Multimedia Streaming Gateway With Jitter Detection

IEEE TRANSACTIONS ON MULTIMEDIA, VOL. 7, NO. 3, JUNE 2005

585

Multimedia Streaming Gateway With Jitter Detection Siu-Ping Chan, Member, IEEE, Chi-Wah Kok, Senior Member, IEEE, and Albert K. Wong, Member, IEEE

Abstract—This paper investigates a novel active buffer management scheme, “Jitter Detection” (JD) for gateway-based congestion control to stream multimedia traffics in packet-switched networks. The quality of multimedia presentation can be greatly degraded due to network delay variation or jitter when transported over a packet-switched network. Jitter degrades the timing relationship among packets in a single media stream and between packets from different media streams and, hence, creates multimedia synchronization problems. Moreover, too much jitter will also degrade the performance of the streaming buffer in the client. Packets received by the client will be rendered useless if they have accumulated enough jitter. The proposed active buffer management scheme will improve the quality of service in multimedia networking by detecting and discarding packets that accumulated enough jitter, such as to maintain a high bandwidth for packets within the multimedia stream’s jitter tolerance. Simulation results have shown that the proposed scheme can effectively lower the average received packet jitter and increase the goodput of the received packets when compared to random early detection (RED) and DropTail used in gateway-based congestion control. Furthermore, simulation results have also revealed that the proposed scheme can maintain the same TCP friendliness when compared to that of RED and DropTail used for multimedia streams. Index Terms—DropTail, jitter detection, multimedia streaming, RED, TCP-friendliness.

I. INTRODUCTION

W

HEN a multimedia stream is transported over packet-switched networks, it assumes a constant network delay to maintain the timing relationship among packets. It is difficult, however, if not impossible, to maintain a constant network delay. As a result, multimedia traffic will usually experience impairments due to network delay variations, or jitters, that may lead to degradation in the quality of service (QoS). The network jitter problem complicates the synchronization problem among packets from a single media stream or between packets that have accumulated too much jitter from different media streams. When the multimedia stream traverses the networks, it will render the stream useless when received by the clients. This is because it will be difficult to re-adjust the timing relationships between multimedia packets from the same or several media streams to assure a synchronized playback of information [5]. In some multimedia streaming systems [7], [8], a streaming buffer is allocated in the client trying to solve the problem. However, a streaming buffer can only help to increase the tolerance of the network jitter for multimedia networking.

Manuscript received December 30, 2002; revised November 20, 2003. This work was supported in part by the Research Grants Council of Hong Kong, China, under Project CERG HKUST6173/00E. The associate editor coordinating the review of this manuscript and approving it for publication was Dr. Hong Heather Yu.. The authors are with the Department of Electronic Engineering, Hong Kong University of Science and Technology, Clear Water Bay, Kowloon, Hong Kong. Digital Object Identifier 10.1109/TMM.2005.843338

The situation will become worse when there is a large amount of jitter-corrupted multimedia packets (multimedia packets that have accumulated jitter larger than the client’s jitter tolerance) in the network. In that case, although jitter-corrupted multimedia packets will be rendered useless when received by the client, they are continuing to consume network bandwidth and thus increasing the congestion of the network for multimedia packets delivery. This new kind of congestion cannot be solved by traditional active buffer management schemes such as random early detection (RED) [1] or DropTail in gateway-based packet-switch network congestion control. RED as proposed by IETF is an active queue management scheme where the incoming packets are dropped randomly with a probability that is related to the current queue length and a queue-size threshold value. Typically, RED does not provide any protection for TCP flows when there is also UDP traffic sharing the queue. UDP traffic will typically lower the throughput of TCP flows. How to achieve TCP-friendliness in RED has been a popular subject of RED related research. In [2], several variations of RED and even packet size is taken into consideration to achieve better loss differentiation and fairness. In [3], the “average delay” is used instead of “average queue length” to calculate the dropping probability in RED. In [4], Preferential Dropping (PD) is proposed to identify individual misbehaving high-bandwidth flows such that the rate of these flows are compressed to certain target values. This paper assumes that multimedia streams are transported by UDP traffics that share the network with TCP traffic. The multimedia packets that have accumulated excessive jitter will become useless to the client. Therefore, it is a waste of network resources to continue to forward these packets in the network. An efficient congestion control for streaming traffic should detect and discard multimedia packets that violate the jitter tolerance. We will further consider the efficiency of the congestion control algorithm in which an estimate of the multimedia packet’s “closeness” to its destination is also used in the discard decision. A novel gateway-based congestion control for streaming traffics known as jitter detection (JD) is proposed, which aims at improving the quality of service by detecting and discarding multimedia packets that have accumulated large jitter so as to maintain a high bandwidth for packets that stay within the multimedia stream’s jitter tolerance. In this paper, we assume that there is a way for the gateway to classify a packet as a TCP or multimedia streaming packet. For example, a packet may be recognized as a streaming packet based on encoding in the protocol field in the IP Version 4 packet header. To keep track of the accumulated jitter in the multimedia UDP packets, a small delay jitter counter is included in the network layer (i.e., IP) header. This jitter counter, in each packet’s

