SEAMLESS MOBILE IP HANDOFF IN ALL-IP WIRELESS NETWORKS Young-June Choi
EECS, University of Michigan Ann Arbor, MI 48109, USA [email protected]
Samsung Electronics Suwon, 443-370, Korea [email protected]
EECS, Seoul National University Seoul, 151-742, Korea [email protected]
Abstract – Mobile wireless networks, including IEEE 802.16e wireless metropolitan area networks, are evolving towards all-IP, and accordingly they face the fundamental overhead problem for mobility management. To reduce the handoff overhead of mobile IP, we design a subnet-based network, instead of a pure all-IP network, that primarily uses layer 2 handoff within a subnet. In such a network, we deploy some dual-linked access points (APs) that are connected to their two neighboring access routers (ARs). Each AR is considered to cover the area over hundreds of APs. Using additional link connections at the APs that are located between two AR regions, our proposed method can smoothly decouple layer 2 and 3 handoff operations. During the layer 3 handoff, either the old or corresponding new AR forwards packets destined for the mobile user through the same AP, so the mobile user experiences low latency and low packet loss. Through simple analysis, we compare our proposed scheme with the conventional approach in terms of latency.
are expected to adopt an all-IP based packet-switched system where IP packets traverse across an access network and the backbone network without any protocol conversion. As a WMAN system, IEEE 802.16d standard was already ratified in 2004, and in addition, a supplementary standard of IEEE 802.16e was issued in 2005 for mobility provision in the IEEE 802.16 systems. Design and implementation of WMANs including the 802.16 systems have been discussed in WiMax group. Based on the IEEE 802.16e, the pioneering system named WiBro has been already deployed in Korea for the 2.3GHz band. On the other hand, IEEE 802.20 group has introduced its new standard, Mobile Broadband Wireless Access (MBWA), which is possibly competing with the IEEE 802.16e based systems. Table 1 shows the technology comparison among WiBro, 802.16e, and 802.20 systems.
I. INTRODUCTION Third-generation (3G) access networks like WCDMA and cdma2000 have a complicated network structure and define various protocols to support the system structure. Since 3G networks basically have evolved from a circuitswitched cellular network, they have their own gateways to interpret IP from the backbone network, together with their own protocols and interfaces for the communication within themselves. Also, those systems support low data rate while covering wide areas. In contrast, the IEEE 802.11 wireless local area network (WLAN) achieves the system throughput of up to 54 Mbps, but covers the service area of two or three hundred meters, thereby being exposed to the difficulty in supporting high mobility users. To meet the intermediate requirements of data rate and service coverage between existing cellular networks and WLANs, wireless metropolitan area networks (WMAN) have been developed. Their fundamental challenges are twofold: achieving dramatically increased cell capacity with mobility support and designing simple network architecture. To attain high cell capacity, novel multiple access and antenna techniques have been introduced. Those examples are orthogonal frequency division multiplexing (OFDM), orthogonal frequency division multiple access (OFDMA), adaptive antenna array, and multiple-inputmultiple-output (MIMO) antennas. Meanwhile, to make the design of network architecture simple, new WMANs
Table 1. Technology comparison of WiBro, 802.16e, and 802.20. WiBro Multiple access
1.25/ 3.5/ 7/14Mhz
Mobility support Interference
- 60 Km/h Sectorization, Mobile multihop relay
management Handoff delay
- 250 Km/h Fractional frequency reuse
- 150 msec
As shown in Table 1, the IEEE 802.16e systems are designed to support the mobile speed of up to 60 Km/h, but mobility of users essentially incurs IP changes in the all-IP scenario. Although the activity in industry accelerates the trend towards new WMANs, some issues including IP mobility still remain vague because current standards deal with physical and MAC layers only. A potential solution to support IP mobility is mobile IP (MIP). However, it is known that MIP results in high handoff latency that amounts to several seconds . This becomes a serious problem in supporting seamless data service in future wireless networks. As a solution, hierarchical MIP and low-latency/fast handoff mechanisms have been introduced [2-4]. Nevertheless, there are still many issues for practical
