SIP Paging and Tracking Behcet Sarikaya, Xiao (Abbie) Zheng

Computer Science DepartmentUniversity of Northern British Columbia

Prince George, BC V2N 4Z9 CanadaEmail: (sarikaya, zhengx@unbc.ca)

Abstract— The paper introduces a new paging techniqueto track and wake up a mobile node attached to anaccess point in wireless LAN network after a SIP INVITEmessage is initiated by a caller. Tracking Agent keeps trackof the mobiles’ handoffs between the access points. PagingAgent triggers tracking agent to page the mobile whena SIP INVITE is received for one of its users. Contexttransfer feature of our paging protocol allows the pagingmessages to deliver the station context in order to enablefaster session reestablishment. Access point then doesonlink paging in wireless link. SIP extensions are neededto trigger the paging agent to start paging and mobilenodes to notify their dormant status using an extended SIPREGISTER method. Onlink paging can be implementedusing TIM or DTIMs of 802.11. Tracking protocol isanalyzed to compare soft-state and hard-state approachesfor state inconsistency ratio, message rate and the overallcost. Simulation model we developed enables us to evaluatethe traffic introduced by the tracking protocol and thecache (state) size. Paging protocol is analyzed for CPUprocessing times and the transmission delays in SIP sessionsetup with paging. Simulation of the paging with contexttransfer is used to show the gains in reauthentication.

Index Terms— Fluid flow and random walk mobilitymodel, Paging Agent, Tracking Agent, Context Transfer,Onlink paging, Session initiation protocol.

I. INTRODUCTION

Location management which keeps track of thelocation of mobile nodes (MN) for delivery ofinformation constitutes an important component ofmobility management in wireless networks. Vari-ous strategies of location management have beendeveloped to date to locate a dormant MN inside awireless network when incoming calls arrive for theMN. Location management has two tasks: locationtracking, which is initiated by MN, and paging,which is initiated by network entities such as basestations, access points (AP), and access routers (AR)[1].

Public cellular networks such as the third/fourthgeneration (3G/4G) Universal Mobile Telecommu-nications System (UMTS) networks handle dormantmode MNs effectively at the link layer, i.e. Layer 2(L2). MN when dormant constantly monitors thepaging channel and wakes up when it is paged.Location tracking is hierarchical, with cells, the basestation’s coverage area, constituting the lowest level.Several cells inside a radio network controller is aUniversal Terrestrial Radio Access Network regis-tration area (URA) and several URAs constitute aregistration area (RA). UMTS does location trackingof active MNs at URA and cell levels. Dormantmode MNs save battery by registering their locationat RA level only when they move to another RA [2].

The Internet Protocol (IP) Paging (Layer 3 Pag-ing) strategies exist in order to handle dormant MNsroaming in wireless LAN environments such as inWi-Fi hotspots. A L3 paging area is defined asthe coverage area of one access router. MN makeslocation registrations when it crosses L3 pagingareas. Any traffic destined to MN triggers a paging

1

request to wake up MN within a paging area. Insidea L3 paging area there are several WLAN cells(L2 paging areas), coverage areas of APs. Locationtracking of active MNs at the access routers can bedone at cell level and this can be combined with IPPaging in order to efficiently locate MN in its mostrecent L2 paging area. Waking up the mobile nodeis done at the link layer using specialized beacons[3], [4].

The Session Initiation Protocol (SIP) is an ap-plication layer protocol that deals with the cre-ation, modification and termination of multimediasessions [5]. SIP is the protocol for Voice over IP(VoIP) call setup. SIP mobility has extensively beeninvestigated [6]. SIP and Mobile IP both can be usedto deal with mobility issues [7], [8], [9]. Howeverwhen they are used together, there could be asevere redundant registration overhead especiallywhen mobile node frequently roams to new subnets.Besides, the indirect communications and tunnellingcaused by MIP may also increase the end to enddelay which is intolerable for VoIP applications.SIP protocol can be extended to support terminalmobility without MIP [10]. Even though a greatdeal of research has been done on SIP mobility,more remains to be done on how to deal with theissues arising from the mobile nodes in dormantmode roaming in Wi-Fi hotspots which we addressin this paper.

We consider the following scenario. A mobileuser connects his/her SIP-enabled phone to IEEE802.11 wireless LAN network and receives a callfrom another mobile user connected to 3G/4GUMTS network. The callee is in dormant mode andtherefore the call is missed. The solution is to dothe location tracking using SIP Location Trackingand then alert/ wake-up the callee using SIP Paging.

The paper first introduces SIP Paging and Loca-tion Tracking methodology in Section 2. Sections3 and 4 discuss the tracking and paging protocols,respectively. Section 5 and 6 presents an analysisand simulation of SIP tracking and paging protocols,respectively. Finally Section 7 concludes the paper.

II. SIP PAGING AND LOCATION TRACKING

Session initiation in VoIP starts with the callersending the INVITE request to the SIP server(Proxy) in the domain for the callee. The caller andcallee are identified at the application layer using

Uniform Resource Locator/ Identifier (URL/URI)which has a form similar to an email address suchas sip:bob@biloxi.com. The establishment succeedswith 200 OK reply transaction issued by the calleereaching the caller providing IP addresses whichare used to transmit multimedia streams. The call isterminated by either side sending the BYE requestfollowed by 200 OK reply.

As an example, Bob’s computer with anIPv6 address identified with SIP URI ofsip:bob@2001:410:1000:1:3656:78FF:FE9A:BCDEis used to represent location. SIP registrationestablishes a binding between an address-of-record(AOR), e.g. sip:bob@unbc.ca and a contactaddress in IPv4/v6. SIP User Agent (UA) sendsits registration message to the home registrar, e.g.sip.unbc.ca.

SIP Paging is triggered after a SIP INVITE mes-sage is received for a dormant MN connected to aWLAN access point (AP) by the proxy server. If theproxy server forwards this datagram to the accessrouter of WLAN network, AR needs to perform aneighbor discovery that involves all APs under theAR. APs notify all MNs using the delivery trafficindication map (DTIM) beacons. However, MNsin power save mode usually do not receive suchbeacons and thus do not participate in the neighbordiscovery and as a result do not receive the INVITEmessage [4]. Because of this, proxy server has tostart a paging procedure which will wake up theMN and deliver the INVITE message. We call sucha protocol SIP Paging protocol in this paper.

