Seamless Handover of Streamed Video over UDP between Wireless ...

3 downloads 109669 Views 231KB Size Report
technologies have also been evolving. Key wireless technologies, e.g. Wireless LAN (WLAN) and 3G, are reaching a level of maturity such that it is now possible ...
Seamless Handover of Streamed Video over UDP between Wireless LANs Ger Cunningham, Seán Murphy, Liam Murphy

Philip Perry

Department of Computer Science University College Dublin Dublin, Ireland {ger.munningham,[email protected], [email protected]}

School of Electronic Engineering Dublin City University Dublin, Ireland [email protected]

Abstract— The seamless handover of streamed video in a WLAN with UDP as the transport layer is considered. The use of handover for stationary nodes in the case of network congestion is motivated. Next, the relationship between delay and loss in WLANs is studied, with the conclusion that delay can be used as an indicator of when loss is likely to occur. Delay can therefore be used as a basis of a handover scheme which can minimize loss. Such a handover scheme is proposed in which the client makes two simultaneous connections to the same server through two separate WLANs, and it is shown that the client can use the relative packet delay between the two streams to determine which network delivers the best performance; this information is then used to determine when to perform a handover. The proposed scheme is implemented and results are presented that show the successful handover of an RTP over UDP stream using this scheme in a live WLAN environment. Keywords- Seamless Handover, Streamed Video, Wireless LAN.

I.

INTRODUCTION

Many predictions relating to widespread distribution of video content over the internet have not materialised as envisaged. However, over the last few years, there has been a steadily increasing amount of video content distributed via the internet; something that has been, to some extent, stimulated by the increase in penetration of broadband internet connectivity. While internet usage patterns have been changing, access technologies have also been evolving. Key wireless technologies, e.g. Wireless LAN (WLAN) and 3G, are reaching a level of maturity such that it is now possible to credibly state that substantial numbers of users will use wireless technologies for internet access in the near future. These users will desire similar behaviour from their wireless internet access to that which is delivered via a wired medium. It is interesting to consider how well wireless access technologies are suited to streaming video content, as it is quite conceivable that significant numbers of users may desire to stream video content to their devices via such technologies. Clearly, it is undesirable for technical considerations such as signal strength or congestion in the access network to impact the quality of the video being viewed by the user. Seamless network connectivity can limit the effect of network issues on user perceived quality, to some extent, but, of itself, does not

necessarily ensure that the quality of experience of the end user is optimal. In the case in which a number of wireless access networks are available to the user, it is interesting to consider which wireless access network should be used at any point in time to deliver the best quality to the user. In this paper, the case in which a number of WLAN access networks are available is considered. The issue is to determine which WLAN can deliver the best video quality to the user. Such a scenario could easily be envisaged in an office environment or in places of high so-called footfall1, such as airports or hotel lobbies. The focus of the paper, then, is on handover issues when multiple WLAN access networks are available. More specifically, the problem is to devise an approach to realising handover which is as close to imperceptible as possible to a user streaming video content while also ensuring that the best available WLAN access network is always used. As such, it is desirable to have no (or minimal) period in which the user has no network connectivity. In this work, UDP is used as the transport protocol of choice for streamed video content. TCP is often used as the transport protocol for streamed media, but it requires quite large buffers. In small footprint mobile devices, resources are limited and the use of a large buffer for streaming is not so desirable. Hence, it is reasonable to consider UDP for streaming in this scenario. Historically, handover has been considered mainly as an issue related to mobility. Experiments were performed to assess the suitability of the work described here to the mobility context, but it was found that UDP – without higher layer error correction – was unable to operate well in the presence of data loss which arose when the mobile node moved away from the WLAN access point (AP). Consequently, the focus of this work is on the case in which handover can be desirable due to network congestion. The paper is laid out as follows. Section 2 discusses handover techniques that have been proposed in the literature. The following section then presents the rationale for using 1

Places of high footfall are places in which there are high concentrations of people, often with some time on their hands.

measured delay as the basis for deciding when to perform a handover. The proposed handover scheme is then described in detail in section 4. In the following section results are presented from a live WLAN environment. Conclusions and future work are discussed in section 6. II.

