MANET as demonstrated in [83]. This has several advantages as discussed below: • By routing part of the intradomain traffic over the high-capacity backbone subnet, the traffic load in the MANET subnet can be alleviated. Thus we may regard transit routing as a kind of load balancing for intradomain traffic. • For certain source and destination pairs in the MANET subnet, e.g. in the case from S to D, transit routing makes it possible to achieve a considerably higher end-to-end throughput, since the wired backbone has much higher bandwidth than wireless links in the MANET.
• Transit routing also provides a higher probability for successful transmis- sions, since the wired links are much more reliable compared to wireless links.
• Wireless communication over multihop is often error prone and instable. By routing over the more reliable wired backbone, it is easier to maintain a more stable traffic stream between mobile nodes separated by many hops.
In order to support transit routing, there are a number of issues and challenges that need to be addressed. Many of these issues and challenges are common to the case of gateway load balancing, which include characterization, selection of default gateways/access points, traffic allocation, packet reordering etc. The main challenge is however to complement the routing protocol with the functionality for transit routing. Most MANET routing protocols such as OLSR or AODV are by default based on the shortest path metric, and as discussed previously, without modifications they are incapable of performing advanced tasks as transit routing, since it in many cases incur routing over longer path. To determine whether it is beneficial to perform transit routing or not for a certain source and destination pair,
the routing protocol basically needs to consider the costCi for routing the traffic
over the “ad hoc” path (i.e. internally within the MANET subnet), and the cost
Cii for the “wired path” (i.e. over the wired backbone subnet). The cost metric
algorithm must be designed such that transit routing is favored only in situations when there is a potential benefit in terms of increased throughput and/or reduced traffic load in the MANET. The work in Paper B is an effort to address this issue, where the focus is to investigate the situations in which transit routing is benefi- cial or not, and thereby develop an appropriate cost metric algorithm in order to facilitate transit routing. A more detailed description of our work related to transit routing can be found in Section 6.3 and in Paper B.
50 Load balancing
S D
Figure 5.3: The concept of multipath load balancing
5.4
Multipath Load Balancing
Multipath routing is a routing approach that allows for load balancing of intrado- main traffic, and therefore it is also called multipath load balancing. The concept is to distribute the traffic between a source and destination pair over multiple al- ternative disjoint/semi-disjoint paths as illustrated in Figure 5.3. Alternatively, multipath routing can also be used for other objectives such as to increase the re- liability and confidentiality of data transmission, optimize energy consumption in the network, and to improve QoS in the network. A considerable number of previ- ous work have investigated and proposed different schemes for multipath routing. The study in [84] provides an overview of the diversity of multipath routing proto- cols for MANETs. The majority of these proposals are based on reactive routing protocols such as AODV and Dynamic Source Routing (DSR), since the reactive approach provides an easier way to discover all possible disjoint paths between a source and destination pair.
Although the concept of multipath routing may contribute to improved perfor- mance, there are a number of issues and challenges that need to be considered as discussed in [85] and [84]. One of the issues is related to route discovery and route maintenance. These tasks are generally more costly to perform in terms of overhead and storage in multipath than in single path routing protocols. Further- more, in sparse networks it can be difficult to find multiple disjoint paths (either node or link disjoint). Even in densed networks the number of node disjoint paths may be limited. Disjoint paths offer certain advantages over non-disjoint paths. For instance, non-disjoint paths are less resilient against link breaks and may have lower aggregate throughput. Another issue is related to route selection. If there are multiple paths between the source and the destination, the question is how many of these paths should be used for data transmission. One may either select all the paths or a subset of them, and the paths may be selected based on certain criteria
5.4. Multipath Load Balancing 51
such as link quality, delay, and bandwidth in order to achieve optimization. Once the source node has selected the paths, it must also decide how to allocate traffic to these paths. Traffic allocation may be performed at different granularity. For instance, a per-connection granularity can be used to allocate all traffic from one connection to a single path. Alternatively, a more fine-grained per-packet granu- larity may be used to distribute packets from a connection amongst the selected paths. It is reported that per-packet granularity results in the best performance since it allows for fine tuning of the traffic distribution in the network [86]. How- ever, the disadvantage of the per-packet granularity is the increased risk for packet reordering at the destination. The last issue is when to trigger route discovery. It can either be triggered each time one of the paths is broken or only when all paths are broken. The first may incur a considerable amount of control traffic overhead while the latter may result in performance degradation. A possible good compro- mise may be to initiate route discovery when a certain percentage of the paths are broken.
