Introduction
The Local Multipoint Distribution Service (LMDS) is a terrestrial cellular broadband technology operating in millimeter-wave bands, such as the 28 GHz band in the United States or the 26 GHz band in Europe. In a 1 GHz band, 850 MHz are used for 40 MHz downstream channels and 150 MHz for 2 MHz upstream channels. The actual bit rate of the channels depends on the adopted modulation technique. The LMDS system differs from mobile cellular systems in that it uses fixed links between a multidirectional hub or head-end, and a number of dispersed fixed subscriber terminals. It also requires line-of-sight (LoS) communications since the millimeter waves form pencil-like beams that can easily be attenuated and/or reflected by physical obstructions. An LMDS network typically comprises multiple overlapping cells of 2–6 km diameter each. A cell basically consists of a head-end with an omnidirectional or sectorized antenna and a number of subscriber terminals, each with a directional antenna.
This technology provides a low-cost, easily scalable solution for rapid introduction of broadband communication services in residential and business areas.
In this article we investigate the potential of the LMDS system to support voice and IP traffic over asynchronous transfer mode (ATM), which are representative forms of stream and elastic traffic profiles, respectively. In particular, we consider an LMDS system implemented according to the Digital Audio Video Council (DAVIC) 1.4 specification [1]. The specification contains definitions of the radio interface and an ATM-oriented medium access control (MAC) protocol for LMDS.
The MAC Protocol
In DAVIC's specification, the head-end is referred to as the air interface unit (AIU) and the subscriber terminals as network interface units (NIUs). Both the downlink from the AIU to the NIUs and the uplink in the reverse direction are partitioned into a number of frequency channels. Downstream channels are slotted time-division multiplexed (TDM) point-to-multipoint channels managed by the AIU. The slots are arranged into MPEG-2-compatible frames, which may encapsulate ATM cells. On the other hand, upstream channels -- shared by the NIUs in the LMDS cell -- are slotted time-division multiple access (TDMA) multipoint-to-point channels, where the frames are formed by sequences of a fixed number of slots, each carrying an ATM cell. Downstream and upstream frames have the same duration in order to ensure correct synchronization between NIUs and the AIU.
Due to the shared nature of upstream channels, a MAC protocol is necessary to coordinate the transmissions of NIUs on these channels. The protocol must provide collision-free channels to higher layers with support for classes of traffic with different timing requirements.
The MAC protocol proposed in the DAVIC specification defines a bandwidth request and allocation mechanism based on a hybrid contention/reserved slot assignment scheme. The protocol defines contention slots that carry MAC messages or higher-layer data, and reserved slots that carry operation, administration, and maintenancy (OAM) cells or higher-layer data. Since contention slots can be randomly accessed by NIUs sharing an upstream channel, collisions may occur when more than one NIU attempt to use the same slot. In particular, contention slots are used by NIUs to send dynamic bandwidth requests of their active ATM connections to the AIU. Upon successful reception of a bandwidth request, the AIU allocates bandwidth capacity to the requesting NIUs in the form of reserved slots in the upstream frame. The distribution of contention and reserved slots in each upstream frame is communicated by the AIU to all the NIUs in the cell via a downstream broadcast channel.
A MAC Protocol Simulator
We developed a simulator to study the performance of DAVIC's LMDS MAC protocol. The objective of this simulator is to determine the expected quality of service (QoS) guarantees for different types of traffic, mainly in terms of average access delay and throughput in upstream channels.
The Communication Model
We adopted a communication model based on the evidence that upstream channels are the main bottleneck of the LMDS system due to their shared nature. Consequently, the simulator considers only upstream channels. Furthermore, assuming that the performance of each upstream channel is independent of the others, we further simplified our model by simulating a single upstream channel. For the sake of simplicity, an ideal channel with negligible loss ratio is considered.
All the NIUs share the same upstream channel and are located at the same distance from the AIU, so they are ideally time-synchronized. Each NIU may have one or more ATM connections with the AIU to transmit messages obtained from independent traffic sources (i.e., there is one source per connection). The sources represent high-level user traffic and work according to models described latter in this article.
The DAVIC specification only defines the type and structure of MAC messages used to establish, maintain, and tear down connections, whereas the choice of some algorithms and techniques needed in the protocol is open to equipment manufacturers and/or operators. Consequently, we selected specific implementation options for these algorithms and techniques.
Contention Resolution Algorithm -- Different contention resolution algorithms (CRAs) have been proposed in the literature as part of MAC protocols to maximize the network throughput and limit access delay by resolving collisions efficiently (e.g., Binary Exponential Backoff and Ternary Tree [2]). We propose to use the CRA presented in Fig. 1. The algorithm is executed by an NIU when it is involved in collisions of contention slots in order to spread out in time the transmission of concurrent bandwidth requests.