RELATED WORK

In order to comment on and relate work to that described here, it is useful to introduce the concepts of soft handover and hard handover. While these terms have been used elsewhere in the literature in some senses, some specific meaning is required here. In this context, soft handover is a handover in which the same data is delivered to the mobile device via the two access networks simultaneously. Clearly, this can be resource intensive, but it does make the possibility of data loss during the handover very small. In contrast, a hard handover is one in which data is streamed via one network at any time: thus at some specific time, a decision is made to receive the data through one network rather than the other. While hard handover is more parsimonious with network resources, it can result in data loss at the user terminal, depending on how quickly the handover is effected and how much data the terminal is transmitting/receiving. The IEEE 802.11b standard for WLANs allows for handover between overlapping WLAN cells at the link layer [1]. Since this only permits connection to one WLAN at a time, it falls into the category of hard handovers. In [2] it was shown that the handover latency can be significant and subject to large variations. During this latency period the station is unable to send or receive traffic and packets queued at the old WLAN will be lost, making it unsuitable for the handover of multimedia traffic. An approach which has received most interest within the research community is that of Mobile IP [3], which enables a Mobile Node (MN) to receive IP packets through a packet forwarding procedure. This approach is well suited to solving the problem of locating a mobile that may be attached to one of a number of networks. However, handovers in Mobile IP are slow and packets can be lost during the handover procedure [4], making it unsuitable for the handover of video traffic. Extensions to the mobility support provided in the Session Initiation Protocol (SIP) [5] has also been proposed for delaysensitive communication [6,7]. SIP is an application layer protocol for establishing and tearing down multimedia sessions. It can help provide personal mobility, terminal mobility, and session mobility. SIP uses hard handover and “in-flight” packets can be lost during the handover period. In [6] it was estimated that packet loss can occur for up to 50ms during the handover period. This estimate did not take into account network initialisation procedures such as AAA or address assignment, and this can be a major part of the overall handover delay. SIP therefore is not good at providing seamless handover of streamed video in networks. Clearly, the above solutions are inadequate for handover of streamed video, primarily because they break the old connection before they make the connection to the new cell. As loss of video data can result in video artefacts, soft handover is

the preferred option and forms the basis of the approach proposed here. In [9] a modified transport layer protocol used path delay to initiate a soft handover with a voice application. The work presented here considers the handover of video in the application layer. Providing handover functionality in the application layer means changes do not have to be made to operating systems and protocol stacks. An approach to realizing handover based on a new socket abstraction was proposed in [11]. While this is an interesting contribution it did not address when to handover satisfactorily. III.

DELAY-CENTRIC HANDOVER

In order to evaluate the use of delay as the basis of a handover mechanism, experiments were performed to investigate the relationship between loss and delay in WLANs. The following test was performed to obtain the delay and loss characteristics of a congested WLAN. A video server and MN were connected to an 802.11b WLAN as shown in figure 1. The server streamed RTP packets over UDP to the MN via the AP. This RTP stream was used to model CBR video traffic. The packets were transmitted at 30 packets per second and contained 1000 bytes of RTP data each. Six background computers (BG) were used to generate background traffic on the WLAN. The background computers were stationary, and were positioned away from each other and well within range of the AP. The background computers sent UDP packets at 1.2 Mbit/s each to an Iperf2 Sink via the AP. The MN remained stationary throughout the test. The background computers contended for the medium with the traffic from the server. Only traffic from the server enters the AP’s queue. At the beginning of the test only the video server transmitted packets. After 30 seconds two of the background traffic generators were activated, and the remainder were activated in succession, with a minute between each activation. Two minutes after the last one was activated, all of the background traffic generators were deactivated, leaving only the video server transmitting packets. Figure 2 shows a graph of the packet delay and packet loss experienced by the RTP packets that were transmitted by the video server to the MN. It shows that the packets experienced some minor jitter when two background computers were transmitting. When three background computers were active, delay was relatively small. Thereafter, delay began to increase substantially, and with six background computers transmitting some packets were lost. Table 1 shows the packet numbers of the packets that were lost. The losses were in clusters and they occurred when the packet delay was very high suggesting the losses were due to queue overflow, rather than by unsuccessful retransmissions at the link layer. The test was repeated a number of times with similar levels of delay and packet losses observed as above. Increasing the 2

Iperf is a tool for measuring maximum TCP and UDP bandwidth.

traffic rate of the background computers lead to greater packet loss. More background stations in the test outlined above would also likely result in greater packet loss. The above results clearly demonstrate that there is a significant correlation between delay and loss in WLAN cells: when delay becomes large – of the order of some 100’s of milliseconds – the system exhibits signs of congestion and is liable to suffer packet loss. Hence it is reasonable to use delay in the cell as an indicator of the likelihood of experiencing packet loss. BG BG

Iperf Sink

BG

BG BG

AP BG Video Server

MN Wire Wireless

Figure 1. WLAN Setup.

2BG

3BG

4BG

5BG

6BG

Figure 2. Packet Delay and Loss in a WLAN.

IV.

