IEEE/ACM TRANSACTIONS ON NETWORKING, OCTOBER ...IEEE/ACM TRANSACTIONS ON NETWORKING, OCTOBER 2008 2...

IEEE/ACM TRANSACTIONS ON NETWORKING, OCTOBER 2008 1

A Comprehensive Analysis of Mobility Managementin MPLS-Based Wireless Access Networks

Rami Langar, Nizar Bouabdallah and Raouf Boutaba

Abstract—Efficient mobility management is one of the majorchallenges for next-generation mobile systems. Indeed, a mobilenode (MN) within an access network may cause excessivesignaling traffic and service disruption due to frequent handoffs.The two latter effects need to be minimized to support quality ofservice (QoS) requirements of emerging multimedia applications.In this perspective, we propose in this paper a new mobilitymanagement scheme designed to track host mobility efficiently soas to minimize both handoff latency and signaling cost. Buildingon and enhancing Mobile IP and taking advantage of MPLStraffic engineering capability, three mechanisms (FH-, FC- andMFC-Micro Mobile MPLS) are introduced. In order to assessthe efficiency of our proposals, all protocols are compared.To achieve this, we develop analytical models to evaluate thesignaling cost and link usage for both two-dimensional and one-dimensional mobility models. Additional mathematical modelsare also provided to derive handoff latency and packet lossrate. Numerical and simulation results show that the proposedmechanisms can significantly reduce the registration updates costand provide low handoff latency and packet loss rate undervarious scenarios.

Index Terms—Mobile IP, MPLS, mobility models, micro-mobility, fast handoff, forwarding chain, residing area, perfor-mance analysis.

I. INTRODUCTION

FUTURE wireless networks are expected to provide IP-based coverage and efficient mobility support with end-

to-end QoS guarantees. Two enabling factors are considered ascrucial: (i) maintaining the network connectivity during nodemobility and; (ii) provisioning the network resources requiredby the Mobile Node (MN) in all the visited subnetworks.

Mobility management protocols are key for service continu-ity in mobile networks. Mobile IP [1], the Internet EngineeringTask Force (IETF) standard, can serve as the basic mobilitymanagement in IP-based wireless networks. According toMobile IP, a MN can change its point of attachment withoutchanging its IP address. To do so, a MN is assigned with apermanent home address in its home network, and will borrowa temporary care-of address (CoA) in each foreign network.The CoA is the foreign agent (FA) IP address of the currentlyvisited foreign network. In this case, the home agent (HA),residing in the MN’s home network, will maintain the mappingbetween the home address and the CoA.

Specifically, packets sent to a MN are first intercepted byits HA, and then tunneled to the current serving FA usingthe MN’s CoA. The FA then decapsulates the packets andforwards them to the MN. In turn, packets transmitted by aMN are directly routed toward their destinations, without theneed to pass through the sending MN’s HA. Clearly, suchrouting method may increase considerably the packet delivery

Manuscript received January 30, 2006; revised February 25, 2007; approvedby IEEE/ACM TRANSACTIONS ON NETWORKING Editor S. Das.

R. Langar and R. Boutaba are with the Computer Science Department,University of Waterloo, ON, N2L3G1 Canada (phone: 519-888-4820; fax:519-885-1208; e-mail: [email protected] ; [email protected]).

N. Bouabdallah is with INRIA/IRISA, 35042 Rennes, France (e-mail:[email protected]).

cost, notably if the receiving MN node is far away from itsHA. This issue is well-known as the triangular routing problem[2]. In addition, Mobile IP induces long handoff latency andlarge signaling load when handoffs occur frequently [2]. Inthis regard, many enhancements to Mobile IP for MNs withfrequent handoffs have been proposed in the literature [3]–[15] to ensure service continuity.

The second major enabling factors identified as crucialfor the evolution process of next-generation mobile networksis efficient network resources provisioning. This issue hasbeen largely studied in both wired and wireless environments.For instance, MultiProtocol Label Switching (MPLS [16])addresses today’s network backbone requirements effectivelyby providing a standardized solution that improves packetforwarding performance and designs high performance QoSguaranteed paths. MPLS strength lies in its ability to accom-modate different application requirements and to engineer ef-ficiently traffic tunnels, avoiding thus congestion and enablingan efficient use of the available bandwidth. As a matter of fact,there is an increasing trend towards the introduction of MPLSin wireless environments [17]– [22].

To meet the requirement of next generation mobile network-s, we propose in this paper a new protocol called Micro MobileMPLS which alleviates the limitations of Mobile IP and inthe same time benefits from MPLS resource provisioningcapability. We propose three mechanisms that enhance bothmobility and QoS support capabilities for next-generationcellular networks. In the first one called FH-Micro MobileMPLS, we consider the fast handoff (FH) mechanism, whichanticipates the LSP procedure setup with an adjacent subnetthat an MN is likely to visit. This mechanism is proposedto reduce service disruption by using the L2 functionali-ties. In the second one called FC-Micro Mobile MPLS, asimple forwarding chain (FC) concept with loop removalis introduced to track efficiently the host mobility within adomain. The forwarding chain can reduce registration updatescost (i.e., volume of signaling messages exchanged duringhandoff operations) and provide low handoff latency and QoSsupport thanks to MPLS capabilities. FC-Micro Mobile MPLSis appropriate for wireless environments with high mobilityrate, where packets must be quickly redirected to their newlocations. In the third mechanism called M(Master)FC-MicroMobile MPLS, we define a new attribute, called residing area,for each MN to achieve further reduction of the signalingcosts. This mechanism can be seen as an extension of theprevious one in a two-dimensional mobility model. To gaugethe effectiveness of our proposed mechanisms, we compare theregistration updates cost, the link usage cost and the handoffperformance metrics (i.e., handoff latency and packet loss rate)using both two-dimensional (2-D) and one-dimensional (1-D)mobility models. Simulation results show that our proposalscan improve significantly the network performances whencompared to existing schemes (FMIP [10], MIP-RR [12],Mobile MPLS [17], H-MPLS [19]) under various scenarios.

The remainder of this paper is organized as follows. In sec-


tion II, we discuss related works and position our own. SectionIII presents our proposed mobility mechanisms. Section IVdescribes the system model used to evaluate the performanceof our proposals. In section V, we develop analytical modelsto derive the signaling cost function of registration updates,the link usage and the handoff performance for all underlyingprotocols using the 2-D and 1-D mobility models. Numericaland simulation results are given in section VI. Finally, sectionVII concludes this paper.

II. RELATED WORK

As explained before, the standard network layer solution,Mobile IP [1], has several shortcomings, such as: triangularrouting, high handoff latency, high global signaling load andscalability issues. These issues are more pronounced in micro-mobility (intra-domain mobility) environment, where handoffoperations are much more frequent than in the macro-mobility(inter-domain mobility) case. We note that the Mobile IPprotocol is, in essence, designed to deal with macro-mobilitymanagement rather than micro-mobility management. To copewith Mobile IP limitations, several interesting solutions havebeen proposed in the literature [8]– [22]. In the following wediscuss some of the most significant ones.

To avoid the triangular routing problem, [8] proposes theMobile IP route optimization scheme, which binds directly thesending node to its destination, avoiding thus to pass throughthe receiving MN’s HA. To achieve this, each sending nodeneeds to maintain a binding cache that contains the CoAs of allthe potential receiving MNs, which may become challengingespecially in dense networks.

To alleviate the high handoff latency, a typical issue in Mo-bile IP, a fast handoff scheme for Mobile IP is proposed in [9][10] using both link layer triggers and buffering mechanisms.

Specifically, authors in [9] propose to reduce the delay andthe data loss during handoff in two ways: either by using thepre-registration or the post-registration handoff methods. In theformer case, layer 3 (L3) handoff is initiated prior to layer 2(L2) handoff occurrence thanks to L2 triggers. The L2 triggeris a signal from L2 to inform L3 of an imminent L2 handoff.Accordingly, the MN initiates the Mobile IP registration toits HA through the old FA before the L2 handoff occurrence.In the latter case (i.e., post-registration), the L3 handoff isinitiated after the L2 handoff occurrence. Specifically, the L2trigger is first used to setup a bi-directional tunnel between theold and the new FAs. Such a tunnel remains active until theMN completes the registration updates with its HA through thenew FA. This however requires some modifications to MobileIP. [9] proposes also a systematic way to choose between thetwo registration methods. The old FA attempts first the pre-registration method before the L2 handoff occurrence, and ifit fails, it attempts the post-registration method.

A second alternative to introduce a fast handoff scheme forMobile IP is proposed in [10]. This scheme, called FMIP,enables a MN to quickly discover that it has moved to anew subnet (by using a router discovery protocol or somelink-specific event) and receive data as soon as its attachmentis detected by the new access router. This approach can beseen as a post-registration handoff method. It has a significanteffect on the performance of real-time and QoS sensitiveapplications. However, the use of a discovery protocol impliesmodifications to Mobile IP in order to support the discoveryof neighboring FAs. As such, extra information exchangesbetween access routers are needed. In addition, the registration

updates cost in FMIP can be excessive, especially for highlymobile nodes and those located far away from their HAs.

In order to tackle the inherent problem of Mobile IPregarding the high signaling cost, authors in [11] proposea distributed dynamic location management scheme. Theyassume that every Foreign Agent (FA) has the functionalityof a FA and Gateway Foreign Agent (GFA). The key ideais to introduce a new level of hierarchy to the basic MobileIP architecture and to adjust the regional network size basedon the user’s current traffic load and mobility information.This scheme can be seen as an extension of the IETF regionalregistration protocol (MIP-RR [12]) in order to improve itsflexibility and adaptability. However, the main difficulty whenrunning these schemes [11] [12] is the computation of theoptimal size of the regional network. Moreover, these schemesdo not support QoS provisioning, a vital requirement for nextgeneration mobile networks.