Burst Merging -- The AIU reserves slots to an NIU upon successful reception of a bandwidth request. The number of reserved slots per upstream frame is calculated according to the negotiated rate of the requesting connection (sustainable cell rate for variable bit rate, VBR, connections; minimum cell rate for guaranteed frame rate, GFR, connections; explained later). Bandwidth reservation is carried out until the AIU receives an unutilized reserved slot from the connection to which the slot was allocated.
Bandwidth requests use contention slots to reach the AIU, thereby being subject to collisions and retransmissions. Intuitively, the average access delay of cells to the upstream channel may be lowered if the number of contention-based accesses is somehow reduced.
In the simulator, when a new burst of cells is generated by a source and the previous burst has not been fully transmitted (i.e., there are still queued cells), the cells of the new burst are merged in the buffer with the cells of the previous burst, so the new cells continue to use the bandwidth allocated by the previous burst without sending a new reservation request. We call this technique burst merging.
Number of Contention Slots in the Upstream Frame -- There are several possible strategies to determine the number of contention slots in the upstream frame, for example:
- A constant number of slots
- A dynamically calculated number of slots according to the number of collisions detected by the AIU
- A number of slots proportional to the number of NIUs in the system
The first alternative is clearly inefficient when dealing with bursty traffic because it does not take into account the dynamic nature of bandwidth requests generated by the sources. In the second case, the strategy is a function of the number of collisions such that the proportion of contention slots under very bursty traffic would be systematically high regardless of the number of actual connections. Furthermore, this approach would have a negative impact on the frame space available for reserved slots. We chose the last alternative as a trade-off between the two previous strategies. The proportion of contention slots was fixed to 15 percent of the number of NIUs (i.e., there is 1 contention slot per 6.66 NIUs in each upstream frame).
Traffic Sources
Different ATM bearer services may be used for voice transmission depending on the type of codecs used. Constant bit rate (CBR) service is well suited for pulse code modulated (PCM) voice, whereas variable bit rate (VBR) service is more convenient when compression, silent suppression, and/or other voice processing techniques are used by the codec, as in adaptive differential PCM (ADPCM) or low-delay code excited linear prediction (LD-CELP). In the simulator, we consider ADPCM-coded [3] voice sources using VBR bearer service. We are not interested in PCM-coded sources using CBR service since bandwidth in this case is statically allocated, so the influence of the sources on the dynamic behavior of the MAC protocol is marginal.
The traffic generated by voice sources is transported over VBR connections, which are modeled in the simulator with on-off sources. The connections are described by a sustainable cell rate (SCR), a peak cell rate (PCR), and a burstiness parameter B = (Ton + Toff)/Ton ≥ 1, where Ton is the average time during which ATM cells are generated at PCR and Toff the average inactivity time. The set of cells generated during a Ton period is called a burst. Ton is an exponentially distributed random variable with mean 1/
= A*min(Ton), where min(Ton) is the time necessary to generate an ATM cell (i.e., 48 bytes of payload) at PCR. The constant A = 60 is empirically introduced to avoid a potentially overwhelming number of burst merges.
The average rate of a VBR connection is in the worst case equal to its SCR, so the relationship among the connection's parameters is SCR = PCR/B = PCR* Ton /(Ton + Toff). We deterministically deduce from this relation the value of Toff for a given Ton: Toff = (PCR * Ton /SCR) – Ton. The idea behind this calculation of Ton and Toff times is to model ill-behaved VBR sources.
With respect to IP traffic, it is well known that ATM's support of IP is not an easy task. It appears to be impossible to support IP traffic with open-loop congestion control schemes. The definition of the available bit rate (ABR) ATM service was an initial response to this evidence. Based on a closed-loop scheme, ABR provides explicit network feedback able to influence the traffic generator. However, this service is very complex to implement in practice; therefore, a new service tailored to the elastic nature of IP traffic (i.e., the source can increase/decrease its rate) is proposed. The service, GFR [4], is frame-oriented and requires minimal interaction between terminals and the ATM network while providing a certain level of service guarantee. Its essence is to guarantee to elastic connections a minimum cell rate (MCR) with fair access to spare bandwidth. In the simulator, we model a GFR bearer service for IP traffic.
Our IP traffic source is based on an empirical model used in [2]. The packets' interarrival times are assumed to be exponentially distributed with mean 1/, and packet lengths distributed as in Table 1. The average packet size is 368.1 octets.
The IP packets generated by the source are encapsulated into link layer control, LLC/SNAP, frames and then in ATM adaptation layer type 5 (AAL5) frames. The resulting frames are segmented into ATM cells.