OUTLINE OF HANDOVER SCHEME

The use of packet delay in deciding when to handover was considered. It is assumed that the end-to-end delay from the server to the MN via the paths used by the MN are approximately equal under uncongested conditions, and that the server is not overloaded. Further, it is assumed that for a WLAN that is experiencing congestion, the delays between the server and the APs is of a lower order than the delays experienced by packets in getting through the WLAN. Such a scenario could be envisaged where the server and client are in the same metropolitan area. Figure 3 shows content being streamed through 2 access networks during a soft handover in the above scenario. The MN has two WLAN interfaces, so that it can handover from one WLAN to another. The MN uses soft handover in the application layer, enabling the video decoder in the MN to present an uninterrupted video stream to the user. In the experiments it was necessary to use two WLAN cards in the MN to access two WLANs simultaneously. However, Microsoft researchers in conjunction with some academics have recently proposed a mechanism by which a single physical WLAN interface can be used to simultaneously access multiple WLANs [10]. Such functionality is likely to exist in future devices. Each of the two WLAN interfaces in the MN has a unique IP address, so that the two streams are independently routed through the networks. Most of the time the MN has one connection to the server streaming the desired video stream, while the second wireless interface looks for another AP. When it discovers another WLAN, it registers with the network and performs any necessary initialisation procedures. When the current WLAN is becoming congested the MN makes another connection to the server and the server initiates a second, identical video stream to the MN via the new WLAN. The MN then decides which is the least congested network, and then drops the stream passing through the more congested network.

Video Server

WLAN A MN WLAN B Wire Wireless

TABLE I. Cluster Number

LOST PACKETS ID of Lost Packets

1

12378,12380 12381,12386,12388 12391,12393 12395 12401

2

12510, 12511, 12512, 12513, 12518

Since one of the key objectives of this work is to ensure that there is minimal data lost, it is reasonable to measure packet delays and to attempt to use these to detect incipient congestion. Further, a handover decision mechanism based on measured packet delays has the potential to minimise data loss as the content is streamed to the client.

Figure 3. Soft Handover

Deciding when to request a second stream is for further study and is outside the scope of this paper. Here, deciding which network is the least congested given two identical streams is considered. When one WLAN is more congested than the other due to traffic load, the packets of the stream passing through the more congested WLAN will arrive at the MN later than the packets of the other stream. In the proposed handover scheme, when the MN is receiving two identical streams the MN records the arrival time of each packet using the MN’s internal clock. The MN subtracts the arrival time of each packet from the arrival

time of the corresponding packet in the other stream. This yields a positive or negative time, called the delay difference, that indicates the difference between the delays experienced by the two packets as they passed through the WLANs. Because the delay in a congested WLAN varies greatly from zero up to a number of seconds, as shown in figure 2, a single delay difference sample in the handover scheme outlined here would not reliably indicate the difference in congestion between the two cells. By averaging the delay differences over a period of time, however, the MN is able to improve the likelihood of correctly determining the difference in congestion between the two cells. The test performed in the previous section was repeated, with the background stations transmitting at 1 Mbit/s each. Figure 4(a) shows a graph of the resulting packet delay. Figures 4(b), 4(c), 4(d) and 4(e) show the moving average of the delay over 200, 400, 600 and 800 packets respectively. These graphs suggest that if the MN were to increase the number of delay difference samples it uses in the calculation of the average in the handover scheme described here, it would improve the likelihood of correctly determining the difference in congestion between the two cells. However, increasing the number of delay differences used in the average also increases the time it takes for the MN to decide which is the least congested network and hence increases the duration that two identical streams are streamed to the MN. In the case of figures 4(b), 4(c), 4(d) and 4(e), this duration is about 6.67s, 13.34s, 20s, and 26.67s respectively. So, clearly there is a tradeoff between increased likelihood of correctly indicating the difference in congestion between two cells and the duration that two identical streams need to be streamed to an MN.

6BG

2BG

3BG

4BG

5BG

Figure 4. Comparison of different Moving Averages of Delay

The above handover scheme is effective at distinguishing between congestion levels that are quite low, and at distinguishing between congestion levels that are significantly different, but it is less successful at identifying which cell is the least congested when the congestion in both cells is very large. However, if both cells are heavily congested then there is little clear benefit of handing over.

V.

WLAN HANDOVER RESULTS

An application layer agent was created to implement the handover scheme described above and was used for gathering the results presented below. Tests were performed to demonstrate a handover of streamed video in a congested WLAN by a stationary MN, in the scenario described above. A video server was fitted with two LAN cards and was connected to two 802.11b APs as shown in figure 5. The MN was multihomed and had two WLAN cards.

full duration of the test using 800 samples was later calculated, and is shown in figure 6(b). The test above shows that the traffic caused by the background computers resulted in substantial packet delays in AP1. It shows that averaging the delay over 800 samples enables the MN to accurately distinguish congestion levels that are quite low, and congestion levels that are significantly different. It shows that packet delay can be used to decide which is the better network when performing a handover.

