May 2002

Challenges In Providing Carrier-Grade Telephony Over Broadband Wireless Networks

BY TOM FLAK & MICHAEL STUMM

Voice telephony is really a killer IP application � especially as it pertains to the newest generation Broadband Wireless Access (BWA) systems. If implemented properly, voice telephony can generate excellent revenue: With toll quality and standard CLASS features, telephony can double the $40-$60 revenue obtainable from offering broadband IP service. At the same time, voice telephony will consume only a relatively small amount of the scarce and expensive wireless spectrum resources. It is for these reasons that service providers considering deployment of new broadband wireless access technology are increasingly viewing voice as critical to a profitable business case.

Since many of the new generation BWA systems were primarily designed to support broadband IP data, the idea of supporting voice by running a standard VoIP protocol, such as SIP or H.323, over a wireless data transport layer seems compelling and obvious. However, VoIP over data-oriented broadband wireless exhibits a number of problems in practice, including unacceptable latency and jitter, as well as unacceptable overhead. Simply layering a VoIP protocol over a data-oriented wireless IP pipe will result in an uneconomical system because firstly, voice will be of sub-standard quality, in which case it will be difficult to generate adequate revenue from the service, and furthermore, because the system will need to be over-provisioned to attain even reasonable quality, in which case deployment costs will be too high. Supporting quality voice in a BWA system is challenging because it needs to be engineered into (and affects) every part of the system and, in particular, all layers of the communication protocol stack.

NEW BWA SYSTEMS AND VOICE
The newest generation BWA systems have a number of characteristics that make them much more attractive to service providers than previous generation systems: they offer multi-megabit peak IP connectivity; they have no line-of-site requirements, eliminating the need for truck-rolls and potentially allowing nomadic usage; they are self-provisioning, significantly lowering cost of operations; and they have reasonably low-cost customer premises equipment (CPE). As a result, roll-out costs are much lower, making these systems much more viable for mass deployment in the residential market. By offering broadband IP connectivity at high data rates, these systems allow service providers to compete with existing DSL and cable infrastructure and can offer a solution for areas where DSL or cable is not available or economically viable.

These systems make use of two hardware components: CPE or wireless gateways located in the home or business to which a local-area network or PC is connected; and a base station (similar to those found in cellular mobile systems). The CPE communicates with the base station over wireless communication channels often referred to as the air interface, and the base stations are connected with each other through an IP backbone network, to the PSTN through a special purpose PSTN gateway, and to the public Internet through an IP router.

A key challenge in supporting quality voice is keeping the end-to-end delay below the acceptable 200ms, taking into account vocoder processing delays, packetization delays, queuing delays, modem delays, radio frame wait time, serial transmission time, IP router and switch forwarding delays, receive jitter buffer delays, and PSTN delays. Jitter must be kept to a minimum so that jitter buffers can be kept small, limiting jitter buffer delays.

One of the reasons it is more difficult to support toll-quality voice over IP-based BWA systems is that the wireless component of the IP network has vastly different characteristics than typical IP networks. The wireless link has far higher error rates and entails larger delays than direct, land-line links. Because the base station can transmit at much higher power levels than the CPE, these systems are asymmetrical in that more bandwidth is available on the downlink than on the uplink. In either case, the wireless link has limited bandwidth compared to the rest of the network � less than the home LAN and far less than the core IP network � and it is by far the most expensive part of the network. (Obtaining spectrum licenses is extremely expensive and unlicensed spectrum is unsuitable for applications requiring QoS guarantees.) For these reasons, the wireless link tends to run at much higher utilization, and contention is often the norm, not the exception.

OPTIMIZATION FOR VOICE OVER BWA
Limiting end-to-end delay to acceptable levels requires specific optimizations at each layer of the communication protocol.

At the physical, air interface (AI) layer, it is advantageous to use a separate, dedicated air interface channel for each voice connection, yet also support Broadcast Data Channels (BDCs), or �fat pipes,� for IP data transmission. Having a separate AI channel per voice is the traditional, connection-oriented approach used by system designers from a voice-centric background. Using separate channels for voice allows for lower delays in transmitting voice packets and guarantees QoS for the voice connections, but systems that only support voice channels do not support the transmission of IP data well, because voice channels do not have sufficient capacity for broadband data and because they do not fit the connectionless IP model. On the other hand, systems with only BDCs do not support voice very well in part because they are likely to operate under heavy contention and introduce extra delays, and because they do not support some of the optimizations necessary.

