Implicit hop-by-hop congestion control in wireless multihop networks

Abstract

It has been shown that TCP and TCP-like congestion control are highly problematic in wireless multihop networks. In this paper we present a novel hop-by-hop congestion control protocol that has been tailored to the specific properties of the shared medium. In the proposed scheme, backpressure towards the source node is established implicitly, by passively observing the medium. A lightweight error detection and correction mechanism guarantees a fast reaction to changing medium conditions and low overhead. Our approach is equally applicable to TCP- and UDP-like data streams. We demonstrate the performance of our approach by an in-depth simulation study. These findings are underlined by testbed results obtained using an implementation of our protocol on real hardware.

Introduction

It has become more and more apparent that wireless multihop networks are much more prone to overload-related problems than traditional wireline networks like the Internet. Appropriate congestion control is thus vital to ensure network stability and acceptable performance.

TCP congestion control, which is one of the major foundations of today’s Internet, has proven to be highly problematic in wireless multihop networks [1], [2], [3], [4], [5]. Severe fairness problems, sub-optimal throughput and throughput stability issues have been reported. Such effects have also been observed experimentally in real multihop wireless networks, e.g. in [6]. Recently it has even been shown that TCP can generally not work as well in those networks as it does in common wired networks, because the rates of multiple TCP flows do not necessarily converge to a fair sharing of the bandwidth due to the shared medium [7].

All this is not too surprising, considering the fundamentally different properties of wired networks and multihop wireless networks. In a wireless network, the medium is a locally shared broadcast medium. Thus, congestion is a spatial phenomenon: neither nodes nor links, but geographical regions of the network are overloaded. The impact of this can be demonstrated by a very simple simulation experiment. In Fig. 1a, ns-2 simulation results of a bidirectional 10-hop chain topology are shown. In the simulation, static routing and the IEEE 802.11 MAC protocol are used. UDP constant bitrate traffic is injected with increasing data rate at both ends of the chain, traveling towards the opposite end. It can clearly be seen that the obtained total throughput drops rapidly once an optimal load is exceeded. This is due to an increasing number of collisions, leading to more and more packet drops. In a wireline network, throughput degrading effects for a too high UDP load are also well-known; they led to the development of TCP congestion control, which is able to deal with these problems very well. However, the problem observed here is of a completely different nature, which becomes immediately clear if we set up an equivalent wired topology in ns-2 and measure the throughput in the same scenario, as done in Fig. 1b—due to the duplex connections used between the routers on the Internet, the two packet streams do never even share a link or queue, and thus of course maximum throughput is achieved and maintained.

Note that the observed throughput drop is not a routing effect, since we use static routing without routing overhead and with no link breaks. Enabling RTS/CTS does not improve the situation significantly, and is thus not of help here. It is also not a problem of TCP congestion control, since TCP is not used. But TCP is also not an appropriate solution to the problem, as explicated at the beginning of this section: TCP-like congestion control does not behave well in wireless multihop networks. The results of the experiment demonstrate the spatial nature of congestion in wireless multihop networks. The problem is fundamental, and needs to be taken into consideration by any wireless multihop network design. Thus, the congestion problem in wireless multihop networks deserves to be reconsidered, and there is a need to search for better suited ways to perform congestion control.

Here, we propose a novel congestion control concept for wireless multihop networks, and a concrete protocol design realizing this concept. Our approach takes the local broadcast property of the wireless medium into account. It considers radio interference and can deal with an interference range that is—as it is the fact in real networks—much larger than the possible transmission range. It applies equally well to both TCP- and UDP-like traffic. It is a cross-layer approach, recombining functionality from different traditional layers, while still preserving an overall separation of functionality.

Our congestion control mechanism is a hop-by-hop approach. Traditionally, hop-by-hop congestion control means that local feedback on the sustainable rate for each node is transmitted to the respective upstream node, in order to establish some kind of backpressure towards the source. Here, we go one step further and actively exploit wireless multihop medium properties. We do not just try to cure the symptoms the traditional protocols exhibit when used in a environment they have not been designed for. In fact, we perceive the properties of the medium not primarily as a handicap, but instead also as an opportunity–an opportunity to solve problems like congestion control in new, different ways.

We call our congestion control approach implicit hop-by-hop congestion control, because its foundations are the hop-by-hop nature and implicit feedback, i.e., information gained by observing the transmissions of other nodes in the neighborhood. Even more central, the rate regulation at the source also happens implicitly, just by obeying some simple packet forwarding rules. The protocol design proposed here that realizes the concept of implicit hop-by-hop congestion control is called cooperative cross-layer congestion control (CXCC), because it relies on the cooperation of multiple layers.

Various factors can influence the congestion situation in a network. However, the most fundamental ones stem from the medium properties. The effects that we have shown above are not caused by, e.g., routing or mobility. There is a great variety of different routing paradigms that have been proposed for wireless multihop networks. By using static routing tables for the evaluations presented here, we stay neutral with regard to the routing protocol, and we avoid effects that might occur only in conjunction with a specific protocol, and thus cannot be unrestrictedly generalized.

The remainder of the paper is structured as follows. In Section 2 we will review some related work. Thereafter, in Section 3, the behavior of wireless multihop networks in congestion situations will be analyzed and the principles of implicit hop-by-hop congestion control will be deduced from these insights. Following that, we will apply the developed principles in a basis variant of a concrete protocol design, the CXCC protocol, in Section 4. In Section 5, we will analyze the performance of this basic approach, point out some remaining problems and show how they can be overcome. How we envision the integration of our approach into the protocol stack is detailed in Section 6. We will then assess the performance of our approach in detail by a simulation study in Section 7. To show that our concepts also work on real hardware, we present a real-world implementation and some experimental measurements in Section 8. Finally, we conclude our paper in Section 9.