1520-9210/$20.00 © 2005 IEEE

586

Fig. 1.


System block diagram.

header, is updated based on the previous jitter value plus the estimated jitter experience by the multimedia packet in the current gateway, measured in four quantized regions. The jitter counter is implemented in the type-of-service field in the IP Version 4 packet header. The type-of-service field itself contains (from left to right) a three-bit Precedence field, three flags (D, T, and R), and two unused bits. Those two unused bits in the field in the header can be used to represent all the four jitter regions. A simple discard algorithm is applied to each streaming packet using the jitter counter to select the discard threshold. For the purpose of this study, we further investigate the notion where a measure of how far or close the packet is from its destination is taken into consideration for the discard decision. The idea here is that a streaming packet that has accumulated a large jitter value should not be dropped if it is close to the client. We assume that the TTL field in the IP packet header can be used as a measure of this “residual distance.” To use the TTL field for this purpose will require the initialization of the TTL field [10]. Instead of using a fixed default “safe value” for all destinations, the sender of a streaming packet needs to set a TTL value that reflects the actual hop-count to the client. More generally speaking, this initial TTL value should be set based on an end-to-end delay budget instead of hop count because the delay on each hop is a function of the link bandwidth, which is nonuniform. To use the TTL, servers and clients would be required to exchange hop-count/delay-budget information during media session setups and when the hop-count changes due to routing updates. A safety margin may be needed in the initial TTL value for protection against various scenarios that cause unnecessary packet discard. The remainder of this paper is organized as follows. In Section II, we specify the two JD schemes, the first without the residual distance consideration and the second with the residual distance consideration. In Section III, we present the simulation environment and simulation results that shows the effectiveness of the two JD schemes. The paper is concluded in Section IV.

II. JITTER-AWARE QoS FOR MULTIMEDIA TRAFFIC Fig. 1 illustrates the operation of a gateway where two types of traffic classes are assumed: UDP and TCP. The UDP traffic carries constant sending rate multimedia traffic, and TCP traffic carries variable bitrate services, such as FTP. Input traffic to the gateway is classified into TCP or UDP streaming traffic. TCP traffic is subjected to the RED scheme while UDP streaming traffic is subjected to the JD scheme before being queued into the output queue. To simplify our discussion, the output queue is assumed to be first-in-first-out (FIFO). Also, the packet size

Fig. 2. Details of JD scheme 1 for delay jitter-sensitive traffic.

of the streaming traffic is assumed to be fixed. In reality, multimedia traffic, such as MPEG-2 VBR audio traffic, is usually chopped down to constant packet size to fit into the network environment. In that case, a simple multimedia frame (either video frame or audio frame) would require different numbers of packets in VBR multimedia bitrate condition. Further, note that losing some of the packets will still allow partial decoding of the multimedia data (e.g., in MPEG audio and video coder). This justified our assumption that the queue management scheme can consider each packet separately. A. Jitter Detection Scheme 1 Fig. 2 presents the pseudocode of JD scheme 1 used in the gateway congestion control. The variable delay is the time delay encountered by a streaming packet in the gateway. Delay can

CHAN et al.: MULTIMEDIA STREAMING GATEWAY WITH JITTER DETECTION

587

Fig. 4. Details of JD with TTL contraint (scheme 2). Fig. 3. Dropping threshold of the JD algorithm.