implementation, and furthermore some packets still experience dropping even under low latency handoff . A possible approach is to reduce the number of MIP triggers by introducing subnet-based all-IP networks instead of pure all-IP networks . In the pure all-IP networks, an access point (AP) acts a role of access router (AR), which causes a frequent MIP handoff operations. In this case, whenever a mobile node (MN) changes its corresponding AP, MIP handoff must be carried out. Unlike this, in the subnet-based networks, an AR covers a region (or subnet) consisting of hundreds (or tens) of APs (i.e., cells). Within a subnet, only link-layer (layer 2) handoff is triggered, so the number of MIP handoff (i.e., layer 3 handoff) operations is reduced significantly. In this paper, we overview the existing approaches for smooth MIP handoff. Then, to achieve smooth MIP handoff, we develop a novel network architecture that has some dual-linked APs in the subnet-based network. When an MN crosses the boundary cell in an AR region towards a different AR, layer 2 (L2) handoff and layer 3 (L3) MIP handoff occur at the same time, which delays the service. To make the L3 handoff smooth, we design some APs to be able to access multiple ARs where each AR covers a different subnet, thereby enhancing the existing handoff mechanism. Since our scheme implements L2 and L3 handoff operations separately, an MN experiences very low latency and no packet drop during the handoff process. We demonstrate the performance of our method by simple analysis and simulation, and present a scenario of deployment in Seoul city.
II. OVERVIEW ON MOBILE IP HANDOFF Supporting IP mobility has been investigated widely. A simple way for the suitable behavior of IP is updating the routing table: when an MN moves into another network, the routing tables of involved routers are updated. However, it requires high overhead and latency, and furthermore the problem becomes serious with the number of MNs as well as the distance between home and visited networks. To overcome this problem, cellular IP and HAWAII  manage micro-mobility in which a network is deliberately divided by micro-mobility and macro-mobility. Cellular IP maintains two caches for local mobility and an MN refreshes the caches periodically for location update. To improve the performance of cellular IP, HAWAII uses a path setup scheme in the local network. A survey including the comparison of these protocols is given in . In the IETF, MIP has been discussed as a powerful and implementable technique because the table update for every involved router seems impossible. In MIP, only home agent (HA) needs to update the table since HA tunnels packets destined for the MN. To do so, each MN has two addresses: IP address and care-of-address (CoA)
that the MN obtains in a visited network. Whenever the MN moves into another network, it only has to configure a CoA and send this information to its HA that keeps the mapping table between its original IP address and CoA. The configuration of CoA has two scenarios: with and without a foreign agent (FA). In the normal case of mobile IPv4 (MIPv4), if a visited network has a FA, the FA allocates a CoA for a visiting MN. Hence every packet destined to the MN is encapsulated by the HA and forwarded to the FA. Then the FA decapsulates and transfers the packet to the destined MN. On the other hand, in the case of no FA, the MN attains its CoA either by dynamic host configuration protocol (DHCP) in IPv4 or by auto-configuration in IPv6. Especially the mobile IPv6 (MIPv6) enhances the MIPv4 by adding route optimization, dynamic HA discovery, and other functions. See RFCs 3344 and 3775 for more details. The HA always plays the role of binding update in both of MIPv4 and MIPv6, so the binding update at an HA is essential for MIP handoff. This makes the handoff delay long when the visited network is far from the home network. Furthermore the MIPv6 involves IP address configuration as well as location update and movement detection whenever the MN changes the subnet. Thus the handoff latency may not be acceptable for real-time applications. To overcome the MIP handoff problem, IETF discusses some techniques such as