Another approach to reduce the signaling cost is the “pointerforwarding” technique used in [13] [14] and [15]. Accordingly,pointers are setup between the old and new subnets by the MN,reducing hence the local registration cost. However to achievethis, authors in [13] assume that there is a correlation betweenthe communication cost and the geographic distances. As such,the coordinates of each FA along the formed forwarding chainneed to be known by the MN.

To achieve the pointer forwarding, authors in [14] and [15]use a caching method and a cascaded tunneling scheme toconnect the MN to its HA. However, the authors limitedtheir studies to the simple one dimensional random walkmodel when analyzing their protocols. Indeed, the use ofsuch schemes in a two-dimensional case will result in theestablishment of non-optimal forwarding chains with loopproblems leading thus to unacceptable end-to-end delay. Inaddition, the QoS support has not been addressed. To by-passsome of the above-mentioned problems, [15] introduced anadditional mechanism to remove loops from the MN’s trajec-tory. However, this technique turns out to be high consumingin terms of memory and bandwidth. The scalability of thistechnique becomes questionable in dense mobile networkssince the system needs to maintain a list of all visited subnetsof all the MNs located in the domain.

So far, the discussed works have focused mainly on improv-ing the Mobile IP performance by proposing new extensions.We believe that the role of an efficient mobility managementprotocol must not be limited to the basic operations regardingthe MN’s connectivity maintenance during its travel. Instead,mobility management must also be able to provision efficientlythe network resources. In this perspective, several works [17]–[22] propose the use of MPLS in IP-based wireless accessnetworks to benefit from its QoS, traffic engineering andreliability (e.g., path protection and restoration mechanisms)capabilities.

Specifically, [17] proposes a scheme to integrate Mobile IPand MPLS protocols. This scheme, called Mobile MPLS, aimsat improving the scalability of the Mobile IP data forwardingprocess by removing the need for IP-in-IP tunneling fromthe Home Agent to the Foreign Agent using Label SwitchedPaths (LSPs). However, such a scheme suffers from the non-applicability to micro-mobility, as the scope of Mobile IP ismore tailored to global mobility (i.e., macro-mobility).

In [18], a signaling framework for intra-domain mobilityusing LSP redirection in a traffic engineered network hasbeen proposed. An enhanced LER called the label edgemobility agent (LEMA) is introduced to support chained LSP-


redirection. The MPLS domain is augmented to support micro-mobility by adding the LEMA functionality to a subset ofexisting MPLS nodes. The scheme has been shown to bescalable and suitable for QoS support [18]. However, thealgorithms for choosing the LEMAs for a particular MNappear to be complex, which affects the reliability and thecost of the proposed scheme.

Authors in [19], [20], [21] and [22] attempt to improvethe performance of Mobile MPLS [17] using different ar-chitectures. Commonly, a Foreign Domain Agent (FDA) isintroduced into each MPLS domain to support intra-domainmobility. However, these works have not taken into accountthe fact that the signaling delay for registration updates couldbe very long, which may cause service disruption for real-time services and will result in increasing the registrationupdates cost, the loss of a large amount of in-flight packets andQoS degradation. Note that in-flight packets are the packetspossibly lost during the handoff period. Indeed, with a highmobility rate, the system performance is critically affectedby frequent registrations and LSPs setup procedure with theFDA, resulting in excessive signaling traffic and long servicedelay. Moreover, most of these works have assumed that allbase stations are MPLS capable which may not be alwaysdesirable as this implies a significant increase in terms of costand complexity.

In this paper, we propose a new micro-mobility manage-ment scheme for MPLS-based wireless access networks thatcombines the benefits of previous works on mobility man-agement with MPLS capabilities. Our scheme integrates threemobility mechanisms. The first one, called FH, is a hybridsolution combining the pre-registration and post-registrationmethods with QoS provisioning and aims at achieving furtherhandoff optimization (reduction of registration updates cost,handoff delay and packet loss rate) compared to [9] and [10].The second mechanism, called FC, is a new implementationalternative of the pointer forwarding concept in an MPLSenvironment. It overcomes the limitations of [13] and [14]by avoiding the geographical constraints and loop issues,respectively. The FC protocol provides therefore a loop-awaremobility mechanism with QoS support for highly mobile users.It also resolves the scalability issue of [15] since only the MNneeds to keep a cache of all its visited subnets (i.e., a cachecontaining the visited FAs’ IP addresses). As such, there isno need to communicate this information to each visited FAas in [15]. In addition, this mechanism reduces the resourcereservation cost by allowing packets to be forwarded throughexisting LSPs. Finally, the third mechanism, called MFC, aimsto optimize the forwarding chain length and thus the end-to-end delay when MNs move in a two-dimensional space. To thebest of our knowledge, this is the first micro-mobility protocolthat takes into account the specific mobility propriety of usersin a two-dimensional space.

III. MICRO MOBILE MPLSIn this section, we describe our proposed Micro Mobile

MPLS scheme. As mentioned earlier, our ultimate aim is toovercome the limitations of Mobile IP and in the same timebenefit from the QoS support capability of MPLS networks.Micro Mobile MPLS is based on the integration of MIP-RR [12] and MPLS [16] protocols. A typical architecture forMicro Mobile MPLS networks is shown in Fig. 1. We assumethat an MPLS access network exists between the Label EdgeRouter Gateway (LERG) and the Label Edge Router/ForeignAgents (LER/FAs). The network architecture is based on a

two-level hierarchy. At the higher level is the LERG thatperforms the role of an edge Label Switching Router (LSR)filtering between intra- and inter-domain signaling. At thesecond level is the LER/FA connected to several access points(APs) that offer link-layer connectivity. We distinguish herebetween link-layer functionalities of the air interface, whichare handled by the AP, and IP-layer mobility (L3 handoff),which occurs when the MN moves between subnets served bydifferent LER/FAs. Note that an LER/FA is the first IP-capablenetwork element seen by the MN.

Fig. 1. Architecture of a Micro Mobile MPLS wireless access network.

In this paper, the handoff latency is defined as the timeelapsed from the moment the handoff event is detected to themoment the first packet is received from the new subnet. Thereare two types of handoff in our scheme: Intra-LER and Inter-LER handoffs [23]. An Intra-LER handoff occurs when theMN moves between two APs managed by the same LER/FA.This kind of handoff is basically L2 (link-layer) handoff. Onthe other hand, an Inter-LER handoff occurs when a new APand the old AP are under different LER/FAs. This kind ofhandoff is typically L3 (network-layer) handoff. In this work,we focus on Inter-LER handoff since it has the most importanteffect on the handoff performance.

To reduce the packet loss during handoff, our schemerelies on the L2 trigger [9]. Indeed, the old LER/FA willbe notified to buffer in-flight packets at each handoff occur-rence. Specifically, once the received signal strength from thecurrent AP falls below the threshold level, the MN sends a“Movement signaling” message to the current LER/FA, tonotify an imminent L2 handoff occurrence. As soon as thecurrent LER/FA receives this message, it initiates the bufferingmechanism. In the following, we present the operation of ourproposed mechanisms, namely, FH-, FC-, and MFC-MicroMobile MPLS.

A. The Fast Handoff mechanism: FH-Micro Mobile MPLS

The main idea behind FH-Micro Mobile MPLS is to antici-pate the L3 handoff using the L2 functionalities and to setup anLSP before the MN moves really into a new subnet to reduceservice disruption. In this context, our proposed FH mechanis-m can be considered as a hybrid solution combining the pre-registration and post-registration methods. Indeed, FH relieson two main operations to accomplish the handoff process:(i) the pre-establishment of a new LSP between the LERGand the new subnet prior to the L2 handoff occurrence (i.e.,pre-registration) subject to the desired QoS requirements. Assuch, the QoS continuity of the communication in progress isensured ; and (ii) the initiation of the L3 handoff immediatelyafter the L2 handoff completion through the new FA evenbefore receiving the new Mobile IP advertisement messagefrom the new subnet (i.e., post-registration).

In what follows, we describe the basic operations of ourFH mechanism. We consider two types of LSP: active LSP


and passive LSP. The active LSP is the one from the LERGto the current serving LER/FA. This LSP is currently usedto transfer data. On the other side, a passive LSP is the onefrom the LERG to the neighboring subnet to which the MN ismoving to. This LSP is not currently used until its activation.

The FH mechanism employs a link-layer movement detec-tion scheme to predict the possible MN’s next location. Inother words, as can be seen in Fig. 2a, once the MN entersan overlapped area of the boundary cells of two subnets, itreceives a L2 beacon from the possible new AP (step 1).Immediately, the MN notifies the current LER/FA for possiblehandoff by sending a Handoff Initiate (HI) signaling message,which contains the MAC address of the new AP (step 2). Notethat in practice, the MN is able to sense a set of candidate APs.In this case, it is assumed that the MN checks continuouslythe signal strength of these APs and then selects the one withthe strongest beacon signal.