SIP Paging and Location Tracking architecturecontains several entities: a SIP proxy server, a Loca-tion Server (colocated with Registrar) (LS), a PagingAgent (PA) and a Tracking Agent (TA) (Fig. 1). MNis identified by LS through SIP registration withits MAC address and its IP address which mighthave changed. LS asks PA to page the MN. PAstarts the paging procedure by sending the PagingRequest message to the tracking agent (TA). TAfinds the corresponding entry in the paging cacheby the matching MAC address and replaces it withMN’s IP address. SIP Paging and Location Trackingsystem is composed of two protocols, the trackingcomponent called SIP Tracking (discussed in Sec.3) and the Paging component called SIP Paging(discussed in Sec. 4).

SIP Paging is hierarchical. At the top level is theSIP Proxy with its PA/TA. The next level is ARs

2

with each AR in charge of its Layer 3 paging area.The lowest level is APs with each AP in charge of itsWLAN cell. TA sends the Paging Request messageto the AR of the most recent L3 paging area. ARstarts the paging with the AP to which MN wasmost recently associated with. Paging request/replymessages are UDP messages. Hierarchical pagingrequires the MNs in dormant mode to be tracked atthe routing area level, to know MN’s latest Layer 3and Layer 2 attachment points, i.e. the most recentAR that serves MN and the most recent access point(AP) MN has associated with. This requires MN toregister its dormant mode status which is possibleby SIP dormant mode registration.

TAs can be replicated at each AR. Such anapproach would reduce tracking message delays asTA is closer to MNs but the cost increases. InSection V, we present an analysis of the centralizedTA versus replicated TAs.

A. SIP Dormant Mode Registration

MNs going into the dormant mode need to signalthis to SIP Paging and Location Tracking system.For this purpose, we extend the SIP protocol byadding dormant mode registration messages to sup-port paging. SIP REGISTER Request message isextended to indicate dormant mode registration withfive new contact header parameters -mode, AR,MAC, ReceiveDTIMsand idletime.

When MN detects prolonged inactivity, it sendsa dormant mode REGISTER request message tothe SIP Proxy. SIP Proxy responds with a dormantmode response message (200 OK) and updatesits location service database. When the idle timeexpires, the entry in the location service databasewill not be deleted, but the value ofmode willbe changed to active. Theidle time of MN canbe extended by another dormant mode REGISTERrequest message.

The other three new contact header parametersareAR, MAC, ReceiveDTIMs. AR field contains theIP address of AR that MN is attached,MAC fieldhas MN’s MAC address andReceiveDTIMsfieldindicates whether its receiveDTIMs bit is set.

The extension header fields are used as follows:When mode field is set to idle, SIP Proxy thenknows that the MN is in dormant mode and theexpected duration is taken fromidletime headerfield. The other fields are sent in PAGE request

message to PA and are used by Paging and Trackingprotocols. SIP Proxy keeps track of the last twovalues of AR in new ARand old AR and sendsboth values to PA. PA usesAR information ofMN to perform paging context transfer to reducethe paging delay. We discuss the paging contexttransfer in Section IV-A.MAC header field is usedfor mapping MN’s MAC address to IP addressby TA. TA passes onreceiveDTIMsto the accesspoint which uses this information to determine themethod to perform on-link paging. AP uses DTIMbeacons ifreceiveDTIMsis set otherwise AP usesTIM beacons.

When MN moves to another subnet in dormantmode, it must send a normal registration messageand a dormant mode registration to SIP Proxy.An example dormant mode REGISTER request/response exchange is shown below:

Register sip:registrar.unbc.ca SIP/2.0From:Alice <Alice@unbc.ca>;tag=456248To:Alice <Alice@unbc.ca>Call-ID:123456@998sdasdh09Cseq: 1 RegisterContact:<sip:Alice@192.0.2.4>;mode=idle;AR=192.0.2.1;ReceiveDTIMs=true;idletime=300;MAC=00-11-11-80-38-36Expires: 3600Content-Length:0

SIP/2.0 200 OKFrom:Alice <Alice@unbc.ca>;tag=456248To:Alice <Alice@unbc.ca>;tag=2493k59kdCall-ID:123456@998sdasdh09Cseq: 1 RegisterContact:<Alice@192.0.2.4>;mode=idle;AR=192.0.2.1;ReceiveDTIMs=true;idletime=300;MAC=00-11-11-80-38-36Expires: 3600Content-Length:0

B. SIP PAGE Request Message

We extend SIP protocol with a new PAGEmethod. PAGE method is used to initiate paging.SIP Proxy server sends this message to the PagingAgent (PA) in the administrative domain. PA isimplemented in a SIP User Agent Client (UAC)module in order to process SIP methods like PAGE.When a PAGE request is received, PA starts paging

3

Fig. 1. SIP Paging Operation

Fig. 2. Paging Agent Receiving PAGE Method from SIP Proxy

of the mobile nodes whose MAC addresses aregiven as parameters to PAGE. PA’s task is to mapthis request down to the IP layer by communicatingwith the tracking agent.

Fig. 2 shows the implementation of PA. A cen-tralized TA can be colocated with PA as shown inthe figure. PA operates at SIP level. It receives aPAGE request from a SIP server and then it usesthe collocated TA to consult its paging cache andto initiate a paging on the wireless link at the AccessPoint (AP) to which the paged dormant mobileis presently associated. At the end of paging, PAreplies with a 200 OK message.

The header fields in a PAGE method are MACaddress of the mobile node, the old and new ARIP addresses and ReceiveDTIMs for one or moremobile nodes to be paged. PA will compare theold and new AR IP addresses. If they are different,

context transfer will be triggered (we will discussthe context transfer with paging in Sec. IV-A indetail).

III. SIP TRACKING

IEEE has defined the inter access point protocol(IAPP) for station (mobile node) handover signalingand station security context exchange [11]. SIPTracking uses IAPP handover messages but extendsthe operation from APs to ARs and TA. The mes-sages are transported using UDP. IAPP requires APsto send Add-mobile message to all other APs whenMN associates with the AP and to send Add-notifyto the old AP when MN reassociates with the AP.

