A Comparison of Mechanisms for Improving Mobile IP Handoff Latency for End-to-End TCP

Robert Hsieh and Aruna Seneviratne

School of Electrical Engineering and Telecommunications The University of New South Wales Sydney, NSW, Australia 2052

roberth@ee.unsw.edu.au, a.seneviratne@unsw.edu.au ABSTRACT

Handoff latency results in packet losses and severe End-to-End TCP performance degradation as TCP, perceiving these losses as congestion, causes source throttling or retransmission. In order to mitigate these effects, various Mobile IP(v6) extensions have been designed to augment the base Mobile IP with hierarchical registration management, address pre-fetching and local retransmission mechanisms. While these methods have reduced the impact of losses on TCP goodput and improved handoff latency, no comparative studies have been done regarding the relative performance amongst them. In this paper, we comprehensively evaluated the impact of layer-3 handoff latency on End-to-End TCP for various Mobile IP(v6) extensions. Five such frameworks are compared with the base Mobile IPv6 framework, namely, i) Hierarchical Mobile IPv6, ii) Hierarchical Mobile IPv6 with Fast-handover, iii) (Flat) Mobile IPv6 with Fast-handover, iv) Simultaneous Bindings, and v) Seamless handoff architecture for Mobile IP (S-MIP). We propose an evaluation model examining the effect of linear and ping-pong movement on handoff latency and TCP goodput, for all above frameworks. Our results show that S-MIP performs best under both ping-pong and linear movements during a handoff, with latency comparable to a layer-2 (access layer) handoff. All other frameworks suffer from packet losses and performance degradation of some sort. We also proposed an optimization for S-MIP which improves the performance by further eliminating the possibility of packets out of order, caused by the local packet forwarding mechanisms of S-MIP.

Keywords
Mobile IP, Hierarchical Mobile IP, Fast-handover, S-MIP, Handoff Management, Seamless Handoff

1. INTRODUCTION
In the Internet (IP) environment, when a mobile node moves and attaches itself to another network, it needs to obtain a new IP address. This changing of IP address means that all existing IP connections to the mobile node need to be terminated and then reestablished. This is essential as the IP routing mechanisms rely on the topological information embedded in the IP address to deliver the data to the correct end-point. Mobile IP (MIP) [14] describes a global solution that overcomes this problem through the use of indirection provided by a set of network agents. It does not require any modifications to existing routers or end correspondent nodes. With MIP, each mobile node is identified by an address from its home network, regardless of the point of attachment. While a mobile node is away from its home network, it obtains an IP address from the visiting network and registers it with a home agent within its home network. The home agent intercepts any packets destined to the mobile node, and tunnels or explicitly routes (source routing) them to the mobile nodes current location. Thus, initiating this indirection requires a timely address reconfiguration procedure and a home network registration process. The time taken for a mobile node to configure a new network care-of address in the visiting network, and the time taken to register with the home agent together constitute the (overall) handoff latency. Handoff latency is the primary cause of packet loss and resulting in performance degradation, especially in the case of reliable endto-end communication. As a result, numerous methods of minimizing the handoff latency have been proposed in the literature. The proposed schemes can be broadly classified into those that operate above the IP layer [4], [5], [2], [1] and those that operate at the IP layer [20], [16], [17], [9]. In general, the solutions that operate at the IP layer are regarded as being more suitable as they do not violate any of the basic Internet design principles [3], and more importantly because they do not require any changes to the protocols at the corresponding nodes. We focus on the IP layer solutions in this paper. Essentially, the IP layer solutions attempt to minimize the registration delay by introducing a hierarchical structure, and lower the address (re)configuration delay through advanced

Categories and Subject Descriptors

C.2.1 [Computer-Communication Networks]: Network Architecture and Design - Wireless communication; C.4 [Computer Systems Organization]: Performance of Systems

General Terms
Performance, Design, Experimentation

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MobiCom03, September 14-19, 2003, San Diego, California, USA. Copyright 2003 ACM 1-58113-753-2/03/0009$5.00.

configuration. A number of ways of introducing hierarchy into IPv4 as well as IPv6 networks, and realizing the advanced configuration have been proposed in the last few years [17], [11], [10], [9], [20], [16], [8]. Analysis of these schemes have shown them to significantly improve the handoff latencies experienced and result in better end-to-end performance for reliable communication in Mobile IP systems [7], [6]. However, despite the several approaches in optimizing advanced configuration, namely how packet forwarding should be carried out to minimize the effect of mobility, that have been proposed, the optimum methods of doing this is still not clear. The main cause is the timing ambiguity [11] problem, i.e. the exact moment when a mobile node switches network is non-deterministic a priori, therefore it is difficult to determine the correct time to start the packet forwarding mechanism. Forwarding too early or forwarding too late will cause packet losses [17], negating the advantages of advanced configuration. In this paper, we address the above by analyzing the performance of five individual handoff latency reduction methods, namely, i) Hierarchical Mobile IPv6 [17], ii) Hierarchical Mobile IPv6 with Fast-handover [17], iii) (Flat) Mobile IPv6 with Fast-handover [9], iv) Simultaneous Bindings [11], and v) Seamless handoff architecture for Mobile IP (S-MIP) [8], through simulation. We examine the handoff latency caused by the mobile node moving linearly and moving in a ping-pong fashion. We show that, proactive advanced configuration, i.e. S-MIP, where a mobile nodes movement pattern is augmented with a specialized forwarding algorithm when performing a handoff, provides the best result. Also, we propose further optimization to this proactive scheme resulting in improved performance. The rest of the paper is organized as follows. Section 2 briefly describes all the handoff latency reduction schemes mentioned in this paper. Section 3 outlines the motivation and our optimization technique over SMIP. The experimental methodology and implementation extension to ns-2 are presented in Section 4, while results of the simulation are shown in Section 5. Section 6 discusses some performance issues relating to handoffs and summarizes our view on each of the latency reduction frameworks examined in this paper. We conclude and outline future research directions in Section 7.

Anchor Point) entity and the protocol for micro and macro mobility is achieved using the features of Mobile IPv6. The MAP in HMIPv6 intercepts all packets on behalf of the MN it serves and tunnels them to the MNs on-link1 care-of address (LCoA). When a MN moves into a new MAP domain, it acquires a regional address (RCoA) and an on-link address (LCoA). In the simplistic case, the address of the MAP entity is used as the RCoA, while the LCoA address is formed using the Stateless Address Autoconfiguration [19] protocol. Other methods of obtaining the RCoA and LCoA are described in [17]. After obtaining these addresses, the MN sends a Binding Update (BU) to the MAP, which will bind the MNs RCoA to the LCoA. If successful, the MAP will return a binding acknowledgement (BAck) to the MN indicating a successful binding (registration). In addition, the MN must also register its new RCoA with its home agent by sending another BU that specifies the binding between its home address and the RCoA, i.e. the MAPs address. When the MN moves to a new access router within the same MAP domain, it simply acquires an LCoA, and updates the MAP.