low latency handoff  for MIPv4, and hierarchical MIP  and fast handoff  for MIPv6. A low latency handoff scheme targets on reducing the delay incurred during the registration process. Considering the layer 2 trigger, it has three methods: pre-registration, post-registration, and combined handoff. The preregistration exploits an early notice of upcoming change in the layer 2 point of attachment and it begins registration before the layer 2 change. The post-registration reversely performs it after completing the layer 2 change. The combined handoff attempts the two methods jointly, so an MN tries the post-registration when the pre-registration fails. While packets arriving at a time between registration request and response may be lost due to the latency in the conventional MIP handoff, the low latency handoff prevents loss. Similarly, a fast handoff scheme tries to decrease the address resolution delay by address pre-configuration in MIPv6. It also uses the layer 2 trigger indicating that an MN will be handed over. On receiving the indication, the old AR begins forwarding packets to a new AR. Hierarchical MIP introduces a mobile anchor point (MAP) to cover a group of routers that also maintain an addressing mapping table. The MAP assigns and manages another IP address named regional CoA (RCoA) in addition to a local CoA (LCoA). When an MN travels a network under the same MAP, it just changes LCoA with the aid of the MAP
without referring to the HA. The HA cares RCoA changes only. Although these algorithms improve handoff performances, there remain some issues. First, the lower layer may not know which network the MN moves into when the low latency handoff or fast handoff runs, so every neighboring AR has to prepare pre-configuration. Second, packets still experience some dropping probability during the handoff between two ARs. These problems have been solved by simultaneous binding or simulcast [8,9]. The simultaneous binding maintains multiple care-of-addresses to which packets are transmitted. Similarly, multicast or simulcast sends packets to multiple candidate routers. However, these approaches show the disadvantage of consuming too much bandwidth for smooth handoff. Without having redundant transmission, the approach in  inserts an MAC bridge between subnets to solve the problem of wasting bandwidth. As the MAC bridge keeps MAC address information, it can relay packets after the layer 2 handoff even though the MIP handoff is not completed. This method is similar to our approach because the dual-linked AP in a WMAN can play the role of MAC bridge in a WLAN.
III. NETWORK MODEL In a pure all-IP scenario, an AP (or base station) acts the role of an AR. It incurs high overhead, however, especially when an MN configures the MIP addresses for handoff. As it takes several seconds to execute MIP handoff , the MIP hinders an MN from carrying out smooth handoff. In addition, WMANs are expected to have a small cell radius due to the use of high frequency band, which possibly results in a short cell residence time. To solve the fundamental problem of the pure all-IP cell architecture, most developers and researchers take an approach of separating the functionality of an AR from that of an AP in order that each undertakes L3 and L2 protocols, respectively. Fig. 1 shows an example of a simple network where an AR manages several APs. This relation is similar to that between base station controller (BSC) and BS in an existing cellular network. A subnet consisting of an AR and many APs can be configured by Gigabit Ethernet. An AR has a lookup table that records the location of each MN. In this way, an AR forwards a packet to the corresponding AP. Then, an MN traveling within the subnet while changing APs performs L2 handoff only without changing the MIP attachment. When it moves into another AR area, it experiences L3 handoff. The subnet-based all-IP network has a hierarchical architecture, so it is possible to perform efficient resource management in spite of inflexibility in configuration. On the other hand, the pure all-IP network suffers from long handoff latency and high signaling overhead since it incurs L3 protocol at each handoff. Its
advantage is in simple and cost-efficient implementation. Further comparisons are presented in . IP Network Access Router
AAA server IP-based Access Network
Fig. 1. The subnet-based access network that has a dual-linked AP.