Each LER/FA has a Neighbor Mapping Table (NMT) thatbinds between the IP and MAC addresses of all neighbor-ing APs. Hence, when the current LER/FA receives the HIsignaling message, it looks into its NMT table to get the newLER/FA’s IP address and then informs the LERG (respectivelythe new FA) in the case of downlink communications (re-spectively uplink communications) for the possible L3 handoff(step 3a). A passive LSP with the desired QoS requirementswill be then established between the LERG and the newsubnet (i.e., pre-registration) using either the RSVP-TE [24]or CR-LDP [25] protocols (step 4). At the same time as step3a, the current LER/FA informs the MN of the new RegionalCare-of-Address (RCoA: the IP address of the new possibleLER/FA) by sending a Neighbor Advertisement signalingmessage (step 3b). This allows the MN to be able to startregistration process with the LERG before receiving the newMobile IP advertisement message (L3 beacon) from the newFA. Specifically, the MN’s registration to the LERG will be

(a) Before handoff: pre-registration step

(b) During handoff: post-registration step

Fig. 2. Operation of the FH-Micro Mobile MPLS scheme in the case ofdownlink communications

initiated immediately after the L2 handoff completion throughthe new FA (i.e., post-registration). Indeed, the L2 handoff islaunched by the MN when the received signal strength fromthe current AP falls below the threshold level. In this case,the MN notifies the current LER/FA to initiate the bufferingmechanism by sending a Movement signaling message (seeFig. 2b, step 1). Once the L2 handoff is accomplished (i.e.,the MN is attached to the new AP), the L3 handoff is initiatedby the MN since it is already aware of the new RCoA (step2). The new FA forwards the Mobile IP registration requestmessage to the LERG (step 3a) and notifies the old LER/FA forthe handoff event (step 3b). As soon as the LERG receives theMobile IP registration request, it activates the pre-establishedpassive LSP and traffic will be delivered through the activatedLSP (step 4a). On the other hand, once the old LER/FA isnotified (step 3b), the in-flight packets are forwarded to theMN through the new FA along a static MPLS tunnel (step4b). By using the fast handoff mechanism, we can improvethe handoff performance of Micro Mobile MPLS and reduceservice disruption.

B. The Forwarding Chain mechanism: FC-Micro Mobile M-PLS

This approach is based on the forwarding chain concept(i.e., set of forwarding paths) and can be considered as a newalternative to implement the pointer forwarding technique inan MPLS environment. Indeed, each time that the MN movesto a new subnet, it registers with the old LER/FA instead ofthe LERG, as shown in Fig. 3. Consequently, the old LER/FAwill send a label request message to the new LER/FA, whichresponds with a label mapping message back to the old FA.This way, the existing LSP (with QoS requirements) betweenthe LERG and the old subnet will be extended to the newone (see Fig. 4). As a result, a forwarding chain of FAs willbe constructed for each MN. To do so, each MN keeps abuffer storing IP addresses of the visited LER/FAs. Packetstraveling towards this MN will be intercepted by the first FA inthe chain (called master FA), taking advantage of the existingLSP between the LERG and the master LER/FA, and thenforwarded along the chain of FAs to the MN. It is easy tosee that such a scheme may cause unacceptable delays dueto long chains. To avoid long forwarding chains, we set athreshold on its length denoted by

��(in terms of the number

of movements). When the threshold is reached, the MN willregister to the LERG and delete all the addresses in its buffer.That is, a new LSP between the LERG and the new subnetwill be established and the new visited LER/FA becomes thenew master FA. This scheme enables a significant reduction ofboth the registration updates and LSP setup signaling messagessent by the MN to the LERG. These LERG registrations arereplaced by simple forwarding chain updates called here localregistrations. As a result, the resource reservation cost will bereduced since packets can be forwarded through the existingLSPs benefiting from existing resource reservation on the oldpath. Moreover, this scheme is appropriate for MNs with highmobility rate, since data packets must be forwarded quicklyto their new locations with incessant QoS support.

The basic operation of the FC-Micro Mobile MPLS schemeis depicted in Fig. 3, where the MN moves from subnet 1 tosubnet 4. We assume that the threshold on the forwarding chainlength is equal to three. When the MN moves to subnet 2, itchecks first the presence of the new subnet’s IP address in itscache memory. Since it is the first time that the MN visits this


Fig. 3. Operation of the FC-Micro Mobile MPLS scheme

(a) Label table of LER/FA1 before handoffIn. port In. label FEC Out. port Out. label

1 20 fec2 - -1 25 fec1 - -... ... ... ... ...

(b) Label table of LER/FA1 after handoffIn. port In. label FEC Out. port Out. label

1 20 fec2 - -1 25 fec1 2 30... ... ... ... ...

(c) Label table of LER/FA2In. port In. label FEC Out. port Out. label

2 30 fec1 - -... ... ... ... ...

(d) Label table of the revisited subnet (LER/FA1)In. port In. label FEC Out. port Out. label

1 20 fec2 - -1 25 fec1 - -... ... ... ... ...

Fig. 4. Label tables of FA nodes when a MN moves between the LER/FA1and the LER/FA2, where fec1 corresponds to the five TCP/IP header fields:MN’s IP address, MN’s port number, protocol number, correspondent node’sIP address and correspondent node’s port number

subnet, it registers with the previous LER/FA1, which is themaster LER/FA. Thus, the existing LSP between the LERGand the LER/FA1 nodes will be extended to the LER/FA2 nodeas shown in tables (a), (b) and (c) of Fig. 4. Likewise, whenthe MN moves to subnet 3, it notifies the previous LER/FA2.In this case, data packets destined to the MN are forwardedalong the existing LSP between the LERG and the LER/FA2and then delivered to the mobile user through the new LSPestablished between the LER/FA2 and LER/FA3 nodes. Thethreshold of the forwarding chain length (

� ��) is reached when

the MN enters subnet 4. At this time, the MN will register tothe LERG and updates its address to the root of the domaindirectly. At the same time, the new LER/FA4 becomes themaster FA of the next forwarding chain.

It is worth noting that according to our FC mechanism,the loops possibly formed during the MN’s movement will besystematically removed. Indeed, each time the MN moves toa new subnet, it checks first the presence of the new FA’s IPaddress in its cache memory. If the FA is already in the MN’sbuffer, the MN deletes the generated loop. In this case, asshown in table (d) of Fig. 4, the revisited FA updates its labeltable by associating null values to both the outgoing port andoutgoing label of MN’s packets, which means that the currentFA becomes the egress node of the new LSP.

Note that our proposed FC mechanism avoids also thegeographical constraints observed in [13] and alleviates the

%Location registration proceduresInitialize �� ;IF (MN enters a new subnet)

compare the address of the new LER/FA to the addresses in buffer;IF (the new address already exists in buffer)

Loop detected: Extract the rank (rg) of the subnet from buffer;�� ;MN deletes the subsequent FA addresses in buffer;New LER/FA updates its label table;

ELSErecord the new LER/FA address in buffer;�� ;IF ( � �� )

MN registers to the old LER/FA: Old LSP extended;ELSE

MN registers to the LERG: New LSP creation;MN notifies the old LER/FA about the handoff;MN deletes all the addresses in buffer;�� ;

ENDIFENDIF

ENDIF

Fig. 5. FC-Micro Mobile MPLS scheme

scalability issue of [15] since only the MN needs to keepa cache of all its visited subnets (i.e., a cache containingthe visited FAs’ IP addresses and not FAs’ coordinates asrequired in [13]). As such, there is no need to communicatethis information to each visited FA as in [15]. The FC-MicroMobile MPLS scheme is described by the pseudocode in Fig.5.

C. The Master Forwarding Chain mechanism: MFC-MicroMobile MPLS

The third mechanism that we propose to handle efficientlylocal mobility is called Master Forwarding Chain (MFC)-Micro Mobile MPLS and can be seen as an extension of theFC-Micro Mobile MPLS in a two-dimensional space [26].

In this mechanism, the MN registers to the LERG onlywhen its distance, in terms of number of hops (instead ofmovements) from the master FA, reaches a threshold � . Toillustrate the difference between FC and MFC mechanisms, weconsider the simple example presented in Fig. 6. Assume thatthe MN’s trajectory is �� !��"��#%$ . Note that ��is the master FA. According to the FC scheme, the forwardingchain is ��&�� !��"��#%$ . Whereas, in the MFC case,the resulting forwarding chain is ��#%$ , which is theshortest path between the master FA and the MN. It is easy tosee that the MFC scheme reduces the forwarding chain size.In other words, considering the same threshold ')(+* ,!* �.- ��0/

,the registration updates with the LERG are more frequent inthe FC case. In addition, the end-to-end delay is reduced withthe MFC scheme thanks to the relatively shorter forwardingchain.

The second main difference between the two mechanisms isrelated to the local registration. Considering the MFC scheme,each time that the MN moves to a new subnet, the new RCoAwill be registered at the master FA instead of the old LER/FA,as long as the threshold is not reached. Henceforth, this kindof registration will be called local registration as opposed tothe LERG registration. Specifically, the local registration isalways done through only one hop in the FC case, whereas itcan be done through up to �2143 hops in the MFC case.

According to the MFC mechanism, as long as the MNremains in a residing area around the master FA, regardless ofthe taken trajectory, it carries out a local registration (see Fig.6). Once it leaves this area, it performs a LERG registrationand the new serving LER/FA becomes the new MN’s master


FA. In the example of Fig. 6, the threshold � is set to 4.This figure shows the MN’s trajectory. The LERG registrationwill be performed by the MN at location � since its distanceto the master FA reaches the threshold � . Note that, unlikethe MFC case, the LERG registration in the FC case dependsmainly on the MN’s trajectory.

It is worth noting that distances between FAs can be com-puted using two methods. We can either setup these distancesin advance or use the coordinates of FAs. In the former case,as the FAs are not mobile components, each one maintains astatic table containing the distance that separates it from eachof the remaining FAs of the same micro-mobility domain. Inthe latter case, each FA keeps only the information regardingthe coordinates of the remaining FAs of the micro-mobilitydomain. Accordingly, it computes the distance (in terms ofnumber of hops using the shortest path) to its destination. Inour work, we have adopted the second method since it is moreflexible. Notably, in case of FA failure, the identity of thefailed FA needs to be notified to all the FAs only, whereasusing the first method, all the distances should be calculatedagain according to the new topology and notified to each FA.Thus, using the first method may results in an increase of thetotal signaling cost.

Fig. 6. Movement path of an MN with � �� and � � � ��The basic operations of the MFC mechanism are listed

below. When the MN enters for the first time into an MFC-Micro Mobile MPLS domain, it performs the same operationsas in FH and FC mechanisms. When the MN moves to a newsubnet (L3 handoff), it proceeds as follows.

