Scafida: A Scale-Free Network Inspired Data Center Architecture

László Gyarmati and Tuan Anh Trinh
Appears in: 
CCR October 2010

Data centers have a crucial role in current Internet architecture supporting content-centric networking. State-of-theart data centers have different architectures like fat-tree, DCell, or BCube. However, their architectures share a common property: symmetry. Due to their symmetric nature, a tricky point with these architectures is that they are hard to be extended in small quantities. Contrary to state-of-the-art data center architectures, we propose an asymmetric data center topology generation method called Scafida inspired by scale-free networks; these data centers have not only small diameters and high fault tolerance, inherited by scale-free networks, but can also be scaled in smaller and less homogenous increments. We extend the original scale-free network generation algorithm of Barabási and Albert to meet the physical constraints of switches and routers. Despite the fact that our method artificially limits the node degrees in the network, our data center architectures keep the preferable properties of scale-free networks. Based on extensive simulations we present preliminary results that are promising regarding the error tolerance, scalability, and flexibility of the architecture.

Public Review By: 
S. Agarwal

Engineers and researchers design the switch-level topologies of datacenters in deliberately symmetric patterns. In sharp contrast to this active area of research, this paper proposes designing datacenter topologies in an asymmetric fashion, inspired by the scale-free Barabási and Albert topologies. In such topologies, the degree distribution of nodes follows a power law distribution, which can be achieved by preferential attachment.
The hope is that a datacenter designed this way will have fewer hops between servers, be more resilient to link failures, and allow the datacenter to grow organically instead of in fixed chunks dictated by current symmetric designs. The challenge is that servers and switches have limited numbers of ports and hence a true power law distribution cannot be achieved.
This paper provides an analytic and simulation study of what could happen if we built a datacenter in this way. The topology construction algorithm, Scafida, limits the number of server and switch connections to realistic values and yet almost achieves the characteristics of scale-free topologies. The paper finds that average path lengths are short, the topologies are resistant to link failures, and they provide large bisection bandwidths.
Unfortunately, the diameter or maximum path length is rather large compared to standard topologies. It is not clear how the lower average path length impacts latency within a datacenter in a meaningful way. The paper does not compare link failure resistance to other topologies. The evaluation assumes perfect routing between and through servers.
All the reviewers agree that the paper asks a compelling and timely question. What if we were to design datacenter networks to be scale-free? The answers provided are not perfect, but certainly shed light on the problem. Surprisingly such topologies provide properties quite similar to carefully designed topologies. The authors have worked quite diligently with the reviewers in improving the paper for publication in CCR.