Video Communication with QoS Guarantees over HIPERLAN/2

Video Communication with QoS Guarantees over HIPERLAN/2 Ji Wang, Ashfaq Khokhar, and Vijay Garg Department of Electrical and Computer Engineering University of Illinois at Chicago, 60607 ABSTRACT This paper examines the impact of different physical modes on network QoS parameters such as throughput, bit error rate, and retransmission delays. An efficient technique for the transmission of MPEG-2 bit stream over HIPERLAN/2 networks is presented that exploits the hierarchical structure of the MPEG-2 bit stream. We show that for a given video bit stream use of multiple modulation schemes chosen based on the characteristics of underlying video data yields superior throughput and better picture quality in terms of PSNR. 1. INTRODUCTION The last two decades have witnessed a tremendous growth in the use and deployment of wireless devices and networks. In this regard, wireless LAN (WLAN) has gained popularity for its capability to interconnect portable or mobile devices and provide access to the Internet. The trend for further growth seems to be within the concept of mobile Internet. This means users can be attached to Internet anywhere at any time. This trend has also driven integrated wireless networks to become a reality. The integration of WLANs, WWANs (Wireless Wide Area Networks), the wired core, and cellular networks under a uniform IP framework offers a wide range of solutions for today and the future wireless market demands. Multimedia communication over such integrated networks will need to support voice, data and video services at different quality of service levels. Prior to connection establishment, such services need to specify end-to-end connection requirements in terms of network resources, and the duration for which these resources are needed. In this paper, we study the communication of MPEG-2 format compressed video data over integrated networks. In particular we investigate efficient solutions for communicating video data over HYPERLAN/2, a WLAN standard introduced by ETSI (the European Telecommunications Standards Institute). Similar to IEEE WLAN standard 802.11a, HIPERLAN/2 operates in an unlicensed frequency band at 5GHz and offers a data rate up to 54Mbps. While 802.11 enables limited QoS management through its PCF (Point Coordination Function) mode, the connection-oriented nature of HIPERLAN/2 makes it possible to implement comprehensive Quality of Service (QoS). Each connection can be assigned a specific QoS in terms of bandwidth, delay, jitter, bit error rate and so on. This can be achieved by operating the physical layer in multiple modes. There are seven types of physical layer modes (PHY modes) available; each of them has different modulation and channelcoding schemes. Any of these schemes can be adapted dynamically by Link Adaptation techniques. The physical layer of HIPERLAN/2 uses Orthogonal Frequency Division Multiplexing (OFDM) techniques. Providing QoS guarantees for real-time data streams such as voice and video over best effort networks is a challenging task. Several techniques based on Diffserv protocol have been investigated. However, most of them treat all portions of an MPEG-2 bit stream uniformly regardless of the importance of the each portion and how it contributes over all towards application specific QoS parameters such as jitter, PSNR, (pixel signal noise ratio) etc. In reality, each bit, depending on which macro block and frame it belongs to, contributes differently towards such application specific QoS parameters. For example if a bit belongs to an intra-coded macro-block from an I- frame, its contribution to the over all PSNR is much more significant compared to a bit belonging to a backward motion compensated macro block in a B- frame. This is because video data is highly temporally correlated and small errors in the bit streams can be easily compensated via online concealment techniques during the decoding phase. Based on the temporal and spatial locality characteristics of video data, and using the flexibility available in the physical layer of HIPERLAN/2, we present an efficient scheme to ensure any application specific QoS for video data. For a given QoS, our scheme provides higher throughputs and utilizes far less network resources compared to the approaches that don t consider application characteristics. Using a marking algorithm, we divide the bit streams into

several portions based on their significance from QoS point of view. The most significant portions of the bit stream are transmitted using PHY modes that yield low bit error rate (BER) and provide higher protection such as ½ BPSK at the expense of low throughput. On the other hand, less significant portions are transmitted using PHY modes that yield high throughput but are also prune to higher BER. An error concealment algorithm is applied to recover from errors. The simulation framework has been designed in MATLAB. For certain values of E b /N 0, the proposed transmission scheme provides higher throughput while providing better picture quality as well. The rest of the paper is organized as follows. In Section 2, a brief description of the HIPERLAN/2 standard is presented and simulation results of different PHY modes in terms of throughput and BER are discussed. For the sake of completeness, the hierarchical structure of the MPEG-2 video bit stream is briefly described in Section 3. This section also contains a marking algorithm that identifies the significance of different portions of an MPEG-2 bit stream and relates it to a corresponding PHY mode of HIPERLAN/2. The simulation results of transmitting video data over HIPERLAN/2 using the proposed marking algorithm are also presented in this section. Conclusions are presented in Section 4. 