Routing Internet traffic in heterogeneous mesh networks: Analysis and algorithms

Abstract

In practical wireless mesh networks (WMNs), gateways are subject to hard capacity limits on the aggregate number of flows (in terms of bit rate) that they can support. Thus, if traffic is routed in the mesh network without considering those constraints, as well as the traffic distribution, some gateways or intermediate mesh routers may rapidly get overloaded, and the network resources can be unevenly utilized. To address this problem, in this paper we firstly develop a multi-class queuing network model to analyze feasible throughput allocations, as well as average end-to-end delay, in heterogeneous WMNs. Guided by our analysis, we design a Capacity-Aware Route Selection algorithm (CARS), which allocates network paths to downstream and upstream Internet flows so as to ensure a more balanced utilization of wireless network resources and gateways’ fixed connections. Through simulations in a number of different network scenarios we show that the CARS scheme significantly outperforms conventional shortest path routing, as well as an alternative routing method that distributes the traffic load on the gateway nodes to minimize its variance.

Introduction

Wireless mesh networks (WMNs) are increasingly deployed to provide cost-effective ubiquitous access to the Internet [1]. Normally, in WMNs a set of stationary wireless mesh routers form a multi-hop wireless backbone, where a small subset of these routers act as gateways being connected to the Internet through high-speed fixed lines [2]. Typically, it is also assumed that the link capacity in the Internet is much larger than the wireless channel capacity. However, this vision is rapidly changing. Real-world mesh networks are frequently used to share a potentially large number of low-speed Internet connections (i.e., DSL fixed lines) available at the customers’ premises. Examples of such networks are Meraki-based deployments in urban areas [3], or the Ozone’s network in Paris, which is composed of 400 mesh routers, most of them using standard DSL links as Internet backhaul, while only ten gateways are provided with an ISP-owned fiber link [4]. In a broader sense, wireless mesh networks are evolving into a converged infrastructure used to share the Internet connectivity of sparsely deployed fixed lines with heterogeneous capacity, ranging from ISP-owned broadband links to subscriber-owned low-speed connections [5].

WMNsMesh being primarily used for Internet access. Therefore, both traffic routing and Internet gateway selection play a crucial role in determining the overall network performance and in ensuring optimal utilization of the mesh infrastructure [6]. Indeed, depending on the location of the mesh nodes and the gateways, some of the mesh nodes may obtain substantially lower throughput than others. Similarly, if many mesh nodes select the same gateway as egress (ingress) point to (from) the Internet, congestion may increase excessively on the wireless channel, or the Internet connection of the gateway can get overloaded. This problem is particularly relevant in the heterogeneous WMNs considered in this study, because low-speed Internet gateways may easily become a bottleneck, limiting the achievable capacity of the entire network, while most of the available studies have assumed that bottlenecks appear only in the wireless network. Finally, a load-unaware gateway selection can lead to an unbalanced utilization of network resources.

To improve load balancing and increase capacity of WMNs, previous studies suggested to use balanced tree structures rooted at the gateways, and to route the traffic along the tree paths. For instance, a heuristic algorithm for calculating load-balanced shortest path trees taking into account flow load is proposed in [7]. In [8], approximated solutions are defined for calculating load-balanced trees that allocate the same bandwidth to all the nodes, using both single-path and multi-path approaches. An alternative strategy is proposed in [9], where the complexity of finding optimal routes is mitigated by considering only delay optimal routing forests, i.e., unions of disjoint trees routed at the gateway nodes. However, tree-based routing structures are less reliable to link failures than mesh-based structures. Furthermore, the admission of a new flow usually triggers complex reconfiguration procedures for the entire tree in order to maintain the load balancing properties.