Hence a hybrid approach, supporting both dedicated voice channels and BDCs, is the best choice. The downside of a hybrid approach is that a higher modem density is required at the base station to support the extra channels, and a much more sophisticated Radio Resource Management (RRM) subsystem is required to (1) set up and tear down voice channels when calls are made and terminated, and (2) to continuously dynamically resize channels to adjust the size of the BDCs to use up all of the extra capacity for data and to accommodate different vocoders.

Having separate, dedicated channels at the physical layer for voice allows adjustment of the bit error rate by customizing ARQ (automatic repeat request) and FEC (forward error correction), two common parameters used in air interfaces. Because voice has higher tolerance for bit errors and dropped packets than data, these parameters can be tuned so as to minimize overhead. (For voice, it is better to tolerate occasional bit errors than to incur extra delay due to retransmissions.) Systems that do not distinguish between voice and data at the physical layer must apply the same coding level for all packets and therefore incur significantly higher overheads.

At the link layer, a number of optimizations are necessary. One example is the need for packet preemption, fragmentation, and reassembly for over-the-air transmission, since large IP packets can unacceptably delay voice packets. For instance, if 200 Kbps uplink capacity is available, then one 1,500 byte packet can delay voice packet by 60ms, which when added to a 60 ms jitter buffer would be two-thirds of the 200ms end-to-end delay budget. Hence, when a voice packet is available for transmission, it must be able to preempt an ongoing transmission of a large IP packet.

Another example of an optimization needed at the link layer is packet header compression. To achieve good quality, typical voice packets contain 10�20 bytes of payload, but standard RTP/UDP/IP packet headers add up to about 40 bytes. This results in 66�80 percent overhead. Standards exist for compressing headers down to a few bytes, but if voice packets are treated separately, then it is possible to compress the header down to a few bits for over-the-air transmission.

At the IP layer, using an appropriate queuing policy is critical. Weighted fair queueing is a reasonable strategy for data packets, but it is not good for voice. Voice packets need to have absolute priority over all data packets. A workable solution is to use absolute priority for voice packets, and a WFQ policy for data packets, but this requires differentiating between voice and data at the IP layer.

Another issue at the IP layer is the treatment of the TOS-byte in the IP header. If a SIP phone or H.323 phone is used to connect (via IP) to the CPE device, then the TOS-byte is often set by the phone to identify the higher-priority voice packets. However, using the TOS-byte to identify higher-priority voice packets is problematic in real world situations, since other applications can also set this byte to obtain high priority. For example, it is not difficult for hackers to be TOS-byte cheaters, setting the TOS-byte for all packets sent from their application (or computer) to obtain better response times for the distributed game they are playing or even for Web browsing. Since this can be detrimental to voice quality, being able to differentiate between voice and data at the IP layer allows the monitoring and control of the TOS byte to guarantee voice the right level of service.

TYING IT ALL TOGETHER
Besides requiring optimizations at all levels of the communication protocol, it is also necessary to have various layers communicate with each other to affect policy. For example, when a call first gets set up, air interface resources are typically reserved assuming an 8 kbps vocoder. But the application may notice that a fax or a modem is being used instead of a telephone and the 8 kbps vocoder, which is not suitable for fax or modem transmission, is dynamically replaced with, say, a 64 kbps vocoder. In that case, the air interface must be notified so that it can resize the channel for increased capacity. As another example, to ensure that a 911 call can get through, it may be necessary to preempt an existing voice connection.

In the other direction, the air interface layer may notice that overall capacity is reduced, say due to the outage of a radio unit. In this case, it needs to be able to inform the application (voice) layer that all connections need to reduce their vocoder rates so as to continue allowing phone calls, albeit with reduced quality. Similarly, the air interface layer will also need to inform the application layer to reduce vocoder rates when the system is running at overcapacity, for example, during holidays or regional emergencies.

CONCLUSION
Voice, if done right � that is, toll-quality with all of the expected CLASS features � can generate tremendous revenue, while consuming only minimal over-the-air bandwidth. However, doing voice right is extremely challenging, especially in keeping end-to-end delays down to the accepted 200ms. Using a standard VoIP protocol can work well in land-line IP-based communication systems in part because these systems tend to be over-provisioned. However, when wireless links are involved, numerous optimizations specific for voice (as described) are required at all layers of the system.

Michael Stumm is co-founder, CTO, and senior vice president at SOMA Networks, and professor of computer engineering at the University of Toronto. Thomas Flak is vice president of marketing at SOMA Networks. Founded in 1998, Soma Networks offers the latest advances in fixed wireless, distributed computing, and Internet technologies in a unique system designed to allow anyone to become a full-service telecommunications service provider offering a feature-rich package of carrier-grade voice and broadband Internet services to the residential and small-office market. For more information, please visit the company online at www.somanetworks.com.

[ Return To The May 2002 Table Of Contents ]