2. HIPERLAN/2 (H/2) HIPERLAN/2 defines three basic layers, namely: Physical layer (PHY), Data Link Control layer (DLC), and Convergence layer (CL). The protocol stack is divided into a control plane and a user plane [1]. The user plane includes functions for transmission of traffic over established connections, and the control plane includes functions for control of establishment, release, and supervision. The transmission format on the physical layer is a burst, which consists of a preamble part and a data part, where the latter could originate from each of the transport layers within DLC. A key feature of the physical layer is to provide several modulation and coding alternatives. This is to adapt for current radio link quality and to meet the requirements for different physical layer properties as defined by the transport channels within DLC. The Data Link Control (DLC) layer constitutes the logical link between an Access Point (AP) and the mobile terminals (MTs). The DLC includes functions for medium access and transmission as well as terminal/user and connection handling. The DLC layer consists of a set of sub layers: medium access control, error control, and radio link control. Compare to WLAN, in HIPERLAN/2, the medium access is based on a TDMA/TDD technique and uses a MAC frame with duration of 2 ms. The centralized control is provided by an AP which informs the MTs at which point in time in the MAC frame they are allowed to transmit their data. Time slots are allocated dynamically depending on the need for transmission resources. HIPERLAN/2 operates as a connection oriented wireless link, supporting the differentiated QoS levels required for transmission of the various media. Convergence Layer (CL) lying between data link layer and network layer provides QoS. The role of CL is two fold: (1) to map the service requirements of the higher layer to the service offered by the DLC layer, and (2) to convert packets received from the core network to the format expected at the lower layer (see Figure 1). There are two types of convergence layers. One is cell based and the other is packet based, we focus on the packet based CL. The CL can be further divided into a common part and a service specific part. Packet based service specific convergence sublayers (SSCS) exist for switched Ethernet and IEEE 1394 Fire wire. The architecture of the CL makes HIPERLAN/2 suitable as a radio access network for different fixed networks, e.g. Ethernet, IP, ATM, UMTS, etc. The main role of the common part is to segment packets received from the SSCS, and to reassemble segmented packets received from the DLC layer before they are handed over to the SSCS. The Ethernet SSCS makes the HIPERLAN/2 network look like wireless segment of a switched Ethernet. CL SAP Packet based CL Part 2 SSCS part Ethernet SSCS Part 1 Common part IP network layer IP SSCS 1394 SSCS Common part Convergence sublayer Segmentation and re-assembly DLC SAP Figure 1: Convergence layer model in the HIPERLAN/2 standard.

2.1 QoS in HIPERLAN/2 Quality of Service (QoS) refers to the ability of a network to provide service for specific network traffic over underlying wire line or wireless technologies. HIPERLAN/2 supports QoS in terms of bandwidth, delay, jitter, bit error rate, etc. Two QoS schemes supported in CL are: best effort and IEEE 802.1p priority schemes. Compared to end-to-end IP network QoS, wireless network may provide QoS only for one or two hops of an end-to-end connection. The cellular radio access in the European Broadband Radio Access for IP based Networks (BRAIN) uses HIPERLAN/2 and can support QoS on a per connection basis. An IP CL is used to provide functions required for mapping the QoS requirements of the individual connection to the QoS parameters available for data link control (DLC) connections [2]. In the DLC layer, HIPERLAN/2 uses centrally controlled TDMA/TDD scheme. Medium access control (MAC), radio link control (RLC) and error control (EC) are the three main functions provided by the DLC layer. MAC enables full rescheduling in every 2 ms and dynamic adjustment of uplink and downlink capacity. RLC provides establishment of connection-oriented secured links service to the CL. There are up to 63 unicast data connections per terminal, which support various QoS parameters. EC provides selective repeat ARQ mode for each connection. Alternatively, for delay intolerant services and multicast services a repetition mode can be used. Therefore, DLC provides means for executing several IP QoS techniques such as prioritization, on demand based bandwidth reservation, and delay guarantee. DLC also provides dynamic frequency selection (DFS), link adaptation (LA), power control, and power saving. PHY mode adapts to current radio link quality and satisfies the requirements of different physical layer as defined by transport channels within DLC. TABLE I provides different PHY modes and their transmission rates [3]. TABLE I: PHY MODES SUPPORTED IN HIPERLAN/2 Seven PHY modes use BPSK, QPSK and 16-QAM as mandatory sub-carrier modulation schemes; the 64- QAM is optional. Forward Error Correction (FEC) is achieved by a convolutional code of rate r = ½ and constraint length of seven. Other code rates of ¾ and 9/16 are obtained by puncturing. The link adaptation scheme in the DLC layer assigns a specific PHY mode to the packet data units (PDUs) dedicated to one connection based on the current radio link conditions. Each connection and its direction can be addressed individually and the assignment may vary from one MAC frame to another. Thus, link adaptation scheme adapts the PHY robustness based on link quality measurements. Total system throughput, transmission delay and BER are the important parameters in determining the performance of the HIPERLAN/2 radio access. Since there are strong interaction between PHY modes and these parameters, it is important to determine which values constitute the basis for the choice of PHY modes. 