IBCube: An Economical and Incremental Datacenter Network

1. Introduction

With the emergence of cloud services and applications, how to efficiently build datacenters becomes an important issue (Liu et al. 2016). Maximizing the profits of datacenters is a major concern of datacenter operators for economical consideration (Zhan et al. 2016). Among the budget of a fully functional datacenter, a fraction of 15% goes to networking, i.e., the network equipments and the wires (Greenberg et al. 2008). In recent years, there have been a number of proposals for efficient datacenter networks (Lee et al. 2016; Ports et al. 2015; Zhu et al. 2015; Perry et al. 2014). Existing systems adopt quite different approaches for constructing datacenter networks (Al-Fares et al. 2008; Guo et al. 2009; Guo et al. 2008).

Existing datacenter network architectures can be classified into three styles, switch-centric, server-centric and hybrid designs (Popa et al. 2010). The switch-centric architecture utilizes switches for packet forwarding (e.g., the fat-tree based architecture (Al-Fares et al. 2008) uses switches for packet forwarding; Arjun et al. (2015) propose to use Clos topologies for connecting the switches to achieve good scalability). Different from the switch-centric design, the server-centric architecture relies on servers for packet forwarding, i.e., packets are forwarded between servers instead of switches. For instance, Abu-libdeh et al. (2010) design an architecture which connects servers using a 3D torus structure. By considering hybrid architectures, Guo et al. propose the DCell (2008) and BCube (2009) structures. DCell (Guo et al. 2008) and BCube (Guo et al. 2009) architectures forward packets using a combination of switches and servers. To evaluate the cost efficiency of existing diverse datacenter networking designs, Stoica et al. (2010) propose a high-level model to quantify and compare the cost of a datacenter network. Stoica’s results show that BCube achieves the lowest cost among those architectures (Popa et al. 2010).

Figure 1.

The port utilization in partial BCube_k with n=8 (N ∈ [2, 512])