Scheduling and Buffer Management
We implemented a scheduling and buffer management architecture in the simulator aimed to support specific QoS levels. A scheduler at the AIU allocates contention, reserved, and free slots (i.e., spare bandwidth) to NIUs in each upstream MAC frame according to the following policy:
- A fixed number of contention slots is first allocated.
- Reserved slots are then assigned to individual NIUs to satisfy the bandwidth requests of their active connections.
- Finally, if there are still free slots (i.e., not allocated as either contention or reserved slots) available in the upstream frame, they are proportionally distributed among the active GFR connections according to their MCR (i.e., connections with higher MCR get more free slots).
It must be pointed out that, even if the AIU is able to discriminate individual connections coming from the same NIU, it is unable to allocate slots on a per-connection basis. The reason is that the MAC message defined in the DAVIC specification to notify the NIUs of slot assignments does not contain a field to specify the connection ID.
Figure 2 shows the scheduling and buffer management mechanisms implemented in an NIU. Per-VC queuing is used for both VBR and GFR connections. This approach allows better traffic control, so rate and delay constraints can be enforced.
The scheduler at the NIU implements a service discipline to distribute the slots allocated by the AIU among its active connections. If the NIU has slots allocated by the AIU but all its queues are empty, the NIU sends empty cells in those slots to the AIU on behalf of its connection with the most allocated bandwidth. Upon reception of these cells, the AIU reduces the bandwidth allocated to the NIU by an amount equal to the bandwidth of the connection selected by the NIU.
Results
The goal of our simulations is to determine the number of simultaneous voice or data connections that can be supported in an upstream channel at particular QoS levels. QoS is considered in terms of the average access delay of cells to the upstream channel and the aggregated throughput of the connections in the upstream channel. For each simulation, performance was measured according to a fixed number of connections during the entire simulation cycle. All the results were calculated from simulations of 2 min of real behavior of the system. We adopted as a reference an upstream channel rate of 4124 cells/s, that is, 1.58 Mb/s of payload. We defined independent scenarios for voice and IP traffic. An ITU-T G.726/G.727 ADPCM codec [3] was considered for voice sources. With this codec the source, after AAL2 encapsulation, generates an average rate of 14.7 kb/s and a peak rate of 36.8 kb/s.
Figure 3 shows the average cell access delay to the upstream channel as a function of the number of voice connections, assuming a single connection per NIU. A strict two-way end-to-end delay bound of 
15–30 ms is required for voice calls using analog terminals, which includes 95 percent of all phones [5]. Considering an upper bound of 12 ms for the cell access delay in the LMDS segment of an entire ATM network, we note that an upstream channel may support up to 50 voice calls.
For the performance evaluation of IP traffic using GFR service, we defined a reference IP source that generates an average rate of 79.188 kb/s (i.e., 5 percent of the upstream channel capacity). The average rate of an IP source is the ratio of the average packet size to its average packet interarrival time.
We first evaluated allocation of spare bandwidth among connections by measuring the throughput attained by the sources when different values of MCR are used. We defined a first scenario where the MCR guaranteed to each GFR connection was 39.594 kb/s (i.e., 50 percent of the source's average rate). In a second scenario, the MCR was set to 79.188 kb/s, 100 percent of the average rate. In Fig. 4 we observe the evolution of the aggregated throughput of a set of connections as a function of the number of connections, assuming a single connection per NIU. Of the two lines with the plus symbol (+), the dotted line represents the sum of MCR when 50 percent of the sources' average rate is guaranteed and the solid line the throughput actually attained by the set of connections (first scenario). The lines with empty circles represent the sum of MCR and throughput when 100 percent of the average source's rate is guaranteed (second scenario). The lower throughput attained when only 50 percent of the sources' rate is guaranteed is explained by the fact that free slots (i.e., spare bandwidth) are allocated with lower priority than reserved slots. Recall that the AIU allocates reserved slots to a connection until it receives a slot unused by the connection. These unused reserved slots represent wasted bandwidth that cannot be "recycled" as spare bandwidth. Clearly, the wasted bandwidth increases proportionally with the number of GFR connections.
Figure 5 shows the average cell access delay as a function of the number of connections when transporting IP traffic. The larger values observed in this figure than those in Fig. 3 are explained by the fact that a relatively large number of the packets generated by the IP source are small (60 percent are 64 octets long). Since there is a bandwidth request for each generated packet (if the packet is not burst-merged with the previous one), a large number of bandwidth requests -- subject to contention -- must reach the AIU. This increases the average cell access delay.
Finally, we investigated the effect of multiplexing several connections in an NIU. We defined a scenario with eight NIUs generating an aggregated traffic equal to 20 percent, 40 percent, and 80 percent of the upstream channel capacity. The traffic of each NIU was uniformly distributed among its multiplexed connections, so the negotiated rate of each connection (the SCR for VBR connections and MCR for IP connections) is C**/(NNIU*Ncnx), where C is the upstream channel capacity (1584 kb/s), * the aggregated traffic (20 percent, 40 percent, or 80 percent of C), NNIU the number of NIUs sharing the upstream channel (fixed to 8), and Ncnx the number of connections per NIU. For VBR connections, burstiness was fixed to B = 2.
