OverQoS Architecture

Figure 1 illustrates an OverQoS network with overlay nodes spanning different routing domains and flows routed within this network. We make no assumptions about the placement of overlay nodes in the Internet. Rather, we assume that a placement of overlay nodes is pre-specified. In this paper, we will assume that the end-to-end path on top of an overlay network is fixed and we will attempt to enhance the QoS along this path in the presence of varying levels of network congestion. We can use existing approaches like RON [32] to determine the overlay path between a pair of end-hosts.

In the remainder of this paper, we will use the term virtual link to refer to the IP path between two overlay nodes and bundle to refer to a stream of application data packets carried across the virtual link. A bundle typically includes packets from multiple transport-layer flows across different sources and destinations. The following constraints and requirements make the design of any overlay-based QoS challenging:

1. Node Placement and Cross Traffic: Overlay nodes will usually span different routing domains and will not be directly connected to the congested links. Hence, one cannot avoid losses or delays along virtual links. Additionally, the losses incurred due to cross traffic is time-varying and can be hard to predict.

2. Fairness: Overlays should not offer QoS at the expense of hurting cross traffic. Therefore, the overlay traffic at an aggregate level should be congestion sensitive and not use more than its fair share. One standard metric for determining fair share is based on TCP-friendliness [27].

3. Stability: Multiple overlay networks independently offering QoS with many virtual links overlapping on congested physical links in the underlying network should be able to co-exist.

=

Figure: The OverQoS system architecture. OverQoS nodes in different AS's communicate with each other over virtual links using the underlying IP paths.

To address these challenges we propose a solution that builds on two design principles:

Bundle loss control: Overlay nodes should bound the loss rate experienced by a bundle along a virtual link in the presence of time-varying cross traffic. We propose a controlled-loss virtual link (CLVL) abstraction to achieve this loss bound and characterize the service received by a bundle.

Resource management within a bundle: An overlay node can control the loss and bandwidth allocations of each flow and/or application within a bundle.

These design principles enable OverQoS to provide a range of useful services to Internet applications. Example of such services are: (1) packet prioritization, (2) smoothing losses (i.e., eliminate the bursts of losses by spreading losses in time), and (3) statistical bandwidth and loss guarantees, though this service can be typically offered only to a small fraction of a bundle's traffic. We next elaborate on our two design principles.