A simpler approach to improve network performance is to define routing metrics for shortest path first (SPF) routing that determine high-throughput paths and/or facilitate load balancing. Initially proposed metrics (e.g., ETX [10] and ETT [11]), focussed only on link characteristics (e.g., frame loss and transmission rates), thus they cannot balance the load. Recent studies proposed to introduce in the metric computation estimates of inter-flow and intra-flow interference (e.g., IRU [12]), location-dependent contention (e.g., CATT [13] and ETP [9]), or load-dependent cost (e.g., the queue length in WCETT-LB [14], or the number of per-link admitted flows in LAETT [15]). Although these metrics have been demonstrated to work quite well in mesh networks, and to provide higher throughput performance than simple hop count, they are completely unaware of the available resources at the gateways. On the contrary, significant performance improvements may be obtained by considering residual capacity of the links between the gateways and the Internet, as well as the load distribution, when routing traffic flows. However, there is a complex interdependence between the way traffic flows are routed in the network and the utilization of network resources, which makes defining simple heuristics to estimate the remaining capacity of a network path or a gateway quite difficult.

To address the above problems, in this paper we make the following contributions. We develop a multi-class queuing network model for heterogeneous WMNs and used it to determine if a given allocation of flows on a set of network paths is feasible. More precisely, our model characterizes the network performance as a function of the traffic pattern, the distribution of gateways and mesh routers in the WMN, the heterogeneity of link capacities, as well as the location-dependent contention on the wireless channel; then, given the routing strategy used to allocate the flow demands on the network paths, we exploit our model to establish if the resulting flow allocation does not violate the network capacity constraints. Moreover, we also mathematically characterize the average packet end-to-end delay, defined as the average time taken by a packet to reach the Internet after it is generated. To validate our modeling methodology, in this study we consider a basic CSMA-based MAC protocol, which implements an idealized collision avoidance mechanism that can always detect if the medium is busy or free before a transmission attempt. The primary goal of this study is not to accurately model the performance of specific standard MAC protocols, but to investigate the impact on system performance of the location-dependent contention inherent to multi-hop environments, due to differences in the number of contending nodes at both endpoints of each communication link.

It is important to point out that several previous studies have proposed to use queuing models to investigate system performance of CSMA-based ad hoc networks. However most of these studies have applied queuing theory to the analysis of single-hop ad hoc networks [16], [17], [18]. To the best of our knowledge, in the literature a few examples exist which deal with the multi-hop case. In [19], the authors model random access multi-hop wireless networks as open $GI/G/1$ queuing networks to analyze the average end-to-end delay and maximum achievable per-node throughput. However, the formulation proposed in [19] can be applied only to random networks, and it does not incorporate flow-level behaviors. Our objective is different from [19], because we consider arbitrary topologies and routing strategies, and we focus on per-flow performance. In our previous paper [20] we have developed a single-class queuing model to analyze the network capacity of heterogeneous WMNs, however, the analysis in [20] is valid only for upstream Internet traffic, which is a somehow unrealistic traffic model for typical WMNs. To go further, in this paper we extend our previous analysis to incorporate generic traffic distributions, which motivates the use of a novel modeling methodology based on multi-class queuing networks.

Guided by our analysis, in this paper we propose a Capacity-Aware Route Selection algorithm (CARS), which integrates traffic routing with gateway selection. Instead of using SPF routing, CARS scheme determines the set of optimal routes from the mesh node that originates a new flow, and the available gateways. It is important to note that any cost function can be used to determine the initial set of optimal network paths. However, isotonic routing metrics are preferable because they permit efficient and loop-free computation of minimum cost paths [21]. Then, CARS allocates the new flow to the best network path that has enough residual capacity (as predicted by our model) to satisfy its bandwidth demands. As a result, a mesh node can discard paths or gateways that cannot accept additional demands. This facilitates load balancing in the network by avoiding the rapid exhaustion of the link capacity of disadvantaged mesh nodes or gateways, leading to a more efficient utilization of both wireless and wired network resources. Through simulations performed in network scenarios with different numbers of gateways and link capacities, we show that CARS scheme results in significant throughput improvements over SPF routing using IRU metric [12], which captures only inter-flow interference (i.e., mutual interference between adjacent flows). Furthermore, the simulation results confirm the accuracy of the proposed modeling methodology.

The remaining of this paper is organized as follows. Section 2 introduces the network model. In Section 3 we develop the capacity analysis. Section 4 describes the proposed CARS algorithm. In Section 5 we present simulation results to validate the analysis, and to compare CARS performance against two other routing algorithms. Finally, conclusions and future work are discussed in Section 6.