Performance Considerations for Efficient Multimedia Streaming in Wireless Local Area Networks Dilip Krishnaswamy1, Robert J Stacey2, Ryan van Alstine3, William J Chimitt4
[email protected],
[email protected],
[email protected],
[email protected]
ICG-CTO-Architecture-Lab, Intel Communications Group, Intel Corporation, USA ABSTRACT This paper presents investigates multimedia streaming over wireless local area networks. Physical layer sigmoid analytical models are presented for 802.11a/g and for 2x3 MIMO 802.11n MIMO-based systems are presented. Performance results in a wireless LAN environment are presented for traffic using UDP and TCP transport mechanisms. Packet losses are observed in WLAN environments which affects the overall throughput available. Possibilities for performance improvements with the use of 802.11e and MIMO technologies are discussed. System platform architecture performance issues for wireless video conferencing between Intel® PXA27x processor-based handheld platforms are presented and results with retry-limit adaptation are also presented. Keywords: Sigmoid Modeling, Throughput, Performance, Streaming Multimedia, Wireless LAN, PHY & MAC layers, MIMO, UDP, TCP, 802.11a, 802.11g, 802.11n, 802.11e, WSM, WME, Platform architecture, Video conferencing
1. INTRODUCTION 1.1. Multimedia streaming over wireless LAN networks Wireless Local Area Networks offer several challenges with regard to multimedia streaming [1, 2]. Dynamic variation in channel conditions with noise and interference impact performance. Dynamic changes in the number of users in the network with their varying data rate requirements resulting in a varying degree of contention and collision in the network impact performance. Real-time adaptation at the MAC-layer is required to adapt to varying conditions. The choice of the transport layer such as TCP or UDP is also of concern. The multimedia application by its very nature has the ability to scale and adapt to varying wireless network conditions, so that the overall impact to the user is minimized. Platform architectural constraints can also impact performance. This paper will present several performance considerations for efficient multimedia streaming in wireless LANs. Section 2 discusses sigmoid modeling for physical layer throughput for WLAN protocols. Section 3 discusses MAC layer tradeoffs. Sections 4 and 5 discuss possibilities for performance improvement with 802.11e and MIMO based techniques, respectively. Section 6 discusses performance results with wireless video streaming with TCP and UDP transport protocols. Section 7 presents platform architecture considerations for wireless video streaming between Intel® XScale-based handheld platforms, and presents performance results with retry-limit variation. Section 8 concludes the paper.
2. PHYSICAL LAYER PERFORMANCE 2.1. Sigmoid throughput modeling Several modulation and coding schemes are available to a wireless station in a wireless data network. Modulation schemes that allow a larger number of bits per symbol, have symbols closer to each other in the constellation diagram, and small errors can result in erroneous decoding. Varying code rates can be employed within each modulation scheme to adapt to changing channel conditions by allowing more bits for coding (lower code rate k/n) as conditions 1
1900 Prairie City Road, M/S FM7-252, Folsom, CA 95630 2111 NE 25th Ave, M/S JF3-248, Hillsboro, OR 97124 3 77 Reed Road, M/S HD2-330, Hudson, MA 01749 4 1900 Prairie City Road, M/S FM7-270, Folsom, CA 95630 2