With respect to performing load balancing, previous work have proposed differ- ent multipath load balancing schemes in order to maximize throughput, and at the same time minimize packet delay and route failure. Multipath load balancing has proven to be efficient in wired networks [86] [87] [88] [89], however, the same effect is difficult to achieve in wireless networks. In contradiction to wired net- works, the challenge with multipath load balancing in wireless networks is the interfering nature of the wireless medium. The interference that occurs among the paths limits the achievable gain in performance. The work in [1] refers to this interference as route coupling, and showed that this problem is especially severe in single channel networks. They showed that the performance gains provided by multipath load balancing is only negligible compared to single path routing. Route coupling is less severe in multi-channel networks, due to locally unique channel as- signments. However, they showed through simulation results that route coupling still exists even in multi-channel networks. Furthermore, by routing over more spatially separated paths, the effect of route coupling is reduced since the level of spatial reuse is consequently higher. However, this often requires that the traffic is routed over longer paths in terms of hop count, causing more network resources to be consumed. The results in [2] also confirmed the conclusion above. They showed through analysis and simulations that the gain provided by multipath load balancing is negligible unless the traffic is distributed over a huge number of paths. Based on these reports, multipath load balancing has therefore not been considered in this thesis.
Chapter 6
Summary and Contributions
6.1
Summary of the Work
The work in this thesis addresses two selected issues in the context of ad hoc network for emergency and rescue operations, i.e. mobility and load balancing. The aim is to investigate the shortcomings of current solutions with respect to these issues, and bring forth new solutions to improve the performance of the network as a whole or for the individual nodes. Thus, our work can be regarded as an effort on the way to realize an emergency and rescue communication system based on the ad hoc network technology. While the work in paper A addresses the issue of mobility, paper B to E address the issue of performing load balancing.
With respect to the issue of mobility, the focus is to investigate the factors that affect the rerouting time in proactive routing protocols. That is, the time duration needed to reroute and restore a broken communication path due to node mobility. The aim is to provide solutions for minimizing the rerouting time such that the packet loss can be minimized and the performance in the network is maximized. As previously discussed in Section 4.4, the rerouting time can be reduced by re- ducing the link break detection time, for example by applying alternative methods for link break detection such as the Fast-OLSR scheme or the link layer feedback mechanism. However, the drawback is that these solutions either incur increased control traffic overhead or the potential for erroneously declaring a link as invalid. The work in paper A shows that besides the impact of the link break detection mechanism, other factors such as the queue length, the input packet rate and the re- transmission limit at the MAC-sublayer can affect the rerouting time significantly. To solve the rerouting time with respect to these factors, either the queue length
54 Summary and Contributions
or the number of retransmissions can be reduced. The problem with reducing the queue length is the potential risk for packet loss due to buffer overflow at higher packet rates. Therefore, paper A proposes a solution based on Adaptive retry limit, where the retransmission mechanism at the MAC-sublayer is gradually decreased in the event of a link break. The reason for this is to minimize the number of transmissions wasted on stale packets, i.e. packets with invalid next hop address. Simulation results show that the proposed solution is very effective. In fact, as long as the data rate into the queue is safely below the capacity of the MAC, the solution eliminates the queueing problem associated with the rerouting time. With respect to the issue of load balancing, the focus is to investigate and ex- plore the feasibility as well as the potential benefits of performing load balancing in MANETs. The aim is to bring forth solutions that can optimize the usage of network resources and to improve the network performance in terms of increased throughput.
