VIEWS: 17 PAGES: 7 CATEGORY: Consumer Electronics POSTED ON: 10/1/2011
Mobile IP is to move the mobile node to maintain its connectivity and design. There are two versions of Mobile IP, namely Mobile IPv4 (RFC 3344, replaces RFC 3220, RFC 2002) and Mobile IPv6 (RFC 3775). Is still widely used in Mobile IPv4.
A PREDICTIVE HANDOFF APPROACH FOR MOBILE IP Da-Wei Zhang(1), Yi Yao(1) (1)State Key Lab. of Software Development Environment, BeiHang University Room 514, YiFu Building, 37 XueYuan Road, Haidian district, Beijing, 100083, China Phone:86-010-82317614,E-mail: dawei@nlsde.buaa.edu.cn Abstract: We present a predictive handoff mechanism is higher than given threshold, MN start L3 handoff. that can predict the occurrence of L2 handoff well. It can We compare our prediction method with scheme of ZB. be used as a trigger for Mobile IP to start network layer By mobility prediction on Zhou[6] ‘s “current” adaptive handoff ahead exactly or pre-allocate resources. Using model and getting position information by nonlinear least this approach link layer of MN detects signal changes square directly, our proposed handoff trigger scheme can based on mobility information to calculate the probability reliably predict future L2 handoff and improve the of coming handoff and trigger network layer. In this handoff performance of Mobile IP. paper we illustrate the theory of this approach in detail This rest is organized as follows: part II provides and provide experiments to evaluate the performance. related work and part III describes mechanism of this The experimental results show that this approach can approach. Performance evaluation is presented in part IV provide an exact trigger with low false alarm and miss and we provide conclusions in part V. detection and ensure the seamless handoff of Mobile IP. Keywords: Mobile IP; link layer trigger; mobility prediction; handoff; kalman filter; 2 ． L2 Handoff Probability Estimate by Mobility Information 1．Introduction In approaches of Mobile IP which uses L2 trigger, it is necessary to start L3 handoff before L2 handoff. In order In standard Mobile IP the handoff of MN between FAs to reduce effect of handoff we must ensure whole brings on communication interruption and packets loss. handoff completed in overlay area of preceding and To shorten handoff latency there are some rapid handoff subsequent subnet, i.e. we must get a proper time to start approach such as fast handoff [1] ,pre-registration handoff L3 handoff. If MN starts handoff later, there may be and post-registration handoff. To decrease packets loss packet lost because signal quality of preceding subnet has there are some smooth handoff approaches such as been unusable before handoff completes. If MN starts packet buffering transmitting mechanism[2] or dual-link handoff earlier, an unnecessary handoff maybe occurs registration[3]. In these approaches a proper L2 trigger when preceding subnet is usable all the while. So L2 time is very important. In rapid handoff, some handoff must trigger handoff to L3 ahead. We should gain L2 procedures should be initiated in advance of actual L2 handoff time by prediction, which lies on change of handoff, so MN must get a proper L2 trigger time. In signal quality in future. smooth handoff old FA must know start time to buffer In wireless network, L2 handoff in general depends on packets. But research about trigger time now is only the signal strength of current BS (base station) and restricted in “Sufficiently”. Literature [4] proposes to start