The Vault

Adaptive HARQ and Scheduling for Video over LTE
Research Paper / Feb 2014

Adaptive HARQ and Scheduling for Video over LTE Avi Rapaport, Weimin Liu, Liangping Ma, Gregory S. Sternberg, Ariela Ziera, and Anantharaman Balasubramanian InterDigital Communications, LLC, USA Abstract—This paper proposes a cross-layer approach to the delivery of real-time video streams over LTE cellular systems. The objective is to maximally improve video quality by adapting the wireless system (specifically, HARQ retransmission and scheduling rules) to video packet importance. Complying with the LTE framework, video packets are given priorities and assigned to different radio bearers or logical channels based on their importance. Applicable to both uplink (UL) and downlink (DL), the adaptive HARQ minimizes the effect of transmission errors by performing potentially more HARQ retransmissions for high-priority video packets while the scheduling algorithm deals with congestion by giving high priority to important packets to minimize their transmission timeout. We show that the combination of adaptive HARQ and priority-based scheduling offers significant gains in video quality by enabling the wireless system to handle congestion and transmission errors, two of the main challenges to cellular delivery of real-time video. Keywords—video; HARQ; QoS; scheduling; proportional fair; MLWDF; 3GPP; LTE; cross-layer; eNodeB; priority I. INTRODUCTION In recent years demand for wireless video applications has grown steadily and is expected to continue to grow with the rapid adoption of smart phones and tablet computers capable of generating and displaying video and transporting video over wireless networks. Although state-of-the-art wireless networks, such as LTE (Long Term Evolution) and LTE-Advanced, offer significantly higher data rates, the growing demand from video applications has proven to be challenging to cellular networks. Video applications are characterized by intensive use of network resources, variable packet importance, a tolerance to some loss of data, and sometimes stringent latency requirements. We consider real-time video applications such as video conferencing, video streaming, and cloud gaming. There has been significant research on cross-layer optimization of video delivery over wireless networks. The work in [1] proposes a weighted round-robin scheduling policy to achieve the best user-perceived video quality under the given delay constraint. Research in [2] proposes a gradient- based scheduling scheme in which user data rates are adjusted dynamically based on channel quality as well as the gradients of a utility function representing the distortion of the received video. Research in [4] shows how to estimate the packet delays in the uplink (UL) direction via the Buffer Status Reports (BSRs) and proposes a scheduling policy based on packet delays. The Hybrid Automatic Repeat reQuest (HARQ) controller in the LTE system plays a key role in achieving the required LTE performance. HARQ is generally optimized for uniform transport block priority. However, video packets are generally not equally important, and they demand different priorities on the wireless network. Figure 1. Video conferencing scenario over LTE showing two wireless hops. The uplink (UL) wireless hop is from the source User Equipment (UE) to the source base station (eNodeB), and the downlink (DL) wireless hop from the destination eNodeB to the destination UE. In this paper, we combine adaptive HARQ parameters and a scheduling rule based on video packet priority to form a coherent scheme to tackle the two key challenges in multiuser latency-limited delivery of video streams over cellular, namely, congestion and transmission errors. The remainder of this paper is organized as follows. Section II describes the proposed approach. Section III discusses the experiment setup and results. Finally, Section IV concludes the paper. II. PROPOSED APPROACH The proposed adaptive HARQ and scheduling scheme divides video packets from a video source into multiple streams or logical channels (LCs) according to packet importance and controls the number of HARQ retransmissions and the scheduling rule at the LTE Media Access Control (MAC) layer in order to optimize video quality. The HARQ mechanism in LTE improves integrity and robustness by providing physical layer retransmission based on feedback from the receiver and by using different Turbo code redundancy versions (RVs) and incremental soft combining. The soft combining operation improves retransmission performance by preserving the information in incorrectly decoded transport blocks and