1) The MN sends a registration message to the newLER/FA. This message contains the IP address of theassociated master FA.

2) The new LER/FA checks the existence of the masterFA’s address. If it exists, the new FA computes theshortest distance to the master FA. According to thisdistance, it performs either a local registration or aLERG registration. Otherwise (if the address does notexist), it performs a LERG registration.

3) Finally, the MN receives a registration reply messageeither from the LERG or from the master FA, accordingto the registration type.

It is worth noting that with our proposed MFC mechanism,loops can not be formed during the MN’s movement since theshortest path between the Master FA and the new FA nodesis selected each time the MN moves to a new subnet.

IV. SYSTEM MODEL

In this section, we study the MN’s mobility. Our aim is todetermine the position of an MN with respect to its master FAin order to be able to predict the MN’s evolution. To do so, we

develop Markovian models. The obtained results will be used,in a later stage, to derive the protocol performance metricssuch as the registration updates cost and the link usage.

The wireless network is divided into subnetworks. Eachsubnet is covered by one LER/FA, called base station incellular networks. Moreover, each LER/FA manages severalaccess points (APs), which offer link-layer connectivity. Themobility model commonly used when planning the LER/FAsdeployment is either the 2-D mobility model or the 1-D mobil-ity model. The first (2-D) model fits those applications, wherethe MN can travel in various directions within the coveredregion, such as urban areas. The second model (1-D or linearmobility model) is more adapted for those applications wherethe MN’s mobility is limited to a pre-specified unidirectionaltrajectory. This model is commonly used in access applicationsalong roads, tunnels and trains railways. In this work, weconsider both mobility models.

A. One-dimensional mobility modelIn this model, each LER/FA has two neighbors. As shown in

Fig. 7(a), an MN which is in subnet ( , can only move to twoneighboring subnets (�� 3 or ( 1 3 with equal probability� (� - �� ); otherwise the connectivity is lost. Note that,considering the linear model, the MN’s behavior is the sameaccording to the MFC or FC schemes (when � - � ��

).In other words, the decision to carry out a local or LERGregistration remains the same in both cases. However, we pointout that the local registration cost is not the same.

(a) Linear mobility model (b) Two-dimensional model

Fig. 7. 1-D and 2-D mobility models

Let '� / be the distance (in terms of number of hops)between the MN’s location at time (current serving LER/FA)and the master LER/FA. The residence time of the MN ineach subnet is assumed to be exponentially distributed with themean 3 �� . �� '� / ��$ forms a Continuous-Time MarkovChain (CTMC) with state space � 3�-�� 3�� * * * � '��1 3 / asdepicted in Fig. 8, where � is

� ��in the FC case and � in

the MFC case.

Fig. 8. CTMC for the linear model

Let �� -! #"%$ ��&('*),+ -/. 0 '� / -21/3 �41657� 3 , be thestationary probability distribution of '� / . Based on Fig. 8,


the balance equations can be derived as follows:�� - � � � � � � � �� - � � � �� - � � � � �� 1� '��1 � /� � �� - � �� - 3

(1)

Solving these equations, we obtain:�� -3�

� � - ��' � 1 1 /� � � 3� 1� '�� 143 /

(2)

B. Two-dimensional mobility model

In our study, we consider polygon-based 2-D model. Typ-ically, each LER/FA covers an hexagon area. This model isbroadly used in the literature. In this case, each subnet issurrounded by six neighbors (see Fig. 7(b)). The MN canmove to one of the neighboring subnets with equal probability� (� - �# ). For simplicity, we only present in this section theMarkovian chain corresponding to the mobility behavior ofan MN in the MFC-Micro Mobile MPLS case. Hence, in theremainder of this section we only consider the MFC case.

Figure 7(b) represents the residing area of an MN for thecase �.-�� . The residing area contains the master FA subnetsurrounded by � - � 1 3 rings of subnets. Each subnet isreferenced by the ring label and its position inside that ring,which determines the exact MN’s position with respect to thecurrent master FA. For example, subnets belonging to ring 1are referenced by �� 3� �� , those belonging to ring 2are referenced by �� 3� �� 3�� , and so on and so forth. Togeneralize, let (! ( - �� 3�� * * *�� ' � 1 3 / designate the ( th ringaway from the master FA. The master FA subnet is denoted by� �� . Subnets belonging to ring ( are referenced by ��" � 3� �� ( . Note that the ring label represents the distance between theMN and the master FA.

Let '� / be the MN’s location within the residing areaat time . The residence time of an MN in each subnet �#�"is assumed to be exponentially distributed with the mean3�� . �� '� / �� !� $ is therefore a Markov process withcontinuous time and finite state space � � - � ��"%$ �& (' ' � 1 3 / � 3� %�( )� (�$ . Recall that our main objective isto determine the MN’s position within the residing area inorder to predict its evolution. According to its next location,the MN can perform either a local registration or a LERGregistration. In the latter case, the master FA will be updatedand its associated residing area will be created.

The resolution of the Markovian chain, as defined above, istime-consuming. Moreover, this chain suffers from the statespace explosion problem, mainly when the threshold � takeshigh values. To avoid this issue, we extract a new chain* '� / from '� / by aggregating its states. In other words, allthe sates where the MN exhibits exactly the same behaviorwill be aggregated. Hence, the size of the state space � �will be drastically reduced. To achieve this, we profit fromthe symmetric property of the 2-D model. Let us revisit theexample of Fig. 7(b). The six states of the first ring aresymmetric. Wherever the MN’s location is in the first ring,it exhibits the same behavior (i.e. its evolution is the same). Itmoves to the master FA subnet with probability � , remains in

the same ring (i.e. at the same distance from the master FA)with probability � � , and goes to ring 2 with probability � � . Inthis case, all these states can be grouped into one aggregatestate. In the second ring, we have to segregate between twocases. The first aggregated state contains the following subnets�� "� ��#+� ��#,� �� $ , whereas the second case comprises�� #� ��#-� �� .�� +� � �� $ subnets. Indeed, the MN exhibitsdifferent behavior according to each aggregated state. Forinstance, in the first case, the MN leaves the residing areawith probability � � instead of � � in the second case. Inwhat follows, we describe the algorithm to perform stateaggregation.

1) As before, ��" denotes the MN’s position in the residingarea. As presented in Fig. 7(b), the state � �" is chosen tobe the one at the top of state � �" � � . Afterwards, each ring( consists of � ( subnets labeled in a clockwise directionas � �" , * * * , � # "" . Let ��/" denote the new aggregatedstate, where ( always designates the ring reference, and�10 the state label inside the ring.

2)

Start with i = 1 until ')( - �21 3 /Repeat �

set � � /" - � " - 2�436573 " � 598" �"

If ')(;:- 3 / �set

� -=< (&1 3�?>

For @ - 3 to @ - �

set �#A /" -B2�C3D573�" � 598" A �" �E2

�F36573 # � 598" � A �"

$( - ( � 3

$Where <HG > denotes the smallest integer greater than or equalto G .

For instance, for �.-�I , we obtain the following aggregatedstates.

� � /� - �J� - �� $ ,� � /� - � � - �� +� �� "� �� #� $ ,� � /� - � � - �� +� �� "� ��K+� ��K,� �� $ ,� � /� - �� +� �� +� #� �+�#-� �+� �L�� $ ,� � /� - � � - �� +� �� K+� �� L�� +� � #� $ ,� � /� - �� +� �� +� "� �+� #� �+�#-� �+�#,� �+� � �� +� � �� "� �� +� �+� � -� $ ,� � /� - �� - �� +� "� ��K,� �� +� �+� �+�� $ ,� � /� - �� +� �� +� #� �+� -� �+� �L�� +� � �� #� �� -� �+� �M�� +� � �� $ ,� � /� - �� +�#+� �+� �� "� �+� � ,� �� $ ,Theorem:Let N -PO0�D��+�&��+� " �� /" � * * *�� A /" � * * *��#Q /" �+� " �� * * *SR

')( - � �T� � * * *�� 1 3 / designate the state space of the newchain

* '� / obtained by aggregation of the initial Markovianchain '� / . The resulting aggregated process

* '� / is alsoMarkovian.

Proof:For convenience, we denote by N " each state of the set N and


by � - � �� " � � < (&1 3

� > the set size. Let � " 5 " � designate the

generator matrix of the initial Markov chain '� / . We arrangethe states of � � according to the space N partitions � i.e.,� ��- O � �� , � �� +� �� +� �� +� �� "� �� #�� , � �� +� �� +� "� �+� +� �+� ,� �+� � �� ,

� �� +� �� +� #� �+� -� �� L�� * * *SR�� . In this case, the infinitesimal

matrix � " 5 " � can be written as a �� symmetric matrix

�� - 3 �� " � � � ( � and has the following form:

� " 5 " � - �� A 5�� F3 A�� 573 �Where � A 5 is the block matrix corresponding to the transitionprobabilities between each element of the set N A and theset N 5 . In addition, these blocks verify the constant-row sumproperty [27]. In other words, we have:

� ( � � � ' � A 5 / " � is a constant, denoted by ! A 5�'"! A 5��*� / *Thus, according to [27], the resulting aggregated process isMarkovian. To illustrate this result, let revisit the example ofFig. 7(b) where � -)� . As explained before and accordingto the state aggregation algorithm, the state space � � of theinitial Markov chain 4'� / can be arranged with respect tothe partition N . The generator matrix � " 5 " � of the process '� / can be written as in (3) (see the top of the page), where# - � � and

. - 1�� . It is easy to see that each block matrixjustifies the constant-row sum property. The new generatormatrix �%$'&'&�( of the aggregated Markov chain

* '� / is thereforegiven by (4) (see the top of the page). In Fig. 9, we representthe state transition diagram of this aggregated process (i.e.,for � - � ). Figure 10 represents the case where �.- I . Forease of use, the aggregate states were assigned numbers, asfollows:

� �*) ��+� �+) 3!�� *) �� /� ) � / �+� �+) � �� /� )� / �� )-, �� /� )-, / �+� � /� )-, /�/

.

