follow a single principle for establishing all of its flyways and consequently, each part of the network can profit from all local, faraway, and neither-local-nor-faraway types of connection. Doughnutie decreases both network diameter and APL, which in turn leads to higher communication efficiency and lower network latency. Moreover, an exclusively designed loop-free load-balanced adaptive routing algorithm, named Doughnuter, is proposed.Employing the Doughnuter routing algorithm results in more efficient utilization of all the Doughnutie network resources. Theoretical analysis and simulation results also prove the superiorities of Doughnutie in throughput, load balancing, fault tolerance, and energy-efficiency.

Study of data center communication network topologies using complex network propagation model

“…Network topologies are selected in light of their use in engineering practices. In particular, switch-centric network topologies mainly include Fat-tree [8], Spine-leaf [9], etc., while servercentric network topologies mainly include Bcube [10], Dcell [12], etc [13][14][15][16][17][18][19][20]. Consequently, in our analysis, we concentrate on these three categories of topologies.…”