deteriorate. As the code rate decreases, the effective data rate reduces, and hence the achievable throughput reduces. We will refer to the term “mode m” to refer to a specific choice of a modulation and coding scheme. The probability Pem (L) , of error in a packet of length L bytes (also referred to as the physical layer packet loss rate PLR), for a given mode m, as a function of the bit error probability pb is given by equation (1), where the inequality represents the fact that one can recover from bit errors in a packet, due to the coding scheme used at the packet level. (1) Pem ( L) ≤ 1 − (1 − pbm )8 L The effective physical layer throughput can be then expressed as TPHYm(x) = A(1 − Pem ( L)) . For a given mode m, TPHYm(x) and Pem (L) can be approximated with sigmoid functions [3] of the form (2) TPHYm(x) = A / (1 + e–λ(x– δ)) λ(x– δ) ) (3) Pem (L) = 1/(1 + e where x is the SINR in dB, and y= TPHYm(x) is the throughput in Mbps. The factor A represents the maximum achievable data rate in Mbps for each mode. The sigmoid function has asymptotes y=A, and y=0, and is symmetric about x = δ. The factors λ and A determine the maximum slope (Aλ/4) of the sigmoid curve at x= δ. Let Km be the upper knee of TPHYm(x), (maximum curvature point of TPHYm(x)), where the curvature function C(x) is defined as '' ' (4) C ( x) =| TPHYm ( x) /(1 + TPPHYm ( x))3 / 2 | Game theoretic formulations could be used to develop optimization techniques for multimedia adaptation using such sigmoid mathematical modeling as described in [4]. PHY-level throughput versus SINR curves are shown for 802.11a/g networks and 802.11n networks in sections 2.2 and sections 2.3. 2.2. 802.11a/g PHY throughput Figure 1 shows the sigmoid PHY layer throughput curves obtained with modeling packet transmissions over an 802.11a WLAN network as a function of the SINR. A 75ns exponentially fading channel [5] was used, and 1000 byte packet lengths were used. The PHY layer throughput L(x) is plotted as a function of the SINR, x. It should be noted that the BPSK, rate ¾ mode is not particularly desirable since the QPSK, rate ½ mode is always better from a throughput perspective for the same SINR. Sigmoid modeling (proposed in [3] for GPRS/EGPRS networks) is used for the throughput, to obtain an analytical expression based on the matlab simulations for packet transmission over the physical layer channel for the 802.11a modulation and coding schemes. For the sigmoid throughput functions in Figure 1, the corresponding sigmoid parameters (A, λ, δ) for the 8 modes are provided in Table 1, along with the Km values.
PHYThroughput (Mbps)
60
64QAM, 3/4 64QAM, 2/3 16QAM, 3/4 16QAM, 1/2 QPSK, 3/4 QPSK, 1/2 BPSK, 3/4 BPSK, 1/2
50
40
30
20
10
0
0
5
10
15
20
25
SINR (in dB) Figure 1 - Sigmoid PHY throughput functions for 802.11a/g networks
30
Mod Scheme, code rate (k/n) 64-QAM, 3/4 64-QAM, 2/3 16-QAM, 3/4 16-QAM, 1/2 QPSK, 3/4 QPSK, 1/2 BPSK, 3/4 BPSK, 1/2
A (Mbps) 54 48 36 24 18 12 9 6
δ (dB) 21.2 18.2 15.1 10.9 9.3 5.3 6.1 2.3
λ (dB-1) 0.419 0.625 0.352 0.375 0.444 0.461 0.417 0.640
Km (dB) 29.45 24.18 23.27 16.87 14.76 9.84 10.40 5.15
Table 1 - Sigmoid parameters for 802.11a/g PHY-throughput as a function of the SINR
2.3. Increased Throughput with Multiple Antennas Multiple transmitting antennas [6] can help increase the data rates in the same channel. Multiple receiving antennas can help in efficient recovery of the transmitted data. Multiple antennas can also be used to increase range and reliability of data transmitted in a channel without increase in data rates. Alternatively, one could increase data rates using multiple antennas by transmitting data in multiple channels simultaneously. Increasing the channel bandwidth from 20MHz to 40MHz can also increase data rates. Higher throughput could be achieved with coding schemes such as LDPC (Low Density Parity Check) codes and Turbo codes. Matlab modeling of increased PHY-level throughput with a 2x3 MIMO (Multiple Input Multiple Output) WLAN network, with 2 Transmitters using SFC and 3 Receivers using MMSE are shown in Figure 2 with a 75ns exponentially fading channel. The BPSK, rate ¾ mode continues to be undesirable in the MIMO configuration with the QPSK, rate ½ mode always better from a throughput perspective for the same SINR. Table 2 shows the sigmoid parameters (A, λ, δ) and the Km values for the 8 modulation and coding schemes. It can be seen that the maximum available throughput has doubled compared to the traditional 802.11a physical layer.