Fig. 9. State transition diagram of the aggregated Markov chain for � �/.

Fig. 10. State transition diagram of the aggregated Markov chain for � �/0Steady state probabilities

Based on the state transition diagram of the aggregatedMarkov chain, we can obtain the steady state probabilityfor state N " � ')(2- 3�� * * *��'� /

. Denote by � " and �21 A43"5( - '�� 3!� * * * � � 1 3 / and @ - ' 3�� * * *�� /76 the stationary

probability of the system for the aggregated state � " and� A /" , respectively. The balance equations for the aggregatedMarkov chain are obtained recursively as follows:

��

� � - � � � � � � � � �� 98;:=< ?>�

� � � �21 � 3� � �� -�� %1 � 3�� - � � � � � � � � � � � 1 � 3� � � � 1 � 3�� - � � � � � � � � 1 � 3� � ��=� (! 4�21 �

� " - � � " �� " � � � �21 � 3" � � �21 � 3" �

(5)


��

�21 � 3� - � � � � � � � � � � � �%1 � 3�� 1 � 3� - � � � � � � � � � � � � � 1 � 3� � � � 1 � 3� � � � 1 � 3� � � � � 1 � 3�� - � � � � � � � � � � � � � �� " � � � �"�)I ( �2143

�21 � 3" - � � � " � � � � " � � � � �21 � 3" � � �21 � 3" �� 1 � 3" � �� 1 � 3" �

(6)

Where � -� 3 if I� (K �21 �� if ( - �21 3

�� )� ( 4�21 3 and � ��& < " �� > 143� 1 � 3" - � � 1 � �� 3" �� 1 � � 3" � � � 1 � �� 3" � � � � � 1 � 3" ��

�� 1 � 3" � �� 1 � � 3" �(7)

Where �� -�� 3 if ( is odd3 if ( is even and �� '4< " �� > 1 � /� if ( is even and � - ' < " � �� > 143 /

and � � - �3 if � ( �21 �� if ( - �21 3

�� 21 3��%1 5 3�T5 - � �21 51�� 3� 5 � � �%1 57� � 3�T51�� %1 5 3�T5 � (8)

Where � - � � if � - � ��3 otherwise

�� 21 �� 1 5 3�T5 � - � � 1 57� � 3� 5 � � � � 1 5 3� 5 � � � � 1 57� � 3� 5 �

�,� � �21 5 3� 5 �� 21 5 3� 5 � �� %1 5 � 3�T5 �(9)

Where � - � � if � - � ��3 otherwise

� �� " � � � " �� " � � 8�� <

� >�A � � � 1 A 3" - 3 (10)

Given the balance equations (5-9) and the normalizationequation (10), the steady state probabilities of the aggregatedMarkov chain can be derived. Note that obtained resultswill be used in the next section to derive the signaling costfunction of registration updates, the link usage and the handoffperformance.

V. PERFORMANCE EVALUATION & ANALYSIS

In this section, we develop analytical models to derivethe link usage cost, the handoff performance and the costfunction of registration updates for both the 2-D and the linearmobility models. This analysis includes our three mobilitymechanisms (FH, FC and MFC) and competing schemes(FMIP [10], MIP-RR [12], Mobile MPLS [17] and H-MPLS[19]). The following parameters are used in our analysis.

Parameters:� average session connection time ( average FA resident time� $ � time interval for a FA to send agent advertisements� �average number of L3 handoff during a session (i.e.,� � - � ( )�� bandwidth of the wired link�� bandwidth of the wireless link� � latency of the wired link (propagation delay)� �� latency of the wireless link (propagation delay)) �routing or label table lookup and processing delay�downlink packet transmission rate�� average size of a signaling message for the registra-tion update� � average size of a label message for LSP setup�� average number of hops between G and � in thewired network�� location update cost between an FA and the HA (hop� message size)� � & location update cost between an LER/FA and theLERG (hop � message size)� ��location update cost between two neighboringLER/FAs (hop � message size) � �traffic load related to LSP setup procedure betweenan FA and the HA (hop � message size) � & traffic load related to LSP setup procedure betweenan LER/FA and the LERG (hop � message size) ��traffic load related to LSP setup procedure betweentwo neighboring LER/FAs (hop � message size)� " and " parameters can be written as:! � � � - � �"� � $# �&% # � � - � � � � $# ��% #� � & - � �� $# � Q('*),+ � & - � � � � $# � Q-'*),+� �� - � �"� � $# � $# �� - � � � � $# � $# (11)

A. Link Usage in the MPLS access network: Linear Model

In the linear mobility model, we consider a full binary treewith the LERG as a root, as shown in Fig. 11. The depth . ofa full binary tree with � nodes is /� 1032 � �54 � 3 . That is, a fullbinary tree of depth . � 3 has � - �76 1 3 nodes (includingthe LERG, LSRs and LER/FAs) and the number of subnets orleaf nodes (i.e., LER/FAs) is � 6 �� .

Fig. 11. A full binary tree access network of depth 8�� used for the linearmodel

Let�:9

denote the link usage in the MPLS access network,which is the number of links used for packet delivery betweenthe MN and the LERG. Recall that, in FMIP, MIP-RR, MobileMPLS, H-MPLS and FH-Micro Mobile MPLS schemes, pack-et are delivered using the shortest path routing. Hence, packetsexchanged between the LERG and any FA traverse . 1 3 -<;hops. However, in both FC- and MFC-Micro Mobile MPLS,packets have to traverse the connection binding the LERG to


the master FA and the forwarding chain binding the master FAto the MN. Recall that, in the linear mobility model, the MNmoves similarly in both FC and MFC mechanisms. The meanvalue of

�:9for both FC- and MFC-Micro Mobile MPLS can

thus be given by:

�:9 ' (FC/MFC)-Micro Mobile MPLS/ - ; � �

�� 1 � � �

(12)

Where the first term ; - . 1 3 is the number of links from theLERG to the master LER/FA and the second term is the meanpath length used to forward packets from the master LER/FAto the current serving LER/FA (i.e., average forwarding chainsize). Recall that � is

� ��in the case of the Forwarding

Chain mechanism and � in the case of the Master ForwardingChain mechanism. Note that the stationary probabilities � � arecalculated in subsection IV-A.

B. Link Usage in the MPLS access network: 2-D Model

As before, we assume that the distance between the LERGand any FA is the same and equal to ; . The link usagein FMIP, MIP-RR, Mobile MPLS, H-MPLS and FH-MicroMobile MPLS schemes, is thus equal to ; . In both FC andMFC cases, we have to take also into account the meanforwarding chain size. Based on the analysis of subsectionIV-B, the mean forwarding chain size for the MFC case isgiven by:

� ' � / -�� " � � � " ')( � � � / �� " � � 8 � <

� �>�A � � ( � 1 A 3"

� � � ' �%� 1�� / 8 :=< ?>�

A � � � 1 A 3� �� ' �� 1 , / � � � � (13)

Hence, the mean value of�:9

for MFC-Micro MobileMPLS is given by:

�:9 ' MFC-Micro Mobile MPLS/ -<; � � ' � /

(14)

In the FC case, the movement of MNs in a 2-D area is nota Markovian process as the MN’s evolution depends on itsmobility history. In other words, the MN’s next registration(i.e., LERG or local registration) depends not only on itscurrent position, but also on its entire trajectory since it hasleft the master FA. This increases the complexity of analysis.Therefore, the link usage parameter of FC-Micro MobileMPLS will be evaluated only through simulations.

C. Registration Updates Cost: Linear Model

Let� � denote the signaling cost of registration updates

when a L3 handoff occurs. It is the traffic load of signalingmessages (hop � message size) exchanged in the networkwhen the MN moves to a new subnet. In FMIP, the MN onlyperforms a home registration update with the HA. In MobileMPLS, we have to take into consideration the additional costassociated to the LSP procedure setup with the new FA. InMIP-RR, only a LERG registration update with the root ofthe domain is required. Additional cost, associated to the LSPprocedure setup with the new FA, is to be considered inH-MPLS. In FH-Micro Mobile MPLS, a LERG registrationupdate and a handoff notification to the old LER/FA (inorder to forward in-flight packets) are performed at every

L3 handoff. Note that in this case, the LSP binding theLERG and the new LER/FA already exists and only needsactivation. In both FC and MFC-Micro Mobile MPLS, a localregistration followed by an LSP procedure setup between twoneighboring FAs along the extended path is required as longas the forwarding chain length does not reach the threshold.Otherwise, a LERG registration and a new LSP between theroot of the domain and the new subnet are performed. Thishappens in the linear model when the distance of the currentserving LER/FA from the master FA is ' � 1 3 / , and theMN moves in the direction that increases this distance (whichhappens with probability p). Recall that in the FC case, thelocal registration is always realized with the old FA, whereasit is performed with the master FA in the MFC case. Wesummarize the expression of registration updates cost for allunderlying protocols as follows:�� ' FMIP

/ - �� ' MIP-RR/ - �� &� � ' Mobile MPLS

/ - �� ' H-MPLS/ - � � & � � & (15)

��

� � ' FH-Micro Mobile MPLS/ - � � & � �� ' FC-Micro Mobile MPLS/

- � � �� Q�� ' � � & � � & /

- � ' 3 � �;� 1 � Q�� / �� Q�� ' � � & � � & /� � ' MFC-Micro Mobile MPLS/

- � � �� 5 � � �� '�1,� 3 / � �� 6

� � � � �� ' �� & � � & /- � ' 3 � �;� 1 � � �� / �� 5 � � ��

�� 1 � � 6� � � � �� ' �� & � � & /(16)

Where� �� - � �� .

