A Load Balancing and Multi-Tenancy Oriented Data Center Virtualization Framework

• Published in: IEEE transactions on parallel and distributed systems Vol. 28; no. 8; pp. 2131 - 2144
• Main Authors: Duan, Jun, Yang, Yuanyuan
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.08.2017; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Architecture; Bandwidths; Blooms (metal); Cloud computing; Communications traffic; Computer centers; data center; data center networks; Data centers; distributed virtual switch; Hardware; Load balancing; Load management; Marketing; Matrices (mathematics); multi-tenancy; OpenFlow; Protocol (computers); Resource management; Servers; Switches; Virtual environments; Virtualization; Voltage control
• ISSN: 1045-9219; 1558-2183
• DOI: 10.1109/TPDS.2017.2657633

Virtualization is an essential step before a bare-metal data center being ready for commercial usage, because it bridges the foreground interface for cloud tenants and the background resource management on underlying infrastructures. A concept at the heart of the foreground is multi-tenancy, which deals with logical isolation of shared virtual computing, storage, and network resources and provides adaptive capability for heterogeneous demands from various tenants. A crucial problem in the background is load balancing, which affects multiple issues including cost, flexibility and availability. In this work, we propose a virtualization framework that consider these two problems simultaneously. Our framework takes advantage of the flourishing application of distributed virtual switch (DVS), and leverages the blooming adoption of OpenFlow protocols. First, the framework accommodates heterogeneous network communication patterns by supporting arbitrary traffic matrices among virtual machines (VMs) in virtual private clouds (VPCs). The only constraint on the network flows is that the bandwidth of a server's network interface. Second, our framework achieves load balancing using an elaborately designed link establishment algorithm. The algorithm takes the configurations of the bare-metal data center and the dynamic network environment as inputs, and adaptively applies a globally bounded oversubscription on every link. Our framework concentrates on the fat-tree architecture, which is widely used in today's data centers.