On Constructing Novel Data Center Networks and Related Flow Scheduling Problem
July 13, 2018
Light Engineering room 250
Advisor: Prof. Yuanyuan Yang
Server-centric network structures have been studied for almost a decade. Experimental evaluation shows their significant advantage over switch-centric network. However, most existing data center products still organize servers in a Clos-based structure, although an immediately emerging issue is the over-subscription in links among higher levels. As pointed out, server-centric network has two critical drawbacks that hinge them from being products, namely expandability issue and availability issue. Hence, server-centric networks still behave less empirical than switch-centric networks. On the other hand, wireless links and other novel technology are introduced to switch-centric networks to alleviate the over-subscription problems.
The purpose of this dissertation is to attack the challenges in server-centric networks thus that they could better serve today’s cloud computing environment. We propose novel interconnection network structures to separately address expandability and availability respectively. The following parts are included in this dissertation. First, to make the designated structure empirical and expandable, we restrict the servers in hand to be with only two network interface card (NIC) ports, or any fixed number of NIC ports. Given this setting, we design structures BCCC and GBC3 respectively. Second, we propose novel structures called RCube and RRect that significantly enhance the availability of cloud computing by introducing server redundancy. Third, in addition to improve the performance of server-centric network, we also investigate the performance of legacy Clos-based switch-centric network with emerging wireless links, and propose a flow scheduling algorithm that jointly considers virtual network embedding and wireless antenna switching. Finally, we evaluate the performance of the proposed structures and algorithms in extensive criteria such as average path distance, aggregate throughput, power consumption, etc., which show the great advantages of the proposed structures and algorithms.