Figure 6 shows the average cell access delay as a function of the number of VBR connections multiplexed per NIU. Considering again an upper bound of 12 ms for the cell access delay in the LMDS segment of the entire network, we observe that the system supports a large number of multiplexed VBR connections per terminal when the total offered traffic in the channel is 20 percent. However, with offered traffic of 40 percent or more, the average cell access delay increases at levels that render the transport of voice traffic infeasible.
Figure 7 shows the average cell access delay for multiplexed IP connections. In this case, average cell access delay is even higher than in Fig. 5 because the number of contention slots per upstream frame (those used to send bandwidth requests) is proportional to the number of NIUs, not to the number of connections. As more connections are multiplexed per NIU, the share of contention slots per connection decreases. Consequently, its chances to successfully transmit bandwidth requests decrease too.
Summary and Conclusions
We present the results of our investigations concerning the potential of the LMDS system to support voice and IP traffic over ATM. Concretely, we evaluate the performance of the ATM-based MAC protocol proposed in the DAVIC specification for the LMDS system when transporting this type of traffic. Considering an LMDS system with 75 upstream channels per cell, our results show that up to 3750 simultaneous voice calls can be supported in an LMDS cell. On the other hand, we observed that the proposed MAC protocol is not particularly well suited to IP traffic. Actually, the protocol can be improved to support the more dynamic bandwidth requirements of IP traffic by incorporating efficient dynamic bandwidth request mechanisms like piggybacking and minislots.
Efficient scheduling algorithms, buffer management techniques, and slot allocation strategies adapted to multiplexed voice and IP traffic are the subject of ongoing work. In the case of stream traffic (e.g., voice), particular attention is given to the control of second-order performance metrics such as delay variance, generically denoted as jitter. Our ultimate goal is to optimize upstream channel utilization while guaranteeing stringent QoS constraints of inherently different traffic types.
Acknowledgments
This article is based on our previously published material [6] from WAS 2000 organized by the DELSON GROUP. The work was partially supported by CONACyT grant # 122688. The authors acknowledge the valuable support of Guest Editor Willie Lu.
References
[1] DAVIC Spec., vol. 1.4, part 8, "Lower Layer Protocols and Physical Interfaces," http://www.davic.org/down1.htm
[2] D. Bonnarigo, M. de Marco, and R. Leonardi, "A Comparison of Back-off and Ternary Tree Collision Resolution Algorithms in HFC Access Networks," Proc. GLOBECOM '98, Sydney, Australia, 1998. pp. 45–50.
[3] ITU-T Rec. G.726, "40, 32, 24, 16 kbit/s Adaptive Differential Pulse Code Modulation (ADPCM)," 1990.
[4] ATM Forum Tech. Spec. af-tm-0121.000, "Traffic Management 4.1," 1999; ftp://ftp.atmforum.com/pub/
approved-specs/af-tm-0121.000.pdf
[5] H. Le Pocher et al., "An Efficient ATM Voice Service with Flexible Jitter and Delay Guarantees," IEEE JSAC, vol. 17, no. 1, Jan. 1999, pp. 51–62.
[6] J. Kuri and M. Gagnaire, "ATM Traffic Management in a LMDS Wireless Access Network," Proc. WAS 2000, San Francisco, CA, 2000, pp. 170–77.
Biographies
Josué Kuri [S] received his M.S. in computer networking from Ecole Supérieure d'Electricité, France, in 1999. From 1996 to 1998 he was with InfoSel SA as a software engineer working on the design and development of real-time systems. He is currently pursuing his Ph.D. at Ecole Nationale Supérieure des Télécommunications (ENST), Paris, France. His areas of active research are the design and dimensioning of core optical networks, and modeling and performance evaluation of access methods for broadband access networks.
Maurice Gagnaire [M] is an associate professor at the Ecole Nationale Supérieure des Télécommunications (ENST), Paris, France. He graduated from the Institut National des Télécommunicationsm Evry, France. He received his Diplôme d'Etudes Approfondies from the University Paris 6, his Ph.D. from ENST (1992), and the Habiliation from the University of Versailles, France (1999). His current research activities are focused on the design of IP over WDM backbone networks and the performance evaluation of MAC protocols for high-speed local loops (PON, LMDS). He is co-author of a book on high-speed networks and author of a book on new broadband access networks (in French). He is on the program committee of various IEEE and IFIP conferences. He is appointed as an expert by the Flemish Government.
