MAC Scheduling Scheme for VoIP Traffic Service in 3G LTE

Sunggu Choi∗ , Kyungkoo Jun† , Yeonseung Shin∗ , Seokhoon Kang† , Byoungjo Choi†
Telecommunication Research Lab. Electronics and Telecommunications Research Institute, Korea Email: {sguchoi, shinys}@etri.re.kr † Dept. of Multimedia Systems Engineering University of Incheon, Koea Email: {kjun, hana, bjc97r}@incheon.ac.kr
Abstract—3G Long Term Evolution, which aims for various mobile multimedia services provision by enhanced wireless performance, proposes the VoIP-based voice service through the PS domain. When delay and loss-sensitive VoIP traffic flows through the PS domain, more challenging technical difficulties are expected than in the existing 3G systems which provide the CS domain based voice service. Moreover, since 3G LTE, which adopts the OFDM as its physical layer, introduces Physical Resource Block (PRB) as the unit for the transmission resources, it becomes necessary to develop new types of resource management schemes. This paper proposes a MAC layer PRB scheduling algorithm for the efficient VoIP service in 3G LTE and shows the simulation results regarding its performance. The key idea of the algorithm consists of two parts; dynamic activation of a VoIP priority mode for the voice QoS satisfaction and adaptive adjustment of the VoIP priority mode duration in order to minimize the performance degradation induced by its priority mode application.
∗ Mobile

I. I NTRODUCTION 3G Long Term Evolution (3G LTE) is a mobile cellular communications technology aiming for maximum 100 Mbps downlink and 50 Mbps uplink speed when using 20 MHz bandwidth, thus enables diverse mobile multimedia service provision [1]. However, we expect many technological challenges for successful 3G LTE based services. One of such difficulties is regarding voice service provision. In existing CDMA and WCDMA based cellular systems, the voice service is provided separate in Circuit Switching (CS) domain unlike other data services which are provided in Packet Switching (PS ) domain, but 3G LTE is designed to provide all services including the voice service in the IP based PS domain. The QoS provision for the VoIP traffic in the PS domain is hard because the VoIP service is sensitive to packet delay and loss. Nevertheless, users expect higher voice service quality from 3G LTE than that of CS domain based voice services of existing cellular systems. Therefore, it is pivotal to satisfy the QoS of VoIP based voice services in 3G LTE. Other cellular systems such as CDMA2000 1xEV-DO propose, in addition to various network QoS provision schemes, voice processing techniques like smart blanking and adaptive anti-jitter playback buffering to support the voice service of which quality is not poorer than that of CS domain based voice services [2]. But, there is a worry that, unless the fading effect

of wireless links is carefully considered, the performance of VoIP based voice service may be optimistically overestimated [3]. When taking into account both wireless specific fading and the delay and loss sensitive VoIP characteristics, the successful voice service implementation in the PS domain is challenging. Prioritization of VoIP packets simply for the QoS satisfaction over other traffic packets is undesirable because it may degrade the overall system performance and it may harm the QoS satisfaction of other multimedia calls. Therefore, it suggests that a sort of adaptive prioritization is more preferable than just granting absolute priority to VoIP traffic. Another difficulty in the voice service of 3G LTE stems from a new transmission resource modeling adopted in 3G LTE. Unlike existing CDMA based cellular systems, 3G LTE takes Physical Resource Block (PRB) as its transmission resource unit; the PRB is a basic transmission unit which has both frequency and time aspects. For example, the MAC scheduling in 3G LTE is to assign PRBs to ongoing calls at every Transmission Time Interval (TTI). Evolved NodeBs (eNodeBs), which are the base stations of 3G LTE, have a fixed number of available PRBs according to their bandwidth and are responsible for allocating PRBs repeatedly at every TTI [4]. In this paper, we propose a MAC layer scheduling scheme for the QoS satisfaction of VoIP based voice service in 3G LTE. The goal of our proposed scheme is to support the QoS of VoIP traffic but minimize possible negative effect on the overall system performance and efficiency, particularly, the downlink throughput. The key ideas of our scheme are a VoIP priority mode and its adaptive duration management. Since the VoIP priority mode assigns PRBs first to VoIP calls, it is able to minimize VoIP packet delay and loss, but the adaptive duration management is able to prevent the overall system performance degradation, which is a possible negative effect of the VoIP priority mode. In our scheme, the duration of VoIP priority mode is dynamically adjusted according to VoIP packet drop rates. As a result, we are able to achieve both the QoS satisfaction and the minimization of the negative effect. Our contributions presented in this paper are as follows. • an eNodeB resource management framework of 3G LTE