2.2 Local Handoff Latency Reduction Protocols

In this section, we illustrate the differences between Low Latency Handoff, Fast-handover, Hierarchical Mobile IPv6 with Fast handover, and Simultaneous Bindings, from the IETF mobileip Working Group. Low Latency address configuration is about configuring an address for the MN in a network that it is likely to move to, before it moves. The Low Latency handoff proposal [10] describes two methods of achieving this, namely Pre-registration and Postregistration. With Pre-registration handoff, the MN is assisted by the network to perform L3 (layer-3) handoff before it completes the L2 (layer-2) handoff. It uses L2 triggers, which arises as a result of beaconing signals from the network the MN is about to move to, to initiate an IP layer (L3) handoff. Its design however, diverges from the clean separation of L2 and L3 of the base Mobile IPv4 scheme. With Post-registration handoff, L2 triggers are used to setup a temporary bi-directional tunnel between the oFA (old Foreign Agent) and the nFA (new FA). This allows the MN to continue using its oFA while performing the registration at the same or later time. A combined method is also possible where, if the Pre-registration does not complete in time, the oFA forwards traffic to the nFA using the Post-registration method in parallel. Fast-handover [9] is the Low Latency handoff equivalent for Mobile IPv6 network. It is similar in concept to the combined method described prior, and consists of three phases: handover initiation, tunnel establishment and packet forwarding. The handover initiation is started by the L2 trigger based on certain policy rule (unspecified by IETF at the time of writing). This is done by the MN sending a Router Solicitation Proxy (RtSolPr) message to the PAR (Previous Access Router) indicating that it
1

2. BACKGROUND 2.1 Hierarchical Structures and Protocols

Hierarchical schemes separate mobility management into micro mobility (intra-domain) and macro mobility (inter-domain). They introduce a Mobility Routing Point (MRP) [6] that separates micro from macro mobility. The MRP entity is normally placed at edges of a network, above a set of access/edge routers which constitute the MRPs network domain. The MRP intercepts all packets on behalf of the mobile node (MN) it serves and redirects them to the MN. This enables MNs, which move between access networks that are within the same MRP network domain, to register with the MRP, thus avoiding potential lengthy run-trip delay associated with registration to its home agent. This type of intra-domain mobility is managed by IP Micromobility management protocols, such as [20] and [16], while inter-domain or macro mobility is almost exclusively managed using Mobile IP. In the context of Hierarchical Mobile IPv6 (HMIPv6), the Mobility Routing Point is equivalent to the MAP (Mobility

Three different addressing scopes exist in IPv6. A global address uniquely identifies a node on the Internet. A regional address is a global address that is specific to a particular region/domain on the Internet. An on-link address is an address local to a domain. It is only a unique identifier inside the specific domain and may not be uniquely identified on the Internet.

wishes to perform a fast-handover to a new attachment point. The RtSolPr contains the link-layer address of the new attachment point, which is derived form the NARs (New Access Router) beacon messages. The MN will receive, in response, a Proxy Router Advertisement (PrRtAdv) message from the PAR, with a set of possible responses indicating that the point of attachment is i) unknown, ii) known but connected through the same access router or iii) is known and specifies the network prefix that the MN should use in forming the new CoA. Subsequently, the MN sends a Fast Binding Update (F-BU) to the PAR using its newly formed CoA based on the prior PrRtAdv response, as the last message before the handover is executed. The MN receives a Fast Binding Acknowledgement (F-BAck) either via the PAR or the NAR indicating a successful binding. The tunnel establishment phase creates a tunnel between the NAR and the PAR. To establish a tunnel, the PAR sends a Handover Initiation (HI) message (containing the MNs requesting CoA and the MNs current CoA) to the NAR. In response, the PAR receives a Handover Acknowledgement (HAck) from the NAR. If the new CoA is accepted by the NAR, the PAR sets up a temporary tunnel to the new CoA. Otherwise, the PAR tunnels packets destined for the MN to the NAR, which will take care of forwarding packets to the MN temporarily. Finally, the packet forwarding phase is performed to smoothen the handoff until subsequent registration by the MN to the home agent is completed. The PAR interacts with the NAR to facilitate the forwarding of packets between them, through the previously established tunnel. The initiation of the forwarding is based on an anticipation timing interval heuristic, that is, the network anticipates as to when a mobile node is likely to handoff and therefore infers the appropriate packet forwarding moment based on the anticipation timing interval. Such an interval is however extremely difficult to generalize, and forwarding too early or too late will result in packet losses, negating the purpose of packet forwarding. Once arriving at the new access network, the MN sends the Fast Neighbor Advertisement (F-NA) message to initiate the flow of packets (to itself) from the NAR. Hierarchical Mobile IPv6 (HMIPv6) with Fast-handover [17] is another attempt to further reduce the overall handoff latency from what Fast-handover can offer alone. By combining HMIPv6 with Fast-handover, latency due to i) address configuration and ii) the subsequent home network/agent registration, can both be reduced. The MAP can be viewed as the local home agent, and in most cases, it is located closer to the MN than the Home Agent (HA). Therefore, the signaling cost saved is the difference between the roundtrip time of the MN to the MAP and the roundtrip time between the MN to the HA, assuming that message processing time within a network node is insignificant in comparison. This combination requires minor modification to the standard HMIPv6 protocol and the Fast-handover protocol, i.e., relocating the forwarding anchor point from the PAR to the MAP as outlined in [17]. An alternative to the packet forwarding scheme has also been proposed, namely, the Simultaneous Bindings framework [11]. It proposes to reduce packet losses at the MN by n-casting packets for a short period to the MNs current location and to n-other locations where the MN is expected to move to. The n-casting can be carried out by the PAR, the MAP or the HA. The Simultaneous Bindings scheme recognizes the problem of not knowing when

Linear Movement Scenario

MAP

Stochastic Movement Scenario (inc. Ping-pong)

MAP

Stationary Scenario
MAP

PAR

NAR

PAR

NAR

PAR

NAR

MN MN MN

Center dividing boundary of 2 network coverage areas

Figure 1. Movement Scenarios the MN is likely to move the timing ambiguity and attempts to remove it by careful packet duplication to multiple access networks. It also claims to be able to address the problem associated with ping-pong movement of MNs between two access routers by this packet duplication process, as it is not necessary to re-configure the MNs CoA during ping-pong movement (rapid back and forth movement between two access routers/points).