selectable base stations，which is shown in Figure 1. L3 handoff together with L2 handoff. But this needs that There are four rules for selecting: L3 handoff must complete before L2 handoff, which is 1) Comparative signal strength rule: select the achieved difficultly. Zainab R. Zaidi and Brian L. strongest signal in any time Mark(ZB) [5] proposed a handoff trigger scheme based on 2) Comparative signal strength rule with relative autoregressive(AR-I) model of mobility, but because of threshold: MN only starts handoff when current signal deviation of extend kalman filter, it must get accurate below some given threshold and signal strength of position information from GPS or other orientation selectable BS stronger than signal strength of current BS. equipment. 3) Comparative signal strength rule with hysteresis We proposed a predictive handoff approach based on margin: MN only starts handoff when signal strength of mobility prediction for L2 of Mobile IP to exactly selectable BS is much stronger than current BS. calculate handoff probability and pass to L3. When MN 4) Comparative signal strength rule with hysteresis entered an overlapped wireless area, it keeps on detecting margin and relative threshold: This rule combines 2 rules the RSSI of APs and calculate the handoff probability by above. Only when signal strength of current BS is under mobility prediction model. Once the handoff probability given threshold and signal strength of selectable BS is hysteresis margin stronger than current BS, MN starts mean equal with current acceleration and is considered for handoff. This rule has been adopted in GSM. obeying the Rayleigh distribution. The continuous-time dynamic equation can be expressed as follows: ⎡. ⎤ ⎢x(t) ⎥ ⎡. ⎤ ⎢ .. ⎥ ⎢x(t) ⎥ ⎢x(t) ⎥ ⎡0 . 1 0 0 0 0 ⎤⎢ .. ⎥ ⎡0 0⎤ ⎡0 0⎤ ⎢ ⎥ ⎢ x(t) BS1 BS2 ⎢... ⎥ ⎢0 .. 0 1 0 0 0 ⎥⎢ ⎥ ⎢0 ⎥⎢... ⎥ ⎢ 0⎥⎥⎡ − ⎤ ⎢0 ⎢ 0⎥⎥ ⎢x(t) ⎥ ⎢0 0 α 0 0 0 ⎥⎢x(t) ⎥ ⎢α 0 ⎥⎢ax ⎥ ⎢α 0 ⎥⎡wx (t)⎤ ⎢. ⎥=⎢ ⎥ +⎢ ⎥ +⎢ ⎥⎢ ⎥ ⎢ y(t)⎥ ⎢0 0 0 0 1 0 ⎥⎢ . ⎥ ⎢0 0 ⎥⎢ − ⎥ ⎢0 0 ⎥⎣wy (t)⎦ ⎢ y(t)⎥ ay ⎥ ⎢ ⎦ ⎣ ⎢ ⎥ ⎢0 0 0 0 0 1 ⎥⎢ .. ⎥ ⎢0 0⎥ ⎢0 0⎥ ⎢ .. ⎥ ⎢ . ⎥⎢ ⎥ ⎢ ⎥ ⎢ ⎥ ⎢ y(t)⎥ ⎢0 ⎣ 0 0 0 0 α ⎥⎢ y(t)⎥ ⎢0 ⎦ ⎣ α⎥⎦ ⎢0 ⎣ α⎥⎦ ⎢ .. ⎥ ⎢...(t)⎥ Figure 1 Predictive Handoff ⎢... ⎥ ⎣y ⎦ In all, in whichever rule adopted, L2 handoff is ⎢ y(t)⎦ ⎣ ⎥ decided by signal strength, so we can get L2 handoff (3) time by studying the change of signal strength with time. . .. . .. In existing cellular systems, MN measures RSSI Where x(t ), x(t ), x(t ), y (t ), y (t ), y (t ) denote (received signal strength indication) to describe signal position, velocity and acceleration in x and y dimension strength. and α is the reciprocal of the random acceleration time Measured in decibels RSSI can be modeled as the sum − − of two terms: one due to path loss, and another due to constant. a x , a y are current acceleration in x and y shadow fading. Fast fading is neglected assuming that a low-pass filter is used to attenuate Rayleigh or Rician dimension. wx (t ), w y (t ) are white Gaussian noises and fade. Therefore, the RSSI from a particular cell, can be uncorrelated. It can be discrete as: formulated as [7] − RSSI (dB ) = R0 − ε lg d + ξ (1) X (k + 1) = Φ (k + 1, k ) X (k ) + U (k ) a + W (k ) (4) Where R0 is a constant determined by transmitted power, T wavelength, and antenna gain of cell . ε is a slope index ⎡ . .. . .. ⎤ Where X ( k ) = x( k ), x( k ), x ( k ), y ( k ), y ( k ), y ( k ) ⎢ ⎣ ⎥ ⎦ (typically, ε = 20 for highways and ε = 40 for microcells in a city), and ξ is the logarithm of the Besides