A Framework for supporting VoIP Services over the Downlink of an OFDMA Network Patrick Hosein Huawei Technologies Co., Ltd. 10180 Telesis Court, Suite 365, San Diego, CA 92121, US Tel: 858.882.0332, Fax: 858.882.0350, Email: phosein@huawei.com Abstract The fourth generation of wireless networks (4G) are presently being discussed in standards bodies (3GPP2, 3GPP, 802.16, 802.20) and one common conclusion is that the most appropriate technology for the forward link is Orthogonal Frequency Division Multiple Access (OFDMA). In addition, it is believed that all services (i.e., both real-time services like voice and also data) can be supported over such a channel using an all- IP network. However, even for 3G networks, the efficient support of services such as Voice over IP (VoIP) over wireless packet switched channels is still to be verified in the field. In this paper we look at the issues involved in transporting VoIP over the forward link of an OFDMA network and propose a framework that can address many of these issues. We support our proposals with numerical results. I. INTRODUCTION The next generation of wireless networks (4G) is being developed to support resource intensive applications (high speed data, real time services etc.) over a wider range and with greater spectral efficiency than the present generation [1], [2]. In order to efficiently utilize the finite radio resources available in such networks, advanced control and resource management structures are needed. Both forward and reverse link resources are managed by the Base Station (BS) and signaled to the Subscriber Station (SS). The rate and granularity of this signaled information determines how efficiently the valuable radio resources are controlled. However, they also determine the overhead penalty paid on the forward link. In addition to the signaling efficiency we need to address the efficiency of the traffic channels. In this paper we investigate these issues for the case of VoIP services over an OFDMA forward link. We propose a framework that can address these issues. Although many papers have been focused on efficiently supporting data over OFDMA channels [3] [7], there are fewer results on VoIP performance. The capacity of VoIP services over CDMA networks have been addressed in [8]. In this paper we plan to address similar issues for OFDMA networks. We provide the framework for our discussion and and for each component of the framework we provide supporting analysis or numerical results. One of the most difficult applications to support over a packet switched channel is voice. An interactive voice stream exhibits an ON/OFF nature, ON while the person is speaking and OFF (or low bit rate) when the person is listening. During the ON period the rate requirements do vary but is limited by the full rate (typically 9.6 kbps) offered by the codec. In the OFF period sufficient information must be conveyed to maintain the background noise information. Studies have shown that full rate frames occur roughly 30% of the time while one eighth rate frames (silence period) occurs 60% of the time with the other 10% being a combination of one half and one quarter rates. Hence sufficient resources must be allocated during both ON and OFF periods for acceptable performance but also no more than necessary must be provided in order to to achieve maximum capacity. In traditional CDMA networks, voice channels are code division multiplexed with each channel individually power controlled. Hence, statistical multiplexing in the power domain allowed reasonable radio resource efficiency. In 1xEV-DO, which was originally designed for data applications, multiple users must share a slot for efficient channel utilization. However this introduces additional complexities and inefficiencies (all served users must be in similar radio conditions and the packet must be retransmitted until all users correctly decode it). Furthermore, the delay constraints that are imposed by the VoIP QoS requirements further restricts the diversity gains that can be achieved. In this paper we try to more closely match the resource needs of VoIP with the resources available in the OFDMA forward link (power, bandwidth and time). We make use of the fact that OFDMA can take advantage of frequency diversity gains and so need not depend on user diversity gains (opportunistic scheduling) as is done in 1xEV- DO. This difference is especially important for delay sensitive traffic. The issues involved in supporting VoIP can be summarized as follows (a) the packets are delay sensitive and hence must be transmitted within an acceptable delay period (and this means that limited user diversity gains are achieved), (b) the voice packets are small and hence (unless frame sharing is used) results in small coding gains, (c) the required throughput is low and hence the user capacity is large (compared to data) and thus requires significantly more signaling overhead. In the rest of the paper we provide some ideas for addressing some of these issues. At present, three standards bodies (3GPP2, 3GPP and WiMAX) are investigating the support of VoIP over OFDMA. Unfortunately none of these are complete and hence performance results are not yet available in the literature. Throughout the document we provide analytic arguments for the potential performance gains of each concept that is proposed. We also