2.2 Simulation of PHY modes We used the MATLAB communication block set to simulate the seven PHY modes and their effects on QoS parameters [4]. The details of simulation process and results are given in the following sections. To simulate different PHY modes, we construct simulation models for each of the modes. Figure 2 shows the interconnection of different processes involved in the simulation. The details of each process may vary depending on the PHY mode being simulated. The input PDU is from the MAC layer. The MAC protocol is used for organizing the access and transmission of data on the radio link. There are two kinds of PDUs, one is long PDU (LCH PDU), which is 54 bytes long, and another is short PDU (SCH PDU) that is 9 bytes long. The PDU error ratio (PER) refers to the error rate of LCH PDU [5]. Input PDU train from DLC layer Convolutional coding Puncture Error Rate Calculation Viterbi Decoder Gaussian White Noise Insert 0 Modulation Scheme Demodulation Scheme Figure. 2: Different simulation modules.

Figure 3 shows the influence of PHY modes on HIPERLAN/2 Bit Error Rate (BER) performance. Note that for a given E b /N 0 BPSK and QPSK have lower BER than M-QAM when we add the same amount of channel noise. In general, BPSK with r = ½ is the most reliable PHY mode, followed by BPSK with r = ¾, QPSK with r = ½, QPSK with r = ¾, 16-QAM with r = 9/16 and 16-QAM with r = ¾. 64-QAM with r = ¾ has the worst BER performance. When we use 16-QAM with r = ¾, the system capacity increases but the connections experience a high retransmission delay. The high delay may be tolerable for the best effort traffic, but not for the real time traffic. Since BPSK with r = ½ achieves the most reliable service due to its low BER, user will pay the highest price to use this service. 64-QAM with r = ¾ can tolerate more errors and may be cheaper. Every user wants to use the cheapest service that satisfies the quality requirements. This makes it important to find out how to improve the quality of service with the least cost. Figure 3: BER performances for different PHY modes. System throughput is another important parameter that can be used to determine the HIPERLAN/2 performance. It can be calculated by considering the protocol overhead introduced by MAC and ARQ retransmissions. MAC layer throughput can be expressed by: Throughputs_MAC = [L LCH /(54/BpS LCH )]*48*8/2ms (1) where L LCH is the number of OFDM symbols available in the MAC Frame for data PDUs (LCH-PDUs). BpS LCH is the number of bytes to be transmitted per OFDM symbol [6]. Because different PHY modes have different L LCH and BpS LCH, MAC layer throughput is determined by the selected PHY mode. DLC layer throughputs can be obtained by multiplying the MAC throughputs with (1-PER PHY mode ). This is because DLC throughput is also affected by ARQ. PER PHY mode in the equation is the current PDU error rate of selected PHY mode [7]. Therefore, the DLC throughputs can be expressed by: Throughputs_DLC = Throughputs_MAC * (1-PER PHYmode ) (2) Figure 4 shows the system throughputs of different PHY modes. Figure 4: System throughput for different PHY modes. 3. MPEG-2 VIDEO TRANSMISSION OVER HIPERLAN/2 For the sake of completeness we first briefly describe the hierarchical structure of the MPEG-2 bit stream. 3.1 MPEG-2 data stream hierarchy An MPEG-2 video sequence is an ordered stream of bits, with a special bit patterns marking the beginning and ending of a logical section. Each video sequence is composed of a series of Groups of Pictures (GOP's). A GOP is composed of a sequence of pictures (frames). A frame is composed of a series of slices. A slice is composed of a series of macroblocks, and a macroblock is composed of 6 or fewer blocks (4 for luminance and 2 for chrominance) and possibly a motion vector. As shown in Figures 5 and 6, the format of MPEG-2 video bit stream is based on a hierarchy structure. It starts with arbitrary number of bytes followed by a sequence header, followed by zero, one or more alternating sequence of GOP header and GOPs, followed by a sequence end marker. A GOP is a series of pictures (frames). Each picture consists of a picture header and the actual picture data. The GOP is