2.3 The S-MIP Protocol

The Seamless Handoff architecture for Mobile IP (S-MIP) [8] provides a different approach to solve the timing ambiguity problem. Unlike the Hierarchical Mobile IPv6 with Fast-handover approach (which uses the L2-trigger) or the Simultaneous Bindings (which uses the L2-trigger and n-casts packets to multiple destinations), in S-MIP, the network uses mobile node location and movement patterns to instruct the MN when and how handoff should be carried out. It is a structured approach amalgamating mobile node location tracking, movement patterning and handoff algorithm(s) into an integral unit, providing smarter handoffs. The mobile node initiates a handoff while the network determines the handoff decision. S-MIP uses signal strength together with triangulation to track the MNs location [8], and determine its movement pattern. The entity which stores the history of the locations and determines the movement pattern, referred to as the Decision Engine (DE), is similar to a MAP, and covers a group of access routers and the associated MNs. The DE makes the handoff decision for the MN as described below. Once a decision is made, it is relayed to the access routers (PAR and NAR) and eventually to the MN as part of the Fast-handover PrRtAdv message2. From the movement pattern, the DE decides whether i) the MN is moving linearly, ii) the MN is moving stochastically (including ping-pong), or iii) the MN is stationary and near the center dividing boundary of two network coverage areas (see Figure 1). If determined to be moving linearly, every subsequent packet from a corresponding node (CN) arriving at the MAP, after the handoff decision is made, will be duplicated and send to both the PAR and the NAR simultaneously. This is referred to as the SynchronizedPacket-Simulcasting (SPS) [8] process. These packets will be marked with a S bit, as an option parameter in the IP header. The NAR maintains an f-buffer, containing packets forwarded from the PAR (f-packets) and an s-buffer, containing packets that are marked with the S bit (s-packets). Upon reception of the F-NA
2

S-MIP builds on the structure of HMIPv6 with Fast-handover and operates similar to that of the Mobile Node Initiated Fasthandover.

message from the MN, the NAR will start emptying the f-buffer followed by the s-buffer. Meanwhile at the PAR, SPS specifies that only those packets (duplicated copy) which do not have the S bit marked will be forwarded to the NAR. In addition, all packets (s/f packets) will be sent on the wireless channel by the PAR. In the case that the MN does not switch networks immediately, it will still be able to receive packets from the PAR. If determined to be moving in a stochastical manner (including ping-pong movement), the PAR and the NAR will be put into anticipation-mode where even through a MN might no longer wish to be associated with an AR, the AR will still maintain the MNs binding, in preparation for the MN returning. This avoids the unnecessary re-setup resource and time costs. In this scenario, the MN is able to switch networks freely, using an FNA message, once the signal strength has decreased to a certain predefined (by the DE) threshold. Lastly, if determined to be in a stationary state near the center dividing boundary between two network coverage, the MN will establish multiple bindings between itself and the ARs, i.e. the MN uses more than one careof address simultaneously. In short, S-MIP is able to differentiate the movement type of the MN and act accordingly using different handoff strategies. Moreover, it is able to determine the likelihood of MNs next move, inferring from the history of recent mobile node movements. Coupling with an advanced local forwarding management scheme, namely the SPS, the S-MIP framework uniquely differentiates itself from other frameworks described prior.

Figure 2. S-MIP Handoff Deficiency receiving the F-BU (which is then forward to the MAP, same as specified in [8]). Therefore, the PAR only needs to send the FBAck message to the NAR. With this modification, the possibility of receiving duplicated packets is reduced to zero. Furthermore, the only possible way that packet losses may occur, if this optimization is enforced, is through channel error during transmission. The probability of this happening is no greater than in the original S-MIP. We argue that it is unnecessary for the PAR to send f-packets onto the wireless channel. Unlike schemes such as Simultaneous Bindings, in S-MIP, once a handoff decision has been received by the MN, it must take place. The second part of the optimization is an IP packet filtering mechanism at the NAR. This is to proactively prevent erroneous packet forwarding by the PAR. One possible cause is when the PAR incorrectly forwards IP packets with the S bit set as f-packets to the NAR. Essentially, a matching algorithm that compares IP packets within the s-buffer and the f-buffer at the NAR is required to discard any identical packets in the s-buffer which have already been sent by the f-buffer. One way to proceed in such comparison is to examine the 16 bit IP identification, fragment offset and flag (used in conjunction with the fragment offset) fields in the IP packet header. In this way, one can be sure that even if IP packets have been fragmented on the traversing path, they can be uniquely identified and therefore be compared and discarded where appropriate. The actual implementation method is not within the scope of this paper. Furthermore, as such matching is only necessary when there are packets in the f-buffer, the computation resources required are minimal as the f-buffer size is small. This is because the size of the f-buffer depends on the time it takes a mobile node to perform an L2 handoff, and the relay time between the two access routers, which amount to tens of milliseconds at most. It must be stressed that this mechanism need not be performed by default at the NAR as the likelihood of packet duplication error, after the first part of our optimization, is very small.

3. A CASE FOR AN OPTIMIZED S-MIP FRAMEWORK

While it has been shown in [8] that S-MIP provides a well composed solution in providing a seamless handoff, we believe that there still exist some deficiencies within the framework. Consider the following (see Figure 2), during a linear handoff, the PAR forwards packets which are not marked with an S bit in the IP option parameter to the NAR (f-packets), as well as sending a copy (f-packets) onto the wireless channel, as part of the SPS process. The sending of f-packets on the PARs wireless channel is to allow the MN to still receive these packets, in case it does not switch networks immediately [8]. Therefore the worst case scenario is one where the MN receives all the f-packets while still in the access network of the PAR, but again receives all the forwarded f-packets (from PAR to NAR), when arriving at the NARs access network. In the case of reliable end-to-end communication, received packets that are out of sequence (in reality duplicated packets) will be interpreted as errors. In terms of TCP, duplicated acknowledgements will be sent by the receiver causing source throttling and maybe even retransmission. A simple solution to this is simply not to send the f-packets onto the wireless channel. In what follows we elaborate on our solution. Assuming that the access technology in use is 802.11b where the message/link corruption in the effective coverage area is zero [8], we can eliminate the out of sequence behavior by simplifying the handoff procedure as follows. Upon sending the F-BU message to the PAR, the MN must immediately switch to the NAR. Furthermore, the PAR must immediately forward packets to the NAR after

Since our optimization deals with reducing the possibility of

packets out of sequence during the forwarding (SPS) phase, the actual handoff latency remains the same when compared with the original S-MIP. Thus, in the simulation experiment that follows, we will initiate the optimization by default. In our comparison, we will illustrate the impact of out of sequence packets in the S-MIP

framework, in other words, what might happen without such optimization.

4. SIMULATION MODEL
The goal of our simulation is to examine the effectiveness of various Mobile IP extension architectures in L3 (IP layer) handoff latency reduction, over reliable end-to-end communication, namely TCP. In particular, we are interested in examining the bulk data flow rather than the interactive data flow scenario, since bulk data flow is more prone to disruption during a handoff.