D. Registration Updates Cost: 2-D Model

The expressions of registration updates cost for the FMIP,MIP-RR, Mobile MPLS, H-MPLS and FH-Micro MobileMPLS schemes remains the same as in the linear model. InFC-Micro Mobile MPLS, the registration updates cost will bederived through simulations. In MFC-Micro Mobile MPLS,the MN reports its location to the LERG only when it crossesthe residing area boundary (i.e., the distance between the MNand the master FA reaches the threshold � ). Using the resultsof subsection IV-B, the expression of registration updates costfor MFC-Micro Mobile MPLS can be written as (17), shownat the top of the next page.

E. Average handoff time

For convenience, let ' � � � � �� / denote the time that takesa message of size � to be forwarded from G to � via boththe wired and wireless links. ' � � � � �&� / can be expressed asfollows:

' � � � � �� / - ! � � � �&� � '�� / � ' � � �&� � 3 / � ) �


� � ' MFC-Micro Mobile MPLS/

- ��J� � � �� " � � � " � ��5 � ')(�143 / � � � ( � � � ')( � 3 / 6 �

� � �� " � � 8 � <� �>�

A � � � 1 A43"� �� 5 � � ' (&1 3 / � � � ( � � � ' ( � 3 / 6

� 8 :=< >�

A � � �� 1 A 3� � � 5 ' � 1 � / �� ' �2143 / �� ' � � & � � & / 6 � � � �� 5 � ' �21 � / �� ' � 1 3 / �� ' � � & � � & / 6

- ��' �� / �� " � � ��' �"� � � � / ')( � � � / � " �� " � � 8 � <

� �>�A � � �%( '

�� / � 1 A 3" � , � ' �� / '��%�21 � ��; / 5 8 :=< >�

A � � � 1 A 3� �� 6� � � � � �� ' � � � � � / ' �!� 1 , � � ; / (17)

Where ! -� �� if G�- MN� otherwise

The average handoff time (� �

) can be expressed as the sumof two terms: disruption time (

� � ) and completion time (��

).

Disruption time (� � ): It is the average time that spends an

MN without connection to any LER/FA during the handoffprocess. In other words, it is the time between the momentthat the MN disconnects from the old FA to the moment thatit connects to the new one. It is easy to see that the disruptiontime becomes null when the overlapping area is large enough.The worst case value for this quantity is equal to the L3 beaconperiod (

� $ � ). � � can be given by the following expression:

� � - ! � if�� ( � $� � � $ � ��

� � �� 1 � �� ( � $� otherwise(18)

Where�� ( � $� denotes the time spent by the MN in the

overlapping area. Note that with FMIP and FH-Micro MobileMPLS schemes, the disruption time (

� � ) corresponds to thephysical disconnection from the old AP until the connectionto the new one. As soon as the MN establishes a physicalconnection with the new AP, it starts receiving in-flight packetsthrough the new FA.

Completion time (��

): It is the time to complete the regis-tration update. We summarize the

��value corresponding to

each scheme as follows:�� ' FMIP

/ - �� ' �� &% # /�� ' MIP-RR/ - �� ' �� ;� + $# /�� ' Mobile MPLS

/ - �� ' �� ;��% # / � �� ' � � � � $# �&% # /�� ' H-MPLS/ - �� ' �� Q-'*),+ / � �� ' � � � � *# � Q-' ),+ /

(19)��

�� ' FH-Micro Mobile MPLS/

- �� ' � � � � ��;� Q-'*),+ � � *# � $# /� � ' FC-Micro Mobile MPLS/

- �� ' � � � � ��;� $# � � $# � $# / � �� ' � � � � *# � $# /� �

� � ' � ��!/ �� ' �� ;� Q-'*),+ / � ' � � � � $# � Q-' ),+ / �� ' MFC-Micro Mobile MPLS

/- �� ' �� ;� $# � � $# � A $# / � �� ' � � � � $# � A $# /� �! ' � / � � ' �� ;� Q-'*),+ / � ' � � � � $# � Q-' ),+ / �

(20)

Where,� � ' � ��0/ -#"�' � ��!/ denotes the average number of L3

handoffs that takes place between two consecutive renewals ofthe forwarding chain ; and ! ' � / -$" ' � /

denotes the averagenumber of visited subnets (L3 handoffs) during a cycle (i.e.,before that the master FA and its associated residing area arerenewed).

F. Total packet loss during a session

The total packet loss ()&% - �� ) during a session is defined

as the sum of lost packets per MN during all handoffs. InMIP-RR, Mobile MPLS and H-MPLS, all in-flight packetswill be lost during the handoff time due to the lack of anybuffering mechanism. In FMIP, FH, FC, and MFC, in-flightpackets would be lost till the buffering mechanism is initiated.Note that this mechanism is initiated at the old FA level in ourmobility mechanisms while it is initiated at the new FA level(or new access router) in the FMIP scheme.

)&% - �� for eachscheme can be expressed as follows:�� )&% - �� ' FMIP

/ - ' �� $# � � $# � $# / � � �/� �)&% - �� ' MIP-RR/ - � � ' MIP-RR

/ � � �/� �)&% - �� ' Mobile MPLS/ - � � ' Mobile MPLS

/ � � � � �)&% - �� ' H-MPLS/ - � � ' H-MPLS

/ � � �/� �)&% - �� ' (FH/FC/MFC)-Micro Mobile MPLS/

- ' �� ;� $# / � � � � �

(21)

G. Buffer size requirement

According to our mobility mechanisms, a buffer is requiredat the old LER/FA to store in-flight packets during eachhandoff operation. As stated before, the buffering mechanismis activated when the current LER/FA receives the Movementsignaling message. This message notifies an imminent L2handoff occurrence. On the other side, the buffering mechanis-m is disabled when the current LER/FA is notified by the MNthrough the new FA to forward in-flight packets. In both FH-and MFC-Micro Mobile MPLS, the signaling message usedto notify the old FA corresponds to the handoff notificationmessage. In FC-Micro Mobile MPLS, it corresponds to theforwarding chain update message. The buffer size requirement( �&' � � (�(0, ) for FMIP, FH-, FC- and MFC-Micro MobileMPLS is listed as follows:! �&' � � ( (0, ' FMIP

/ - 0 � � � ' �"� � � �� $# / 3=� ��&' � � ( (0, ' (FH/FC/MFC)-Micro Mobile MPLS/

- 0 � � � ' �"� � � �� $# � � $# � $# / 3 � � (22)


VI. NUMERICAL & SIMULATION RESULTS

In this section, we compare all underlying protocols usingboth analytical and simulation approaches. The parametersettings in our experiments are listed in table I.

TABLE IPARAMETER SETTINGS

Parameter Value Parameter Value��1000 sec �� 4��

5 � 50 sec (default 20) � � 10 �� sec��1 sec �� 100 Mbps

L2 Beacon 100 msec � �� 11 Mbps�! 48 bytes � � 1 msec� � 28 bytes � �� 2 msec"2 � 16 (default 9) 8 3 � 17#

( � �� or � ) 1 � 15 (default 4) $ 64 Kbps

Figure 12 represents the link usage cost of all underlyingprotocols for both the linear and the 2-D mobility models.We can see that FC- and MFC-Micro Mobile MPLS have thesame cost in both mobility models when �.- ��

. This costis higher compared to the remaining protocols. This slightdifference is due to the additional cost introduced by theforwarding chain.

2 4 6 8 10 12 14 162

4

6

8

10

12

14

16

18

Hop distance, δ

Link

usa

ge

Link usage vs. δ

FMIP / MIP−RR / Mobile MPLS / H−MPLS (anal)FH−Micro Mobile MPLS (anal)(FC/MFC)−Micro Mobile MPLS (anal, K=Lth=4)FC−Micro Mobile MPLS (simu, Lth=4)MFC−Micro Mobile MPLS (simu, K=4)

2 4 6 8 10 12 14 1610

0

101

102

Hop distance, δ

Link

usa

ge

Link usage vs. δ

FMIP / MIP−RR / Mobile MPLS / H−MPLS (anal)FH−Micro Mobile MPLS (anal)FC−Micro Mobile MPLS (simu, Lth=4)MFC−Micro Mobile MPLS (simu, K=4)MFC−Micro Mobile MPLS (anal, K=4)

(a) Linear model (b) 2-D model

Fig. 12. Link usage cost

2 4 6 8 10 12 14 1610

1

102

103

104

Hop distance, δ

Reg

istr

atio

n U

pdat

es C

ost (

Byt

e)

Registration Updates Cost vs. δ

FMIP (anal)MIP−RR (anal)Mobile MPLS (anal)H−MPLS (anal)FH−Micro Mobile MPLS (anal)FC−Micro Mobile MPLS (simu, Lth=4)FC−Micro Mobile MPLS (anal, Lth=4)MFC−Micro Mobile MPLS (simu, K=4)MFC−Micro Mobile MPLS (anal, K=4)

2 4 6 8 10 12 14 1610

2

103

104

Hop distance, δ

Reg

istr

atio

n U

pdat

es C

ost (

Byt

e)

Registration Updates Cost vs. δ

FMIP (anal)MIP−RR (anal)Mobile MPLS (anal)H−MPLS (anal)FH−Micro Mobile MPLS (anal)FC−Micro Mobile MPLS (simu, Lth=4)MFC−Micro Mobile MPLS (simu, K=4)MFC−Micro Mobile MPLS (anal, K=4)


