Dynamic Bandwidth Allocation for Preventing Congestion in Data Center Networks

Abstract

Running multiple virtual machines over a real physical machine is a promising way to provide agility in current data centers. Virtual machines belong to one application are striped over multiple nodes, and the generated traffic often shares the substrate network with other traffic of other applications. In such conditions, clients can experience severely degraded performance, such as TCP throughput collapse and network congestion due to competing network traffic. The basic cause of this problem is that network traffic from multiple sources which shares the same network link can cause transient overloads in the link. In this paper, we make the case that network virtualization opens up a new set of opportunities to solve such congestion performance problem. We present an architecture which compartmentalize virtual machines of same application into same virtual networks by network slicing, and divides the role of the traditional ISPs into two: infrastructure providers and service providers to achieve more commercial agility needed by cloud computing in particular. We also present a dynamic bandwidth allocation mechanism, which can prevent congestion and maximize utilization of substrate networks. Experimental result shows that the network slicing mechanism and bandwidth allocation algorithm can prevent network congestion significantly.