4.1 Implementation Details

This section describes the protocols we have implemented as extensions to the Network Simulator version 2 (ns-2). These protocols include Mobile IPv6, Hierarchical Mobile IPv6, Fasthandover, Simultaneous Bindings and S-MIP. The base ns distribution (ns-allinone2.1b7a) [18] was patched with the ns wireless extension module [21], allowing basic Mobile IP(v4) protocol operations. It was further extended with our implementation. For the Mobile IPv6 protocol suite, we did not implement a complete set of proper IPv6 features as this is unnecessary for our simulation purposes. It must be noted that we did not implement all the protocols mentioned in this paper in full. Rather, we only implemented what is necessary for our simulation purposes. In respect to our simulation, the key differentiator between Mobile IPv4 and Mobile IPv6 is the Binding Cache Management. This includes the IP Destination Option, which is necessary to support the Home Address Option. We implemented these on top of the existing registration, packet encapsulation/decapsulation mechanisms of the ns wireless extension module. No security mechanism was implemented for the binding cache update. For Hierarchical Mobile IPv6 protocol suite, a MAP Agent was implemented to provide the MAP registration functionalities. The MAP agent only supports the use of its address as the RCoA for the MN. A simplified MAP discovery mechanism was created to enable the MAPs RCoA to be discovered by MNs. The Mobile IP router advertisement (beacon) message was modified to include the MAP advertisement option. The schematic of a MAP node is depicted in Figure 3. For the Fast-handover protocol suite, we made a slight modification to the ns Node entity to facilitate the use of the encapsulator/decapsulator provided by the wireless extension module. This enables tunnel mechanisms (IP encapsulations in IP [13]) to be setup between all types of node in ns, namely, wired, wireless and hybrid wired-wireless nodes (i.e. the BaseStationNode entity). Further, the new messages required by the Fast-handover, that is, PrRtAdv, RtSolPr, HI, HAck, F-BU, FBAck and F-NA were also added. The BaseStation Node in ns-2 was further modified to handle these messages. For the Simultaneous Bindings protocol suite, we are only interested in the case where the MN binds with the MAP entity. Other cases, binding with the HA or the PAR, are not considered in our comparison. Therefore we implemented purely the operations required for the MAP binding case. The bicasting/ncasting mechanism was added to the MAP Node.

Figure 3. Schematic of a MAP Node

Figure 4. Schematic of Mobile Node For the S-MIP protocol suite, we implemented the Decision Engine Agent which collects the MN position information and makes handoff decisions based on movement pattern identification. The positioning in this case can be easily calculated using the Node classs getLoc() support (not necessarily with the signal strength value as described in [8]). Also, all new messages associated with the S-MIP protocol were implemented in full. The support for simulcasting (SPS) was also included, consisting of the s-buffer, the f-buffer, and the forwarding techniques. The optimization techniques described in Section 3 were implemented and executed as default for the S-MIP framework. Apart from these protocol extensions, two other important modifications were necessary. Firstly, the WaveLan implementation of the current ns-2, was conceived through the Monarch project [12], and was mainly developed for the simulation of wireless ad-hoc networks (broadcast mode). What is required for our purpose is the support for the infrastructure mode. While a new implementation of the complete 802.11b standard would be ideal, we opted for a simpler emulative solution. This is made possible by a new Connection Monitor (CMon) entity. It is inserted between the port classifier (dmux_) and the receiving agent(s) (e.g. TCP/UDP) connecting to specific port(s) of the port classifier (dmux_). The CMon controls the MNs reception of packets of a communication flow, e.g. TCP. When the MN starts a L2 handoff to the new access network, CMon is set to drop any received packets until the L2 handoff is

the PAR while others to NARs, all with 2ms delay, 1M/s links. All links use the RED (Random Early Detection) queue, except links from the intermediate nodes to the ARs (both NAR and PAR), which are Droptail (FIFO) queues. The access routers are set to be 70 meters apart with free space environment in between. This reduces the complexity of results analysis, as we only need to consider signal interference. We assume the use of 802.11 as our access technology and the effective coverage area is set to 40 meters in radius. A ns TCP source (Tahoe) agent is attached to the CN and a ns TCP sink agent is attached at the MN. The TCP packet size is set to 512 bytes and window size is 32. A bulk data transfer application is attached on the established TCP link that transfers packets from the CN to the MN, 5 seconds after the start of the simulation. The total duration for the simulation is 80 seconds. We model two movement scenarios, namely the MN moving linearly between two access networks and MN moving in a back and forth (ping-pong) fashion near the midway between two access network coverage areas. For the linear case, the MN starts moving towards the NAR from the PAR 10 seconds into the simulation, at the speed of 1 meter/second (approximating human walking speed). It stops at the NAR when simulation time reaches 80 seconds. The corresponding handoff algorithm used in this scenario is the priority handoff, where the MN switches from one AR to another AR if the priority (contained in the beacon message) of the new AR is higher than that of the current one. We favor this handoff strategy over midway handoff to minimize unexpected interferences and to maintain the consistency as to when a handoff will take place, for comparison purposes. We set the beacon priority of the NAR to be higher than the PAR, hence the handoff should occur when the MN receives the first few beacon messages from the NAR at around 40 seconds into the simulation. (The priority handoff is included with the ns wireless extension module.) For the ping-pong case, we define the movement of the MN to be similar to that of the linear case, except that at 46 seconds into the simulation (just past midway) the MN reverses its direction of movement back towards the PAR. It reverses its direction every 2 seconds thereafter, until finally heading towards the NAR again, 52 seconds into the simulation. (See the stochastic scenario in Figure 1 for the illustration of this ping-pong movement.) The corresponding handoff algorithm used in this scenario is the midway handoff algorithm. Finally, in all simulations, we consider the simplistic case with only one MN initiating one single connection at a time.

Figure 5. Simulation Network Topology complete. With this, we are able to treat the receptions of control messages, (e.g. periodic beacons from ARs which arrive at the registration agent (regagent_) bypassing the CMon) as if operating in the periodic channel scanning mode. On the other hand, CMon emulates the channel change, which occurs when the MN is switching between two access networks, by blocking the agents from receiving any packets during this period. We assume that adjacent access networks use different frequency channels. (See Figure 4 for the schematic of the Mobile node.) Another advantage of using the CMon is the flexibility to identify communication flows. For instances, if operating in the route optimized mode where the MN is required to perform a binding update to the corresponding node, the MN is able to query the CMon about the identity of any corresponding node per communication flow. Secondly, we implemented an additional handoff algorithm/strategy, namely Midway handoff, to facilitate the comparison experiments. Since we know that for S-MIP, a MN will perform the handoff around the center dividing boundary between two access networks [8], the purpose of the Midway handoff algorithm is to ensure that all other frameworks handoff near the time vicinity of S-MIP. This is to provide a more accurate and consistent comparison. The Midway handoff is defined as such that a MN must handoff to an access router closer to itself, determined via the reception of the beacon message. This is especially tailored for the ping-pong movement experiments purposes.

5. EXPERIMENTAL RESULTS
The experiments were separated into two groups, namely, the linear movement experiments and the ping-pong movement experiments. We first simulated all the frameworks mentioned, namely, MIPv6 [15], MIPv6 (flat) with Fast-handover, HMIPv6, HMIPv6 with Fast-handover, Simultaneous Bindings, and S-MIP (both optimized and non optimized), with the linear movement scenario. Following this, we simulated all frameworks with the ping-pong movement scenario. We set the L2 handoff time to 20ms and address resolution time to 100ms for all experiments3.

4.2 Simulation Scenarios

Figure 5 illustrates the network topology used for the simulation experiments. This topology reflect the setup of an open space local environment (where the MN is situated) connecting to a distant home network. Both Corresponding Node (CN) and the Home Agent (HA) are connected to an (dummy) intermediate node (N1) with 2 milliseconds (ms) delay, 100 Megabits/s (M/s) links. The link between N1 and the MAP is a 100M/s link with 50ms delay. This simulates the distant home network (macro mobility). Below the MAP is considered to be the local network (micro mobility). The MAP is connected to 3 intermediate nodes (N2, N3, N4) with 2ms delay, 10M/s links. N2 is connected with

Currently, we lack statistical L2 handoff values to formulate a L2-handoff delay distribution. Therefore we are unable to

The address resolution time here refers to the time taken for a MN to obtain a new care-of address. We measured the TCP goodput, delay and the cwnd (TCP congestion window) values for all simulations.

