Memory requirements for future Internet routers with essentially-perfect QoS guarantees

The theory of a future Internet network which achieves essentially-perfect QoS guarantees for all QoS-enabled traffic flows for all loads ≤ 100% of capacity has recently been established. A scheduling algorithm with a bounded normalized service lead/lag (NSLL) is used to schedule traffic flows within the routers. An 'Application-Specific Token-Buffer Traffic Shaper' is used at the traffic sources, to achieve a bounded NSLL on incoming bursty traffic flows. An 'Application-Specific Playback Queue' is used to perfectly regenerate the original busty traffic flows at every destination. Under these conditions, it has been established that every QoS-enabled flow: (i) is delivered with essentially-perfect end-to-end QoS guarantees, and (ii) buffers O(K) cells/packets per router, where K is the bound on the NSLL. In this paper, we reduce the router buffering requirements significantly, so that each router buffers ≤ one cell/packet per QoS-enabled traffic flow, a reduction of up to 1K-10K over existing technologies. The proposed technology can be incorporated into new routers with negligible hardware cost, and is compatible with existing IntServ, DiffServ, MPLS and RSVP-TE protocols.