B. Jitter Detection Scheme 2 be estimated from the current buffer occupancy (CBO) of the output queue together with the estimated available output link capacity (LC), such that (1) The average delay experienced by the incoming packets is estimated as an exponentially weighted moving average (EWMA) of the delay (2) By means of EWMA, short-term increases of average delay can be prevented by the smoothing effect of the weighted average. The jitter experienced by the multimedia streaming packet is given by (3) The end-to-end delay jitter counter that tracks the accumulated jitter of the multimedia packets is defined in Figs. 2 and 3, where is updated by quantizing the accumulated jitter stored by plus the estimated jitter experience by the multimedia packets in the gateway into one of the four regions, as shown in Fig. 2. We to define the region, use a bound which is equal to as shown in Fig. 3. and each region has a width of Finally, the dropping threshold (threshold) is determined by (4) The multimedia packet will be sent to the output queue if and only if

in JD scheme 1 determines the allowable jitter The tolerance for each multimedia packet. It must be chosen with great care because the jitter accumulated at each gateway is is used, nonuniformly distributed. As a result, if a tight it will be highly probable that the multimedia packets will be dropped early in the gateway path. On the other hand, if a loose is used, it will be highly probable that the multimedia packets received by the client will be useless due to jitter tolerance violation. Optimal performance can be achieved only if ’s are used at the gateway path. In particular, different as the multimedia packets enter the gateway that is closer to the client, the more loose, i.e., the bigger the value, of should be used. As discussed in the Introduction, such adaptation can be achieved by extracting the hop-count knowledge carried in the TTL field. To simplify our discussion, we assume that the value in the TTL field in the packet header is the remaining hop-count. The JD schemes 1 and 2 therefore differ by the additional residual hop-count information. It should be noted that the actual residual distance estimation as discussed in the Introduction is difficult to obtain. The assumption that the TTL field carries the residual hop-count allows us to understand the gain in performance achievable with the estimated residual distance information. The pseudo code of the modified JD scheme 2 is shown in Fig. 4, where the TTL value is first set to equal to the hop-count at the server and is decremented by one every time the packet passes through a gateway. is adapted by adding , where is the TTL The value, such that the adapted is given by . should be chosen such that , where is the total hop-count in the gateway path. The rest of the JD algorithm remains the same.

(5) The proposed JD scheme deals with the jitter control problem when the channel under concern or the receiver buffer size is limited. In order to ensure that the client buffer never underflows (or overflows), the inflow rate must match that of the outflow rate, as stated in [6], by feedback control or by dropping those early arrived packets. Discarding those packets with large accumulated jitter can help to better control the inflow rate into the receiver buffer.

C. Performance Bounds and Analysis of Dropping Decision The performance of the JD can be examined by measuring the delay jitter value of the received packet using the sample mean delay due to the use of the finite number of sample packets . By using the Chebyshev inequality, an upper bound on the and the initial value of probability of dropping packet can be derived. Assume that the delay values of the received multimedia packets are independent and are iden-

588


Fig. 5. Simulation topology.

tically distributed with a standard deviation of . The sample mean delay . The Chebyshev inequality can be applied to obtain a loose bound that helps to find the relationship between the probability in the estimation of delay jitter of dropping packets, of the received packets using and , the number of sample . The dropping packets , and the dropping threshold is defined as where threshold is the maximum delay jitter value allowed for the desired application and .

(6)

(7)

(8) Therefore, given the number of packets to estimate the sample mean delay during the initial state, it can be used to and . define the values of the parameter Also, (8) provides a performance bound of the dropping probawith . bility This packet dropping decision model of JD schemes shows a good analytical reason on changing the ‘‘ ’’ according to the ‘‘closeness’’ to the destination in order to vary the probability of the dropping packets. It has been analytically showed ’’, the lower the probability of that the larger the ‘‘ dropping the packets. This explains why JD Scheme 2 allows

more packets to pass through the gateway when compared to that of JD Scheme 1. III. SIMULATION RESULTS Simulations are performed using Network Simulator [9] to investigate the performance of the proposed algorithms. Fig. 5 shows the simulation topology. There are two TCP sources (FTP traffic) and two UDP sources (Multimedia traffic, i.e., MPEG-2 audio coder with encoded/multimedia bit rate 0.8 Mb/s) connecting to the streaming gateway R1 by the links with capacity of 10 Mb/s and a propagation delay of 3 ms. The MPEG-2 audio sequence test sample is encoded in a variable multimedia bit rate (VBR) format and is carried by constant channel bit rate (CBR) UDP traffic in this simulation. For simplicity, we use “UDP” packets to represent the audio streams in the following discussion. Fig. 5 details the settings of the traffic through the network. The packets will go through a chain of 12 gateways (R1–R12) before arriving at the client. Those gateways (R1–R12) connect to each other by the links with capacity of 1.5 Mb/s and propagation delay of 20 ms. The bottleneck of such a network topology will be in the chain of those 12 gateways. RED and JD are implemented in the gateways for congestion control of the TCP and UDP packets, respectively. Fig. 5 also details the parameter settings of the simulations. Several simulations have been performed to investigate the TCP-friendliness of the system, the delay jitter, and the useful throughput of the multimedia traffic. Two sets of additional simulations are performed using the DropTail and the RED congestion control schemes on the multimedia traffic for performance comparison with that of the proposed JD algorithms. A. TCP-Friendliness It is well known that UDP does not respond to packet loss. Greedy high-bandwidth UDP traffic will typically lower the throughput of TCP flows. Therefore, we proposed to classify