1-4244-0264-6/07/$25.00 ©2007 IEEE

1441

This design is advantageous particularly in the sense that the QoS provision. like HSDPA based cellular systems. MAC scheduling. average queue length of each transmission queues. which are the transmission resources. we introduce the characteristics of the PRBs. The amount of data bits that can be carried by one PRB is determined by wireless channel environment between eNodeBs and user mobile terminals. 1442 .• a MAC layer scheduling scheme to support the QoS of VoIP traffic • a simulation of the proposed scheme to evaluate the performance regarding the QoS support and analyze the effect on the system throughput II. and the state information management function collects and store data such as wireless link conditions. the role of the MAC scheduler can be rephrased as the PRB allocation to the transmission queues. the PRB is the smallest transmission unit in Orthogonal Frequency Division Multiple Access (OFDMA) based wireless systems. eNodeBs have separate transmission queues for each downlink user call. on overall system performance. The PRB has both frequency and time aspects: it consists of m subcarriers in frequency while n subframes each of which is 0. per terminal Channel Quality Indicator (CQI) information showing downlink wireless link conditions. which will be discussed shortly. PRBs The structure and allocation of the eNodeB transmission resource. The eNodeB structure with separate queues for user calls We also assume that. As aforementioned. The state information management function collects and stores data necessary for the RAC and the MAC scheduling operations. as shown in Figure 2. The downlink MAC scheduler determines how many PRBs should be assigned to each ongoing call at Transmission Time Fig. Radio Admission Control (RAC). VoIP packet drop ratio due to excessive delay. The RAC determines whether to admit new call requests. queue lengths. the shortcomings are that scheduling itself becomes complicated and the signaling overhead increases. as shown in Figure 2. one PRB is defined as 12 subcarriers and 2 subframes [5][6]. minimize negative impact. and the resource management structure of the eNodeBs. 1. which differs according to call traffic types. Frequency Intervals (TTIs).Queue length . and other related information.VoIP packet drop rate 12 Subcarriers Allocated PRB 0. because the multiplexing is able to improve the efficiency of the PRB allocation. OFDM based wireless transmission Figure 1 shows the structure of the PRBs and an example of the PRB allocation schemes by the eNodeBs. However. III. at the same time. In this paper we assume that the allocated PRBs are able to carry only data bits from the corresponding calls. becomes simple.5 ms 1 ms 0. packet drop and loss rates. QoS Scheduler QoS Scheduler PRB allocation . consists of three functions. uses Adaptive Modulation and Coding (AMC) which changes modulation and coding schemes depending on wireless link conditions. In the case of multiplexing in which it is feasible for a PRB to contain data bits from different calls. In the current 3G LTE standard recommendation. and the resource management for downlink transmission is the main target of our proposed scheme. The resource management structure of eNodeBs. which is induced when prioritizing the VoIP calls. S YSTEM M ODEL User mobile terminals and service providing eNodeBs are within the scope of this paper. and state information management. the MAC scheduling allocates the PRBs over ongoing calls.5 ms Time eNodeB Available PRBs TTI TTI TTI TTI Fig.5 ms long. This separate queue structure enables the following: one user mobile terminal has more than one ongoing downlink calls and each call is independently scheduled. User call Downlink Traffic New Call Admission Request Radio Admission Control Bandwidth Free PRB per Call flow traffic eNodeB and Downlink Status Info.CQI . In this section. Thus. P ROPOSED A LGORITHM The purpose of our proposed algorithm is to satisfy the QoS of VoIP calls and. It is because 3G LTE. 2. the discussion about the support of multiplexing by using Virtual Resource Block (VRB) is active in 3GPP meetings [4]. We assume that this function stores the following data.