provide numerical results but due to the lack of published simulation results no comparisons are performed. II. FRAME BUNDLING VoIP traffic consists of periodic, small payload packets. Transmission of such small payloads over a wireless packet switched channel tends to be inefficient. In the case of 1xEV- DO, voice frames from multiple users are combined and served in a slot (multi-user packets) to overcome this problem. However this method itself introduces some problems as mentioned previously. Simulations have shown that, under heavy loading, the average number of users served in a packet is between 2-3. In the case of OFDMA, one can instead take advantage of the frequency diversity and so be able to provide periodic service to a user. However, the small payload once again becomes an issue and so multi-user packets have again been recommended. We propose an alternative to avoid some of the issues with multi-user packets. From our experience with providing VoIP service over 1xEV-DO we found that a scheduling delay on the order of 100ms or more was necessary to achieve sufficiently high user diversity gains. This corresponds to a scheduling delay of 5 voice frame periods. We can get similar delay performance if we instead combined 5 consecutive voice frames and served this to the user during the time slot. This will provide sufficiently large payloads for acceptable transmission coding gains. Naturally one can trade framing efficiency and delay by varying the number of frames in a bundle. By bundling a single user s frames and serving the bundled packet in a single 20ms period, we lose user diversity gains (i.e., scheduling each user during their positive fades). However, we instead serve a user s bundled packet over a 20ms period (voice frame duration) and take advantage of frequency and time diversity in this service period. We will denote the bundling period (in voice frame durations) of user i by τ i so that every 20τ i ms the scheduler must serve the τ i accumulated frames in the subsequent 20 ms period. We recommend a default value of τ i = 2 during a talk spurt and τ i = 8 during a silence period. However, we allow this parameter to be more flexible. For the active period this introduces an additional 20ms in the delay which is acceptable if we can show sufficiently large capacity gains. For each user we can vary τ over time. As the number of users increases, we can increase τ for all users. This results in increased end to end delays for users but increases the bundling efficiencies and hence the system capacity. We can also monitor the number of bundled bits per packet. When this exceeds some threshold (determined by the packet size required for reasonable coding gains) we can reduce τ. If however, the number of bundled bits per packet falls below some minimum threshold (this is an indication of a silence period and hence delay is less of a concern) then τ should be increased. Note that for 1xEV-DO, it is recommended that every 8th or 12th voice frame be transmitted during the silence period. The bundling of voice frames of a single user has Full Rate 1 2 3 20 ms Fig. 1. Half Rate Voice Frame Arrivals Repeated transmissions of bundled frame 5 ms varying power Bundled Frame (with voice frame repetition) 1 2 2 1.25 ms Bundling and transmission of voice frames a similar effect as multi-user packets but with less overhead (albeit with increased delay). The silence period is detected when the total number of bundled bits falls below some threshold for some number of consecutive bundled frame periods. When the connection goes into the silence period, resources allocated to the connection are reduced and signaled to the SS. The bundling period can be increased (since delay is less of an issue) and/or the resources allocated to the connection can be reduced. These will be described in more detail when we discuss resource allocation/deallocation. In the proposals presently being discussed in standards bodies, users are served in a periodic fashion as described above. Therefore, bundling can easily be incorporated in such standards since it just implies that service to a user is delayed until sufficient VoIP frames accumulate. Hence there are no specific performance advantage of the method proposed in this section. If a particular VoIP frame is received at less than full rate, then this data is repeated within the encoded packet (two repetitions for half rate, four for quarter rate and eight for eighth rate). Due to the nature of voice, multiple consecutive frames tend to be coded at the same rate and so in most cases the repetition rate will be the same for all voice frames within the packet. The transmission power of the first transmission of the encoded packet is further adjusted based on the least repeated voice frame in the packet. Larger reductions in the initial transmission power are made for larger repetition rates (due to the expected repetition gain). In Figure 1 we pictorially depict the single user bundling and fractional rate repetitions proposed. III. DYNAMIC BANDWIDTH ALLOCATION If we perform bundling, each user will be serviced in a bursty manner. Whenever a new user has to be allocated resources, the BS should assign its servicing period during