combining it with the retransmitted transport block. The combined information from retransmissions has a better chance of being decoded correctly than that from individual retransmissions. In a typical LTE system, the maximum number of HARQ retransmissions is set per UE by the Radio Resource Control (RRC) layer to a constant number (typically 4) and is used for the entire wireless link without consideration for the type or the importance of the transmitted data in the MAC packet data unit (PDU). In the proposed scheme, classification of video packets into priority groups is performed during video encoding according The 47th Asilomar Conference on Signals, Systems and Computers, 2013 to the severity of the impact that the loss of each video packet would have on video quality. Radio bearers or logical channels (LCs) are created to match these priorities. Based on this plurality of logical channels, two techniques are applied – adaptive HARQ and scheduling rules. Higher-priority logical channels are assigned a larger number of potential HARQ retransmissions. At the same time, a scheduling rule gives preference to higher-priority LCs. This differentiation results in unequal error protection and possibly out-of-order delivery of the video packets. A. LTE System Model for Real-Time Video Applications A simplified LTE architecture for real-time video application, emphasizing the UE-eNobeB data-plane connection, is illustrated in Figure 2. The foundation of the proposed scheme is that multiple Evolved Packet System (EPS) bearers with their assigned QoS Class Identifiers (QCIs) are established on either the source or destination cellular hops, or both, to transport a video stream over the cellular network. For the n levels of video packet priority assigned by the video encoder, there will be n EPS bearers created. The EPS bearers are then mapped to radio bearers and logical channels in a one- to-one manner, following the LTE framework. Each logical channel belongs to a QoS group with a specific priority according to the initial setup of the EPS bearers (e.g., LC1: highest priority, LCn: lowest priority). eNodeB UE Video Codec MAC PHY Internet ... Application Signaling (SIP/SDP) Application Function (IMS CSCF) Policy Control (PCRF) P-GWS-GW PDCP1 RLC1 IP UDP PDCPn RLCn RTP1 RTPn... MAC ...PDCP1 RLC1 IP PDCPn RLCn EPS Bearer Establishment Peer User Wireless LinkHARQ Feedback, New/Retransmission Logical Channel #1 Scheduling PHY Figure 2. Simplified LTE architecture for real-time video applications, emphasizing connections between a UE and an eNodeB through logical channels while showing their connection to the Internet through the Serving Gateway (S-GW) and the Packet Data Network Gateway (P-GW). Video data are delivered using Real-time Transport Protocol (RTP) over User Datagram Protocol (UDP). For clarity, only one logical channel is drawn. On the source hop, the packets generated by the video encoder in the source UE are directed to their corresponding logical channels to match the priority of the LC with that of the video packet. Similarly on the destination hop, the downlink (DL) EPS bearers are established in the same manner as those on the source hop by the IMS PCRF (IP Multimedia Service Policy and Charging Rules Function). The video packets that reach the destination eNodeB are mapped to the corresponding logical channels according to their EPS bearers. The eNodeB sets the maximum number of HARQ retransmissions for the video packets that are associated with a specific logical channel according to its QCI. As a result of EPS bearer differentiation, packets assigned a higher priority will be given a potentially larger number of HARQ retransmissions. In addition, the proposed scheme also directs the scheduler in the eNodeB MAC to send higher-priority packets first and lower-priority packets next that have not reached their delay limits. This operation may result in out-of-order packet transmission over the cellular link with higher probability of dropping lower-priority packets due to transmission timeout. B. Video Encoding with Packet Priority Video encoders such as H.264 exploit the similarity between video frames to achieve high-efficiency and high- quality compression. Highly compressed video streams are vulnerable to even a small number of lost packets depending on the location of the packets within the dependency scheme. The 47th Asilomar Conference on Signals, Systems and Computers, 2013 The priority assignment to a video packet would be optimal if it can accurately reflect the impact of error propagation on the received video quality. Modern video encoders have the capability