Fig. 13. Registration updates cost at every L3 handoff

Figure 13(a) plots the different registration updates costsat every L3 handoff for the linear mobility model. The FH-Micro Mobile MPLS scheme exhibits a lower registration costthan FMIP, Mobile MPLS and H-MPLS protocols, since therequired LSP between the LERG and the new LER/FA alreadyexists. However, it has a higher registration cost than MIP-RR,due to the extra signaling messages sent to the old subnet toforward in-flight packets. FC- and MFC-Micro Mobile MPLS,on the other hand, provide the lowest registration cost sincesome expensive LERG registration updates are replaced bylow-cost local registrations. Note that the LERG registrationupdates cost increases with the distance ; . In view of this, the

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1510

1

102

103

104

Threshold, (Lth

, K)

Reg

istr

atio

n U

pdat

es C

ost (

Byt

e)

Registration Updates Cost vs. φ (Depth l=10)

H−MPLS (anal)FH−Micro Mobile MPLS (anal)FC−Micro Mobile MPLS (simu)FC−Micro Mobile MPLS (anal)MFC−Micro Mobile MPLS (simu)MFC−Micro Mobile MPLS (anal)

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 1510

1

102

103

104

Threshold, (Lth

, K)

Reg

istr

atio

n U

pdat

es C

ost (

Byt

e)

Registration Updates Cost vs. φ (δ = 9)

H−MPLS (anal)FH−Micro Mobile MPLS (anal)FC−Micro Mobile MPLS (simu)MFC−Micro Mobile MPLS (simu)MFC−Micro Mobile MPLS (anal)


Fig. 14. Effect of � �� and � on the registration updates cost in (FC/MFC)-Micro Mobile MPLS

0 5 10 15 20 25 302

3

4

5

6

7

Hop distance, δ

Opt

imal

K

Kopt

vs. hop distance δ

Linear model2−D model

Fig. 15. Optimal value of threshold � in MFC-Micro Mobile MPLS as afunction of the hop distance

"

FC and MFC gain, compared to previous approaches, increaseswith this distance. It is worth noting that the MFC mechanismhas a higher cost than FC. In fact, the local registration in theFC case is cheaper than that in the MFC case. Recall that inFC, the local registration is always realized with the previousFA (i.e. 1 hop), whereas it is performed with the master FA(i.e. up to �21 3 hops) in MFC.

This difference persists when using the 2-D mobility modelas depicted in Fig. 13(b). Indeed, both FC and MFC schemesexhibit the smallest costs always. However, this figure showsthat minimum registration cost is obtained by different strate-gies according to the value of ; . In this particular case (i.e.,� - � �� - , ), the FC scheme stands out as the best choicewhen ; &% , otherwise the MFC scheme provides the best cost.Recall that considering the same threshold (i.e. � - � ��

),the expensive registration updates with the LERG are morefrequent in the FC case. However, the local registration costin the FC strategy is cheaper than the MFC one. In this regard,when ; is large, the LERG registration cost is a dominant cost.Hence, the MFC mechanism stands out as the best choice.Otherwise, when ; is relatively small, the FC mechanismbecomes the best choice. Notice that for both mobility models,analytical results practically coincide with the simulation ones,which illustrates the accuracy of our study.

Figure 14 depicts the cost of registration updates for FC- andMFC-Micro Mobile MPLS as a function of their respectivethresholds and for both mobility models. We can observethat the cost of registration updates for MFC-Micro MobileMPLS is a convex function of � , where the minimum costis obtained for � � � � (in these figures � � � � - , ). In fact, theLERG registration frequency decreases with the increase ofthe threshold � - � , since more and more expensive LERGregistrations are replaced by local registrations. On the otherhand, the local registration cost increases with the threshold,since the average distance between the MN and the master


10−2

10−1

100

101

0

0.5

1

1.5

2

2.5

3

3.5x 10

7

Call−to−Mobility Ratio, (ψ)

Reg

istr

atio

n U

pdat

es C

ost

Registration Updates Cost vs. CMR (Depth l=10)

FMIP (anal)MIP−RR (anal)Mobile MPLS (anal)H−MPLS (anal)FH−Micro Mobile MPLS (anal)FC−Micro Mobile MPLS (anal)MFC−Micro Mobile MPLS (anal)

10−2

10−1

100

101

0

0.5

1

1.5

2

2.5

3

3.5x 10

7

Call−to−Mobility Ratio, (ψ)

Reg

istr

atio

n U

pdat

es C

ost

Registration Updates Cost vs. CMR (δ = 9)

FMIP (anal)MIP−RR (anal)Mobile MPLS (anal)H−MPLS (anal)FH−Micro Mobile MPLS (anal)FC−Micro Mobile MPLS (simu)MFC−Micro Mobile MPLS (anal)


Fig. 16. Effect of CMR on the registration updates cost

5 10 15 20 25 30 35 40 45 5010

0

101

102

103

FA Resident Time, tr (sec)

Tot

al L

ost P

acke

ts (

Kb)

Total Lost Packets vs. FA Resident Time

FMIP (anal)MIP−RR (anal)Mobile MPLS (anal)H−MPLS (anal)FH−Micro Mobile MPLS (simu)FC−Micro Mobile MPLS (simu)MFC−Micro Mobile MPLS (simu)(FH/FC/MFC)−Micro Mobile MPLS (anal)

Fig. 17. Total lost packets during a session

FA increases. In view of this, the optimal cost is a trade-off between these two opposite requirements. Specifically, theregistration updates cost of MFC-Micro Mobile MPLS can bewritten as follows:� � ' MFC-Micro Mobile MPLS

/- � ' � / � (registration cost between two neighboring FAs)

� 3! ' � / � (LERG registration cost)

- � ' � / �� 3! ' � / ' � � & � � & /Where

� ' � / - " ' � /denotes the average distance (in terms

of number of hops) between the MN and the master FA and! ' � / - " ' � /denotes the average number of visited subnets

during a cycle (i.e., inside the residing area). This formulaexhibits clearly the convex behavior of the MFC registrationupdates cost. Finally, we notice that analytical results, for both2-D and linear mobility models, practically coincide with thesimulation results, which illustrates the accuracy of our study.

On the other side, in FC-Micro Mobile MPLS, the cost ofregistration updates decreases with the threshold � - � ��

. In-deed, the expensive LERG registrations become less frequent.They are replaced by low-cost (1 hop) local registrations.However, we note that the threshold value will be limitedby delay constraint. Typically, delay sensitive applications,such as video or voice services, will require small values of� ��

to ensure acceptable end-to-end delay. For instance, inFig. 14(b), if the maximum acceptable value of

��is less

than 4, the MFC mechanism is the best choice; otherwise,the FC mechanism becomes the best. Note that, in the linearmodel, FC-Micro Mobile MPLS is always the best choice.Finally, it is worth noting that the variation of � (i.e., �or

� ��) does not affect the performance of the other studied

protocols (FMIP, MIP-RR, Mobile MPLS, H-MPLS, and FH-Micro Mobile MPLS). Consequently, the results presented infigures 12 and 13 for these schemes are � -independent.

Figure 15 shows the optimal threshold � � � � in MFC-MicroMobile MPLS for different values of the hop distance ; . It isinteresting to note that the optimal threshold � � � � increaseswith the distance ; . This result is expected since the LERGregistration cost increases with this distance.

Figure 16 shows the effect of the call-to-mobility ratio(CMR) on the registration updates cost for different schemesunder both mobility models and using the optimal value� � � � - , . CMR is defined as the ratio of the packet arrivalrate

�to the mobility rate. In this figure, we observe that when

the CMR is small (i.e., when the MN handoffs frequently),the FH-Micro Mobile MPLS scheme generates less signalingtraffic than the FMIP, Mobile MPLS and H-MPLS schemes,which are more suitable for mobile users with high CMR.We notice also that both FC- and MFC-Micro Mobile MPLScan significantly reduce the registration updates cost mainlywhen the CMR value is low. Our results demonstrate thatthese schemes can reduce the MIP-RR signaling cost by 72%and 57% when the CMR is small for the linear and 2-Dmobility models, respectively. Note that, this considerable gainis obtained with respect to the MIP-RR scheme, which exhibitsthe best cost among the existing protocols. According to theseresults, FC and MFC mechanisms are the best strategies forMNs with high mobility rate.

The average handoff time values for different schemes arelisted in table II. Each value was obtained by averaging 100consecutive simulations. The simulated access network con-sists of a full binary tree of depth . - 3 � and � - � �� - , .Every leaf node (LER/FA) is connected to one AP. Duringsimulations, the MN moves periodically between neighboringAPs and receives downlink packets. The simulations are runusing the Network Simulator NS-2 [28]. As can be seen,FH-Micro Mobile MPLS improves the handoff performancewhen compared with FMIP, Mobile MPLS and H-MPLSsince the registrations are performed within a local domainand the required active LSP between the root and the newsubnet already exists. Furthermore, FC-Micro Mobile MPLSprovides the lowest average handoff time, since in this case theregistrations are often carried out with the previous FA insteadof the LERG, which enables shorter delay to complete theregistration updates. The handoff delay increases slightly withMFC compared to FC, since the MN performs registrationswith the master LER/FA instead of the previous FA.

TABLE IIAVERAGE HANDOFF TIME IN MSEC

FMIP 30.198Mobile MPLS 56.256

MIP-RR 22.159H-MPLS 40.199

FH-Micro Mobile MPLS 24.171FC-Micro Mobile MPLS 18.136

MFC-Micro Mobile MPLS 20.145

Figure 17 shows the amount of packet losses during thewhole connection session for different schemes assuming thesame scenario. We can observe that the total packet losses forall approaches increases when the MN handoffs frequently(i.e., when the FA resident time is short). Notice that MobileMPLS has the largest amount of packet losses. In contrast,FH-, FC- and MFC-Micro Mobile MPLS, provide the smallestamount of packet losses thanks to the buffering mechanism.In this case, the maximum buffer size requirement for eachMN is about 4.016 KB. This means that a memory of size128 MB can handle over than 30 thousands of MNs.


