Discussion

End-to-end Recovery vs Overlay CLVL: An alternative to applying the CLVL abstraction on an overlay network is to apply loss control on an end-to-end per flow basis. There are several arguments against end-to-end loss control: First, using FEC to apply end-to-end loss control is far more expensive than applying it on an aggregate level. For example, in order to provide a $0.1\%$ loss guarantee to a $64$ Kbps stream (like game console traffic or IP telephony stream) over a bursty channel with an average loss rate of say $2\%$, the minimum amount of FEC required can be as high as $32$ Kbps. However, if $10$ such flows are aggregated at an overlay node, the per-flow FEC requirement can drop to lower than $5$ Kbps. Second, with a better distribution of overlay nodes, we expect the overlay links to have much smaller RTTs than end-to-end RTTs. Hence, overlay-level recovery using ARQ has better delay characteristics than end-to-end recovery. Finally, aggregation of flows within an overlay provides the ability to trade resources across different flows (or within packets of the same flow) which is fundamentally necessary to provide better QoS properties.

Delay guarantees: Overlay networks have no control over variations in queuing delays along virtual links and hence cannot offer delay assurances. On the other hand, overlay networks have been used to route around congestion [33,7]. Such techniques can be embedded into an overlay to improve the end-to-end delay characteristics of a path.

Over-provisioning: Recent measurement studies have shown that Internet backbones are over-provisioned and have low levels of congestion [19,17]. This questions the basic need for Internet QoS. We contend that over-provisioning is not necessarily a permanent feature of the Internet, but a reflection of the big disparity between the poor connectivity at edges, and the backbone capacity. As more homes and enterprises become connected over faster, multi-megabit/s or higher, links with optical fibers, we expect that at least some parts of the Internet such as small ISPs to become more congested. This trend is already evident in countries like Japan where ISPs offer $10$ Mbps broadband connections to homes [4]. In addition, many ISPs already provide aggregate QoS within their networks using MPLS technologies [26]. We believe that overlays are the right platform for translating these aggregate intra-domain QoS to meaningful end-to-end QoS guarantees.