MNs roam in WLAN cells associat-ing/reassociating with APs. SIP Tracking protocolis used to establish L2 location information calledpaging cache at the Tracking Agent (TA) whichis collocated with a SIP server. ARs discover theTA in the domain using a discovery protocol. SIPTracking operation is shown in Fig. 3.

A new entry to the paging cache is added whenMN associates/reassociates with an AP. The pagingcache contains MAC and IP addresses of MN andAP, a total of 20 bytes per MN in IPv4. ADD-notifyinteraction of SIP Tracking is sent by multicast byan AP when the mobile associates/reassociates withit. ADD-notify packet contains MAC address ofthe mobile in its data field. AR after receiving theADD-notify extracts AP’s IP address and mobilenode’s MAC address and sends an ADD-mobileUDP packet to TA.

TA gets MAC address and AP IP address fieldsfrom ADD-mobile messages received from AR.Note that at the time of an IEEE 802.11 association,the IP address of the mobile is not known. Themobile creates an IP address after it receives routeradvertisement from AR and then goes through du-plicate address detection in IPv6 or makes a statefuladdress configuration with DHCP server. AR knowsIP address of the mobile and sends ADD-mobilemessage to TA to update PC with MAC and IPaddresses. After this message, TA can have thecomplete location information for the mobile.

In soft-state (SS) tracking protocol, the pagingcache is soft-state, i. e. the entries expire and areremoved if not refreshed after timeout. The entriesmay be updated due to mobility of MN reassociatingwith a new AP. The new AP then sends a MOVE-notify UDP packet to the old AP through AR which

4

Fig. 3. Tracking Protocol Operation

in turn updates TA with ADD-mobile (Fig. 3). ARmay refresh entries by periodically sending ADD-mobile messages for all MNs for which AR has aneighbor cache entry. When MN leaves the WLANnetwork, its paging cache entry will remain until theentry times out.

In hard-state tracking protocol (HS), a reliablesignaling protocol such as TCP is used to establishstate at TA. When MN leaves the WLAN do-main, the state is removed explicitly using Remove-Mobile message of hard-state SIP Tracking proto-col.

IV. SIP PAGING

PA starts paging procedure by sending the PagingRequest message to TA. Paging Request messagecontains MAC address, AR address, ReceiveDTIMsand optionally a context data block fields for oneor more MNs. TA finds the corresponding entry inthe paging cache by the matching MAC address andreplaces it with MN’s IP address then sends a PagingRequest message to the AR which in turn sends itto the AP (Fig. 1).

After receiving Paging Request, AP starts on-link paging to wake up the node. On-link pagingis an extension of IEEE 802.11 power save modeprocedures in infrastructure networks. AP uses TIMor DTIM beacons depending on the settings of MN.For this purpose it uses ReceiveDTIMs field ofPaging Request message (if set DTIM otherwiseTIM beacons are used). MN wakes up and checksto see if its address is in the data frame and if it is,then it goes to the active mode in Layer 3.

AP sends a Paging Reply message back to TAand TA sends it to PA. After the reply is received,LS sends the URI and IP address to the SIP serverand then SIP server sends the INVITE message toMN. If Paging Reply is negative, TA sends PagingRequest message to the other APs in the samesubnet and then to APs in the other subnets, i. e. TAis in charge of the whole domain.

If no positive reply comes from MN within atimeout period, AR will send Paging Request mes-sage to the other APs in the distribution system in around robin fashion. If MN can still not be locatedunder this AR, TA will page the other ARs in thedomain. If no positive Paging Reply message isreceived, LS returns no URIs and therefore the SIPserver is required to go to the other administrativedomains [12].

A. Context Transfer with Paging

After waking up, mobile node may need to au-thenticate itself and perform layer 3 registrationsbefore VoIP application could be resumed. This isa time consuming process. If AAA/Mobile IP au-thentication is used, a full authentication could takeabout 52.5 ms [13]. If MN has already authenticateditself in another subnet of the same administrativedomain, we could transfer the MN’s context fromthe old AR to the new AR to reduce this delay.Context information is transfered in Paging Requestmessages extended with a context data block field.

When MN goes to dormant mode, a SIP dormantmode REGISTER request message is sent to the SIPProxy. SIP Proxy gets the information of the AR thatMN is attached from theAR extension header fieldof the dormant mode REGISTER request messageas described in Section II-A. If MN moves toanother AR in dormant mode, a new SIP dormantmode REGISTER request message will need to besent to SIP Proxy. Context transfer will not happenat handoff time because MN is still in dormant modeand it could move to another subnet before wakingup. SIP Proxy keeps track of the new AR and theprevious (old) AR and sends both AR addresses toPA in PAGE request message.

When PA receives a PAGE Request, it comparesthe new and old AR values. If they are different,PA will start the context transfer by sending acontext data request message to the old AR. OldAR responds with the context data message for this

5

Fig. 4. Paging with Context Transfer Interactions

MN as in IETF’s Context Transfer Protocol [14]. PAthen adds the context data field to Paging Requestmessage sent to TA (Fig. 4).

TA processes Paging Request message fields foreach MN to be paged. It sends a Paging Requestmessage for each different new AR under whichthere are MNs to be paged. TA includes the contextdata into the paging request and sends it to the newAR. New AR will extract and install the context andthen send the context in a paging request to the APthat MN is attached.

V. SIP TRACKING PROTOCOL ANALYSIS

Tracking protocol is analyzed and some simula-tion results are presented.

A. Soft-State Model

A continuous time Markov chain model of var-ious signaling protocols (soft-state, soft-state withexplicit removal or reliable trigger messages, re-liable trigger/removal messages and hard-state) isdeveloped in [15] which we adopt in this paper.

AR in Fig. 3 is the signaling sender that installs/updates the state at TA, the signaling receiver andstarts a refresh timer T of 2s, the same value asthe neighbor cache refresh timer. TA, after receivingstate installation/ refresh message enters/ updatesthe paging cache entry and starts a soft-state timeouttimer X of 90s. The entry is removed after X timesout because AR did not update it or because AR’smessage has been lost which is called false removal.