Fig. 6. (a) Average throughput of TCP against different packet sizes. (b) Average throughput of UDP against different packet sizes.

TCP and UDP traffic and apply different buffer management schemes to the two types of traffic before putting them into the FIFO queue. It is important that the proposed JD schemes have certain TCP-friendliness such that it will not lower the throughput of the TCP flows. Simulation results in Fig. 6(a) shows the average throughput of TCP traffic against different packet sizes under different buffer management schemes for the UDP traffic and RED for the TCP traffic. It can be observed that the proposed JD schemes 1 and 2 can achieve the same level of TCP throughput when compared to that of other buffer management scheme. Similarly, Fig. 6(b) presents the average throughput of UDP traffic against different packet sizes from the same simulations as those in Fig. 6(a). The average throughput of UDP in the proposed schemes decrease with increased packet sizes. Also, it is lower than that of DropTail and RED algorithms. This is because UDP packets with large accumulated jitter are discarded. Fig. 7(a) shows the throughput of UDP traffic when it passes through each gateway (R1–R12). The result shows the effect of the TTL counter in JD scheme 2, where a lower packet drop rate is observed when the gateway is “closer” to the client. The TCP-friendliness of the proposed gateway-based buffer management scheme is directly related to the ratio of the UDP throughput to the TCP throughput. Fig. 7(b) therefore shows the degree of TCP-friendliness with different packet sizes for different gateway-based buffer

589

Fig. 7. (a) Average throughput of UDP packets at each router using JD scheme 2. (b) Ratio of average throughput of UDP and TCP.

management schemes. It can be observed that the proposed JD scheme can achieve similar TCP-friendliness as the other simulated buffer management schemes. B. Delay Jitter of Multimedia Traffic Simulations had been performed to investigate the delay jitter of the multimedia traffic. Fig. 8(b) shows the delay jitter pattern of the multimedia traffic using RED. The delay jitter is defined as indicated in Fig. 8(a). The curve in Fig. 8(b) as shows the average delay jitter which is defined as the average value of delay jitter from to . The same delay jitter pattern is found in other schemes including DropTail and JD schemes 1 and 2. The periodic pattern found in all of the delay jitter figures are mainly due to the congestion control mechanism of TCP traffic. The results show that our schemes can achieve lower average delay jitter than that using DropTail and RED, as shown in Fig. 9(a). To investigate the scalability and stability of the JD mechanisms, we further simulated the same network topology with two more identical UDP traffics added (MPEG-2 audio stream) to that of the previous simulation such that it allows us to investigate the performance of the proposed JD algorithms under a heavy traffic network. The simulation results in Fig. 9(b) showed that, even under heavy traffic conditions, the JD schemes still achieve the lowest average delay jitter when compared to that of DropTail and RED.

590


Fig. 8. (a) Packet arrival and delay jitter definition. (b) Delay jitter for some received UDP packets using RED (P acket Size 900 bytes).

=

C. Useful Goodput of Multimedia Traffic As discussed in the Introduction, Multimedia UDP traffic is delay jitter-sensitive. When the jitter carried by the packet is greater than the tolerance threshold, the packet is regarded as useless even if it successfully arrives at the client. Congestion will happen when there is no enough available channel bandwidth. Forwarding a packet that arrives too early or too late will reduce the available bandwidth of the channel and thus reduces the jitter tolerance of the gateways along the path. The reduced jitter tolerance will result in more packets arriving at the receiver with too much jitter that makes it useless or gets dropped when JD is applied. Therefore, it is the “useful” goodput of the multimedia traffic that is of concern and not the throughput. “Useful” means that the multimedia packet arrives at the client within the jitter toleras in Fig. 8(a). The “useful” ance, i.e., multimedia packet can meet its scheduled playout position and be played out to the users. In the simulation, UsefulThreshold is set to be equal to or smaller than one half of the playout interval . Here, UsefulThreshold is to determine whether the received multimedia packets are “useful” or not, while the given Jmax will determine whether the packet in the queue should be discarded or not. They are actually different in meaning while at the same time directly proportional to each other in values. The range of useful threshold values used in the simulation is in the range of 0.05–0.1 s. Such a setting is useful for audio streaming such as MPEG-2 with the audio bit rate ranging from 6 kb/s to 1 Mb/s, as in [5]. This means that the receiver buffer