The subnet-based network reduces the frequency of L3 handoff that is accompanied by relatively long latency. So we consider the subnet-based network as a practical IPbased mobile system. Nevertheless, an MN still experiences a long latency when it performs an L3 handoff. For seamless L3 handoff, we develop a dual-linked AP model where some APs are connected to two neighboring ARs at the same time as shown in Fig. 1. Obviously, our approach can be extended to support the case where an AP is linked with more than two ARs by adding more links as many as neighboring ARs to that AP. In this paper, we will use the terminology of ‘dual’ as the general implementation term. In the conventional subnet-based model, an MN performs L2 and L3 handoffs at the same time when it crosses the boundary of a subnet. This may cause a serious problem of communication blackout because L2 handoff typically exploits a conservative method in preventing the ping-pong effect. It happens like this. An MN starts an L2 handoff when the signal power of the corresponding AP is weak. As an L3 handoff is accompanied by a long latency, the signal may turn too weak during the L3 handoff, resulting in a blackout. In contrast, our network model with some dual-linked APs decouples L2 and L3 handoffs, thereby providing a flexible handoff mechanism. Since each dual-linked AP can access both ARs of new and old, it helps L3 handoff to be performed independently of L2 handoff when an MN stays in its coverage. More explanations are given next.
IV. SMOOTH L3 HANDOFF BY DAUL-LINKED APS
2) Mobile IP handoff Following a proper movement detection, an MN performs MIP handoff. The MN begins MIP handoff by sending a request message. After the MN obtains a CoA for the new subnet, the AR forwards the request message to the MN's HA to update the MN's location information. During this process, packets arriving at the AP via the old AR will be transferred to the MN by using the new CoA. This is possible because the AP can access the two ARs, which is a unique feature in our scheme. In contrast, in a conventional network, some packets arriving at the old AP or AR will be dropped or forwarded via old and new ARs, so the forwarded packets will experience some latency. In our scheme, each dual-linked AP maintains a table for mapping between old and new CoA during the handoff procedure. Indeed, the handoff latency in MIP is mainly incurred by HA registration. In our architecture, however, there exists only a slight handoff latency since every packet arrives at the same AP via either old or new AR during the handoff. Fig. 2 shows an example of our handoff scheme. An MN and its corresponding AP perform L2 handoff whenever it crosses a cell boundary, while performing L3 handoff separately from L2 handoff.
HA IP Backbone Network
AR 2 AP 1
L3 Handoff (AR1-AR2) L2 Handoff (AP2-AP3)
Dual Linked AP
AP 2 L2 Handoff (AP1-AP2)
1) Future movement prediction Generally an MN is able to sense the presence of neighboring APs since each AP broadcasts its pilot signal. When the MN enters the service area of a dual-linked AP, it triggers L2 handoff. Completing the handoff, it predicts the movement by detecting the pilot strengths of neighboring APs. If it is likely to move into some other subnet, it prepares L3 handoff. The handoff can be initiated by either the MN or the AP. If L3 handoff is triggered too early, there exists a possibility of too many L3 handoffs, resulting in the pingponging effect. On the other hand, if too late, L2 and L3 handoffs are incurred at the same time. In this case, we cannot reduce the handoff delay because L3 handoff dominates the overall delay. This motivates us to design an algorithm that initiates early L3 handoff following the concept of the existing L2 handoff algorithm. The graph in Fig. 2 shows an example of L2 and L3 handoff triggers. In this scenario, if the measured pilot signal strength at the MN from a new AP (AP3) exceeds that of the old AP (AP2) by Th1 for the time interval I1, the L2 handoff towards the new AP is triggered . If the new AP belongs to a different subnet, the L3 handoff is initiated according to the thresholds Th2 and I2. In this case, the L3 handoff must start before the next L2 handoff. Deciding the threshold values is an implementation issue, so we do not deal with it here.
AP 3 Time L3 Handoff
Fig. 2. An example of network-aided smooth mobile IP handoff.