VII. CONCLUSION

This paper described a new micro-mobility managementscheme, called Micro Mobile MPLS, that supports both mobil-ity management and QoS resource provisioning in IP/MPLS-based wireless access networks. Micro Mobile MPLS inte-grates three mobility management mechanisms: FH-, FC- andMFC-Micro Mobile MPLS. With FH, we anticipate the LSPestablishment before the MN moves into a new subnet toreduce service disruption by using the L2 functionalities. WithFC and MFC, we track the MN’s movement using the conceptsof forwarding chains and residing area respectively. Theseconcepts limit the range of handoff signaling messages andLSP setup signaling messages to a local area. Doing so, weavoid the relatively long distance negotiations with the root ofthe domain. We demonstrated how our mechanisms reduce thecost of registration updates and provide low handoff latencyand small packet loss rate. In particular, we showed thatthe proposed mechanisms provide short handoff delay, witha slight advantage for the FC protocol. Also we found theFC mechanism to yield the smallest registration updates costalways when considering linear mobile movements. In the two-dimensional mobility case, we found that the MFC protocolstands out as the best choice for delay sensitive applications;otherwise the FC mechanism provides the lowest cost. Usingboth analytical and simulation approaches, we compared ourproposed mechanisms with existing solutions (FMIP, MIP-RR,Mobile MPLS and H-MPLS). We analytically derived theregistration updates cost, the link usage and the handoffperformance for all protocols and we found that our proposedmechanisms achieve a substantial signaling cost reduction andimprove the handoff performance at the price of a slightincrease in link usage cost.

REFERENCES

[1] C. Perkins, IP Mobility Support for IPv4, RFC 3220, January 2002.[2] I.F. Akyildiz, J. Xie, S. Mohanty, “A survey of mobility management

in next-generation all-IP-based wireless systems”, IEEE Wireless Com.Mag., vol. 11, issue 4, pp. 16-28, Aug. 2004.

[3] A. Campbell, J. Gomez, S. Kim, Z. Turanyi, A. Valko and C-Y Wan,“Design, Implementation and Evaluation of Cellular IP”, IEEE Pers.Commun. Mag., August 2000.

[4] R. Ramjee, K. Varadhan, L. Salgarelli et al., “HAWAII: a Domain-basedApproach for Supporting Mobility in Wide-area Wireless Networks”,IEEE/ACM Transactions on Networking, vol. 10, pp. 396-410, June 2002.

[5] S. Das, A. Misra, P. Agrawal, and S.K. Das, “TeleMIP:Telecommunications-enhanced Mobile IP architecture for fastintradomain mobility”, IEEE Pers. Com. Mag., vol. 7, pp. 50-58,August 2000.

[6] A. Grilo, P. Estrela and M. Nunes, “Terminal independent mobility forIP (TIMIP)”, IEEE Com. Mag., vol 39, pp. 34-41, December 2001.

[7] A. Misra, S. Das, A. Dutta, A. Mcauley, and S.K. Das, “IDMP-based fasthandoffs and paging in IP-based 4G mobile networks”, IEEE Commun.Mag., vol.40, pp. 138-145, Mars 2002.

[8] C. Perkins and D. Johnson, “Route Optimization in Mobile IP”, Internetdraft, IETF, draft-ietf-mobileip-optim-11.txt, September 2001.

[9] K. El Malki, “Low Latency Handoffs in Mobile IPv4”, IETF, draft-ietf-mobileip-lowlatency-handoffs-v4-11.txt, October 2005, work in progress.

[10] R. Koodli, Ed., Fast Handovers for Mobile IPv6, RFC 4068, July 2005.[11] J. Xie and I. F. Akyildiz, “A distributed dynamic regional location

management scheme for mobile IP”, IEEE Infocom 2002, vol. 2, pp.1069-1078.

[12] E. Gustafsson, A. Jonsson and C. Perkins, “Mobile IPv4 Regional Reg-istration”, draft-ietf-mip4-reg-tunnel-01.txt, Nov. 2005, work in progress.

[13] Y. Bejerano, I. Cidon, “An Anchor Chain Scheme for IP MobilityManagement”, Wireless Networks, vol. 9, no.5, pp 409-420, 2003.

[14] C. Chu, C. Weng, “Pointer forwarding MIPv6 mobility management”,IEEE Globecom, vol. 3, pp. 2133-2137, November 2002.

[15] W. Ma and Y. Fang, “Dynamic Hierarchical Mobility ManagementStrategy for Mobile IP Networks”, IEEE J. Select. Areas Commun., vol.22, pp. 664-676, May 2004.

[16] E. Rosen, A. Viswanathan, and R. Callon, “Multiprotocol Label Switch-ing Architecture”, Internet IETF RFC 3031, January 2001.

[17] Z. Ren, C. Tham, C. Foo, and C. Ko, “Integration of Mobile IP andMulti-Protocol Label Switching”, IEEE ICC’01, vol. 7, pp. 2123-2127.

[18] F.M. Chiussi, D.A. Khotimsky, and S. Krishnan, “Mobility Managementin Third-Generation All-IP Networks”, IEEE Commun. Mag., pp. 124-135, September 2002.

[19] T. Yang, Y. Dong, Y. Zhang, and D. Makrakis, “Practical Approachesfor Supporting Micro Mobility with MPLS”, IEEE ICT’02, China.

[20] H. Kim, K.D. Wong, W. Chen and W.C. Lau, “Mobility-Aware MPLS inIP-based Wireless Access Networks”, IEEE Globecom, vol. 6, pp. 3444-3448, November 2001.

[21] K. Xie, V. Wong and V. Leung, “Support of Micro-Mobility in MPLS-based Wireless Access Networks”, IEEE WCNC, vol. 2, pp. 1242-1247,March 2003.

[22] V. Vassiliou, H. L. Owen, D. Barlow, J. Sokol, H-P. Huth and J. Grim-minger, “M-MPLS: Micromobility-enabled Multiprotocol Label Switch-ing”, IEEE ICC, vol. 1, pp. 250-255, May 2003.

[23] R. Langar, S. Tohme and N. Bouabdallah, “Mobility managementsupport and performance analysis for Wireless MPLS networks”, ACM-Wiley Int. J. Network Management, vol. 16, issue 4, pp. 279-294, July2006.

[24] D. Awduche, L. Berger, T. Li, V. Srinivasan,and G. Swallow, “RSVP-TE: Extensions to RSVP for LSP Tunnels”, Internet IETF RFC 3209,December 2001.

[25] B. Jamoussi, L. Anderson, R. Callon, R. Dantu, et al. “Constraint-BasedLSP Setup using LDP”, Internet IETF RFC 3212, January 2002.

[26] R. Langar, N. Bouabdallah, S. Tohme, R. Boutaba, “Mobility Modelingand Handoff Analysis for IP/MPLS-Based Cellular Networks”, in proc.of the 49th IEEE Globecom, San Francisco, USA, November 2006.

[27] G. Rubino and B. Sericola, “A finite characterization of weak lumpableMarkov processes. Part II: the continuous time case”, Stochastic Processesand their Applications, 45, 1993.

[28] ns2 home page, http://www.isi.edu/nsnam/ns

Rami Langar received the B.S. degree in telecom-munications engineering from the Ecole Superieuredes Communications (Sup’Com), Tunis, Tunisia,in 2001. He received the M.S. degree from theUniversity of Paris VI, France, in 2002, and thePh.D. degree from the Ecole Nationale Superieuredes Telecommunications de Paris, France, in 2006,both in Network and Computer Science. In 2006,he was also with the Ecole Normale Suprieure deCachan, France, as teaching and research assistant.He is currently a Postdoctoral research fellow inthe School of Computer Science at the University

of Waterloo, Canada. His research interests include 4G visions, mobilitymanagement, resource management, wireless mesh networks, performanceevaluation and quality of service support.

Nizar Bouabdallah received the B.S. degree intelecommunications engineering from Ecole Su-perieur des Communications (Sup’Com), Tunis,Tunisia, in 2001, and the M.S. and Ph.D. degrees innetwork and computer science from the Universityof Paris VI, Paris, France, in 2002 and 2004, re-spectively. He joined Alcatel Research Laboratories,Marcoussis, France, in 2002, while working on hisPh.D. degree. In 2005, he was with the North Caroli-na State University, Raleigh, NC, USA, as a Postdoc-toral Fellow. He is currently a researcher at INRIA(Institut National de Recherche en Informatique et

en Automatique). Since February 2007, he has been a Visitor Researcherat the School of Computer Science, University of Waterloo, Canada. Hisresearch interests include optical networking, wireless and sensor networks,performance evaluation, network planning and modeling, as well as controland management architectures.

Raouf Boutaba received the MSc. and PhD. De-grees in Computer Science from the UniversityPierre & Marie Curie, Paris, in 1990 and 1994respectively. He is currently a Full Professor ofComputer Science at the University of Waterloo. Hisresearch interests include network, resource and ser-vice management in multimedia wired and wirelessnetworks. Dr. Boutaba is the founder and Editor-in-Chief of the IEEE Transactions on Network andService Management and on the editorial boards ofseveral other journals. He is currently a distinguishedlecturer of the IEEE Communications Society, the

chairman of the IEEE Technical Committee on Information Infrastructureand the IFIP Working Group 6.6 on Network and Distributed SystemsManagement. He has received several best paper awards and other recognitionssuch as the Premier’s research excellence award.

IEEE/ACM TRANSACTIONS ON NETWORKING, OCTOBER ...IEEE/ACM TRANSACTIONS ON NETWORKING, OCTOBER 2008 2...

Documents

Transcript of IEEE/ACM TRANSACTIONS ON NETWORKING, OCTOBER ...IEEE/ACM TRANSACTIONS ON NETWORKING, OCTOBER 2008 2...