The work in paper B addresses load balancing for intradomain traffic. The assump- tion behind this work is that there exists a wired high capacity backbone subnet part of a MANET. Traditionally, such a backbone subnet is exclusively used to provide Internet connectivity to wireless nodes and to extend coverage area. However, the work in paper B demonstrates that it is also possible to exploit the capacity of this backbone subnet to alleviate the load in the MANET. This is achieved by rout- ing part of the intradomain traffic, i.e. traffic between nodes in a MANET, over the backbone subnet. We refer to this kind of load balancing as transit routing. Transit routing does not only allow for load balancing, but for certain source and destination pairs, the throughput can also be considerably increased. Paper B thus proposes a cost metric algorithm in order to facilitate transit routing. This algo- rithm is designed to commence transit routing only when appropriate. This means that when there is a performance gain in terms of throughput by using the alter- native path through the backbone subnet, the cost metric algorithm will favor this path. Simulation results show that by using the concept of transit routing, it is possible to enhance the throughput by 50% on the simulated topologies.
The work in paper C to E addresses load balancing for interdomain traffic. The assumption is that there exist multiple gateways in the network that provide con- nectivity to external network domains, such as the global Internet. The advantage of having multiple gateways in the network is among others the increased capacity for interdomain traffic. To exploit and optimize the usage of the increased capacity, gateway load balancing needs to be performed in order to distribute interdomain traffic between the gateways more evenly. This is a rather challenging task espe- cially when talking about wireless networks. Thus the work in paper C investigates the feasibility of performing gateway load balancing and explores the factors that
6.1. Summary of the Work 55
affect the potential gain of it. It is shown that a number of factors can affect the efficiency of load balancing as listed below:
• Level of asymmetry • Offered load
• Level of spatial reuse • Sensing range
• Shape and size of the network • Location of gateways
However, these factors alone cannot explain why the performance of load balanc- ing is high for certain topologies while it is poor for others. Obviously, the specific layout of the topology is also an important factor.
The work in Paper D therefore focuses on investigating the importance of the lay- out of a topology. Using the congestion maps, a number of interesting character- istics are discovered, which contributes to explain why the layout of the topology has significant impact on the performance of load balancing. Based on these re- sults, two different load balancing schemes are proposed in Paper D and E. These schemes demonstrate that radio load information can be used to perform load bal- ancing in MANETs. The load balancing scheme in Paper D is based on a deter- ministic gateway selection algorithm, and is proven to perform well for topologies with higher asymmetry level (with respect to node distribution and traffic load). However, for moderate asymmetry level, simulation results show that the solution is inefficient. This is due to what we refer to as the synchronized rerouting prob- lem, and is in fact a consequence of the distributed nature of MANETs. Thus the focus in paper E is to solve this problem by applying a randomized gateway selec- tion approach instead of a deterministic approach as is previously done in paper D. In addition to performing load balancing, the proposed scheme in paper E also per- forms admission control in order to prevent the network load in reaching a critical high level. Simulation results show that the new scheme is indeed more efficient than the scheme in paper D.
The work in all the above papers is tightly related to the preceding chapters. The IEEE 802.11 MAC-sublayer and the OLSR routing protocol presented in Chapter 3 and Chapter 4 represent the two fundamental technologies that are used throughout the work in this thesis. IEEE 802.11 is the dominating technology in the research related to ad hoc networks. OLSR is one of the most popular proactive routing protocols for MANET, and it is a natural choice in our research due to several rea- sons. First, an implementation of OLSR exists for the ns-2 simulator. Second, the
56 Summary and Contributions
advantage of using a proactive routing protocol is that it can quickly and dynam- ically adjust or rebalance traffic load to the changing conditions in the network. Third, routing metric parameters can be easily integrated with the routing protocol in order to achieve QoS routing.
Finally, the proposed solutions in our papers involve modifications or the integra- tion of new mechanisms into these protocols. The Adaptive retry limit solution in paper A is an example of such a modification to the MAC-sublayer in order to reduce the rerouting time. In addition, paper D and E demonstrate how the MAC-sublayer can be modified to provide radio load information to the routing layer. Similarly, the work in paper B, D and E, show that with new functionality, it is possible to increase network performance through transit routing and gateway load balancing.