Figure 2 - Sigmoid PHY throughput functions for 2x3 MIMO (SFC Tx and MMSE Rx)
Mod Scheme, code rate (k/n) 64-QAM, 3/4 64-QAM, 2/3 16-QAM, 3/4 16-QAM, 1/2 QPSK, 3/4 QPSK, 1/2 BPSK, 3/4 BPSK, 1/2
A (Mbps) 108 96 72 48 36 24 18 12
δ (dB) 18.8 17.4 13.2 9.4 7.1 3.1 3.5 0.7
λ (dB-1) 0.854 0.833 0.889 1.258 0.907 1.006 0.815 3.158
Km (dB) 24.5 23.1 18.3 12.9 11.3 6.6 7.2 2.0
Table 2 - Sigmoid parameters for 2x3 MIMO 802.11n PHY-throughput as a function of the SINR
3. MAC-LAYER PERFORMANCE IMPACT The effective throughput at the top of the MAC is further reduced due to a number of factors [7][8][2][1] such as the number of current users in the network medium, user requirements, priorities, retry-limits, and link adaptation schemes used, channel conditions based on noise and interference, backoff counter depths, backoff stages, protocol timing and header overheads, and also, the amount of additional time that the medium is unused/idle. The overall throughput is also affected by the transport mechanism used (TCP/UDP/UDP-lite), and whether there is additional application-layer redundancy such as FEC across packets being used. In general, the overall throughput as a function of the SINR continues to assume a sigmoidal form with a reduced maximum asymptotic value for the throughput.
3.1. Packet Errors and Retransmissions An ARQ mechanism is used at the MAC such that a retransmission is attempted in the absence of an acknowledgement. Packets may be in error if the acknowledgement is not received from the destination in the SIFS duration period after transmission. Either the packet may be in error when it reaches the destination or the return acknowledgement might itself be in error. In both cases, a packet is retransmitted by the MAC, as long as the retry-limit is not reached. During retransmissions, link adaptation may be performed to attempt sending a packet in a more robust modulation and coding scheme. Queue backup in the transmit queue at the MAC layer could result in packets being dropped which can also impact performance. Jitter and delay requirements may impact how many retransmissions should be attempted before further packet retransmission attempts are discarded. If a reliable transport mechanism such as TCP is used, then retransmissions at the MAC layer should be preferred as MAC-level retransmissions have significantly less overhead compared to retransmissions attempted from the transport layer. Additional overhead at the MAC layer occurs due to collision with other users in the CSMA/CA protocol. To avoid collision, an exponentially increasing backoff counter is used, which can cause increased overhead. 3.2. Link Adaptation
Figure 3 - Hysteresis Loop for Adaptation One can apply a penalty function [3] given by P(x) = α (x – Km) to the throughput for x > Km. The desired behavior is depicted in the “hysteresis loop” in Figure 3 where OOP represents the optimal operating point [3] for a given choice of a modulation and coding scheme (this is a figure of the throughput functions zoomed in for two modes). The throughput improvement becomes marginal beyond the OOP. The term ASP denotes the Adaptation Switching Point where it can
become advantageous to switch to a different modulation and coding scheme. As the network conditions change, the OOP will dynamically vary for each wireless station. It can be difficult to distinguish between foreign interference, collisions and noise. The effectiveness of the adaptation is limited by the rate at which the channel changes (including activity of other devices) and how fast the algorithm adapts (limited by the integration period over the metrics used for adaptation).
4. PERFORMANCE IMPROVEMENTS TO CONSIDER WITH 802.11e/WME/WSM The 802.11e protocol is designed to optimize QoS. Two service classes are suggested – QoSNoAck and QoSAck. The QoSNoAck mechanism may not be very useful as it is desirable to attempt retransmissions at the MAC layer. There are 4 access categories (Voice, Video, Best Effort, and Background) and 8 user priorities (2 for each access category). The access point configures the MAC-level parameters such as the contention window size (CWmin/max), the interframe spacings (AIFS), and the transmission opportunity duration (TxOP). Reducing the contention window size reduces the overhead of backoff so that a transmission can be attempted quickly. Reducing the interframe spacings also reduces the additional protocol timing overhead relative to the time taken to transmit data. Increasing the transmission duration (TxOP) reduces contention overhead by allocating a fixed period during which an STA can transmit more than one MDSU. Block acknowledgements can be used to reduce overhead of acknowledgements. A block acknowledgement can indicate the received MDSUs so the MDSU’s in error can be retransmitted during the TxOP duration. The WME (Wireless Multimedia Enhancements) specification implements the EDCA contention-based access mechanism suggested in 802.11e with the access priorities. The WSM (WiFi Scheduled Multimedia) specification provides support for HCCA-based centralized scheduling in addition to EDCA. The HCCA mechanism allows for bandwidth to be dedicated to a specific transmission in a contention-free period. All stations admitted with dedicated bandwidth in the contention-free period complete their transmissions and then the rest of the stations contend for access using EDCA during the contention-period.