a lightly loaded 20ms period. Any additional resources that are subsequently assigned to this SS are done so in the corresponding servicing period. The initial resources allocated to an SS are determined as follows. The encoder packet size is the smallest one that can hold τ full rate frames. Using the most recently reported channel quality information (CQI) we can then determine the transmission power necessary to deliver the packet with the required Frame Error Rate (FER) over a single VoIP channel (this is described later but is simply a specified number of diverse frequency/symbol cells). In present standards proposals, the resources (bandwidth dimension) for a VoIP channel is fixed and the appropriate power needed to provide the required FER is used when transmitting to the user. However, it is also possible to optimize resources in the frequency dimension (i.e., find the optimal number of subcarriers that should be used to serve the user). Because the subcarrier resources are allocated in groups then the best we can do is determine the optimal number of groups to assign to the user. The modeling and solution of this optimization problem is beyond the scope of this paper. However, we instead provide an intuitive argument. Suppose that we assign a fixed amount of bandwidth resources to each user and adjust their power for the required FER. If all users are at the edge of the cell then they each would require significant power. If we try to serve all users that need to be served we run out of power. If on the other hand all users are close to the antenna then they each require little power for service and so the capacity is limited by the number of VoIP channels supported. The system is now bandwidth limited. In the former case (power limited) one can show that capacity can be increased by increasing the bandwidth resources assigned to each user while in the latter case the opposite is true. An in-depth discussion of this optimization problem can be found in [9]. The above implies that users with low channel gains should be provided more bandwidth resources than those with good channel gains. If this is done, then to achieve the same rate the power required for users at the cell edge will decrease while that for users close to the antenna will increase. In other words, the power spread among users provided with the same rate is decreased. We can therefore monitor the power required for a user and increase the allocated bandwidth when the power exceeds some threshold and decrease the allocated bandwidth when the allocated power falls below some lower threshold. The potential gain in performance from this dynamic bandwidth allocation method can be found in [9]. IV. POWER CONTROLLED FRAME TRANSMISSIONS As described above, if the required power exceeds the maximum allowed VoIP subchannel power then an additional VoIP channel is added and the procedure is repeated. As the number of channels allocated increases, the required power is reduced. When it first falls within the allowable power range the corresponding number of channels is noted and assigned for the first transmission to the SS. For subsequent bundled frame transmissions the initial power is adjusted based on the success/failure of previous transmissions (outer loop power control). However the same Modulation and Coding Scheme (MCS) and slot resources are used (and hence need not be signaled). If the required power falls outside the allowed power range then the most recently reported CQI is used to determine the appropriate number of VoIP channels and this information is signaled to the SS. Note that the reported CQI is only used when allocated resources must be adjusted. Furthermore, the CQI is averaged over the entire band and over several physical layer frame periods and hence it tracks the slow fading of the SS. This means that the CQI reporting rate can be significantly reduced (compared to that used for data applications). The inner loop power control (which uses ACK/NAK information) is instead used to compensate for fast fading. As described above, during the servicing period, a packet containing a payload of approximately τ full rate voice frames must be transmitted. We spread this transmission in both the frequency and time domains as follows. For simplicity, assume a frame length of 1.25ms and suppose that four slot Hybrid ARQ (HARQ) interlacing is performed so that retransmissions are delayed by 5ms. This would allow up to four transmissions of the frame during a 20ms period. Let us consider one of these transmissions. By using HARQ we target a maximum of three retransmissions and experience an average of say 2.5 transmissions. So the ACK/NAK of the sub-packets frees the resources of 1.5 transmissions for use by other users. However by spreading this transmission in the frequency domain we increase the accuracy of the estimate of the received signal and hence we can more closely approach the target number of transmissions while maintaining the specified FER. Therefore by using frequency diversity together with power control the average number of transmissions needed can be made close to the target amount (i.e, the variance in the number of transmissions required per frame can be made small) and so the average value approaches the maximum value. If this was done perfectly (i.e. exactly 4 transmissions are always needed) then the ACK/NAK information is no longer necessary since we can assume that four transmissions are always needed. We instead use the ACK/NAK bit for controlling the transmission power and will instead refer to them as the UP/DOWN power control bit. First note that four transmissions are always made (so no early termination gain). When the BS receives an UP it increases the transmission power of the subsequent transmission by some specified factor. If a DOWN is received then it decreases the transmission power by some specified factor. The final UP/DOWN bit is used to report success or failure of the transmission. This information is only used for the outer loop power control algorithm since we do not retransmit the bundled frame at a higher Radio Link Protocol layer. Therefore, the reliability of the UP/DOWN bit need not be as high as the ACK/NAK bit used for HARQ. The SS determines whether to report UP/DOWN by comparing the total signal energy so far received with some specified target. A different target is used for each transmission