to construct efficient and often repetitive dependency schemes which provide a clear packet priority assignment. For example, in H.264 Hierarchical P video coding [9], video frames can be divided into temporal layers according to their frame reference and thus naturally associated with different priorities. Another approach to assigning video packet priority, which is used in this paper to evaluate the adaptive HARQ and the scheduling rule, is based on the distance (in terms of frame index) of the video frame contained in the packet from an IDR frame. When using the traditional IPPP encoding structure with recurring IDR frame insertion, P video frames closer to the most recently received IDR frames are assigned higher priority levels compared to those further away from the IDR (see Figure 3). The intuition is that an erroneous P frame immediately after IDR will generate error propagation into more P frames until the next IDR whereas an erroneous P frame close to the end of the GoP (group of pictures) will only propagate into a few frames until the next IDR. Figure 3. Video frame priority based on distance from IDR. C. Mapping Packet Priority to Maximum HARQ Retransmissions The LTE layer-2 protocol stack consists of three sub-layers – PDCP (Packet Data Convergence Protocol), RLC (Radio Link Control) and MAC (Medium Access Control). IP packets are mapped into radio bearers according to their assigned QCI. Radio bearers are then mapped to logical channels in a one-to- one fashion after the SAR (segmentation and reassembly) functionality in the RLC sub-layer. For video, the packet priority to logical channel mapping is one-to-one, i.e., priority 0→LC1,…, priority n–1→LCn. Multiple MAC SDUs from several LCs can potentially be multiplexed into a single MAC PDU. Each SDU represents a packet assigned to a single LC with corresponding priority. Since each MAC PDU must be assigned a single maximum number of HARQ retransmissions, there is a need to map the potentially multiple MAC SDU priorities to a single maximum number of HARQ retransmissions at the MAC scheduler for both UL and DL. The priority vector P(i) consists of Ni elements corresponding to the SDUs assembled in PDU i. Each element in the vector P(i) is p(i, j) ∈ {0, 1, …, n-1}, the priority of the SDU j in PDU i, with 0 being the highest. Each priority vector is assigned a single priority value PPDU(i) that is equal to the highest priority among all SDUs in the PDU:    )},({min jipiP jPDU    Assuming the total number of logical channels is NLC, and the upper and lower limits of the maximum number of HARQ retransmission are UH and LH, respectively, the maximum number of HARQ retransmissions for PDU i, NH(i), is calculated as below for NLC > 1:           1LC HH PDUHH N LU iPUiN   where   represents integer flooring. For NLC=1, NH(i) = UH. D. Scheduling Rule In addition to adapting the maximum number of potential HARQ retransmissions, the proposed scheme also uses a scheduling rule that gives higher transmission priority to higher-priority video packets in logical channels that have longer backlogs. The scheduler algorithm used in our scheme is a form of the M-LWDF scheduler [3] modified for multiple logical channels, in which the head-of-line (HOL) delay of a video application is taken to be the maximum HOL delay of all of its logical channels. The algorithm is implemented in two steps. In the first step, at time t for sub-band c, user k is chosen to satisfy:           trtQqtFctk cijijjii ,maxmaxarg,    where Fi(t) is the proportional-fair (PF) weight term of user i which is the reciprocal of its historical average data rate, Qi,j(t) is the HOL delay for logical channel ݆ of user i, ric(t) is the potential data rate that can be achieved for user i on this sub- band during this TTI (transmit time interval), which can be obtained from the channel quality indication (CQI), and qj is the weight for logical channel j calculated as below qj =−log(δj) / Tj (4) where Tj is the largest delay that LC j can tolerate, and δj is the maximum packet loss rate acceptable. In the second step, the allocated resources for user k are distributed among the LCs according to the LCs priority. For video data we use the absolute priority-based method in which the LC with the highest priority is served first, and, if there are resources left, other LCs are served according to their priority. This scheduler algorithm can work with any number of LCs and can be used for both the DL and UL. For the DL, both steps are executed in the eNodeB MAC whereas for the UL, the first step is executed in the eNodeB MAC and the second step is performed in the UE MAC