SHF: Small: Designing Expandable and Cost-Effective Server-Centric Interconnects for Data Centers

点击翻译按钮获取中文摘要

In modern data centers, the performance of data center networks (DCNs) plays a critical role. Currently, DCNs are mainly divided into two categories: switch-centric networks and server-centric networks. In a switch-centric network, switches are responsible for a variety of tasks such as routing and addressing, while servers are only used to send and receive packets in the network. In a server-centric network, the computational intensive tasks like routing are put on the servers, which act not only as end hosts, but also as relay nodes for each other. A significant advantage of server-centric networks is that network hardware cost can be drastically reduced, as inexpensive commodity switches are sufficient given that most complex network tasks have been shifted to servers where computation resources are abundant. Moreover, since servers are much more programmable than switches, server-centric networks can accelerate the process of network innovation. Motivated by the importance and opportunities of deploying server-centric networks in DCNs, this research systematically investigates the fundamental and challenging issues towards building cost-efficient server-centric DCNs with guaranteed performance.The objective of this research is to design novel, extendable and cost-effective server-centric interconnects and associated routing algorithms for data centers. This research explores the unique novel features and techniques in data centers to build cost-efficient interconnects with dual-port servers or any fixed number of NIC ports. More specifically, the research focuses on following closely coupled issues: (1) construct cost-efficient DCN interconnects using commercial off-the-shelf (COTS) switches and dual-port servers that are available in current markets; (2) efficient routing algorithms for unicast, broadcast and multicast that can sufficiently take advantage of the novel interconnects; (3) employ network virtualization technology to offer a more flexible solution to both industry and academic research; (4) construct cost-efficient DCN interconnects using servers with any fixed number of NIC ports that meet the trend of future servers; (5) conduct a comprehensive performance evaluation through extensive simulations and implementation of proposed schemes in a realistic network prototype. The proposed research combines theoretical analysis, algorithm design, network optimization, simulation and prototyping techniques to provide a comprehensive working solution that enables high performance next generation DCNs. This research hopes to impact fundamental design principles of server-centric DCNs. The outcome of this research has the potential to boost the performance of DCNs while keeping the cost low, and facilitate numerous cloud computing applications currently hosted in data centers. A project goal is to train graduate students and promote the participation of female engineering students. The important findings of this project are to be disseminated to the research community by way of conferences, journals and web site access.