attempt. Based on the statistics of the success/failure of the present and past transmission attempts, the SS adjusts the targets. Note that the SS will in effect be re-shaping the target versus time function. The optimal shape will depend on the Doppler of the SS. In addition, the BS adjusts the initial transmission power of the next bundled frame using the success/failure outcome of previous transmissions. The BS adjustment is done to take into account changes in the slow fading (shadow fading). If the FER rises above some threshold then the initial transmission power is increased, while if the FER is below the threshold then the initial transmission power is decreased. In addition, if a new CQI report has been received since the last frame was transmitted then this is also used in adjusting the initial transmission power. Therefore the CQI adjustment will compensate for the average channel conditions while the outer loop power control will will take into account the variation of the fast fading. Also recall that if VoIP frames are repeated within a frame then this is also used to adjust the power. Note that at any point in time the power requirements of each SS will vary. This will provide some statistical multiplexing gains in the power domain. As we will see in a later section, any unused power and bandwidth resources are provided to data users. Unlike the case of 1xEV-DO where the scheduler discards frames that have been in the queue too long, we have no such losses in this case because the queuing (bundling period) and transmission (servicing period) times are deterministic. Hence the voice packet loss rate (PLR) is determined solely by the residual FER of the link. One should however note that a loss of a single physical layer frame results in the loss of multiple consecutive voice frames if bundling is used. V. ALLOCATION OF BANDWIDTH/TIME RESOURCES Next we describe how resources are allocated and deallocated to VoIP users. We first define the basic VoIP channel unit. The number of sub-carrier/symbol pairs (which we call cells) needed for such a channel is determined as follows. We find the least number of cells needed to support a user with the desired FER using the maximum MCS, the minimum allowable power (i.e., p min ) and assuming that τ full rate frames are to be served in four transmissions over a 20ms period. This would be the number of cells needed to serve the best radio condition user during an active spurt. As the user s radio conditions deteriorates, the allocated power will be increased up to the maximum allowable value (p max ) before more cell resources are allocated to the user. Let s assume that M cells are required. A SS will be assigned a VoIP channel which will consist of one or more of these basic VoIP subchannels. A specific region can be reserved for VoIP subchannels or they can be spread throughout the entire frame. We will see that any unused resources can be assigned to non-voip users. If a specific region is reserved then the size and location of the region within the frame must be signaled whenever it is changed (this must be done because VoIP subchannels will be indexed within this region). However, since the size or this region depends on the average number of VoIP users in the sector then it should be changed infrequently (on the order of hours) and the resulting signaling overhead is minimal. The size of this region determines the maximum subchannel index needed and hence the number of bits needed to signal subchannel assignments to the SSs. Assume that a maximum of K VoIP subchannels are supported and hence KM cells are available for VoIP usage. We use log 2 (K) bits to signal the index of the allocated VoIP subchannel. In other words we assume some algorithm that deterministically determines the set of diverse cells making up a subchannel given the subchannel index. The specific algorithm is not important for our discussion (and there are many alternatives) since any algorithm will work with our proposed framework. A VoIP subchannel is allocated to a SS and then signaled using the Signaling Subchannel (SSCH). However, since the subchannel can be shared with τ 1 other SSs, then a time index must also be signaled to let it know which 20ms period must be monitored. Note that this means that a VoIP subchannel can only be shared with SSs using the same value of τ (and typically τ will be limited to a small number of options). Each allocated VoIP subchannel must be signaled to the SS within a SSCH. If the BS signals a subchannel that is already assigned to the SS then this indicates that the SS should drop the subchannel. Note that multiple VoIP subchannels may be assigned to a SS depending on the resource needs of the SS. For simplicity, we assume that multiple subchannels assigned to a single SS all use the same servicing period. The first subchannel allocated to a SS is chosen as follows. The BS finds the lowest subchannel index such that the corresponding subchannel has the same bundling rate and is not full (i.e. at least one