B. Handoff latency 1) Typical mobile IP handoff We now analyze handoff latency that an MN experiences during the handoff procedures. In the case of MIP handoff, L2 handoff occurs at the same time because the MN crosses a subnet as illustrated above. An MN in the conventional MIP handoff without a dual-linked AP suffers from long delay because it begins the L3 registration process after the L2 handoff has been completed. Moreover, the MN is unable to send or receive packets during the L3 registration process. We obtain the total handoff latency T by adding T2 and T3, which indicate handoff latency incurred by L2 and L3 handoff mechanisms, respectively. Investigating T2 depends on the system implementation, so we just observe the components of T3 as follows. • TMN-AP, TAP-AR, and TAR-HA: the time required for a message to pass from MN to AP, from AP to AR, and from AR to HA, respectively. • PMN, PAP, PAR, and PHA: the processing time required by MN, AP, AR, and HA, respectively. Using these parameters, we can express T3 as T3 = PMN+PAP+PAR+PHA + 2(TMN-AP+TAP-AR+TAR-HA). 2) Smooth mobile IP handoff in the dual-linked AP model In our mechanism, the handoff latency an MN experiences is smaller compared to that in the legacy one. Moreover, the MN undergoes T2 and T3 handoff operations separately. During the L3 handoff process at an AP, the MN can send or receive packets although the corresponding AR and HA have not finished the handoff process. This is because the AP can receive packets continually through either old or new AR. Thus, we obtain T3 in our scheme as follows.
T3=PMN+PAP+2TMN-AP. C. Packet loss
of L2 and L3 handoffs in the conventional structure. In contrast, our scheme achieves small latency as shown in Fig. 4 (b), where we clearly observe the separation of L2 and L3 handoffs as well. Both handoff operations have similar latency of several milliseconds. We omit the results for other scenarios because they exhibit similar tendencies.
L2 Handoff 5500
L2 + L3 Handoff 5000
(a) Conventional scheme
L3 Handoff 5000
(b) Our scheme 4000 100
Fig. 3. Packet forwarding comparison.
D. Performance evaluation Using the ns-2, we evaluated the performance of our mechanism in a simple topology where an MN traverses a dual-linked AP that belongs to two different ARs. For simplicity, we assume that old and new ARs play the role of HA and FA, respectively. Each link has the bandwidth of 5Mbps and the propagation delay of 2msec. We suppose an MN generates a TCP traffic with a data rate of 500kbps destined to a corresponding node that is linked to the two ARs. The MN starts its data transmission at the center of the AP and moves towards the center of the other AP with 30 Km/h, where the radius of APs is 1 Km. As the performance is closely related with the network topology and the handoff scheme, we considered a simple scenario that runs MIP only for evaluation. Fig. 4 shows the packet sequence number of which increase indicates how well the MN communicates with the corresponding AP. In Fig. 4 (a), the packet transmission stops for about six seconds, which amounts to the latency
In the conventional scheme, the new AP or AR may receive packets for an MN to which it does not have a direct connection while the MN is performing the handoff between subnets. In this case, the new AP or AR may drop or buffer packets destined for the MN. This case also occurs when the whole handoff process has not completed yet even if the MN has established a new connection already. Some packets may be lost although a low latency handoff mechanism works . Therefore, as shown in Fig. 3 (a), some packets should be forwarded to the corresponding location during the handoff process. Unless the system supports packet forwarding, packets will be dropped. Refer to  for the evaluation of packet dropping. In contrast, our scheme has almost no packet loss during the handoff. As shown in Fig. 3 (b), our architecture does not show any loss because the AP has a connection to each of old and new ARs.
Fig. 4. The packet sequence number; (a) mobile IP handoff, (b) our scheme.
V. DEPLOYMENT SCENARIO IN A WIBRO NETWORK We now show a practical scenario how our proposed APs can be deployed. Fig. 5 exhibits possible locations of dual-linked APs in Seoul city when IEEE 802.16e-based WiBro system begins its service. We expect that four ARs can cover the whole city and each AR is connected to approximately over 100 APs. Then we conjecture that the number of dual-linked APs is approximately 50, and half of them plays the role of changing ARs within the Seoul city. This amounts to about 5% of the total number of APs that will be deployed in Seoul. Since a city is surrounded by mountains that usually become the boundary of the city and only a few roadways connect the city to the suburbs, the number of dual-linked APs can be further reduced. Also a river crosses Seoul city, so dual-linked APs will be probably located on bridges. This means that if an MN travels through a bridge, it will pass a dual-linked AP. Since bridges and roadways are mostly used by fast moving mobile users, movement prediction will be much easier.