intended to assist random access into a sequence. A GOP is independently decodable unit that can be of any size. It always begins with an Intra-coded frame. (There is one caveat here, SEQUENCES are a higher level structure than GOPs, and may contain information about quantization tables. Their information is needed to decode all the following GOPs.) MPEG picture is composed of slices where each slice is a sequence of macroblocks in order. Each slice is an independently decodable unit too. There can be one slice per frame, one slice per macroblock, or anything in between. The slice structure is intended to allow decoding in the presence of errors. It also allows parallel encoding/decoding at the slice level. An I-frame carries most original information and can propagate errors to other frames. P-frame is coded as differences from the last I- or P-frame. Taking the last I- or P-frame and predicting the values of each new pixel predict the new P-frame. P-frame uses Motion Prediction and DCT encoding. A macroblock in a P frame can be encoded either as an Intra-coded macroblock or as a forward predicted macroblock. Different P frames have different abilities to propagate errors. This is determined by the position of a P frame in one GOP. B frame is coded as differences from the last or next I or P frame. B frame uses the same prediction as P frames. But for each macroblock, either the previous I or P frame or the next I- or P- frame is used. B frame also uses Motion Prediction and DCT encoding. A macroblock in a B frame can be encoded as an intra-coded macroblock, a forward predicted macroblock or a Bi-directional macroblock. B frame does not propagate any errors in the transmission. In order to detect the number of the sequence, GOP, picture, slice, frame, or a macroblock in a bit stream, one should know the starting code for each structure. This is shown in TABLE II. Figure 5: MPEG-2 bit stream structure. TABLE II: STARTING CODES OF DIFFERENT HEADERS Byte 0 Byte 1 Byte 2 Byte 3 Used for 00 00 01 Start code prefix Stream ID Figure 6: MPEG-2 video bit stream structure. There are three types of pictures that the MPEG standard defines: Intra-coded pictures (I-frame), Predicted pictures (P-frame) and Bi-directional pictures (B-frame). An I-frame uses Discrete Cosine Transform (DCT) encoding only to compress a single frame without reference to any other frame in the sequence. Only intra-coded macroblocks are available in an I-frame. 00 00 01 00 Pictures 00 00 01 01-AF Slice 00 00 01 B0 Reserved 00 00 01 B1 Reserved 00 00 01 B2 User data 00 00 01 B3 Seq. header 00 00 01 B4 Seq. error 00 00 01 B5 Extension 00 00 01 B6 Reserved 00 00 01 B7 Seq. end 00 00 01 B8 GOP From the table, the starting code for a sequence is 00 00 01 B3, whereas sequence end code is 00 00 01 B7. Each GOP has its own starting code as 00 00 01 B8. GOP consists of several frames whose starting code is 00 00 01 00. From picture header, we can find which kind of frame it is. Finally, slices in one frame have its own sequence number starting from 00 00 01 01 to 00 00 01 AF [8].

3.2 Marking algorithm Since different portions of an MPEG bit stream have different significance in terms of their impact on the picture quality, we can dynamically choose an optimal combination of modulation and channel-coding schemes to satisfy the overall BER and system capacity requirements. The simulation results in previous section show that there is some trade-off between the two requirements: with the lower BER, a modulation and channel-coding scheme provide lower system throughput. We designed a making algorithm to classify macroblocks in a MPEG-2 bitstream. It identifies the modulation schemes for each macroblock depending on its type and the frame it belongs to. TABLE III shows an example marking of the macroblocks. In our future work, we are exploring more detailed marking algorithms that also examine frequency contents of the macroblock to classify its significance. TABLE III: MARKING ALGORITHM USED IN MPEG-2 BIT STREAM. on the frame-level marking only. We have used BPSK with r = ½ scheme to transmit an I-frame, QPSK with r = ¾ scheme to transmit a P-frame, 16-QAM with r = ¾ scheme to transmit the first B-frame and 64-QAM with r = ¾ scheme to transmit the second B-frame. Figures 7 and 8 show the BER performance and system throughput of the bit stream transmission using different modulation schemes. The BER performance of BPSK and QPSK modulation schemes is better than that of the marking algorithm based transmission scheme. The reason is obvious: marking algorithm assigns QAM modulation scheme to transmit B frames which has the worst BER performance and brings a large number of bit errors into original bit stream. On the other hand, I-frame and P-frame contain the original information and are transmitted by the most reliable schemes BPSK and QPSK. Therefore, almost no errors occur in these two frames. Although the BER increases after applying the marking algorithm, the quality of the decoded pictures is still acceptable. Errors caused in the B frame can be corrected with the reference of I frame and P frame using simple error concealment techniques. Figure 7: BER performance for transmission of MPEG bit stream with marking algorithm. 3.3 Simulation Results As discussed in the previous section, combining simulation results of QoS performance in HIPERLAN/2 with a marking algorithm we can optimize the transmission of MPEG bit data. The MPEG bit stream used in our simulations contains one GOP consisting of four frames in the sequence IPBB. The results reported in the following are based The curve with circles in Figure 8 represents the system throughputs of MPEG-2 bit stream that is obtained using the marking algorithm based transmission scheme. When E b /N o increases, throughput keeps on increasing until E b /N o reaches 15dB. At that point, system throughputs are almost 10Mbps. This is much larger than the value obtained by using BPSK or QPSK scheme only.