Fig. 9. Average Delay Jitter for some received UDP packets using different queue management schemes. (a) Original network with two sets of TCP traffic and two sets of UDP traffic. (b) Modified network with two sets of TCP traffic and four sets of UDP traffic.

only needs to store up to 0.1 Mb of data in order to compensate for the effect of delay jitter of the multimedia packets. Simulations were performed to investigate the average useful goodput of multimedia UDP traffic against different packet sizes and different useful threshold. Results are shown in Fig. 10(a) and (b), respectively. The results showed that our schemes can achieve higher average useful goodput of multimedia UDP traffic than that using DropTail and RED. Moreover, the percentage of useless packets among those received packets against different packet sizes and different useful threshold have also been investigated, and the simulation results are shown in Fig. 11(a) and (b), respectively. Again, the results showed that our schemes can maintain a low percentage of useless packets among those received packets when compared to that using DropTail and RED. Notice that the simulation results obtained by Network Simulator have random variation due to the nature of Monte Carlo simulations employed in NS. As a result, there may be inconsistency for individual simulations; however, the overall trend is clearly observable in Figs. 10 and 11, which agree well with the theoretical development. For one or two particular experimental setups, there are small variations from the theoretical predicted results. Since the variation is small, we considered the simulation results to be acceptable and are reported in this paper. The variation in simulation results is more obvious in large packet


Fig. 10. (a) Useful goodput of UDP traffic against different packet sizes (Useful T hreshold 0.1 s). (b) Useful goodput of UDP traffic against different useful threshold (P acket Size 900 bytes).

=

=

591

Fig. 11. (a) Percentage of useless UDP packets against different packet sizes (Usef ul T hreshold 0.1 s). (b) Percentage of useless UDP packets against different useful threshold (P acket Size 900 bytes).

=

=

sizes because the useful goodput is small. As a result, the percentage of variation increases in spite of the magnitude of the simulation variation remaining the same.

D. Audio Quality Comparison Between Different Queue Management Schemes The results in Fig. 6(b) showed that the proposed JD schemes for multimedia streams such as audio achieves relatively lower received throughput of multimedia UDP traffic than that of using RED and DropTail. However, if we consider the “useful” received multimedia packets, there will be more “useless” packets when RED or DropTail is used than when using the proposed modified JD scheme. The important results in Fig. 12 show that the proposed JD schemes can achieve better quality for MPEG-2 audio streams than can RED and DropTail within a given useful threshold. It shows that the JD scheme gives up to approximately 0.2–0.3-dB advantage in SNR over RED and up to approximately a 0.5-dB advantage over DropTail. Here, the results are computed and reported as exponential weighted moving average for every window size of one second. As a result, a smaller fluctuation in SNR is observed in the audio playout obtained by the simulation.

Fig. 12. Audio quality comparison between different queue management schemes.

IV. CONCLUSION A novel streaming friendly gateway-based buffer management scheme known as jitter detection (JD) is proposed in this paper. The JD algorithm can improve the QoS of multimedia networks by detecting and discarding multimedia packets that violate the jitter tolerance measure. We proposed to classify the TCP and UDP traffic, such that RED and JD buffer management schemes are applied to individual traffic respectively before sending it to the output FIFO queue. Furthermore, we inves-