Basically the MAC scheduler allocates the PRBs to ongoing calls in a round robin way. then two or more PRBs are allocated. On the contrary. if there are still available PRBs at the end of VoIP call only scheduling. even in the VoIP priority mode. the earlier a corresponding call is scheduled to have the PRBs. low drop ratio means that the QoS of VoIP calls are satisfied at decent level. which will be discussed shortly. The continuation period of the VoIP priority mode is restricted because it can degrade overall system performance by excessively concentrating eNodeB resources on VoIP calls. the MAC scheduler assigns. By using preset minimum and maximum thresholds for the VoIP packet drop ratio. Particularly in the VoIP priority mode. 3. the earlier the corresponding call is scheduled. The second part also operates at every TTI. The adaptive control of this mode’s duration is to prevent system performance degradation.* SINR Schedule calls with larger factor values first Factor = Queue Len. The limit of consecutive application of the VoIP priority mode is adaptively changed between prespecified minimum and maximum according to VoIP packet drop ratio. Our PRB allocation scheme is designed by making modifications to Channel Adaptive Fair Queuing [7]. the normal mode is set and the count is reset. however it often happens that it is necessary to allocate more than one PRBs to calls. Otherwise. in other words. For example. The maximum count increases or decreases depending on overall VoIP packet drop ratio which is measured at the eNodeB. the remaining PRBs are allocated to non VoIP calls in the same way of the normal mode. the longer the queue is and the better the wireless link state is. in the case of the normal mode. the allocation starts again in the same order and it repeats until there is no available PRB or no call has data to send. the VoIP priority mode is turned on. It continues as long as there are remaining PRBs and there are calls to have data to send. if the VoIP priority mode is used. The activation of the VoIP priority mode is determined at every TTI by considering whether ongoing VoIP calls exist as well as how long the VoIP priority mode has continued. Scheduling algorithm at every TTI where Qlen (i) and γi is the queue length and the SINR value of ongoing call i respectively. Scheduling Start at every TTI Yes Is there VoIP calls? No Apply VoIP priority mode Yes Apply Normal mode Successive count of the VoIP priority mode > Limit Factor = Queue Len. The order calculation is executed at every TTI. After the scheduling order is determined. and the second one is the adjustment of the maximum number of consecutive TTIs to which the VoIP priority mode is applied. The adjustment algorithm for the maximum allowable count of consecutive VoIP priority TTIs is shown in Figure 4. in part. The aforementioned equation 1 implies that. in principle. the size of one packet is larger than the capacity of one PRB but the packet segmentation is not allowed. we use the following equation 1. that there are many ongoing VoIP calls. For example. high drop ratio implies. Df (i) = Qlen (i) ∗ γi (1) at every TTI. if there exist VoIP calls and the count is below the limit. The VoIP priority mode is to schedule only the VoIP packet data over TTI. At the start of the scheduling process at every TTI. it is safe to reduce the duration of the VoIP priority mode to serve more other types calls. One thing to note is that. the queue length and SINR of each call. The first one is the PRB allocation which operates The flow chart of the PRB allocation algorithm is shown in Figure 3. For example.The cores of the proposed algorithm are a VoIP priority mode and adaptive control of the VoIP priority mode duration period. which are usually inefficient in the sense of PRB utilization due to small packet sizes. the activation of the VoIP priority mode is determined by whether there exist VoIP calls or not. one PRB to a call at a time. the maximum consecutive count decreases by one because it implies that either there are enough resources or 1443 . thus it is necessary to increase the limit in order to allow more consecutive TTIs to be dedicated to VoIP calls. and also whether the count of the consecutive VoIP priority mode enabled TTIs exceeds the limit. The details of our proposed scheme largely consist of two algorithms. if the drop ratio is below the minimum threshold. assign PRBs only to VoIP calls in order to satisfy the QoS of the voice service. If there are still remaining PRBs at the end of the round-robin cycle. The larger the factor values are. are allocated one PRB at a time as described earlier. It prevents PRB resources from being wasted. The details of the proposed algorithm are presented later. remaining PRBs after the allocation to VoIP calls are assigned to other types calls in order to avoid the resource waste. In our scheme. * SINR Yes Schedule VoIP calls with larger factor values first Yes Yes PRBs remain & More VoIP calls to send data? PRBs remain & More calls to send data? No No PRBs remain? No End of Scheduling Fig. or all ongoing calls. but we determine the scheduling order of calls according to the sizes of the following factors. in a round robin way but the calls with long queues and a good wireless links are first. Only VoIP calls.