time index is free). The BS may consider more than one 20ms servicing periods in making this decision. If this is done then the servicing period that allows the smallest subchannel index is chosen. Whenever an additional subchannel must be allocated to this SS the BS makes this allocation over the same servicing period that was used for the previous subchannel allocations. In Figure 2 we illustrate some of the features of the proposed VoIP subchannel framework. For simplicity we use illustrative parameters (not practical ones). We assume 1.25ms frames and an HARQ interlacing of four frames. The figure contains frames 1, 5, 9 and 13 of a 20ms period. If a user is allocated to the jth time index of subchannel i we denote his resources with the the tuple (i, j). Therefore user 1 uses (1,1) and user 2 uses (2, 1) and (4, 1). However, frame 1 is the last transmission in user 2 s servicing period and the second subchannel is subsequently dropped. In this particular example, no other VoIP user requires this dropped resource. Therefore, in frame 1 the maximum index is 4 while in frames 9 and 13 it becomes 3. In frame 5 the maximum index is also 4 because the fourth subchannel is not released until frame 9. We will see later that this maximum index can be used to allocate unused resources to other users. In particular we have divided the frame into
VoIP Channel (index 1) Frequency 1 2 Time Index 1 2 hole max index Time Sub band max index max index max index VoIP Users Sub band resources used for Localized Resource Channels we need approximately one SSCH subchannel every second for each user. Note that if resources were allocated for each voice frame transmission then this would have resulted in 50 SSCH channels per second. If the silence period is blanked instead of transmitting oneeighth frames then the same approach outlined here can be used. In this case zero resources are allocated to the SS when its voice session moves into a silence period. This frees up the resources for other VoIP and/or data users. When a new activity period begins the SS is again allocated resources. The operator can decide to blank silence periods only during heavy loading and maintain better voice quality during light loading. Whenever a SSCH is used for a VoIP user, the subchannel index (log 2 (K) bits) as well as the time index (log 2 (τ) bits) must be included. In addition, each time the maximum used index changes (due to a change in the size of the VoIP region) this information must be broadcast to all users so that the data users know which unused VoIP resources are available. Fig. 2. Example of VoIP Subchannel Allocations sub-bands and the leftover resources of each sub-band can be assigned to a data user who can be served based on subband CQI feedback. In frame 5, user 4 uses (1,2) while user 5 uses (2,2) and (3,2). Note that although the resource (4,2) is not allocated to anyone, it still cannot be used because subchannel 4 is still being reserved for user 2 until it is released in frame 9. However, with slightly increased complexity one can also make (4,2) available for data usage. If we assume that user 4 terminates its servicing period in frame 2 and deallocates its subchannel after this transmission then (1,2) is free in frame 13. However unfortunately this hole cannot be used without broadcasting the entire allocation structure to all data users which would require significant signaling overhead. VI. SIGNALING RESOURCE ALLOCATION INFORMATION In the previous sections we described how resources are allocated to the SS and how the allocated resources vary with the needs of the SS (i.e., the offered load) as well as the subchannel conditions of the SS. Due to the large number of VoIP users this signaling must be lightweight. In this section we describe the signaling procedure in more detail and estimate the signaling load. When a SS first requests resources for a VoIP session, the BS uses one or more SSCH subchannels to allocate one or more VoIP subchannels, together with the associated time indices. Each time the bundling rate and/or the resource allocation changes, the new information is again signaled in a SSCH subchannel. Let us consider the case of two bundling rates, one for active periods and one for silence periods. The average duration of these periods is on the order of 2 to 3 seconds and so signaling due to these changes will average at most one per second. In addition, resource allocations are changed based on changes in path loss. Typically, even for vehicular users, this will change on the order of once per 2-3 seconds. So in total VII. SHARING UNUSED DATA SUBCHANNELS Unused VoIP resources (except holes) can be easily assigned to data users. If the largest index of all assigned VoIP subchannels is k then subchannels k+1 to K are free. These free subsubchannels are used as follows. The BS picks a non-voip SS to be served in a sub-band and uses VoIP subchannels k + 1 to K for transmissions to the SS. Since k is known to the SS (it is broadcast each time it changes) then it knows which resources has been assigned to it. In addition to determining the leftover resources in that band available for data transmissions, the BS also needs to know what power resources are available. If we assume a nominal power of p nom per subchannel then Kp nom is available for use by VoIP users. Therefore if we subtract the actual power that is used by the served VoIP users in the band then the remaining power is available for the chosen data user. Note that this power