Markov chain starts at state 0 which is also its ab-sorbing state which models the handover occurring

Fig. 5. Soft-State Markov Chain Model of Tracking

resulting in an ADD/MOVE-notify message sent toAR (Fig. 5). State 5 occurs when AR removed MNfrom its neighbor cache but MN is still in the pagingcache at TA. State 2 is when both AR and TA statesare consistent. The chain transitions to state 1 fromstate 0 if ADD-mobile message is lost. AR willrefresh the state which is represented by a transitionfrom 1 to 2. The chain transitions to state 3 from 2when AR sends an ADD-mobile message to update(refresh) TA. The chain transitions to state 4 whenthe update/refresh message is lost. MN stays in thesystem for a mean session interval of 1 hour afterwhich its state is removed at AR. This is modeledas a transition from state 1 to 0. The transitionsfrom states 2 and 4 to 5 occur when TA has alreadyinstalled state for the MN. Eventually this state willbe removed which is shown by a transition fromstate 5 to state 0. Fig. 5 has transitions that modelthe false removal from state 2 and 4 to state 1.

The probability of loss for AR to TA com-munication is much lower than for MN to APcommunication over 802.11 links. Since state inAR will be triggered by MN’s L3 activity, we willuse 802.11 link loss rate as the probability of loss.Loss rates exceeding 20% have been reported in theliterature [16], [17], so we will use 0.15 as meanvalue since loss rates are much lower in the linksbetween AR and TA. Signaling state update intervaldepends on the mobility pattern of MN. We will usethe time a pedestrian takes to cross 30m WLANcell boundary to define the signaling state updateinterval (1/λu). This value is set to 60s for a slowmove user and 5s for a fast move user (together with

6

5s for T).The false state removal rate at receiver-soft-state

λf is the probability that the signaling messages arelost during the period of the soft-state timeout timervalue and is calculated by:

P⌊X/T ⌋l /X

which yields 0.0005625 for fast move and 0.000225for slow move users.

For D, the signaling channel delay, the number ofhops between AR and TA and association delay ofan MN with its AP are the main contributing factors.Association delay involves the phases of probing,authentication and reassociation each involves theexchange of several frames between MN and APand the total delay values between 100 and 400 msare reported in [16]. Therefore the association delaywill be the dominant factor for D and we use 200msas a typical value.

B. Hard-State Model

As a hard-state signaling protocol, SIP Trackingsends ADD-mobile messages using TCP. Since TCPguarantees message delivery, there is no need torefresh the state established at the signaling receiver,the TA. Also there is no need to keep soft-statetimeout timer at TA. However, a new message calledRemove-mobile is needed for explicit state removalat TA which is sent when AR no longer has MN inits neighbor cache. Remove-mobile messages cannot be triggered by MN disassociating from APbecause IEEE 802.11 does not require MN to senda disassociate request message.

Markov model of the hard-state SIP tracking isshown in Fig. 6. A new state, state 6 is used tomodel the loss of Remove-mobile messages. Weused 0.0001 for the false state removal rate,λw atthe hard-state receiver from [15].

C. Results Of Analysis

We obtained steady-state transition probabilitiesbased on the configuration parameters given in TableI for both soft-state and hard-state Markov chains.We will compare soft- and hard-state approachesfor inconsistency ratio, signaling rate and the costof signaling overhead.

Inconsistency ratio is the fraction of time thesystem spends outside of state 2 where both TA

Fig. 6. Hard-State Markov Chain Model of Tracking

TABLE I

SIP TRACKING PARAMETER VALUES

Name Parameter Valuesignaling state update interval 1/λu 5ssignaling channel delay D 200msloss rate Pl 0.15sender’s mean signaling life-time

1/λd 3600s

soft-state refresh timer value T 2ssoft-state state timeout timervalue

X 6s

false state removal rate atreceiver-soft-state

λf 0.0005625

message transmission timervalue for hard-state

R 800ms

false state removal rate atreceiver-hard-state

λw 0.0001

and AR have consistent values. Fig. 7 shows theinconsistency ratio as a function of the sessionlength.

Average signaling message rate is the total num-ber of messages required during a signaling sessionnormalized over the session length. Fig. 8 showsit versus sender’s mean signaling lifetime. Both ofFigs. 7 and 8 show that with increasing sessionlength, inconsistency ratio and average message ratedecrease. This means that WLAN users stayingactive for short periods of time induce less trackingmessages to be exchanged and also there is lesslikelihood of state inconsistency at AR and TA. Thefigures also show that fast moving users incur highermessage rates in both soft- and hard-state protocols.Hard-state protocol has smaller message rates thansoft-state protocol and inconsistency ratio is smallerfor hard-state than soft-state protocol.

7

0 200 400 600 800 10000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

sender’s mean signaling lifetime(s)

inco

nsis

ten

cy r

atio

SS+FastMoveHS+FastMoveSS+SlowMoveHS+SlowMove

Fig. 7. Inconsistency Ratio

0 200 400 600 800 10000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

sender’s mean signaling lifetime(s)

sig

na

lin

g m

essa

ge

ra

te

SS+FastMoveHS+FastMoveSS+SlowMoveHS+SlowMove

Fig. 8. Signaling Message Rate

The cost of signaling is the sum of the signalingmessage cost and the cost arising from the stateinconsistency. State inconsistency in tracking resultsin increased paging messages in order to correctlylocate MN, i. e. it increases paging latency. Wedefine paging latency as the number of extra sig-naling messages caused by state inconsistency, i.e.the paging request messages from AR to APs andthe paging reply messages from APs to AR. In caseof inconsistency, TA will page all the APs belowan AR, thus assuming 5 APs per AR, the latencywill increase to 10. We use this value to factor theinconsistency ratio in the overall cost formula.