6450

6400

f h

5.1 Hierarchical Mobile IPv6

This section describes the handoff latency comparison between HMIPv6 and MIPv6. In Figure 6, the source_send_mip and the source_send_hmip curves (the lower two curves) illustrate the handoff delay for the MIPv6 and the HMIPv6 respectively from the perspective of sources (CN) sending sequence. In the MIPv6 case, the time between the first retransmitted packet sent by the source and the last time this packet was sent is labeled with a, in Figure 6. It is approximately 747 milliseconds (ms). This must not be used as a measure for the overall handoff delay. We need to add the time required for this packet to reach the MN. At least an additional 60ms is necessary (five 2ms delay link segments, one 50ms link plus wireless link delay, see Figure 5), as we are operating in non route-optimized mode, where packets from the CN are sent to the HA first before being routed to the MN. The total delay is in fact 814ms. In this paper, we use this duration as a standard measure for all frameworks analyzed. We refer to this duration as delay D. In the HMIPv6 case, label b depicts the L2 handoff duration (happening at the MN) while label c depicts the address resolution period. Subsequently, the MN performs the MAP binding update (label d). After the successful binding, the MN receives out of sequence packets, and thus sends acknowledgement (ack) messages, containing the expected sequence number, back to the CN. This period is marked with label e, and consists of the time for ack messages to traverse from the MN to the MAP, the time from the MAP to the HA, and the time from the HA to the CN. After the CN receives three such ack messages, retransmission starts. Note that the TCP enters slow start due to duplicated ack messages received previously. The D value for this case is 326ms. Figure 7 illustrates the effect that address resolution time has on the handoff latency. The source_send_mip, source_send_hmip, and the source_send_hmip_long curves illustrate the handoff delay for MIPv6, HMIPv6 with 100ms address resolution delay and HMIPv6 with 200ms address resolution delay respectively. Labels a through e are the same as in Figure 6. Label g depicts the 200ms address resolution period, while labels f and h represent the L2 handoff and the MAP binding update. Compared with the approximately 60ms delay for the previous case (label e in Figure 6), there is an approximately 450ms delay period (label i) before the retransmission starts. This is because the delay (mainly due to address resolution) in obtaining a care-of address means that the MN is unable to receive any packets, which eventually drives the TCP sources (CN) to reach zero window and stop transmitting. When this happens, the performance is similar to that of the simple MIPv6 case, as all packets sent from the offered window are lost due to handoff. Eventually, TCP enters retransmission mode, similar to that of the MIPv6 framework.

TCP Sequence Number

6350

6300

6250
b c d e

6200
a

source_send_fast source_send_hmip source_send_mip

6150 40 40.5 41 Time (seconds) 41.5 42

Figure 6. Comparison between MIPv6, HMIPv6 and MIPv6 (flat) with Fast-Handover
6450
a

6400
f g h i

TCP Sequence Number

6350

6300

6250
b c d e

6200

source_send_hmip source_send_hmip_long source_send_mip

6150 40 40.5 41 Time (seconds) 41.5 42

Figure 7. Comparison between MIPv6, HMIPv6 and HMIPv6 with longer address resolution (200ms)

5.2 Mobile IPv6 (Flat) with Fast-Handover

We compare MIPv6 and MIPv6 with Fast-handover in this section. Figure 6 illustrates the handoff delay for MIPv6 with Fast-handover with the source_send_fast curve, from the perspective of the sources sending sequence. Label f shows the time taken to set up the fast-handover, that is, measured from the time the MN sends the RtSolPr message till the time it receives the PrRtAdv message, including the intermediate message exchange between the PAR and the NAR (HI and HAck messages). As can be observed in Figure 6, at around the 40.5s mark, the source_send_fast sends a little longer than the source_send_mip before being interrupted by the handoff. This is because the setting up of fast-handover does not prevent TCP from continuing packet transfer. A TCP flow only starts to be disrupted once entering the binding process. The binding update (with HA) duration is depicted using label g, while the L2 handoff is through label h. We switch network (L2 handoff) immediately after sending the binding updates, for all schemes employing the Fast-handover mechanism, in consideration of consistency and easy comparison purposes. Similar to the MIPv6 case (label e),

provide a more realistic simulation except to fix the L2 handoff time. Same applies to address resolution time.

a)

6250

6250

6200

6200

TCP Sequence Number

B A
b d f e c

TCP Sequence Number

6150

6150

6100
a

6100

6050

6050

6000 sink_recv (data) sink_send (ack) 5950 40 40.5 41 Time (seconds) 41.5 42

6000

source_send source_recv sink_recv sink_send

5950 40.2

40.4

40.6 Time (seconds)

40.8

41

41.2

b)
6250

Figure 9. Handoff Behavior for S-MIP in detail

6200

6150
TCP Sequence Number
i

Secondly, the MN starts receiving out of sequence packets after switching to the new access network (box B).
g j h k

6100

6050
cwnd value

30 25 20 15 10 5 0 40.6 40.8 41 41.2 41.4 41.6 41.8 42 41 Time (seconds) 41.5 42

m

6000

5950

source_send (data) source_recv (ack)

5900 40 40.5

Figure 8. Handoff Behavior for HMIPv6 with Fasthandover in detail label i depicts the delay of the ack message traversing from the MN to the CN. The binding update delay is proportional to the round-trip delay between the CN and the MN (traversing via the HA in our case). This is the weakness of operating Fast-handover in the flat Mobile IP environment. Even though that the forwarding mechanism is delivering packets from the PAR to the NAR, the MN is unable to receive these until the binding update is complete, signified by the sending of the F-NA message to activate the data flow at the NAR. The D value for this case is 358 ms.

Figure 8a shows the details of handoff from the TCP receivers (sink) perspective. The receivers receiving sequence is denoted by the sink_recv (data) curve while the sending sequence is denoted by the sink_send (ack) curve. Likewise, Figure 8b shows the details of handoff from the TCP senders (source) perspective. The senders sending sequence and the receiving sequence are depicted by the source_send (data) and the source_recv (ack) curves respectively. The label a in Figure 8a depicts the time vicinity where the first instance of disruption in the TCP communication occurs, namely the L2 handoff. Some packets are lost as a result. Due to the Fast-handover forwarding setup between the PAR and the NAR, the MN is able to receive the forwarded packets (through the use of the F-NA message) at the new access network from the NAR shortly after arriving, as shown near label b. However, these packets contain out of order sequence numbers (higher than expected). As a result, an ack reply, containing the expected sequence number, is sent for each of the (out of sequence) data packet received (label c). This is reflected in the senders receiving sequence (label g). These duplicated acks trigger the TCP sender to enter the congestion mode and slow start occurs (around label h). The TCP slow start process continues until around label d. By this point in time, the missing packets have all been received by the MN. Due to accumulative acknowledgement, the MNs window moves a whole segment and opens abruptly. This is reflected in the TCP sender near label i, as well as the corresponding cwnd curve shown at around 41.2 second (label m). However, packets sent previously by the sender (label j) still arrive at the receiver (label e). These are repeatedly acknowledged with the same expected sequence number (label f) by the receiver. Eventually these ack messages arrive at the sender, but are discarded since the sequence numbers have already been acknowledged. The communication returns to normal operation after this. The D value for this case is 270 ms.