Initially the server streamed two simultaneous streams of RTP packets to the MN via the two APs. The RTP streams were used to model CBR video traffic. The packets were transmitted at 30 packets per second and contained 1000 bytes of RTP data each. The client defaulted to choosing the stream received via AP 1.

2BG

3BG

trigger to start average ti 5BG 4BG

6BG

Six background computers were used as before to generate background UDP traffic at 1Mbit/s each to the Iperf Sink via AP1. The background computers were stationary, and were positioned away from each other and well within range of the AP. The MN remained stationary throughout the test. BG

Iperf Sink

BG

BG

BG BG

AP 1

Video Server

BG MN

AP 2

Wire Wireless

Figure 5. WLAN Setup

At the beginning of the test only the video server transmitted packets. After 30 seconds two of the background traffic generators were turned on, and the remainder were activated in succession, with a minute between each activation. Two minutes after the last background traffic generator was turned on, all of them were deactivated, leaving only the video server transmitting packets. The trace shown in figure 6(a) shows the delay difference between the packets of the two streams. Triggering the MN to perform an average of the delay differences was performed manually, and the time that it occurred is indicated in figure 6(a). In a fully implemented soft handover the MN would generate this trigger having determined that the current network is congested. Figure 6(c) shows that the client performed a handover to the stream received via AP2 a short time after the average was triggered (about 27 seconds). In the above handover, the MN used 800 samples when calculating the average of the delay differences. To show what the average would have been had the manual trigger occurred at other times in the test, the moving average of the delay differences over the

Stream 2

Stream 1

Figure 6. Delay and Handover decision.

VI.

CONCLUSIONS AND FUTURE WORK

In this paper we showed that packet loss and substantial packet delay can occur in a congested WLAN, and that packet delay can be used as an indicator of when loss is likely to occur. We showed that delay difference between packets in two streams in a soft handover can be used to decide which is the better stream for the MN to use from a packet loss perspective. We showed how a stationary MN can apply this in a congested WLAN to determine when to handover. We described a scheme that can be implemented in the MN to achieve this. Finally, we presented results showing a successful handover in a live WLAN environment. Further investigation is planned to address some of the issues mentioned above. These include investigation of issues associated with initiating a second stream and the number of samples in the moving average. Also, an alternative scheme using jitter is being examined to overcome the issue of accurately determining congestion levels when both cells are heavily congested. ACKNOWLEDGMENTS The support of the Informatics Research initiative of Enterprise Ireland is gratefully acknowledged. Thanks also to Andrew Kelly for help in the setting up of the experiments.

REFERENCES [1]

IEEE 802.11b/d3.0, “Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specification,” August 1999. [2] A. Mishra, M. Shin, and W. Arbaugh, “An empirical analysis of the IEEE 802.11 MAC layer handoff process,” ACM SIGCOMM Computer Communication Review, April 2003. [3] C. Perkins, “IP mobility support,” RFC (Proposed Standard) 2002, IETF, Oct 1996. [4] A. Stephane and A. H. Aghvami, “Fast handover schemes for future wireless IP networks: a proposal and analysis” , Proc. IEEE Vehicular Technology Conference, pp. 2046-2050, May 2001. [5] M. Handley, H. Schulzrinne, E. Schooler, and J. Rosenberg, “SIP: session initiation protocol”, RFC 2543, IETF, Mar 1999. [6] E. Wedlund and H. Schulzrinne, “Mobility support using SIP”, Proc. ACM WoWMoM'99, USA, pp 76-82, Aug 1999. [7] E. Wedlund and H. Schulzrinne, “Application layer mobility using SIP”, Mobile Computing and Communications Review, Volume 1, Number 2, 2001. [8] J. Liu, D. Nicol, L. Perrone, and M. Liljenstam, “Towards high performance modeling of the 802.11 wireless protocol”, Proc. Winter Simulation Conference, pp 1315-1320, December 2001 [9] J. Noonan, A. Kelly, P. Perry, S. Murphy, and L.Murphy, “Simulations of multimedia traffic over SCTP modified for delay -centric handover”, Proc. World Wireless Congress, May 2004. [10] R. Chandra, Pr. Bahl, and Pl. Bahl, “Multinet: connecting to multiple IEEE 802.11 networks using a single wireless card”, Proc. IEEE Infocom, 2004. [11] J. Kristiansson and P. Parnes, “Application-layer mobility support for streaming real-time media”, Proc. IEEE Wireless Communications and Networking Conference, 2004.