will vary because of the fluctuating power demands of each VoIP user due to power control. As we see in Figure 2 although all left over power is made available to the data user, not all of the leftover slots are available. This is because of the creation of holes that will exist until a new demand is made for the particular type of resource. The fraction of such holes will decrease as the VoIP load increases because the time between resource demands will decrease and hence the average life of these holes will decrease. In order to make use of these holes, the BS would have to broadcast the set of free VoIP subchannel indices and this overhead can be significant. VIII. SOFTER HANDOFF In this section we discuss how softer handoffs can be performed with this framework. As long as the resource requirements of the VoIP users is small then there is no need to serve the user via multiple sectors. However, when the number of VoIP subchannels allocated to the user reaches some limit then, if additional resources are required for the user, such resources should be provided by other sectors in the forward
1 1 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 cdf 0.5 CDF 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Power Utilization 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Loading Fig. 3. Power Utilization for 460 VoIP Users Fig. 4. Bandwidth Utilization for 460 VoIP Users link active set of the user. We will call the next best sector in the active set the auxiliary sector. Suppose that some subset of subchannels have been allocated to the SS. If the size of this subset reaches some threshold then additional resources are allocated by serving the SS on one or more of the corresponding subchannels in the auxiliary sector if they are available. If they are all in use then we can either wait until one or more become free or add an additional subchannel from the serving sector. When one or more of the concerned subchannels in the auxiliary sector becomes available, the additional subchannel can be deallocated from the serving sector and assigned to the concerned SS. Power control in the serving sector is performed as before. A fixed power is used for the subchannels in the auxiliary sector so as not to exhaust its resources. IX. PERFORMANCE RESULTS We considered the case of 460 VoIP users which was approximately the capacity of the system considered. We developed a model for the system (which was validated with simulations) using typical parameter values as defined in the 3GPP2 simulation strawman. Our two resources of concern are power and bandwidth. In Figure 3 we plot the Cumulative Distribution Function for the power used for transmission. We find that the power is sufficient for this level of loading. In Figure 4 we plot the Cumulative Distribution Function for the bandwidth utilization which is the bandwidth leftover and available for data traffic (note this includes the holes but the percentage of these is quite small due to the heavy loading). Again we find that the bandwidth is sufficient to support the user loading. X. SUMMARY AND CONCLUSIONS We considered the problem of VoIP service over the forward link of an OFDMA network. The issues and the approaches for solving them can be summarized as follows Fast fading is compensated via frequency diversity (instead of user diversity as in 1xEV-DO). Variations in subchannel quality due to slow fading is offset through the use of power controlled subchannels. Path loss variations which require too large a change in power is instead handled through bandwidth adjustments. The forward link signaling is reduced by using power control instead of rate control to combat fast fading. Statistical multiplexing gains in the frequency/time domains is obtained by the sharing of these resources across VoIP users. The leftover resources are made available to delay tolerant traffic. The design is flexible in that it allows the operator to trade voice quality (end to end delay) with system capacity. The subchannel quality reporting rate can be reduced since this information is only used to compensate for slow fading. REFERENCES [1] A. Ghosh, D. Wolter, J. Andrews and R. Chen, Broadband wireless access with WiMax/802.16: Current performance benchmarks and future potential, IEEE Commun Mag. Feb. 2005. [2] IEEE Std. 802.16-2004, IEEE Standard for Local and Metropolitan Area Networks Part 16: Air Interface for Fixed Broadband Wireless Access Systems, Oct. 2004. [3] C. Y. Wong, R. S. Cheng, K. B. Letaief, and R. D. Murch, Multiuser OFDM with adaptive subcarrier, bit, and power allocation, IEEE J. Select. Areas Commun., vol. 17, no. 10, Oct. 1999. [4] G. Munz, S. Pfletschinger and J. Speidel, An efficient waterfilling algorithm for multiple access OFDM, IEEE Globecom, 2002. [5] G. Song and Y. (G). Li, Cross-layer optimization for OFDM wireless network part I and part II, IEEE Trans. Wireless Commun., vol.4, no. 2, Mar. 2005. [6] J. Jang and K. Bok Lee, Transmit power adaptation for multiuser OFDM systems, IEEE J. Select. Areas Commun., vol. 21, Feb. 2003. [7] G. Song, Y. (G). Li, L. J. Cimini and H. Zheng, Joint channel-aware and queue-aware data scheduling in multiple shared wireless channels, in Proc., IEEE Wireless Commun. Networking Conf., Mar. 2004. [8] P. Hosein, Capacity of Packetized Voice Services over Time-Shared Wireless Packet Data Channels, Proc. IEEE INFOCOM, March, 2005. [9] P. Hosein, Optimal Allocation of Bandwidth and Power Resources to OFDMA VoIP Channels Proc. ICWPC, Feb. 2006.