Figure 8: Throughputs for transmission of MPEG bit stream with marking algorithm. Figure 9 shows the quality of pictures obtained after decoding the marking algorithm based MPEG-2 sequence. Note the black streaks in 2 nd and 3 rd pictured. These represent the damaged blocks. Figure 10: Pictures of Figure 11 with an error concealment algorithm at the receiver end. The MPEG-2 bit stream used in our simulations was obtained from www.mpeg.org, namely video.bits. It is an official MPEG-2 video conformance testing bitstream at CSTV. Figure. 10 shows pictures obtained after applying a simple error concealment procedure where a macroblock with error is replaced by a macroblock at the same position from the previous decoded frame. Figure 9: Decoded pictures that were transmitted using the marking algorithm. The quality of the pictures received before and after the error concealment algorithm is presented in terms of pixel to signal noise ratio (PSNR), as shown in Figure 11. For lower values of E b /N 0 (> 7 db), the quality of picture in our proposed scheme is better compared to when only QPSK or QAM are used. While for the

same ranges of E b /N 0 the throughput of the proposed marking algorithm based transmission scheme is also better. 5. REFERENCES [1] Janne Korhonen: HIPERLAN/2, Department of Computer Science and Engineering, Helsinki University of Technology, 1999. [2] ETSI TR 101 957 V1.1.1 Technical Report: Requirements and Architectures for Interworking between HIPERLAN/2 and 3 rd Generation Cellular systems. Figure 11: PSNR of the pictures received in the presence of an error concealment algorithm. 4. CONCLUSIONS We have studied the impact of different HIPERLAN/2 PHY modes on network QoS parameters such as BER transmission delay, and throughput. We have used the MATLAB communication toolbox to build up the simulation models for each PHY mode. Combining these simulation results with a marking algorithm, we can improve the quality of the transmission for MPEG- 2 bit stream in terms of system capacity and Bit Error Rate (BER). The proposed marking algorithm exploits the fact that different portions of an MPEG bit stream have different significance in terms of their impact on the picture quality. This has allowed us an dynamically choose an optimal combination of modulation and channel-coding schemes to satisfy the overall BER and system capacity requirements. We are investigating an improved marking algorithm that produces differentiated macroblocks not only based on the coding schemes used inside the macroblocks but also uses the frequency contents of the macroblocks. [3] J. Khun-Jush, P. Schramm, U. Washsmann, G. Wenger: Structure and Performance of HIPERLAN/2 Physical Layer. pp. 2667-2671 IEEE VTC 99 fall (Amsterdam). [4] The Mathworks home supported documentation: Learning About the Communication Blockset. http://www.mathworks.com/access/helpdesk/help/toolb ox/commblks/commblks.shtml [5] Angela Doufexi, Simon Armour, Andrew Nix, David Bull: A Comparison of HIPERLAN/2 and IEEE 802.11a Physical and MAC layers. Centre for Communication Research, University of Bristol, UK. [6] A. Kadelka, A. Hettich, S. Dick: Performance Evaluation of the MAC Protocol of ETSI BRAN HiperLAN/2 Standard. Proc. of the European Wireless 99, ISBN 3-8007-2490-1, (Munich, Germany), pp. 157-162, Oct. 1999 [7] B. Walke: Mobile Radio Networks. New York, USA: Wiley & Sons Ltd., 1, ed., 1999. [8] MPEG headers Quick reference. http://members.aol.com/mpucoder/dvd/mpeghdrs.ht ml