VI. CONCLUDING REMARK In this paper, we examined a novel network architecture in the view of IP mobility support. The network model provides a seamless MIP handoff mechanism by using dual-linked APs in a subnet-based all-IP network. In a subnet consisting of an AR and hundreds of APs, we located dual-linked APs on the subnet boundary that are able to access more than two ARs directly by virtue of added access links. Therefore L2 and L3 handoff operations can be separated smoothly, thereby reducing the handoff latency significantly. Our scheme gives a
 K. Malki, “Low latency handoffs in mobile IPv4”, draft-ietf-mobileip-lowlatency-handoffs-v4-09.txt, June, 2004.  R. Koodli, “Fast handovers for mobile IPv6”, draft-ietfmipshop-fast-mipv6-03.txt, Oct. 2004.  C. Blondia, et.al., “Performance evaluation of layer 3 low latency handoff mechanisms”, ACM Mobi. Net. App., vol. 9, no. 6, Dec. 2004, pp. 633-645.  Y.-J. Choi, K. B. Lee, and S. Bahk, “All-IP 4G network architecture for efficient mobility and resource management”, to appear in IEEE Wireless Commun. Mag., April 2007.  D. Saha, A. Mukherjee, I. S. Misra, M. Chakraborty, N. Subhash, “Mobility support in IP: a survey of related protocols”, IEEE Network, vol. 18, no. 6, Nov. 2004, pp. 34-40.  A. Stephane, A. Mihailovic, and A. H. Aghvami, “Mechanisms and hierarchical topology for fast handover in wireless IP networks”, IEEE Commun. Mag., vol. 38, no. 11, Nov. 2000, pp. 112-115.  R. Hsieh, Z. G. Zhou, and A. Seneviratne, “S-MIP: a seamless handoff architecture for mobile IP”, Proc. IEEE INFOCOM '03, San Francisco, CA, USA, March 2003.  H. Holma and A. Toskala, “WCDMA for UMTS”, Wiley, 2000.
promising opportunity for seamless mobility support in next generation communication networks. For real implementation, there are some technical issues to be investigated further. First, the dual-linked APs should be designed to handle twice the load as single-linked one, especially in transmitting some control packets. For example, the amount of router advertisement traffic could be doubled in the dual-linked APs. Second, in case multicast service is available, the dual-linked AP should support its function. Lastly, the cost to implement the duallinked APs could be increased more because of the additional functions. ACKNOWLEDGEMENT The authors would like to thank the team members for the 4G project of Samsung Advanced Institute of Technology for their helpful discussions.
REFERENCES  H. Yokota, A. Idoue, T. Hasegawa, and T. Kato, “Link layer assisted mobile IP fast handoff method over wireless LAN networks”, Proc. ACM MOBICOM '02, Atlanta, GA, USA, Sept. 2002.  H. Soliman, C. Castelluccia, K. Malki, L. Bellier,, “Hierarchical mobile IPv6 mobility management (HMIPv6)”, draft-ietf-mipshop-hmipv6-03.txt, Oct. 2004.
2Km Ilsan-gu Goyang city Dobong-gu Namyangju city
Dukyang-gu Goyang city Gangbuk-gu Eunpyeong-gu Seongbuk-gu
Guri city Jungrang-gu
Gyeyang-gu Incheon city
Seongdong-gu Ohjung-gu Bucheon city
Bupyung-gu Incheon city
Wonmi-gu Bucheon city
Sosa-gu Bucheon city
Gwanak-gu Geumcheon-gu Gwangmyung city
Sujung-gu Seongnam city
Jungwon-gu Seongnam city
Fig. 5. A scenario of WiBro deployment in Seoul city. The map shows four ARs and approximately fifty dual-linked APs.