5. PERFORMANCE IMPROVEMENTS TO CONSIDER WITH 802.11n 802.11n will introduce a number of techniques aimed at improving data rates in the PHY and efficiency in the MAC which together will provide significantly higher throughput over the existing 802.11 standards. Among the PHY techniques, spatial multiplexing through MIMO promises to provide a significant boost in throughput as does the use of wider (40MHz) channels. Further incremental improvement will be achieved by increasing the number of data carrying OFDM subcarriers, higher 5/6 or 7/8 rate convolutional coding and/or the use of advanced channel capacity approaching codes such as LDPC. Range and reliability improvements will be achieved with additional transmit and receive diversity gained from the use of multiple antennas. MAC layer efficiency improvements are required since the fixed overhead in the PHY preamble and the inter-frame spacing become more prominent with increased PHY data rates. The primary technique used to increase the MAC efficiency will be aggregation – combining multiple MPDUs into a single PPDU. This necessitates the use of the block acknowledgement protocol introduced in 802.11e. Aggregation also enables data frames and control frames to be piggybacked and advantage may be taken of this to combine data frames with acknowledgements in the return direction. With reverse direction data the holder of a TXOP subleases a portion of it to a peer with an aggregate data transfer so that the peer may combine the returned block acknowledgement with data. This has been shown to improve efficiency, especially with traffic where the bulk of the data flow is in one direction and a small amount of data is returned in the reverse direction, for example video with control feedback or a TCP session where TCP acks make up the reverse direction data. An additional enhancement that improves performance in the presence of interference is closed loop link adaptation. The MAC protocol may be modified to include explicit feedback from the responder to aid the initiator in selecting an optimal modulation and coding scheme. The current implicit scheme based on perceived packet error rate is suboptimal in the presence of interference, the impact of which cannot be distinguished from actual changes in the channel conditions.
6. PERFORMANCE RESULTS WITH VIDEO STREAMING OVER WLANs 6.1. Video Streaming Setup Measurements have been taken to determine the characteristics of streaming media transported over wireless LANs. These measurements were taken in 22 homes in the Portland, OR and San Diego, CA areas using 802.11g devices. Samples of the measurements showing typical characteristics are given in the figures in the discussions that follow. The measured results show throughput and packet loss as seen above the MAC. Packet loss results when either the transmit queue overflows or the maximum retransmit count is reached. IxChariot was used to stream 6Mbps constant bit rate traffic between end-points with AiroPeek used as a sniffer. Figure 4 shows a typical test setup for measurements.
Figure 4 - Video Streaming Setup
6.2. Measured vs Theoretical throughput for 802.11g WLANs The measured maximum throughput of a standard 802.11g setup is given in Figure 5.
Throughput (Mbps)
35.0 G-only (theoretical max)
30.0 25.0
G-only
20.0 15.0
CTS-to-self (9uS, short)
10.0
RTS/CTS (9uS, short)
5.0 0.0 0
500
1000
1500
Packet Size (Bytes)
Figure 5 - Measured maximum throughput with less than 1% loss [Test031804, DI-774, GWL-G810] Throughput was measured using the Ixia test setup using a binary search with less than 1% packet loss as a constraint. At least 3 measurements were taken and the results averaged. These tests used the short (9uS) slot time and the CTS and RTS, when used, where sent with a short preamble and 11 Mbps modulation. Data packets used 54 Mbps modulation and the ACKs used 24 Mbps modulation (using 24 Mbps results in only 4uS of additional overhead over 54 Mbps for these short packets, but greatly increases reliability). The AP and STA were 4 meters apart.