5.3 Hierarchical Mobile IPv6 with Fasthandover

This section details the handoff behavior for the HMIPv6 with Fast-handover scheme. In prior work, Hsieh [7] illustrated the performance of a simple superimposition of HMIPv6 scheme with the Fast-handover protocol. It showed that superimposition results in a complex behavior. In contrast, we present the integrated approach, outlined in the Appendix section of [17]. The main difference being that the forwarding mechanism (part of the Fasthandover) is anchored at the MAP, instead of at the PAR. The integrated approach is an optimized solution in comparison to that of the simple superimposition. Essentially, two notable events occur as denoted with box A and box B in Figure 8a. Firstly, packet loss happens when L2 handoff is initiated (box A).

5.4 Simultaneous Bindings and S-MIP

We compare the Simultaneous Bindings scheme and the S-MIP framework (optimized) in this section. A hypothetical result is also shown for comparison purposes, namely the smip_nonop

a)
6300

Source Sending Sequences

b)
6300

Sink Receiving Sequences

6250

6250

6200

6200

TCP Sequence Number

TCP Sequence Number

6150

6150

6100

6100

6050

6050

6000

source_send_smip_nonop source_send_bicast

6000

sink_recv_smip_nonop sink_recv_bicast

5950 40 40.5 41 Time (seconds) 41.5 42

5950 40 40.5 41 Time (seconds) 41.5 42

c)
6300

Source Receiving Sequences

d)
6300

Sink Sending Sequences

6250

6250

6200 6200

TCP Sequence Number

6150
c

TCP Sequence Number

6150
b

6100

6100

6050

6050 6000 6000 sink_send_smip_nonop sink_send_bicast

5950

source_recv_smip_nonop source_recv_bicast

5900 40 40.5 41 Time (seconds) 41.5 42

5950 40 40.5 41 Time (seconds) 41.5 42

Figure 10. Handoff Behavior for Simultaneous Bindings and S-MIP (non optimized) in detail (non optimized) case, where we deliberately created the situation where out of sequence (duplicated) packets are present at the NAR. As can be seen from the source_send, source_recv, sink_recv and sink_send curves in Figure 9, the S-MIP (optimized) framework performs exceedingly well, considering that the overall handoff perceived by the sender is a mere 100ms in duration approximately (source_send curve). There is no apparent packet loss perceived by the sender and therefore no congestion or flow control is required. The D value cannot be calculated since there are no packets being retransmitted. The emptying of the s-buffer and the f-buffer corresponds to ack message responses near label b (sink_send curve) in Figure 9. In other words, these acknowledgements, for data packets sent by the source (around label a), are simply delayed slightly in reply, due to the handoff. Subsequently, everything returns to normal as if no handoff had taken place. In contrast, the worst case scenario for the non optimized S-MIP case is shown as the nonop (S-MIP non optimized) curves in Figure 10a-d, where packet out of sequence takes place at the NAR in the framework. The source_send_smip_nonop, sink_recv_smip_nonop, source_recv_smip_nonop, and sink_send_smip_nonop, curves illustrate the hypothetical case for the S-MIP non optimized scheme, in Figure 10a-d respectively. A few things can be noted in this scenario. Due to the coarse grain buffering mechanism of S-MIP, this is a case where no packets are lost during the handoff. However, we deliberately induce duplicated packets within the fbuffer and the s-buffer of the NAR, to cause out of sequence behavior. This illustrates the effect of receiving out of sequence packets by the MN and how it may upset the TCP flow within the S-MIP framework. Nevertheless, compared to packet loss, the penalty is not as severe in our experimental context, because of the accumulative acknowledgement mechanism. The two key handoff behaviors are, firstly, the receiving of the duplicated packets by the MN (label a in Figure 10b). Secondly, these duplicated packets cause the MN to send duplicated ack messages (label b). Fortunately, these have already been acknowledged previously, therefore are discarded by the CN (label c). The handoff behavior for Simultaneous Bindings is similar to that of the HMIPv6 with Fast-handover case, as can be seen from the source_send_bicast (Figure 10a), sink_recv_bicast (Figure 10b), source_recv_bicast (Figure 10c), and the sink_send_bicast (Figure 10d) curves, when compared with Figure 8. This is because, the benefit of bicasting/n-casting have been eroded away, as the MN is traversing linearly and switching networks immediately after sending the binding update to the MAP in our simulation. For the Simultaneous Bindings case, as the MN is switching immediately, the only subtle difference between it and

8000 7900 7800

TCP Sequence Number

9000 8800 8600 8400

TCP Sequence Number

7700 7600 7500 7400 7300 7200 7100 45 46 47 48 49 Time (seconds) 50 51 52 53 source_send_fast_pp source_send_miph_pp source_send_mip_pp

8200 8000 7800 7600 7400 7200 7000 45 46 47 48 49 Time (seconds) 50 51 52 53 source_send_hmipfast_pp source_send_bicast_pp source_send_smip_pp source_send_smip_nonop_pp

Figure 11. Ping-pong behavior for MIPv6, HMIPv6 and MIPv6 (flat) with Fast-handover the HMIPv6 with Fast-handover case, is that some packets inside the link segment between the MAP and the PAR might be lost as a result of the network switch. However, since the packet forwarding delay between the PAR and the NAR is much smaller than the time taken to complete a L2 handoff, even if these packets are forwarded from the PAR to the NAR (as in the case for HMIPv6 with Fast-handover), it would nevertheless be dropped by the NAR, since the MN is not likely to have completed the L2 handoff. Even if the MN is to switch slightly later, in order to eliminate segment packet loss [8], some packet losses will still take place due to the L2 handoff, similar to that of the HMIPv6 with Fasthandover case. Although Simultaneous Bindings might cover-up the timing ambiguity, we argue that the performance improvement is minimal, as packet loss is a substantial penalty. In comparison, as shown in Figure 10a-d with the nonop (S-MIP non optimized) curves, the severity of packet loss is much greater than that of packets out of sequence. For this reason, the performance for Simultaneous Binding and HMIPv6 with Fast-handover would be similar. The D value is 268ms for the Simultaneous Bindings framework.

Figure 12. Ping-pong behavior for HMIPv6 with Fasthandover, Simultaneous Bindings and S-MIP Fast-handover and the Simultaneous Bindings schemes are affected to a lesser extent. However, severe throttling (i.e. the decrease in gradient) is still notable for both cases. Remarkably, the S-MIP case illustrates excellent resilience to ping-pong movement. The communication remains virtually unaffected. Essentially, the ping-pong movement can be perceived as multiple individual linear handoffs. The ping-pong movement defined in the experiment is too short for the MIPv6 scheme to complete a single linear handoff therefore the scheme performs poorly. Except for S-MIP, if we shorten the period in-between the switching back and forward (towards the D value), eventually all schemes will break down (smaller than D value), similar to that of the MIPv6 case. The S-MIP guards against such break down as another handoff requested by the MN will not be allowed by the DE until the current one is completed, as specified in [8].