Since we measure the cost versus T, soft-staterefresh timer, the cost is fixed for hard-state signal-ing. The cost versus T is shown in Fig. 9. Soft-stateoverall cost values for slow move users decreaseand approach to the fixed hard-state value up to

0 20 40 60 80 1000

2

4

6

8

10

12

soft state refresh time T(s)

co

st(

C)

SS+FastMoveHS+FastMoveSS+SlowMoveHS+SlowMove

Fig. 9. Overall Cost

0 0.2 0.4 0.6 0.8 10

0.2

0.4

0.6

0.8

1

1.2

1.4

loss rate

inco

nsis

ten

cy r

atio

SS+FastMoveHS+FastMoveSS+SlowMoveHS+SlowMove

Fig. 10. Effects of Loss Rate

some values of T after which the cost starts toincrease. On the other hand, using a low value for Twill increase the message signaling cost. The sameobservation can be made for fast move users. FromFig. 9, we can find optimal values to use for T,between 2 to 7 which are the values we used. Fora loss ratePl of 0.2 the optimal value range forT is between 2 and 5. From this we conclude thatlower loss rate enables longer optimal refresh timervalues.

Fig. 10 shows the effects of loss rate on in-consistency ratio. Hard-state protocol has very lowinconsistency ratio even at high loss rates due to thereduced signaling. For fast move users inconsistencyratio increases for both hard-state and soft-stateprotocols but hard-state ratios are much lower thanthe soft-state rate.

8

D. Effect of TA Placement on Tracking Performance

Tracking Agent could be replicated at each Ac-cess Router instead of a single centralized TA.Markov chain models will stay the same as de-scribed above whether TA is at SIP Proxy or ateach AR but some of the parameters used in solvingvarious performance measures will differ. If TAis placed at each AR, the signaling life time willbe shorter. If there are four ARs in a subnet, thesignaling life time for TA at each AR will be 1/4 ofthe signaling life time where TA is at SIP Proxy. Weused1/λd = 900s for TA at each AR and obtainedthe results for the inconsistency ratio, signalling lifetime and the overall cost.

We can see from Fig. 7 and Fig. 8 that both incon-sistency ratio and signaling message rate decrease asthe signaling lifetime increases. So the centralizedTA architecture will have lower inconsistency ratioand lower signaling message rate. We validated thiswith the analysis results, the resulting figures areomitted to save space.

After the analysis with the parameters describedabove, we also find that with the same soft staterefresh time value, the overall cost is slightly lowerwhen TA is at Proxy. With the same loss rate,the inconsistency ratio for TA at Proxy is slightlylower too. So we conclude that the centralized TAarchitecture will have a better performance than thereplicated TA architecture.

E. Simulation Results

We developed a simulation model of the trackingprotocol using OPNET in order to study the effectsof MN mobility (speed) on the paging cache sizeand the traffic introduced by the protocol. Fig. 11shows the simulation topology.

We placed 10 MNs under each of 4 APs initially(AP2-AP5 in Fig. 11). MNs start to move andchange directions randomly every 30s. For the soft-state protocol, the refresh interval is set to 2s andcache expiry time is set to 6s. We measured thepaging cache size at the TA as a function of thespeed of MN and the results are shown in Table II.We also measured the tracking traffic rate in bits/s.Each update packet is 75 bytes long including MACand IP headers. This would yield an average trafficrate of 75*40/2 or 12 kbps which was validatedusing simulation. Simulation results indicate that thetracking protocol is lightweight because its state size

Fig. 11. Tracking Protocol Simulation Architecture

TABLE II

CACHE SIZE VERSUS SPEED

Speed of MNs (m/s) cache size (bytes)1 8003 8005 8008 796.7558510 786.64437

is O(n) where n is the number of MNs and the trafficintroduced is almost negligible.

We also developed a simulation model to verifyinconsistency ratio analysis results in Fig. 7. Thesimulation topology is the same as in Fig. 11, exceptthat only one MN is in the domain at the right handside. The MN moves between AP2 and AP5. Thechannel loss rate is set to 0.2 and the channel delayis set to 200ms as in Table I. For the fast movescenario, the time for MN to move from AP2 toAP5 is 15 seconds. The refresh interval is 2s andcache expiry time is 6s. For the slow move scenario,the time for MN to move from AP2 to AP5 is 3minutes. The refresh interval is 5s and cache expirytime is 15s.

The results of the inconsistency ratio simulationis shown in Fig. 12. We can see that when thesignaling time is short, the inconsistency ratio forthe slow move is much lower than the theoreticalanalysis in Fig. 7. That is because in our slow movemodel, the first handoff happens after 60 seconds.As the signaling mean time increases, the simulation

9

0 200 400 600 800 10000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

sender’s mean signaling lifetime(s)

inco

nsis

ten

cy r

atio

SS+FastMove, analysisSS+SlowMove, analysisSS+FastMove, simulationSS+SlowMove, simulation

Fig. 12. Inconsistency Ratio Simulation

results get closer to the analysis results.

VI. A NALYSIS OF PAGING

We analytically derive the update and refresh pro-cessing loads for SIP paging induced at SIP servers.We compare the processing loads with another pag-ing system called Mobile IPv6 Hierarchical Paging(MIPv6HP). Hierarchical MIPv6 (HMIPv6) is anextension of Mobile IPv6 to support intra domainmobility [18]. MIPv6 Hierarchical Paging protocol(MIPv6HP) extends HMIPv6 with paging. Subnetsin a network are divided into different paging areas.When a mobile node moves into a new subnet inthe same network, if the old subnet and new subnetare in the same paging area, MN doesn’t perform alayer 3 registration; otherwise it registers with theserver called Mobility Anchor Point (MAP) [4].

Deterministic fluid flow and random walk mobil-ity models are used to model user (mobile node)mobility in WLAN cells [19]. For both models, weassume an hexagonal cellular network architecture.The coverage area of an AP is an hexagonal cellwith a perimeter ofLB and area of

√3L2

B

24. There are

BD APs andRD ARs in the network.For the deterministioc fluid flow model, the rate

of mobile hosts crossing a boundary of perimeterl ata speedv is R(l) = ρvl/π, when the user movementis uniformly distributed over[0, 2π]. The parameterρ represents the active user density.

The rate of layer 2 handoffs isR(l)BD. Sothe processing load for Add-Mobile messages isPTAA

ρvLBBD

π, wherePTAA

is CPU processing timefor Add-Mobile messages.