Notice that the measured throughput is significantly below the theoretical maximum. Packet errors occur even in a very clean RF environment. Packet loss results in retransmission, with the retransmitted packets using additional air time. Packet loss also results in the contention window widening (since packet loss is assumed to result from collisions) and thus an increase in the inter-packet spacing. These results provide realistic upper bound on actual throughput that can be expected on an 802.11g WLAN. 6.3. Impact of interference on wireless multimedia streaming with TCP/IP Figure 6 shows the impact of prolonged interference when using TCP as the transport protocol. TCP provides a reliable transport and will retransmit lost data at the transport layer. Video transported over TCP at low bit rates does not suffer the loss seen with video transported over RTP. TCP does, however, suffer from prolonged interference events. In this experiment, a constant bit rate is streamed using TCP as the transport. Here a prolonged interference event caused the TCP queues to backup applying backpressure to the application. As a result, the throughput drops. The backup in the queues results in considerable delay as shown in the response time. An unhelpful side effect of TCP is showing up here. TCP assumes that packet loss is due to congestion and as a result it drastically reduces its transmit rate on detecting packet loss and then slowly ramps up again. This response is not well suited to the unpredictability and rapidly changing nature of the interference and unnecessarily crimps throughput and adds delay.
Figure 6 - Impact of prolonged interference with TCP
6.4. Wireless Multimedia Streaming with RTP/UDP Figure 7 shows measured throughput and packet loss for a 6Mbps UDP stream between a server in the family room to a client in the home office separated roughly by 10m linear distance and two floors up. Several experiments were performed with streaming done for a duration of 5 minutes in each experiment and the figure shows the outcome from a typical experiment. Included below is the RTP Payload info used for the experiments. – 802.11 MAC Header 24 bytes – 802.2 LLC Header 8 bytes – IP Header 20 bytes – UDP Header 8 bytes – RTP Header 12 bytes – Payload (7 x 188 byte MPEG frame) 1316 bytes – 802.11 MAC FCS 4 bytes Over the 5 minute run, it will be noticed that there are short periods (lasting up to a few seconds) of significant packet loss. During these periods, access to the medium is blocked due to energy detect by the CCA mechanism and throughput is constrained by a large number of retransmits on each packet. As a result, the transmit queue backs up. When conditions improve, a throughput burst occurs following the degradation as the transmit queue empties at the faster rate now supported by the medium. Figure 8 shows measured throughput and consecutive lost datagrams for a 6Mbps UDP stream between two devices in the same room separated by about 3m linear distance. Notice that a low level of packet loss is being experienced. Here, interference is resulting in enough consecutive packets being lost that the retry limit is reached occasionally and the packet discarded.
Figure 7 - Measured UDP throughput showing short periods of packet loss
Figure 8 - Measured UDP throughput showing light packet loss
Figure 9 - Measured UDP throughput showing severe degradation
Figure 9 shows measured throughput and packet loss packet loss for a 6Mbps UDP stream. In this case the streaming is interrupted by an extended period of interference lasting up to a minute. During this period, the high packet error rate results in the rate adaptation algorithm selecting a lower PHY data rate in an attempt to make the transmission more robust. The low PHY data rate and large number of retries per packet constrains available throughput to something lower than that required by the traffic stream. Packet loss results from tail drop as the transmit queue overflows and head drop as the retry limit is reached on many of the packets. 6.5. Overall Application/Transport layer considerations for Wireless Multimedia Streaming In general the performance results suggest that UDP be used as the preferred transport mechanism over TCP for multimedia streaming to reduce impact due to TCP’s congestion-avoidance mechanism. One should attempt MAClevel retries when using UDP, and application-layer FEC between packets over UDP [1][2] could be used to compensate for lost packets at the MAC layer. One needs to exploit scalability in multimedia representation and identify the most important information to communicate given the available conditions. MAC-level optimizations and crosslayer optimizations [2] conditions and joint source/channel coding [9] can help in adapting to optimally transfer the most relevant information over the wireless channel in response to current channel conditions.