6. DISCUSSION 6.1 Performance Comparison

Table 1 illustrates the relative goodput for all frameworks in comparison to that of a scenario where the MN is stationary near the PAR, i.e. no handoff takes place. We measure the time taken for each framework to transfer a 6.5 megabytes (M) file, with the

5.5 Ping-pong Movement Results

In this section, we examine the effect of ping-pong movement for all frameworks which we have examined thus far. We have defined in the previous section how the ping-pong scenario is simulated (see Figure 1 and Section 4.2). Figure 11 depicts the handoff behavior of ping-pong movement for the MIPv6, MIPv6 (flat) with Fast-handover and HMIPv6 frameworks with curves source_send_mip_pp, source_send_fast_pp and source_send_hmip_pp respectively. Similarly, Figure 12 illustrates the handoff behavior for HMIPv6 with Fast-handover, Simultaneous Bindings, S-MIP and S-MIP non optimized schemes with source_send_hmipfast_pp, source_send_bicast_pp, source_send_smip_pp and source_send_smip_nonop_pp curves respectively. (Note that we are only presenting from the perspective of sources sending sequence here.) As can be seen in Figure 11, the MIPv6 completely breaks down during the pingpong movement. The MIPv6 with Fast-handover and the HMIPv6 schemes are also notably disrupted with distinguishable breaks in the communication flow. Figure 12 shows that the HMIPv6 with

Table 1. Performance Matrix Avg. Goodput (Kbytes/second) Frameworks

MIPv6 MIPv6 + FastH HMIPv6 HMIPv6 + FastH Bicast/n-cast SMIP (nonop) SMIP No Handoff Linear 100.847 101.213 101.520 101.593 101.580 102.007 103.106 Pingpong 83.820 90.240 91.587 93.867 92.713 91.113 102.120

Time to transfer 6.5M file (seconds)

Linear 66.001 65.762 65.563 65.516 65.524 65.127 64.554 Pingpong 79.408 73.759 72.674 70.909 71.051 68.538 65.178 64.438

103.293

180 160 140

cwnd cwnd

180 160 140 120 100

TCP Sequence Number

6450 6400 6350 6300

SMIP PP 30 35 40 45 50 55 60 65 70 75 80 Time (seconds)

120 100 80 60 40 30 35 40 45 50 55 60 65 70 75 80 Time (seconds) 140 120 100 SMIP

80 60 40

6250 6200 6150 6100 6050 6000 source_send_mip source_send_hmip source_send_fast source_send_hmipfast source_send_smip_nonop source_send_bicast source_send_smip 40 40.5 41 Time (seconds) 41.5 42

120 100 80
cwnd

cwnd

80 60 40 20 0 30 35 40 45 50 55 60 65 70 75 80 Time (seconds) Bicast

60 40 20 0 30 35 40 45 50 55 60 65 70 75 80 Time (seconds) Bicast PP

5950

Figure 13. A comparison on cwnd value over time for SMIP and Simultaneous Bindings linear and the ping-pong movement scenarios. As shown in Table 1, with the perfect condition (no handoff), it takes 64.438s to transfer 6.5M file from the CN to the MN with the TCP goodput of 103.293 kilobytes/second (K/s). The S-MIP framework compares well with the no handoff case, achieving a goodput of 103.106 K/s for linear movement and requires slightly longer transfer time of 64.554s. What makes S-MIP stand out is its performance during the ping-pong movement scenario, where the goodput is still maintained at over 100K/s, at 102.120 K/s, and the transfer time at 65.178s. The closest goodput is at 93.867 K/s in comparison, showing a significant reduction in performance. The worst performer is naturally the MIPv6 scheme, taking almost the full 80s to transfer the data in the ping-pong scenario, while achieving a 66.001s transfer time (still the worst) for the linear case. The spread in time for the linear scenario between the best and the worst framework is 1.447 seconds. However, this variation is amplified for the ping-pong scenario to 14.23 seconds, an increase of almost 10 folds. In conjunction to Table 1, we illustrate the TCP congestion window (cwnd value) plotted against time for the S-MIP framework in Figure 13. As can be observed, for both the linear (handoff at around t = 40s) and pingpong movement scenarios (handoff between t = 45s and t = 55s), the cwnd values for S-MIP (SMIP/SMIP PP) do not drop significantly unlike the case for the Simultaneous Bindings scheme (Bicast/Bicast PP), which typifies the cwnd behavior for all other frameworks. Figure 14 illustrates the handoff delay (linear) for all frameworks mentioned in this paper, from the perspective of the TCP sources sending sequence. We transpose curves for HMIPv6 with Fasthandover, Simultaneous Bindings, SMIP non optimized and SMIP, in order to view clearly the relationship among themselves, as well as between them and those which possess much longer delay, i.e. MIP, HMIPv6 and MIPv6 with Fast-handover schemes. With our experimental setup, the order of the most effective latency reduction scheme to the least effective is: SMIP, SMIP non optimized (if considered), Simultaneous Bindings (D = 268ms), HMIPv6 with Fast-handover (D = 270ms), HMIPv6 (D = 326ms), MIPv6 (flat) with Fast-handover (D = 358ms), and lastly MIPv6 (D = 814ms). We rank Simultaneous Bindings higher than that of the HMIPv6 with Fast-handover because we believe it can

Figure 14. A comparison between all Frameworks potentially offer better results, even though in our simulation scenario, the two are comparable. The performance of HMIPv6 and MIPv6 (flat) with Fast-handover are dependent on the experimental topology layout. Therefore the ordering between the two is only applicable to this series of simulation experiments. It is well understood that, if applying the normal Mobile IP handoff algorithm, the overall handoff latency is at least 2 seconds, as the mobile node is required to wait for 2 beacon misses before initiating a handoff. In this paper, we look at an optimization of this, where the MN initiates a handoff when it receives an unseen beacon from a new access router4. With this, the handoff latency values ranges from approximately 800ms (for the MIP(v6) case) to approximately 250ms (for the Simultaneous Bindings case), a significant reduction from 2 seconds. More impressively, with the S-MIP scheme, although the actual handoff delay is around 100ms, the latency perceived by the sender and the receiver is zero, i.e. as if no handoff has ever occurred.

