Section 9.5. Other Networking Issues

9.5. Other Networking Issues

In addition to link layer and network layer issues, as discussed previously, higher layer protocols and system issues must also be considered in the design and analysis of any network, including a UWB network. Here, we consider issues related to the performance of the Transmission Control Protocol (TCP), transport layer protocol [29], and QoS, which is a system issue.

9.5.1. TCP Performance in a Wireless Environment

Similar to other types of wireless links, the physical characteristics of the UWB physical layer may impact the performance realized by the transport layer, specifically TCP.

RF interference anywhere in the large amount of bandwidth occupied by a UWB system may lead to signal degradation and, thus, bit and frame errors. Although interference mitigation techniques can be used at the physical and data link layers, some data packets may still be lost. For example, with adaptive techniques it takes time to sense and react to changes in channel conditions. TCP responds particularly poorly to packet loss because it is designed to treat packet loss as an indication of congestion [30]. This is a reasonable assumption in typical wired networks, but is often not valid for wireless links. Multiple packet losses cause TCP to time out before retransmitting lost packets. TCP may reduce its TCP congestion window to as little as the number of bytes in one maximum segment size (MSS). This unnecessarily limits end-to-end throughput because the sender must send packets more slowly. After interference mitigation mechanisms take effect in a UWB system, the signal-to-interference ratio (SIR) may return to a desirable level. However, because the size of the TCP congestion window is reduced, a UWB sender can only send packets at a reduced data rate for a relatively long period of time, which results in wasted bandwidth.

Several improvements can be made to enhance TCP performance in UWB networks. An obvious way is to hide packet errors from TCP. If packet errors can be taken care of at the physical and/or data link layers, TCP will not be aware of the errors and, thus, will not invoke congestion control. However, retransmission at the data link layer poses another problem [31, 32]. TCP estimates the round-trip time (RTT) of a given packet based on previous samples of RTT. The value of the TCP timer, called the retransmission time-out (RTO) interval, is calculated based on the estimated value of RTT. Because errors in a wireless channel tend to be bursty, the estimated RTT may not be a good predication for a sudden increase in RTT due to retransmissions in a data link layer that uses an ARQ scheme for error recovery. In a wired network, a sudden increase in the RTT value is very likely due to congestion at intermediate routers. Therefore, it is reasonable for TCP to act accordingly, specifically by reducing congestion window size. However, in a wireless network, sudden channel degradation that is caused by fading or interference may lead to retransmissions at the data link layer, which affects estimates of the RTT value by TCP. If the errors are not frequent, the RTO value will not account for RTT variations due to data link layer retransmissions. Unaware of what is really happening at the lower layers, TCP may unnecessarily invoke congestion control after a time-out occurs. If the RTO is set to a small value, packet retransmission at the data link layer may trigger congestion control at TCP. Setting the TCP time-out to a larger value, however, may lead to a slow response, that leads to congestion in the network and, thus, slow recovery from congestion losses. Therefore, the value of RTO in a wireless environment should be calculated carefully to account for both interference losses and congestion.

The second way to improve TCP performance is to split each end-to-end TCP connection into two or more separate tandem connections [5, 33]. Because a UWB network is likely confined to a small region, a UWB radio link is likely to be the last (or first) hop in an end-to-end path from sender to receiver. Thus, for a UWB network, a single TCP connection can be split into two tandem connections, one on the wired portion of the route and one over the wireless portion.

By splitting a single TCP connection into two separate connections, each can have different flow control and congestion control mechanisms. Also, maximum packet sizes and time-out values can be different. Therefore, the implementation of TCP or some other transport protocol over the wireless connection can be tuned to perform well in a wireless environment. In addition, error recovery can be much faster due to the relatively short RTT on the wireless link. However, this scheme may require modifications to TCP, which often is not desirable or feasible.

The third scheme is to use explicit notification. Much research has focused on improving the performance of TCP using explicit notifications [34, 35]. If a node determines that the packet loss is due to a transmission error instead of congestion in the network, the node can inform the sender using an explicit notification. Thus, the sender just retransmits the packet without invoking congestion control.

All of these schemes require modifications to the conventional data link layer or to TCP. Because such modification of core network protocols requires updates of base stations and/or routers, all three schemes mentioned have yet to be implemented on a large scale.

9.5.2. Quality of Service Management

A UWB network can be designed to carry different types of data traffic with different QoS requirements. At the application layer, QoS is hard to quantify because it generally refers to the application quality as perceived by the user; for instance, the visual quality and/or the sound quality of streaming video content. QoS provisioning maps application-level QoS requirements onto a unique set of network-level QoS parameters or QoS metrics [36]. Conventional QoS metrics include throughput, packet loss rate, end-to-end delay, and delay jitter. QoS metrics for mobile wireless networks may include more parameters, such as power consumption and network coverage. Power saving is important because a network of battery powered devices will not be able to provide any service if the batteries are exhausted. Similarly, an ad hoc network may want to expand its coverage as much as possible to interconnect more devices. Network-level QoS parameters usually have quantifiable values or bounds that can be determined from the application's technical specifications or from experiments.