The VoIP packet drop rate comparison There may be negative effects when using the VoIP priority mode even though this mode proved to be quite effective for the QoS of VoIP as just shown in the above. 5. when using our proposed VoIP priority mode. of available PRBs Maximum VoIP packet delay at eNodeB Maximum VoIP packet drop rate. axis. we compare the performance of our algorithm with that of a kind of CAFQ algorithm. Finally. when the number of VoIP calls is just 10.Start the adjustment of the limit. when the VoIP priority is not used. the drop rates remain at low level around 1% in spite of the increase of VoIP calls. we measure the average VoIP packet drop rate per unit time and the average throughput of eNodeB as we increase the number of ongoing VoIP calls gradually while there are a fixed number of other type calls. it is observed that. for consecutive VoIP priority mode application TABLE I S IMULATION PARAMETERS Simulation Parameter Values 1 Km 5 MHz 12 subcarriers. if the drop ratio is above the maximum threshold. 16QAM. It is to be conservative in managing resources for VoIP traffic because VoIP packets are sensitive to delay and loss. Cont. 4. if the drop rate stays between the minimum and maximum threshold. To analyze the effect of the VoIP priority mode. which uses the same PRB allocation scheme as ours except that it does not have the VoIP priority mode. Keep the limit smaller than the maximum if (Cont > Cont_max) Cont = Cont_max Yes Fig. IV. S IMULATION FOR THE P ROPOSED A LGORITHM P ERFORMANCE E VALUATION We conduct simulations to analyze the impact of our proposed algorithm on the QoS satisfaction of VoIP calls and the overall performance of eNodeB. the current maximum count is kept. which is the core of our proposed algorithm. We assume that VoIP packets are considered as dropped unless they are retransmitted to user terminals within 20 ms after they arrive at eNodeB. This simulation results shows that the VoIP priority mode plays a significant role in guaranteeing the QoS of VoIP calls. linearly decrease and exponentially increase. are measured and plotted on Y axis as we increase the number of VoIP calls step by step from one to maximum fifty on X ? Yes rd < Tmin No VoIP packet drop rate is larget than the maximum threshold rd > Tmax No Keep the current limit Cont = Cont. it is represented as a separate solid line. 6 20 msec. for example. if (Cont < Cont_min) Cont = Cont_min VoIP packet drop rate is smaller than the minimum threshold Bandwidth PRB structure TTI Num. the drop rate already reaches the maximum allowable rate. less than 1% drop rate even when 50 VoIP calls are in progress. Particularly. 64QAM Calculate VoIP packet drop rate. Fig. As shown in Figure 5. On the contrary. the packet drop rate rises rapidly as the number of ongoing VoIP calls increases.5% 1. since 5% packet drop rate of VoIP traffic is considered as the maximum allowable drop rate to meet the QoS. Tmax Minimum VoIP packet drop rate. Keep the limit larger than the minimum. Especially. d Cell radius r Decrease the limit by one. Cont = Cont 1. as shown in Figure 5. A possible 1444 . 2 subframes 1 msec. 5%. Tmin Modulations for AMC Increase the limit two times.5% QPSK. 4. We adjust the maximum success count asymmetrically as described earlier. Algorithm for determining the allowable count of consecutive VoIP priority mode application VoIP calls are given sufficient number of PRBs. Cont = 2 * Cont. the maximum count is doubled because it means that the PRBs allocated to VoIP calls are not enough. The VoIP packet drop rates. On the contrary. The parameters to simulate the eNodeB and user mobile terminals are as shown in Table I. To analyze quantitatively the QoS satisfaction of VoIP traffic and the performance of eNodeB when using each algorithm.