Fig. 13. CPU Processing Time-Fluid Flow

The number of refresh messages at TA isNT

,where N is the number of users and T is the refreshinterval for soft state protocol. So the processingload for refresh messages isPTAR

√3ρL2

BBD

24T, where

PTARis CPU processing time for refresh messages.

Similarly, the processing load for SIP REGIS-TER request messages isPProxyU

ρvLB

√BDRD

π, where

PProxyUis CPU processing time for SIP REGISTER

messages. The processing load for refresh messagesfor SIP REGISTER requests isPProxyR

√3ρL2

BBD

24TSIP,

where PProxyRis CPU processing time for the

refresh messages andTSIP is the refresh interval.CPUTA or CPUProxy in fluid flow model can

be expressed as:CPUProxy = PTAA

ρvLBBD

π+

PTAR

√3ρL2

BBD

24T+ PProxyU

ρvLB

√BDRD

π+

PProxyR

√3ρL2

BBD

24TSIP

If TA is at AR as in MIPv6HP, the CPU process-ing load will be divided into TA part at AR and PApart at SIP Proxy, where

CPUTA = PTAA

ρvLBBD

π+ PTAR

√3ρL2

BBD

24T

CPUPA = PProxyU

ρvLB

√BDRD

π+ PProxyR

√3ρL2

BBD

24TSIP

As for MIPv6HP, we can derive CPU process-ing time at the MAP similarly as inCPUTA:CPUMAP = PTAA

ρvLBBD

π+ PTAR

√3ρL2

BBD

24T+

PMAPU

(1−γ)ρvLB

√BDRD

π+ PMAPR

√3ρL2

BBD

24TMAPwhere

PMAPUand PMAPR

are MAP’s CPU processingtime for registration messages and refresh messagesrespectively.

Table III shows the configuration parameters andtheir values we use in plottingCPUMAP and

10

TABLE III

CONFIGURATION PARAMETERS FORPROCESSINGLOAD

ANALYSIS

Item Valueρ 1000/km2

LB 0.24kmRD 20PTAA

0.2msecPTAR

0.2msecPProxyU

0.15secPProxyR

0.015secT 2secTSIP 300secTMAP 300secPMAPU

0.5secPMAPR

0.05secγ 0.1

Fig. 14. Random Walk Mobility Model Architecture

CPUProxy. SIP REGISTER request message pro-cessing time is set to 150ms [20]. Numerical CPUprocessing times in seconds for SIP paging andMIPv6HP are plotted in Fig. 13. In the figure, thenumber of APs and velocity are represented witha tuple (AP, v) as one of the coordinates. Fig. 13shows that when v and the number of APs increase,the CPU processing time increases slower in SIPpaging than in HMIPv6 hierarchical paging. This isdue to the smaller processing time value used forSIP REGISTER request message.

For the random walk mobility model, the topol-ogy is shown in Fig. 14. The cell 0 is the center cell,and the cells labeled 1 form the first ring around ring0, i.e.r = 0. Each AR will be in charge of the samenumber of rings, R. Each ring r is composed of 6rcells. p is the probability that MN will remain inthe current cell [21].

If an MN is located in a cell of r(r > 0),the probability that a movement will result in anincrease in the distance from the center cell will be13

+ 16r

, and the probability that a movement willresult in a decrease in the distance from the centercell will be 1

3− 1

6r. The transition probabilitiesαr,r+1

andβr,r−1 represent the probabilities of distance ofMN from the center cell increasing or decreasing.

αr,r+1 =

{

(1 − p) if r = 0(1 − p)(1

3+ 1

6r) if 1 ≤ r ≤ R

βr,r−1 = (1 − p)(1

3−

1

6r) if 1 ≤ r ≤ R

Let πr,R be the steady state probability of state rwithin a subnet consisting of R rings.

π0,R =1

1 +R∑

r=1

r−1∏

i=0

αi,i+1

βi+1,i

πr,R = π0,R

r−1∏

i=0

αi,i+1

βi+1,ifor 1 ≤ r ≤ R

So the probability that an MN performs a SIP reg-istration isπR,RαR,R+1. CPUTA in the random walkmobility model can be expressed as :CPUproxy =πRαR,R+1PproxyU

+(1−πRαR,R+1)PTAA

T̄+

PproxyR

TSIP+

PTAR

T

where T̄ is the average residence time of MNstaying in a cell.

Fig. 15 shows the variation in the CPU processingtime as the average cell residence time is changedin the random walk model. Two values for R, 1 and4 are considered and p is varied from 0.2 to 0.6.

From the figure, we observe that CPU process-ing time decreases as the average residence timeincreases. This is due to the fact that MN has lowermobility as the average residence time increases.When R is larger, each AR will be in charge ofmore APs. When a MN moves from one AP cellto another, the probability of it still staying underthe same AR is higher. So, as Fig. 15 shows, thefrequency of SIP registration will be lower, and theCPU processing time will also be less.

A. Paging Delay Analysis

We will analyze transmission delays in SIP pag-ing. Processing delays and queueing delays willnot be considered. SIP signaling transmission delayanalysis on wireless links was presented for UDP,

11

1 2 3 4 5 6 7 8 9 100

0.02

0.04

0.06

0.08

0.1

0.12

Average Residence Time(sec)

SIP

Pro

xy P

roce

ssin

g L

oa

d

R=1,p=0.2R=1,p=0.4R=1,p=0.6R=4,p=0.2R=4,p=0.4R=4,p=0.6

Fig. 15. SIP Proxy Processing Load-Random Walk

Fig. 16. SIP Session Setup after Paging

TCP and RTP transport in [22]. We apply the UDPtransport analysis of [22] to SIP Paging.

When SIP Proxy recieves a SIP INVITE messagefor a MN in dormant mode, a paging request mes-sage is sent downstream to AP. AP will do the linklayer paging and MN will send a paging reply backto AP. AP forwards the paging reply packet to SIPProxy. Then SIP proxy knows that MN woke up andsends the SIP INVITE message to it. MN sends as200 OK message back and SIP proxy acknowledgesit with an ACK message. All the SIP messages aresent in UDP packets. We assume that the wiredlink is lossless. Let p be the probability of lossin the wireless link. Fig. 16 shows the interactionsinvolved in SIP session setup after SIP Paging.

