Flow adaptation and bandwidth control for IP-over-ATM

We devise a flow-aware traffic management framework for the better-than-best-effort support of long-duration adaptive IP flows over ATM. It is an integrated approach to service concatenation, flow management, and congestion control. Fundamental to this approach is the idea of improved interactions between adaptive applications and the network, and increased coordination between end-to-end and link-level traffic control. On the order of once per second, an application discloses to an ATM gateway in hints its traffic type, flow-rate variability and predictive bandwidth demand information. These feedforward hints assist the gateway in the management of packet flows and the adoption of link-level guaranteed service and flow-controlled service. In return, the gateway conveys to the application in hints ATM bandwidth availability information. These feedback hints help the application to make rate adjustments. Explicit feedforward control improves bandwidth efficiency when the ATM network is not the bottleneck, while explicit feedback control helps combat gateway congestion when the ATM network is the bottleneck.; Underlying this flow-aware traffic management framework are three suites of mechanisms. First, we design a protocol which facilitates the exchange of hints between a host and an ATM gateway, with application hints piggybacked on forward-going data packets and network hints on backward-going feedback or acknowledgment packets. In the host, hints are passed directly to/from the TCP protocol or indirectly to/from a UDP-based adaptive application via a modified socket API. Second, we devise techniques for use by adaptive sources to predict short-term bandwidth demand and monitor flow-rate variability. These techniques are implemented in two UDP-based rate-adaptive applications as well as the window-adaptive TCP protocol. Third, we develop ways in which an ATM gateway uses traffic type and flow-rate variability information to selectively aggregate flows, and uses predictive bandwidth demand hints to renegotiate CBR bandwidth, to interoperate with ABR rate-based flow control, and to enable the fair-sharing of VC capacity among flows.; Via simulations and experimentation with real applications over real networks, we demonstrate the feasibility and effectiveness of the framework for the transport of rate-adaptive flows and TCP traffic in LAN-ATM-LAN environment. Specifically, in the case of using CBR VCs, a good balance between application performance and reservation efficiency can be attained without excessive packet drops and queue buildup in the gateway. In the case of using ABR VCs, our approach leads to improved link utilization, reduced packet losses, and a higher degree of band-width fairness among flows. We also identify scenarios in which selective flow aggregation reaps the benefits of multiplexing gain without undermining flow performance. (Abstract shortened by UMI.)...