592


tigated the performance improvement of the JD scheme by incorporating the knowledge of residual hop-count in the jitter detection and discarding algorithm. Simulation results have shown that the proposed algorithm can maintain the same TCP-friendliness as can the RED and DropTail schemes. At the same time, the JD scheme reduces the average delay jitter of the multimedia packets and thus maintains a high useful throughput for the multimedia traffic when compared to that of the traditional gateway-based buffer management scheme. REFERENCES [1] S. Floyd and V. Jacobson, “Random early detection gateways for congestion avoidance,” IEEE/ACM Trans. Netw., vol. 1, no. 4, pp. 397–413, Aug. 1993. [2] S. De Cnodder, O. Elloumi, and K. Pauwels, “RED behavior with different packet sizes,” in Proc. 5th IEEE Symp. Computers Commun., 2000, pp. 793–799. [3] J. B. Pippas and I. S. Venieris, “A RED variation for delay control,” in Proc. ICC, vol. 1, 2000, pp. 475–479. [4] S. Floyd and R. Mahajan, “Controlling High-Bandwidth Flows at the Congested Router,”, Nov. 20, 2000. Tech. Rep. of ACIRI. [5] K. Shuaib, T. Saadawi, M. Lee, and B. Basch, “Dejittering in the transport of MPEG-2 and MPEG-4 video,” Mulitmedia Syst., vol. 8, pp. 231–239, 2000. [6] S. Chaudhry, M. Raziuddin, and A. Choudhary, “On guaranteed bandwidth channels,” in Proc. ICNP, Nov. 1995, pp. 47–55. [7] S. P. Chan and C. W. Kok, “Protocol and buffer design for multimedia-on-demand system,” in Proc. IEEE PCM, Oct. 2001, pp. 1010–1015. , “Bitrate adaptation flow control for multimedia-on-demand,” in [8] Proc. IEEE ICC, May 2002, pp. 2503–2507. [9] Network Simulator (ns) [Online]. Available: http://www.isi.edu/nsnam/ns [10] Requirements for Internet Hosts—Communication Layers (1989, Oct.). [Online]

Siu-Ping Chan (S’01–M’04) received the B.Eng. and Ph.D. degrees in in electrical and electronic engineering, both from the Hong Kong University of Science and Technology (HKUST), 2000 and 2004, respectively. He is currently a Postdoctoral Research Scholar, Department of Electrical Engineering, University of Washington, Seattle. From 2003 to 2004, he was a Research Associate, Digital Signal Processing Center of the HKUST. During the summer of 2002, he was a Visiting Scholar, Signal and Imaging Processing Institute, University of Southern California, Los Angeles. His research interests include multimedia streaming protocols, rate and congestion control mechanisms, multimedia gateway design, layered multimedia and security issues in multimedia multicast.

Chi-Wah Kok (S’88–M’92–SM’00) received the B.Eng. and M.Phil. degrees from the City University of Hong Kong in 1990 and 1993, respectively, and the M.Sc. and Ph.D. degrees from the University of Wisconsin, Madison, in 1996 and 1997, respectively, all in electrical and engineering. He joined the Department of Electrical and Electronic Engineering, Hong Kong University of Science and Technology, Kowloon, in 1999, where he is currently an Assistant Professor. Prior to that, he was with several industrial corporations and universities, including Sony US Research Laboratory, Stanford University, and Lattice Semiconductor Corporation. He has published more than 100 papers in IEEE journals and conferences. His research interests are in the area of signal processing, with applications in multimedia and communications.

Albert K. Wong (S’83–M’88) received the B.S., M.S., E.E., and Ph.D. degrees from the Massachusetts Institute of Technology, Cambridge, in 1982, 1984, 1984, and 1988, respectively, all in electrical engineering. He is currently Vice President—Wireless Communication of the Applied Science and Technology Research Institute of Hong Kong. From 1988 to 2000, he was with AT&T and Lucent Technologies, Bell Laboratories, in New Jersey, Hong Kong, and China. There he held a series of positions as a Member of Technical Staff, Distinguished Member of Technical Staff, Technical Manager, Director of Technical Marketing, and Director of Sales and Technical Marketing. From 2000 to 2001, he was Chief Operating Officer of Transtech Services Group, a Hong Kong-based company engaged in the building of an optical preform and fiber plant in Hong Kong. From 2001 to 2002, he was a Visiting Associate Professor with the Electrical and Electronic Engineering Department, Hong Kong University of Science and Technology, Kowloon, where he remains as an Adjunct Associate Professor. From 1997 to 1998, he was a Visiting Associate Professor with the Information Engineering Department, Chinese University of Hong Kong. He has also held adjunct faculty positions at the Polytechnic University of New York and at Rutgers University. His study at MIT was in the area of atmospheric optical communications, wavelength division systems, and networking. His work at Bell Laboratories covered basic research and product development in photonic and ATM switching systems, and marketing and sales of data networking and optical networking products and systems. His current work focuses on wireless applications, telecommunication network convergence, and multimedia communications.