For packets involved in SIP session setup afterSIP Paging, the back off timer after the ith trans-mission, Tr(i) doubles after each retransmission. So

TABLE IV

CONFIGURATION PARAMETERS FORPAGING DELAY ANALYSIS

Messages Message size(bytes)paging requset 140paging response 56SIP INVITE 728200OK 898ACK 428

Tr(i) = 2i−1Tr(1). Let us consider the deliveryof the messages such as Paging Request/ Reply,INVITE, etc. in Fig. 16. the original retransmissiontimer, Tr(1) will be Tr(1) = Dmessage + Dreply,whereDmessage is the message transmission delay(Paging Request),Dreply is the reply message trans-mission delay (Paging Reply).

Let Nm be the maximum number of transmis-sions, i.e. the total of the first transmission andany subsequent application level retransmissions.The normalized delayTmessage for the successfultransmission of the SIP message is as follows:

Tmessage = 11−qNm [(1 − q)(Dmessage)

+(1 − q)q(Tr(1) + Dmessage)+(1 − q)q2(3Tr(1) + Dmessage) + ...+(1 − q)qNm((2Nm − 1)Tr(1) + Dmessage)]

= Dmessage − Tr(1) + (1−q)(1−(2q)Nm)(1−qNm)(1−2q)

Tr(1)

where q is the probability of a transaction havingfailed, which means either the message or the replyis lost. q can be expressed as a function of the lossrate as1 − (1 − Pl)

2.The total delay is

D =N

∑

i=1

Tmessage(i)

where N is the number of messages necessary forSIP session setup with paging.

We consider IEEE 802.11b which provides awireless channel of 11Mbps for numerical analysis.The parameters are given in Table IV. The valuesare derived from [22], [23]. We set Nm to 5 and Nto 4.

We assume the wired propagation and link delayare 0.5ms and the wired link bandwidth is 100Mb/s.DTIM is used for link layer paging and beaconperiod is 3 and beacon interval is 100ms. Withthese values, link layer paging takes 300 ms. Fig. 17shows the delay in SIP session setup with paging, Din seconds evaluated at various loss rates between

12

0 0.05 0.1 0.15 0.20.03

0.035

0.04

0.045

0.05

0.055

0.06

wireless link loss rate

Pa

gin

g D

ela

y

3 hops5 hops20 hops

Fig. 17. Average paging delay in 11Mbps WLAN

0-20% on 11Mbps WLAN links. The number ofhops between TA and AP are varied from 3 up to20. The figure shows that SIP session setup delayincreases with both the loss rate and the number ofhops between TA and AP. For an MN connectedto an 11Mbps WLAN link, SIP session setup withpaging takes a minimum of approximately 35msecand this value could go as high as 56msec if TA is20 hops away and the wireless link loss rate is high.

B. Analysis of Context Transfer with Paging

We developed a simulation model of the pagingprotocol using OPNET to study the effects of con-text transfer. The simulation topology is the same asFig. 11 except that only one MN is in the domainon the right side.

In our simulation, the mobile node will movefrom AP3 under AR2 to AP4 under AR3. Twoscenarios are simulated here. In scenario 1, handoffhappens when mobile node is in dormant mode.Context transfer will not be used. Authentication ofMN will take 52.5ms.

In scenario 2, handoff happens when MN is indormant mode too. We will transfer the contextwith paging messages as described above. In bothscenarios, a SIP INVITE message will be sent forthe dormant MN to trigger paging after the handoff.We use PANA context as the context transferred inscenario 2. The elements in the PANA context arelisted in the Appendix [24]. The length of context is76+sizeof (ISP-Name) bytes. In our simulation, thecontext size is set to 100 bytes. We assume DTIM

TABLE V

PAGING RESPONSETIME IN DIFFERENTSCENARIOS

Scenario Paging response time (seconds)Paging without context transfer 0.089069Paging with context transfer 0.032057

beacons are used in layer 2 paging. The beaconinterval is 3 and beacon period is set to 10ms. Theresulting paging response times are shown in TableV. We can see that context transfer is efficient toreduce the paging response time.

VII. CONCLUSIONS

We introduced a new methodology called SIPPaging and Location Tracking to track L3/L2 loca-tion of the mobile VoIP nodes in a WLAN domainand then page and wake up the node when a callcomes. SIP Paging is done with a Paging Agentcomponent to alert a dormant mobile node after aSIP INVITE message is received in the domain. SIPLocation Tracking is done with a Tracking Agent atthe SIP layer which receives location updates fromthe access routers based on associations made tothe access points. Paging is triggered by PA and TAsends the paging request to the access point to whichthe mobile was associated last. The access pointuses on-link paging to wake up the mobile. Mobilenodes are required to register their dormant modestatus with SIP Proxy server using the extensionswe introduced. PA communicates with SIP Proxyusing a new SIP message we introduced.

We analysed two versions of the tracking proto-col, soft-state and hard-state for state inconsisten-cies, signaling rates and the overall cost. Resultsindicate how the parameters of the soft-state proto-col need to be tailored so that the cost is lower thanthe hard-state protocol. The results show that if theloss rates are lower higher values can be selected forsoft-state refresh timer. We simulated the soft-statetracking protocol to determine paging cache sizeand the tracking traffic rate. We determined that ourcentralized TA architecture gives better performanceresults than the replicated TA architecture.

SIP Paging is also analyzed in order to determinethe processing load at SIP Proxy using fluid flowand random walk mobility models and then comparethe results with Mobile IPv6 Hierarchical Pagingwhich uses replicated TAs. Paging delay is analyzedand the delay is shown to increase with the loss rate

13

and the number of hops between the tracking agentand the access points influences the delay strongly.Context transfer with paging analysis shows that thepaging response times can be drastically reduced ifthe context is transferred during paging.