814 v7. [4] 3GPP. Awoniyi. we found that our scheme was able to keep the VoIP packet drop rate below 5%. As a result of applying our scheme. of VoIP calls is low as aforementioned. the fact that. When observing the throughput decrease ratio in Figure 6. ”VoIP over CDMA2000 1xEV–DO Revision A. 6. R EFERENCES [1] 3GPP. schedules with high probability non–realtime calls earlier. Black. The long queue length of non–realtime calls is due to the fact that non–realtime packets.0. Sharp.1. Particularly. we were able to satisfy the QoS of VoIP calls by scheduling VoIP packets first. 2006. when not using the VoIP priority mode in the simulation. Figure 6 shows the overall throughput of eNodeB. decreases as the number of VoIP calls increases. It is due to the fact that VoIP packet sizes are small and the queue length of VoIP calls are short. It is because the PRB utilization. V. which the normal mode schedules earlier than VoIP calls. which is the maximum allowable drop rate to support the QoS of VoIP. As shown in Figure 6. Secondly. we proposed an efficient VoIP packet scheduling scheme for the voice service in 3G LTE. regardless of using the VoIP priority mode or not. 3GPP. C ONCLUSIONS In this paper. we conduct a simulation and measure the downlink throughput of eNodeB under the same environment as Figure 5. at the same time. minimize the throughput decrease.” R1-062714. In summary. why the normal mode schedules non– realtime calls first.3. We discuss in sequence. In addition. R.11a WLANs. is able to satisfy the QoS of VoIP based voice service and. ”Channel adaptive fair queueing for scheduling integrated voice and data services in multicode CDMA systems. By conducting the simulations.” IEEE Communications Magazine.” R1062819. Such minimization of throughput decrease is made possible because the consecutive application of the VoIP priority mode are adaptively restricted and remaining PRBs. but the throughput decrease ratio was less than 3%. ”Effect of Fading on the Performance of VoIP in IEEE 802. have large enough data to fill up allocated PRB because of long queues. Feb. NEC. the decrease amount of the normal mode is smaller than that of the VoIP priority mode because non-realtime calls. which endure relatively long delay. ratio The eNodeB throughput comparison and the throughput decrease negative effect is that our proposed mode might degrade the efficiency of the eNodeB resource utilization because VoIP calls are seldom able to fully utilize the allocated PRB capacity. [6] Nokia. The cores of our scheme are the VoIP priority mode and the adaptive duration adjustment of the mode application. which means how much capacity of an allocated PRB is used. To analyze such negative effect of the VoIP priority mode. [2] M. even when using the VoIP priority mode. The decrease ratio of the VoIP priority mode to the normal mode is just less than 3% even when 50 VoIP calls are ongoing. Panasonic. are able to stay longer in queues. Lau. V. the negative effect of the VoIP priority mode on the overall throughput is negligible. The drop of excessively delayed packets in queues makes the VoIP queue length even shorter.913 v7. ”Further Investigations on Resource Block Sie for E-UTRA. not enough amount to fill up assigned PRBs. highlights the significance of the VoIP priority mode. P. for the VoIP priority mode and the normal mode respectively. At first. [3] O. S. Y. F.” TR 25. Wang. Lau.” in Proceedings of IEEE International Conference on Communications 2004. ”Requirements for Evolved UTRA (E-UTRA) and Evolved UTRAN (E-UTRAN). Kwok. but minimize the negative impact of the VoIP priority mode on the overall system performance by adaptively adjusting consecutive VoIP priority mode application. [5] NTT DoCoMo. Y. Lott. 3GPP. Tokgoz. the scheduler. ”Considerations for minimum TTI size for downlink LTE. thus VoIP calls generally have small amount of data to transmit. the higher PRB utilization of non– realtime calls is due to the fact that non–realtime calls. have generally higher PRB utilization than VoIP calls. our proposed scheme. 1445 . firstly. since the queue lengths of non–realtime calls are usually longer than those of VoIP calls.Fig. W. and secondly why non–realtime calls have higher PRB utilization. M. However. 27(9):809–820. are allocated to non–realtime calls in order to avoid resource waste.” Computer Communications. when scheduled. Diaz. as the number of VoIP calls increases gradually. the VoIP priority mode and the adaptive adjustment of consecutive priority mode application. very few VoIP calls were able to be served in the packet drop rate aspect. Tobagi. Yavuz.” TR 25. ”Physical layer aspects for evolved Universal Terrestrial Radio Access (UTRA). Kapoor. the decrease ratio of the throughput of the VoIP priority mode to that of the normal mode is separately plotted in percentage terms in order to easily observe how much throughput degradation is induced by the VoIP priority mode. [7] L. the throughput of both modes. C. of the normal mode which takes only wireless link condition and queue length into account. June 2004.0. Grob.