The large bandwidth-delay product of wide area gigabit networks poses important questions for network congestion and flow control schemes, since for bursty traffic the transit time of a burst may typically be much smaller than the propagation delay through the network. The small number of nodes in each testbed made it difficult to carry out meaningful experiments for this problem, however, and so it was not a major focus of testbed work.
Some analytical and simulation studies were carried out on selected aspects of the problem by NCSU and Berkeley [3,6]. MIT developed modifications to existing TCP protocol mechanisms to provide improved congestion control properties for large bandwidth-delay products, in particular to avoid long retransmission timeouts when packets are lost and to return to a full throughput state more quickly after lost packet events [2]. The lost packet events are used by TCP in the current Internet to infer the existence of network congestion and invoke a traffic backoff/recovery algorithm.
In contrast to reliance on TCP or other strictly end-to-end mechanisms to deal with network congestion, Wisconsin developed a new scheme involving both source and network node mechanisms. Called Dynamic Time Windows (DTW), their approach is targeted towards wide area gigabit ATM networks with large bandwidth-delay products. It uses a time window at each source to control the source's burstiness, with a larger window duration allowing greater burstiness. Feedback of information from network switches along the source-destination path causes the source to dynamically adjust its window as a function of aggregate switch traffic.
DTW is predicated on the use of fixed routes between each source and destination, as is the case for ATM virtual circuits, and on the use of weighted fair queuing (WFQ) in each switch. WFQ is used to bound additional traffic burstiness introduced by the switches, allowing DTW to bound overall network congestion time through control of the source time windows. The scheme allows network administrators to trade off packet loss due to switch buffer overflow with the time required for the network to return to a stable state after congestion occurs, through definition of a global network constant.
Analysis was used to establish the scheme's general properties under simplifying assumptions, with extensive simulations used to establish its performance for a variety of traffic and network situations. The simulation results showed the system to be stable, with its steady-state behavior characterized by periodic oscillations of source time windows about their average values. In simulation comparisons to a TCP-like end-to-end packet window system, the DTW system had lower average network delays and switch packet losses, along with higher source delays. This reflects DTW's more rapid response to network congestion events, resulting in larger queuing time of original packets by the source and fewer retransmissions due to dropped packets in the network.
Experimental data on DTW was obtained using Blanca testbed facilities between Wisconsin and UIUC. A router at Wisconsin was used as the source node, sending traffic through the Wisconsin, Chicago, and UIUC Xunet ATM switches to a receiving node at UIUC. DTW monitoring and feedback algorithms were implemented in the Wisconsin switch, with a local traffic generator at the switch used to introduce bottleneck traffic. A simple weighted round robin algorithm was used with the switch's programmable queuing hardware to approximate the use of WFQ.
The results showed that, at least for the simple experimental configuration used, DTW adjusted correctly to different values of congestion thresholds, system loads, negotiated source throughput, and changing network state. Although the link data rate was limited to 45 Mbps due to problems with Blanca 622 Mbps equipment at the time of the experiments, the experiments yielded data consistent with the more extensive simulation results and provided valuable experience with the DTW algorithms under real-world conditions.