7. WIRELESS VIDEO CONFERENCING BETWEEN HANDHELDS Finally, we discuss platform architecture issues and experiments performed with wireless video streaming between two handheld platforms based on the Intel® XScale architecture. The Intel® PXA27x Processor Developer Kit (Mainstone) constitutes a software-development kit based on the Intel® PXA270 Processor [10]. This kit provides the system components necessary for Microsoft Windows CE, Palm, Linux and other full-featured handheld operating systems. The Intel® PXA270 Processor is the third implementation of the Intel®XScale microarchitecture family, featuring an LCD controller, expanded card interface, and more conservative power management features. It is also the first processor in the family to include a baseband interface, integrated SRAM, a camera capture interface, scalable core frequencies, USB On-the-Go (OTG), full SDIO support, and an Intel coprocessor. The coprocessor extends the Intel® XScale microarchitecture capabilities by adding MMX and SSE functionality and additional audio and video processing operations [11]. A video conferencing application was used to demonstrate the capabilities of the Intel® PXA270 processor. The application is run under the Montavista distribution of embedded Linux that has been ported to the Intel®PXA27x Processor Development Kit. Montavista has based their distribution on version 2.4.21 of the Linux kernel. Included in this build of the OS is camera support, full duplex audio support, and PCMCIA support and full TPC/IP networking functionality. When running, the video conferencing flow captures frames from a camera that is mounted on the development kit. The captured frames are then processed through an MPEG4 encoder that utilizes Intel’s Integrated Performance Primitives. At the same time, the audio is taken from the microphone and condensed using GSMAMR, which is also a function included with the Performance Primitives. The encoded data is put into packets and sent across the network via TCP/IP. The received data is then decoded on a second development kit and printed to the screen or played on the speaker. This same mechanism occurs on both boards with a two way flow of packets, thus demonstrating video conferencing. From Figure 10, which depicts the video flows during video conferencing on the platform, it can be seen that there are several data flows in progress simultaneously on the platform, with the camera output being used to create an outgoing video stream and a self-preview video stream. In addition, an incoming video stream needs to processed, with encoding and decoding of audio streams also being processed at the same time, to enable video conferencing. The ability to perform all these simultaneous tasks imposes restrictions on the platform which constraints the maximum available performance on the platform. These platform architecture considerations are discussed next. 7.1. Platform Architecture Considerations The frames captured from the camera are in raw YUV 4:2:2 QCIF format (176x144). The maximum frame rate for MPEG4 QCIF encode on a Mainstone platform is approximately 90 frames per second. However, since the camera cannot capture more than 40 frames per second, the performance is limited by the camera capture interface to begin with. When a 12.2 bitrate GSMAMR encode is introduced, the frame rate will drop slightly further due to increased overhead in the platform for the additional processing. When running QCIF MPEG4 encode/decode from the camera and GSMAMR encode/decode from the microphone on one platform without sending any packets over a network, the frame rate is determined to be 28 – 32 frames per second.
Figure 10 - Video conferencing flow on the Intel® PXA27x Processor ([10, 11]) When creating packets with the encoded data and sending over a network via TCP/IP, the frame rate drops slightly. The first test done was using a wired network without any additional traffic. Two Mainstone platforms were connected through a hub and set up to communicate directly with each other. Each video frame could be an I-frame (approx 13Kbytes) or a P or B frame (0.2 – 1.2 Kbytes). Each audio frame is 310 bytes in length. The application was designed to grab the data and create a packet w/ four frames of video and audio before it is sent over the network using TCP/IP. Hence each packet is of variable size and could vary from about 2.1Kbytes to about 19.8Kbytes, resulting in a VBR (variable bit rate) multimedia stream. When using a wired link to communicate data with TCP/IP, a frame rate close to 23 frames per second on average is observed, due to the network interface overhead and driver limitations. It should be noted that the wired network driver under Montavista’s implementation of Linux could be optimized further for the Mainstone platform and better performance could be achieved. A high bandwidth network and better implementation of the drivers could enable increased throughput, and thus increasing the frame rates or frame sizes. Finally, when moving to a wireless networking solution, the bandwidth is considerably lower, so a frame rate from about 14 to 16 frames per second is observed. Wireless driver optimizations could improve performance and this implementation needs further investigation. Overall multimedia processing performance on the platform is primarily limited by the network bandwidth, the peripheral