We designed a new SIP Tracking protocol whichrelies on the access router creating/ refreshing en-tries in the paging cache triggered by layer 2 han-dover events of the mobiles. If IAPP is alreadyimplemented in a WLAN network, SIP Trackingshould be able to interoprate with IAPP. Extensionsrequired for this to the SIP Tracking are minimaland are left for future work.

Three more versions of SIP Tracking protocolcan be defined as a soft-state with explicit re-moval, reliable trigger and reliable trigger/removalmessages. These three versions can be analyticallycompared with the soft-state and hard-state versionswe have in the paper. SIP Paging transmission delayanalysis can be extended assuming TCP transportfor the messages. Replicating the Paging and Track-ing Agents has to be considered for fault tolerantoperation of SIP Paging. Implementation of SIPPaging and Tracking protocols is also left as futurework.

REFERENCES

[1] C. Ng and H. Chan, “Enhanced distance-based location man-agement of mobile communication systems using a cell coor-dinates approach,”IEEE Trans. on Mobile Computing, vol. 4,no. 1, pp. 41–55, January/February 2005.

[2] B. Patil, Y. Saifullah, S. Faccin, S. Sreemanthula, L. Arava-mudhan, S. Sharma, and R. Mononen,IP in Wireless Networks.Prentice-Hall, 2003.

[3] K. Kim, S. Pack, and Y. Choi,On Reducing Paging Cost in IP-based Wireless/Mobile Networks. Lecture Notes in ComputerScience (LNCS), August 2004, vol. 3090.

[4] T. Ozugur and B. Sarikaya, “Combining layer 2-layer3 paging for wireless LANs,” IEEE Transactions onWireless Communications, 2006, to appear. [Online]. Available:http://web.unbc.ca/˜sarikaya/publications/journal/04872.pdf

[5] G. Camarillo,SIP Demystified. McGraw Hill, 2002.[6] “Supporting mobility with SIP,” AR Greenhouse Web

Site, Telcordia Technologies, 1999-2005. [Online]. Available:http://www.argreenhouse.com/SIP-mobile/

[7] S. Zeadally, F. Siddiqui, N. DeepakMavatoor, and P.Randhawa,“SIP and Mobile IP integration to support seamless mobility,”in IEEE PIMRC, 2004.

[8] C. Politis, K. Chew, and R. Tafazolli, “Multilayer mobility man-agement for all-IP networks: Pure SIP vs. hybrid SIP/MobileIP,” in IEEE VTC, 2003.

[9] H. Lee, S. Lee, and D. Cho, “Mobility management based onthe integration of Mobile IP and session initiation protocol innext generation mobile data networks,” inIEEE VTC, 2003.

[10] N. Nakajima, A. Dutta, S. Das, and H. Schulzrinne, “Handoffdelay analysis and measurement for SIP based mobility inIPv6,” in IEEE ICC, 2003.

[11] “Trial-use recommended practice for multi-vendor access pointinteroperability via an inter-access point protocol across dis-tribution systems supporting IEEE 802.11 operation,” 802.11F,IEEE, pp. 1–67, July 2003.

[12] H. Schulzrinne and E. Wedlund, “Application layer mobilityusing SIP,” Mobile Computing and Communications Review,vol. 1, no. 2, pp. 1–9, 2001.

[13] A. Hess and G. Schaefer, “Performance evaluation ofAAA/Mobile IP authentication,” in In Proc. of 2nd Polish-German Teletraffic Symposium(PGTS’02), Gdansk, Poland,September 2002.

[14] J. Loughney, M. Nakhjiri, C. Perkins, and R. Koodli, “Contexttransfer protocol (CXTP),” RFC 4067, IETF, pp. 1–33, July2005.

[15] P. Ji, Z. Ge, J. Kurose, and D. Towsley,A Comparison of Hard-state and Soft-state Signaling Protocols. SIGCOMM 2003,Karlsruhe, Germany, Aug. 25-29, 2003, pp.251-262.

[16] A. Mishra, M. Shin, and W. A. Arbaugh,Empirical Analysis ofthe IEEE 802.11 MAC Layer Handoff Process. ACM ComputerCommunications Review, April 2003, vol. 33, no. 2.

[17] D. D. Couto, D. Aguayo, B. Chambers, and R. Morris, “Per-formance of multihop wireless networks: Shortest path is notenough,” inHotnets, Princeton, NJ, Oct. 2002.

[18] H. Soliman, C. Castelluccia, K. E. Marki, and L. Bellier,“Hierarchical mobile IPv6 mobility management (HMIPv6),”RFC 4140, IETF, pp. 1–28, Aug. 2005.

[19] S. Mohan and R. Jain, “Two user location strategies for personalcommunications services,”IEEE Personal Comm., vol. 1, no. 1,pp. 42–50, 1994.

[20] A. Dutta, K. D. Wong, J. Burns, R. Jain, A. McAuley, K. Young,and H. Schulzrinne, “Realization of integrated mobility man-agement protocol for ad-hoc networks,” inIEEE Military Com-munication Conference (MILCOM), 2002.

[21] S. Pack and Y. Choi, “A study on hierarchical mobile IPv6in IP-based cellular networks,”IEICE Trans. On Comm., vol.E87-B, no. 3, pp. 462–469, March 2004.

[22] H. Fathi, S. Chakraborty, and R. Prasad, “Optimizationof VoIPsession setup delay over wireless links using SIP,” inIEEEGlobecom, 2004.

[23] B. Sarikaya and T. Ozugur, “SIP paging for wireless LAN hostsfor VoIP,” in IEEE VTC 2005-Spring, Stockholm, Sweden, May30-June 1, 2005.

[24] J. Bournelle, M. Laurent-Maknavicius, H. Tschofenig,Y. E.Mghazli, G. Giaretta, R. Lopez, and Y. Ohba, “Use of contexttransfer protocol (CXTP) for PANA,” inInternet-Draft draft-bournelle-pana-ctp-03, June 24, 2005.

APPENDIX

PANA Context

Data Type LengthSession-LifetimeElapsed

unsigned32 Fixed

AAA-Key-int UTF8String Fixed(64octets)

ISP-Identifier Unsigned32 FixedISP-Name UTF8String VariableNAP/ISP SeparateAuthentication

Unsigned32 Fixed

14