On the Design and Analysis of Cloud Data Center Network Architectures

Abstract

Cloud computing has become pervasive in the IT world, as well as in our daily lives. The underlying infrastructures for cloud computing are the cloud data centers. The Data Center Network (DCN) defines what networking devices are used and how different devices are interconnected in a cloud data center; thus, it has great impacts on the total cost, performances, and power consumption of the entire data center. Conventional DCNs use tree-based architectures, where a limited number of high-end switches and high-bandwidth links are used at the core and aggregation levels to provide required bandwidth capacity. A conventional DCN often suffers from high expenses and low fault-tolerance, because high-end switches are expensive and a failure of such a high-end switch will result in disastrous consequences in the network. To avoid the problems and drawbacks in conventional DCNs, recent works adopt an important design principle: using Commodity-Off-The-Shelf (COTS) cheap switches to scale out data centers to large sizes, instead of using high-end switches to scale up data centers. Based on this scale-out principle, a large number of novel DCN architectures have been proposed. These DCN architectures are classified into two categories: switch-centric and server-centric DCN architectures. In both switch-centric and server-centric architectures, COTS switches are used to scale out the network to a large size. In switch-centric DCNs, routing intelligence is placed on switches; each server usually uses only one port of the Network Interface Card (NIC) to connect to the switches. In server-centric DCNs, switches are only used as dummy cross-bars; servers in the network serve as both computation nodes and packet forwarding nodes that connect switches and other servers, and routing intelligence is placed on servers, where multiple NIC ports may be used. This dissertation considers two fundamental problems in designing DCN architectures using the scale-out principle. The first problem considers how to maximize the total number of dual-port servers in a server-centric DCN given a network diameter constraint. Motivated by the Moore Bound, which provides the upper bound on the number of nodes in a traditional graph given a node degree and diameter, we give an upper bound on the maximum number of dual-port servers in a DCN, given a network diameter constraint and a switch port number. Then, we propose three novel DCN architectures, SWCube, SWKautz, and SWdBruijn, whose numbers of servers are close to the upper bound, and are larger than existing DCN architectures in most cases. SWCube is based on the generalized hypercube. SWCube accommodates a comparable number of servers to that of DPillar, which is the largest existing one prior to our work. SWKautz and SWdBruijn are based on the Kautz graph and the de Bruijn graph, respectively. They always accommodate more servers than DPillar. We investigate various properties of SWCube, SWKautz, and SWdBruijn; we also compare them with various existing DCN architectures and demonstrate their advantages over existing architectures. The second problem focuses on the tradeoffs between network performances and power consumption in designing DCN architectures. We have two motivations for our work. The first one is that most existing works take extreme designs in terms of improving network performances and reducing the power consumption. Some DCNs use too many networking devices to improve the performances; their power consumption is very high. Other DCNs use two few networking devices, and their performances are very poor. We are interested in exploring the quantitative tradeoffs between network performances and power consumption in designing DCN architectures. The second motivation is that there do not exist important unified performance and power consumption metrics for general DCNs. Thus, we propose two important unified performance and power consumption metrics. Then, we propose three novel DCN architectures that achieve important tradeoff points in the design spectrum: FCell, FSquare, and FRectangle. Besides, we find that in all these three new architectures, routing intelligence can be placed on both servers and switches; thus they enjoy the advantages of both switch-centric and server-centric architectures, and can be regarded as a new category of DCN architectures, the dual-centric DCN architectures. We also investigate various other properties for our proposed architectures and verify that they are excellent candidates for practical cloud data centers.