interface, and internal software implementation overheads and the performance is not limited by the CPU or system memories. Additional optimization could use light-weight communication information exchange to indicate degradation in battery-availability, and memory-availability with additional tasks in the system. Such information can be useful for the two end points to adaptively reduce processing requirements in terms of frame size and frame rates and hence minimize energy utilization to extend the duration of the video conferencing session longer. To perform some experiments on wireless network traffic, two Mainstone development platforms were used running the Montavista MVCLEE v3.0 implementation of Linux for the Xscale Microarchitecture. This was based on version 2.4.21 of the linux kernel and both the OS image and the application were built with the iwmmxt_le v.3.3 Montavista tools (compilers, linkers, etc). For the wireless connectivity, two Orinoco Wireless 802.11b PCMCIA cards were used with version 3.2.7 of the driver running in the kernel. 7.2. Retry-Limit Variation The wireless Linux implementation of the driver allowed for flexibility to investigate variations in the retry-limit used at the MAC layer. An additional 30fps video stream was transmitted using VideoLAN between two wireless laptops in the same network. Experiments were performed with and without the additional stream. Table 3 shows the results with the packet error rate varying as a function of the retry limit for both cases (with and without the additional contending
stream). Each experiment was done over at least 1000 packets and the experiment was repeated at least 10 times. The results in the table are averaged over multiple runs. In general, the packet error rate dropped with increasing max retrylimit value. It can be observed that the packet error rate increased for same retry limit value with the additional video stream causing collision in the same network between laptops. The observed MAC throughput varied from 1.5Mbps to 1.9Mbps with increasing retry-limit value without additional video stream. The observed MAC throughput varied from 0.7Mbps to 0.8Mbps with increasing retry-limit value with additional video stream. The perceived video quality also improved with increasing retry limit. Retry Limit
PER ( video conferencing) (%)
0 1 2 3
27.22 13.24 6.64 1.83
PER (video conferencing + additional video stream) (%) 28.43 13.74 7.43 2.29
Table 3 - Packet error rate reduction with increase in retry limit at the MAC layer
8. CONCLUSIONS To summarize the contributions of this paper, sigmoid models for physical layer throughput modeling for 802.11a/g and 2x3 MIMO 802.11n wireless LANs were presented. Results from wireless multimedia streaming experiments with RTP/UDP and TCP transport mechanisms were presented. PHY/MAC/Transport/Application layer optimization considerations were discussed. In general, any overall solution for multimedia over wireless networks must have the ability to scale and adapt to dynamic variations in wireless conditions. System architecture and performance issues for wireless video conferencing between two Intel® PXA27x-based systems were presented. Several considerations in the overall platform architecture can constrain the available multimedia performance in a system, and adaptive techniques such as retry-limit variation at the MAC-layer can improve multimedia performance.
REFERENCES [1] K. Stuhlmuller, N, Farber, M. Link, B. Girod, "Analysis of Video Transmission over Lossy Channels", IEEE Journal on Selected Areas in Communication, Vol 18, No 6, June 2000. [2] M. van der Schaar, S. Krishnamachari, S. Choi, X. Xu, “Adaptive cross-layer protection strategies for robust scalable video transmission over 802.11 WLANs”, IEEE Journal on Selected Areas in Communication, Vol. 21, Issue 10 , Dec. 2003, pp.1752 – 1763. [3] D. Krishnaswamy, “Game Theoretic Formulations for network-assisted resource management in wireless networks”, IEEE Vehicular Technology Conference, pp. 1312- 1316, Sep ‘02. [4] D. Krishnaswamy, M. van der Schaar, "Adaptive Modulated Scalable Video Transmission over Wireless Networks with a Game-Theoretic Approach", IEEE Multimedia Signal Processing Workshop, 2004, to appear. [5] N. Chayat, “Tentative Criteria for Comparison of Modulation Methods”, IEEE P802.11-97-96, September 1997. [6] G. J. Foschini, M. J. Gans, “On Limits of Wireless Communications in a Fading Environment when Using Multiple Antennas”, Wireless Personal Communications 6: 311–335, 1998. [7] G. Bianchi, “Performance analysis of the IEEE 802.11 Distributed Coordination Function”, IEEE Journal on Selected Areas in Communications, Vol 18, Issue 3, Mar’00, pp 535-547. [8] D. Qiao, S. Choi, K.G. Shin, “Goodput analysis and link adaptation for IEEE 802.11a wireless LANs”, IEEE Transactions on Mobile Computing, Vol 1, Issue 4, Oct-Dec ’02, pp278-292. [9] L. Cheng, W. Zhang, L. Chen, “Rate-Distortion Optimized Unequal Loss Protection for FGS Compressed Video”, IEEE Transactions on Broadcasting, Vol. 50, No, 2, June 2004, pp. 126-131. [10] Intel ® PXA27x Processor Family, http://developer.intel.com/design/pca/prodbref/253820.htm [11] N.C. Paver, B.C. Aldrich, M.H. Khan, “Intel® Wireless MMXTM Technology: A 64-Bit SIMD Architecture for Mobile Multimedia”, Intl Conf on Acoustics, Speech, and Signal Processing, Apr 03, Vol 2, pp 305-308