6.2 Summary
The performance of the L3 handoff latency reduction for the HMIPv6 framework is bounded by the address resolution time (A), the one-way trip time (T) from the CN to the MN, and the size of the sliding window (W). If time A is greater than the time for the sender to send W packets to a receiver located T seconds away, then the performance degrades to that of the MIPv6 case. For the (Flat) MIPv6 with Fast-handover case, its performance is affected predominantly by the round-trip time between the CN and the MN (the forwarding delay from the PAR to the NAR is very small in comparison). This time determines the number of packet errors (loss, if the NAR not buffering, duplicated ack, if the NAR is buffering) that occur at the MN. A recovery process, i.e. slow start, is almost inevitable, unless no more than 3 errors are detected. For HMIPv6 with Fast-handover and the Simultaneous Bindings frameworks, the key deficiency is the packet loss due to the L2 handoff, as these two frameworks do not explicitly buffer packets

This is only suitable for linear movement or very fast MN movement, i.e. the MN is within a moving vehicle.

during a handoff, triggering congestion control mechanisms. In comparison with this, the subsequent packets out of sequence have a lesser impact on the handoff latency. Even though this should have also triggered the congestion control, due to the setup of the forwarding scheme and the accumulative acknowledgement at the sender, the activation of the congestion control mechanism is avoided. Furthermore, initiating a handoff or packet forwarding based on a pre-determined or on-the-fly anticipation timing interval is undesirable. It is most likely to be sub-optimal, unless the anticipation takes into consideration not just the time MN is likely to switch network, but also the movement direction of the MN. S-MIP is a novel scheme that addresses all of the short comings described prior. It buffers packets explicitly at the NAR and provides an explicit anticipation method through network determined handoff, calculated passively using movement tracking and the MN position identification (via triangular signal strength evaluation). Although in [8], Hsieh illustrated a primitive MN position identification scheme for open space environment, S-MIP is structured in a way where more sophisticated positioning scheme maybe developed and plugged-in to the framework, allowing it to operate in different physical environments. This is possible since the DE messaging (the trigger) mechanism is access network independent. We identified that with the original S-MIP framework, packets out of sequence are still likely to occur if the MN is allow to receive packets from the PAR (i.e. not switching network immediately), after sending the binding update to the MAP. In this case, the emptying of the f-buffer (containing packets which the MN have already received previously at the PAR) will cause the MN to send duplicated ack messages to the CN, resulting in source throttling and performance degradation. Depending on the flavor of the TCP, this error is not as severe as packet loss due to the accumulative acknowledgement mechanism. We presented an optimization for S-MIP which prevents the occurrence of such situation. We argue that by immediately switching network, after sending the binding update, the possibility of such performance degradation can be reduced.

mobile nodes. How a mobile node should switch network is critical when considering stochastic/ping-pong movements. Lack of proper management leads to unnecessarily high setup cost, and may results in severe disruption if the switching frequency is shorter than the minimal time required for the L3 handoff. In future work we plan to study S-MIP under multiple connection scenarios where connections interfere with each other, and examine its performance under such condition, in both linear and ping-pong movement cases. We will examine the scalability of the Decision Engine and the impact of multiple node and multiple connections on the S-MIP architectural design. Furthermore, we plan to design more sophisticated positioning schemes which maybe used to enable the S-MIP framework to operate under more complicated physical environments.

8. ACKNOWLEDGMENTS
The authors would like to thank the anonymous reviewers and Mobqos Group members. Special thanks to Yuri Ismailov, Stephen Herborn, Tim Hu and Eranga Perera.

9. REFERENCES
[1] [2] Bakre, A., and Badrinath, B.R. I-TCP: Indirect TCP for Mobile Hosts. In Proceedings of ICDCS, May 1995. Balakrishnan, H., Seshan, S., and Katz, R. H. Improving Reliable Transport and Handoff Performance in Cellular Wireless Networks. ACM Wireless Networks, Vol. 1, No. 4, December 1995. Blumenthal, M. S., and Clark, D. D. Rethinking the design of the Internet: The end to end argument vs. the brave new world. ACM Transactions on Internet Technology, Vol. 1, No. 1, August 2001. Caceres, R., and Iftode, L. Improving the Performance of Reliable Transport Protocols in Mobile Computing Environments. IEEE Journal on Selected Areas in Communications, Vol. 13, No. 5, June 1995. Caceres, R., and Padmanabhan, V. N. Fast and Scalable Handoffs in Wireless Internetworks. In Proceedings of MobiCom, 1996. Campbell, A. T. et al. Comparison of IP Micro-Mobility Protocols. IEEE Wireless Communications Magazine, Vol. 9, No. 1, February 2002. Hsieh, R., and Seneviratne, A. Performance analysis on Hierarchical Mobile IPv6 with Fast-handoff over End-toEnd TCP. In Proceedings of GLOBECOM, Taipei, Taiwan 2002. Hsieh, R., Zhou, Z.-G., and Seneviratne, A. S-MIP: A Seamless Handoff Architecture for Mobile IP. In Proceedings of INFOCOM, San Francisco, USA, 2003. Koodli, R. (eds.). Fast Handovers for Mobile IPv6. Internet Draft, IETF, September 2002. Work in Progress. Malki, K. (eds.). Low Latency Handoffs in Mobile IPv4. Internet Draft, IETF, June 2002. Work in Progress. Malki, K., and Soliman, H. Simultaneous Bindings for Mobile IPv6 Fast Handoffs. Internet Draft, IETF, June 2002. Work in Progress. Monarch project. http://www.monarch.cs.rice.edu. Perkins, C. IP Encapsulation within IP. RFC 2003, IETF, October 1996.

[3]

[4]

[5]

[6]

7. CONCLUSION
Analyses of the various handoff latency reduction frameworks examined in this paper showed that it is possible to significantly reduce the latency perceived by a mobile node during a handoff. This is especially true for the S-MIP case, achieving packet lossless handovers, where the mobile node and the correspondent node are unaware of the handoff which is taking place. We further optimize the S-MIP scheme by eliminating deficiencies associated with possible packets out of sequence during its local forwarding management (SPS) phase. Not knowing when a mobile node will initiate a handoff and how it is likely to move after a handoff will inevitably result in packet losses. With packet forwarding (as the case for Fast-handovers) or packet n-casting (Simultaneous Bindings), these losses may be reduced. However, this is at the expense of duplicated packets and/or delay in receiving packets. The S-MIP framework specifies, with reasonable accuracy, when a mobile node should switch and how it should switch proactively, through passive movement tracking. It also specifies different techniques in optimizing a handoff by categorizing movement patterns of [7]

[8]

[9] [10] [11]

[12] [13]

[14] Perkins, C. IP Mobility support for IPv4. RFC 3344, IETF, August 2002. [15] Perkins, C., Johnson, D., and Arkko, J. Mobility Support in IPv6. Internet Draft, IETF, January 2003. Work in Progress. [16] Ramjee, R. et al. HAWAII: A Domain-Based Approach for Supporting Mobility in Wide-area Wireless Networks. In Proceedings of ICNP, 1999. [17] Soliman, H., Castelluccia, C., Malki, K., and Bellier, L. Hierarchical MIPv6 Mobility Management. Internet Draft, IETF, October, 2002. Work in Progress.

[18] The UCB/LBNL/VINT Network Simulator ns (version 2). http://www.isi.edu/nsnam. [19] Thomson, S., and Narten, T. IPv6 stateless address autoconfiguration. RFC 2462, IETF, December 1998. [20] Valko, A. et al. Design and Performance of Cellular Access Networks. IEEE Personal Communications, Special issue on IP-Based Mobile Telecommunications Networks, 2000. [21] Widmer, J. Extensions to the ns Network Simulator. http://www.informatik.uni-mannheim.de/~widmer.