following the LTE standard [5]. The 47th Asilomar Conference on Signals, Systems and Computers, 2013 III. PERFORMANCE EVALUATION A. Simulation Setup The proposed scheme is evaluated using an LTE cellular network system-level simulation model compliant to 3GPP 36.814 [6] for both DL and UL. The system simulation model consists of the integration of the LTE DL (UL) multi-user system-level simulator, a packet data flow emulator, and an H.264 video codec for each user. The LTE system simulator consists of a model of the LTE PHY layer with a significant level of abstraction, the MAC-layer HARQ, and the scheduler. The data flow emulator simulates the LTE RLC sub-layer UM (unacknowledged mode) and acts as the interface between the video codec and the LTE MAC, performing IP packet SAR functions. Video coding is implemented using the JM video encoder and decoder [8]. Three levels of video packet priority are assigned (see Figure 3) – priority 0: IDR0-P4, priority 1: P5-P9, priority 2:P10-P14. On the transmitter side, a packet is discarded if the period the packet waits for transmission exceeds the scheduling and transmission timeout period of 150ms. On the receiver side, a packet is discarded if it is received incorrectly after HARQ operation, and packet reordering is performed. The integrated simulation calculates PSNR (peak signal-to-noise ratio) for each video frame and each user after taking into account the dropped video frames and error concealment using frame copy. In the UL direction, the scheduler in the eNodeB does not have direct knowledge about the UL buffer occupancy information on the UE side. Therefore our UL simulation also includes the BSR (buffer status report) message reporting mechanism, accounting for messaging errors and delays and their effect on video quality. Tables 1 and 2 list the key simulation parameters for the video codec and the LTE system. B. Simulation Results and Analysis Using a single video sequence, Figure 4 illustrates the advantage of using adaptive maximum HARQ retransmissions by comparing the per-frame PSNR with a fixed maximum number of HARQ retransmissions of 4 and that with an adaptive maximum number of HARQ retransmissions in the range of 2 to 6. For the adaptive HARQ (dashed line), most of the errors are located just before IDR frames which results in a PSNR loss for a short period whereas for the fixed HARQ (solid line), the errors are located in random locations with respect to IDR frames which results in generally longer periods of PSNR degradation. Figure 4. PSNR per frame comparison between fixed HARQ (4 – solid line) and adaptive (2 to 6 – dashed line) HARQ retransmissions. Circle markers indicate positions of packet errors whereas square markers represent positions of IDR frames. Figure 5. Composite CCDF of video packet PSNR for the top 85% of the users served by the LTE network. Since LTE is a shared network, we examine the video quality experienced by the majority of the users. Figure 5 shows the composite CCDF (complementary cumulative distribution function) of video PSNRs, i.e., the percentage of frames having a PSNR that is greater than a given PSNR threshold, for the top 85% of the users served by the LTE system. In generating the CCDF in Figure 5, we first compute 0 10 20 30 40 50 60 22 24 26 28 30 32 34 36 38 40 42 Frame Index PS N R [d B ] 15 20 25 30 35 40 10 20 30 40 50 60 70 80 90 100 PSNR [dB] C C D F [% ] Fixed HARQ: 4 Adaptive HARQ: 2–6 TABLE I LTE SYSTEM PARAMETERS Parameter Value System bandwidth 10MHz Duplexing mode Frequency division (FDD) Number of users per sector 12 Number of sectors per cell site 3 Number of cell sites 57 CQI feedback model Non-ideal, 6 TTI delay Number of logical channels 3 Duration of simulation 10 seconds Maximum HARQ retransmissions 4 for fixed, 2-6 for adaptive Scheduling algorithm Modified M-LWDF Target block error rate (without HARQ) 10% Scheduling timeout 150ms Payload model Video (non full-buffer) TABLE II VIDEO PARAMETERS Parameter Value Sequence Foreman, Football Error recovery Periodic IDR IDR frequency Every 15 frames Number of priority levels 3 Fixed HARQ: 5 frame errors, PSNR=32.12 dB Adaptive HARQ: 5 frame errors, PSNR=36.07 dB Distribution of PSNR for 85% of All Users The 47th Asilomar Conference on Signals, Systems and Computers, 2013 the CCDF of the PSNR (over time) for each user. Next at each PSNR we select the users with CCDFs in the top 85%. The remaining 15% are removed from the dataset. Finally, the surviving data for all users are merged and the CCDF over time and users is plotted in Figure 5. The combined gain of adaptive HARQ and multiple-LC scheduling over fixed HARQ and single-LC is very significant (1–5dB) for most PSNR values. As discussed earlier, packet loss in a delay-limited system can occur due to packet transmission timeout (congestion) or transmission error. Figure 6 (top) shows the combined packet loss rate as a function of video bit rate which directly leads to system loading. Figure 6 (bottom) illustrates the performance gains that can be attributed to the adaptive HARQ and the priority-based scheduling rule under different system loading conditions. For low and normal system loading, the PSNR gain (~0.7 dB) mainly comes from the adaptive HARQ technique whereas under heavy system loading conditions the main PSNR gain (~3 dB) is achieved by priority-based scheduling. Figure 6. DL video packet loss rate (top) and rate-distortion (bottom) relative to system loading conditions, illustrating the relative benefits of adaptive HARQ and multiple-LC scheduling. Since packet priority is tied to logical channel in the proposed scheme, the single-LC adaptive HARQ scenario is simulated by tagging each packet with a priority, bypassing the logical channel mechanism. Similarly, the 3-LC fixed HARQ scenario is simulated by forcing the maximum number of retransmissions to 4 regardless the LC. Similar results are achieved when these techniques are applied to UL video transmission. Under low and normal system loading, the PSNR gain is 0.8dB due to adaptive HARQ, and under heavy system load, a PSNR gain of 1.5 dB is achieved due to multiple-LC scheduling. IV. CONCLUSIONS This paper proposes a cross-layer approach to improving real-time delivery of video over LTE cellular networks. Assignment of packet priority is performed by an H.264 video encoder to reflect the impact of the packet loss on video quality. To enable differentiation of video packets in wireless transmission, an architecture based on the LTE framework is proposed to map video packets into different radio bearers or logical channels according to their priority. The video-aware HARQ in the LTE MAC adapts its maximum number of retransmissions based on video packet priorities while the scheduler in the base station gives priority to high-importance packets in a multiuser, resource-limited, environment. Simulation results show that the combination of adaptive HARQ and priority-based scheduling methods offer significant gains in video quality by tackling two main problems in multiuser wireless systems, namely congestion and transmission errors. The scheduling rule is shown to be effective (3.5 dB gain in PSNR) in improving video quality when the wireless network is heavily loaded or congested while the adaptive HARQ minimizes the effect of transmission errors on high-priority video packets (0.5~1 dB gain in PSNR). REFERENCES [1] H. Luo, S. Ci, D. Wu, J. Wu, and H. Tang, “Quality-driven cross-layer optimized video delivery over LTE,” IEEE Comm. Mag., Vol.48, No.2, pp.102-109, February 2010. [2] P. Pahalawatta, R. Berry, T. Pappas, and A. Katsaggelos, “Content- aware resource allocation and packet scheduling for video transmission over wireless networks,” IEEE Journal on Selected Areas in Communications, vol. 25, no. 4, pp. 749–759, 2007 [3] M. Iturralde, T.A. Yahiya, A. Wei, and A.-L. Beylot, “Performance study of multimedia services using virtual token mechanism for resource allocation in LTE metworks,” IEEE Vehicular Technology Conference, 2011. [4] A. Baid, R. Madan, and A. Sampath, “Delay estimation and fast iterative scheduling policies for LTE uplink,” 10th Intl Symposium on Modeling and Optimization in Mobile, Ad Hoc and Wireless Networks (WiOpt), May 2012. [5] 3GPP TS 36.321, v10.1.0, Medium Access Control (MAC) protocol specification, March 2011. [6] 3GPP TS 36.814 V9.0.0.0, Further advancements for E-UTRA physical layer aspects, March 2011 [7] 3GPP TS 36.300, v9.8.0, Evolved Universal Terrestrial Radio Access (E-UTRA) and Evolved Universal Terrestrial Radio Access Network (E- UTRAN); Overall description; Stage 2, December 2010. [8] H.264/AVC Software Coordination: [9] D. Hong, M. Horowitz, A. Eleftheriadis, and T. Wiegand, “H.264 hierarchical P coding in the context of ultra-low delay, low complexity applications,” in Picture Coding Symposium (PCS) 2010, pp. 146–149, 2010. 0.1 1 10 0.8 0.9 1 1.1 1.2 Single LC, Fixed HARQ Single LC, Adaptive HARQ 3 LCs, Fixed HARQ 3 LCs, Adaptive HARQ Video Bitrate [Mbps] PE R [% ] 31 32 33 34 35 36 37 0.8 0.9 1 1.1 1.2 PS N R [d B ] Single LC, Fixed HARQ Single LC, Adaptive HARQ 3 LCs, Fixed HARQ 3 LCs, Adaptive HARQ Light Loading Normal Loading Heavy Loading Video Bitrate [Mbps] The 47th Asilomar Conference on Signals, Systems and Computers, 2013