In wired networks, the types of network services can be classified into the following three major categories:

1. Best effort service provides no performance guarantees, and the system treats all traffic equally.

2. Better than best effort service does not have any deterministic guarantees, but makes a best effort to support the requested QoS requirements. For example, an application that belongs to a higher priority class will receive better service than an application that belongs to a lower priority class. Differentiated Service (DiffServ) provides such better than best effort services [37].

3. Guaranteed service delivers the highest quality of service and guarantees network performance metrics in deterministic or statistical terms. For example, the network may guarantee a certain minimum bandwidth provided to an application or guarantee a delay within a specified value. Integrated Service (IntServ) provides guaranteed services [38].

A good QoS management scheme should not only provide QoS, but should also attempt to maximize the utilization of channel resources. QoS management schemes support QoS by considering different aspects of a network, such as resource reservation, admission control, QoS routing, and packet scheduling.

Even though many QoS management strategies have been proposed for wired networks, they may not be suitable for use in a wireless network due to the challenges posed by the wireless environment. Because of the inherent unreliability of a wireless link, it is impossible to guarantee at any time the fulfillment of QoS requirements at the physical layer. Under this condition, adaptation mechanisms need to be implemented at the data link layer or, possibly, also at higher layers to reduce the impact of an unreliable physical layer on QoS as much as possible. In spite of these mechanisms, QoS requirements still may not be guaranteed deterministically in a wireless environment. Instead, QoS will most likely be provided to applications in a qualitative fashion, where applications with higher priority enjoy generally better service than applications with lower priority. Alternatively, the network performance could be guaranteed in a statistical manner. For example, in the long run and under certain channel conditions, packets of a real-time multimedia application can meet their deadlines with a certain probability.

To meet QoS requirements, all protocol layers and network components must cooperate. However in practice, the introduction of QoS in the application layer places the most demands on the data link and network layers. To meet a given QoS requirement, the MAC sublayer needs to solve the problem of medium contention and provide adaptive scheduling and resource allocation, while the LLC sublayer needs to provide reliable communication over the link that can compensate for impairments at the physical layer. Many MAC protocols based on CSMA schemes have been proposed to provide QoS in a distributed wireless network. Examples include the Group Allocation Multiple Access with Packet-Sensing (GAMA-PS) protocol [39] and the Black-Burst (BB) contention mechanism [40]. According to the current IEEE 802.15.3 standard draft, a piconet coordinator (PNC) acts as the central controller and manages QoS requirements within the piconet. Centralized control schemes usually out-perform distributed schemes, because the central control node has a global view of the network. This type of piconet architecture is similar to that of a traditional cellular network where a base station manages scheduling and resource reservation in a cell. However, there is a major difference between a UWB piconet and a cell in a cellular network. In a UWB piconet, a PNC only provides control functions and does not relay data traffic for devices, whereas a base station in a cellular system both controls scheduling and resource allocation and relays traffic for mobile terminals. Due to this major difference, the scheduling and resource allocation protocols for a UWB piconet can be quite different from those designed for cellular networks. For instance, Giancola, et al., investigated dynamic resource allocation schemes for UWB networks [41].

The network layer should be adaptive enough to accommodate different data traffic characteristics and QoS requirements. Much research has focused on QoS routing, which refers to the discovery and maintenance of routes that can satisfy QoS requirements under given resource constraints. Several QoS routing algorithms have been proposed with a variety of QoS requirements and resource constraints. Some examples include CEDAR [42], Predictive Location-Based QoS Routing [43], and Localized QoS Routing [44]. However, despite this research, QoS routing remains a challenging problem. QoS routing requires frequent updates of link state information, such as delay, available bandwidth, and loss rate, to make policy decisions. The frequent updates may result in prohibitively high control overhead for an ad hoc network with limited bandwidth. In addition, the dynamic nature of wireless ad hoc networks makes maintaining precise and consistent link state information extremely difficult [45].

General QoS support in a multihop ad hoc network is difficult, if not impossible, especially if nodes are mobile or the wireless environment is highly dynamic. Much of the recent research in this area has focused on methods where lower layers provide QoS for the application layer. The resulting protocols are often complex, and practical implementation is difficult or even infeasible. A more realistic approach is based on an adaptive QoS model, which requires applications to adapt to the time varying resources provided by the network. Kumwilaisak, et al., proposed a QoS mapping architecture where QoS requirements are adapted to wireless link variations [46]. In addition, QoS requirements at the application layer can be varied according to other indicators from the network, such as congestion levels detected at the link layer or the network layer.