state equation we should get enough shadowing component, which is found to be a zero-mean observation variables to estimate X (k ) . When MN can Gaussian random variable with standard deviation 4–8 dB. get the position coordinate directly by GPS, SOA or TOA d represents the distance between the MN and BS of cell , orientation systems, the observation equation is: which can be further expressed in terms of the mobile’s Z (k + 1) = H * X (k + 1) + V position ( x (t ), y (t )) at time t and the location of ⎡1 0 0 0 0 0⎤ base station ( a , b) : Where H = ⎢ ⎥ (5) ⎣0 0 0 1 0 0 ⎦ d = ( x(t ) − a ) 2 + ( y (t ) − b) 2 (2) Otherwise MN can estimate by RSSI of BS it detects. All in all, to get L2 trigger time we must know d (t ) , Assume that there are at least 3 BS available, so the observation equation is [3] i.e., we must estimate and predict movement state of MN. When different information got, we can predict handoff by Z (k + 1) = H ( X (k + 1), k + 1) + V (k + 1) = 2 method below. ⎡ ( x ( k +1) − x1 ) 2 + ( y ( k +1) − y1 ) 2 ⎤ ⎡ RSSI 1(k + 1) ⎤ ⎢ R1 − ε lg ⎥ ⎢ RSSI 2(k + 1)⎥ ⎢ ( x ( k +1) − x 2 ) 2 + ( y ( k +1) − y 2 ) 2 ⎥ ⎥ = ⎢ R 2 − ε lg 2.1 Handoff Prediction with Movement ⎢ ⎥ Model ⎢ RSSI 3(k + 1) ⎥ ⎢ ( x ( k +1) − x3 ) 2 + ( y ( k +1) − y 3 ) 2 ⎥ ⎢ ⎥ ⎢ R3 − ε lg ⎥ When exact position information can be estimated, ⎣... ⎦ ⎢... ⎥ we can predict future position of MN and then estimate ⎣ ⎦ future RSSI MN detects. Based on theory of maneuvering (6) targets tracking [8] [9], we propose a mobility predictive Where ( x1 , y1 ), ( x 2 , y 2 ), ( x 3 , y 3 ) denote the model: position of BS. Applying linearization method in extended According to “current” model of Zhou [6], we Kalman filter, where the linearization takes place about suppose that random acceleration of MN is one rank the filter’s estimated trajectory, the linearized correlated in time. The acceleration of next time has a measurement equation is: ⎡RSSI1 ⎤ signal strength rule with hysteresis margin K , then future ⎢RSSI2⎥ ^ ^ handoff criterion is: Z (k +1) = ⎢ ⎥ = h( X k +1,k ) + Hk Xk − Hk X k +1,k + Vk RSSI 2 (t 0 + jT , t 0 ) − RSSI 1(t 0 + jT , t 0 ) ⎢RSSI3⎥ ⎢ ⎥ = R 2 − R1 (10) ⎣... ⎦ (( x ( t 0 + t ,t 0 ) − x 2 ) + ( y ( t 0 + t ,t 0 ) − y 2 ) ) 2 2 (7) − ε lg (( x ( t 0 + t ,t 0 ) − x1) 2 + ( y ( t 0 + t , t 0 ) − y1 ) 2 ) ≥K And recursive equation is: ^ ^ − X k +1,k = Φ k +1, k X k + U k a ( k ) We can get L2 handoff probability at t 0 + jT : Pk +1, k = Φ k +1,k Pk Φ T k +1,k + Q k Ph (t 0 + jT ) = ⎛ (( x ( t 0 + t ,t 0 ) − x 2 ) 2 + ( y ( t 0 + t ,t 0 ) − y 2 ) 2 ) ⎞ ^ ^ T R 2 − R1 − K K k +1 = Pk +1,k H k k +1 [ H k +1 Pk +1, k + R k +1 ] −1 (8) ⎜ P⎜ ^ ≤e ε ⎟ ^ ^ ^ ⎜ (( x ( t 0 + t ,t 0 ) − x1 ) 2 + ( y ( t 0 + t ,t 0 ) − y1 ) 2 ) ^ ⎟ ⎟ X k +1 = X k +1,k + K k +1 [ Z k +1 − h ( X k +1, k )] ⎝ ⎠ (11) Pk +1 = [ I − K k +1 H k +1 ] Pk +1,k ( = P X 2 + Y 2 ≤ R2 ) Because of nonlinear disposal in observation equations the Kalman filter has sizeable deviation and = ∫∫ f x ( x ) f y ( y )dxdy x2 + y2 ≤R2 often tends to distortion, so we proposed another method Where to increase precision. That is calculating position ⎛^ sx − x1 ⎞ coordinate first by nonlinear least square such as X ~ N ⎜ x(t 0 + jT , t 0 ) + 2 , Pj ,k (1,1) ⎟ Gauss-Newton Method. Then the observation equation is ⎝ 1− s ⎠ equal to (5). Considering the mean of “current” acceleration is ⎛ ^ sy − y1 ⎞ Y ~ N ⎜ y (t 0 + jT , t 0 ) + 2 , Pj ,k (4,4) ⎟ ^ equal to a k , and sampling interval is T , then predictive ⎝ 1− s ⎠ step is j , and optimal fixed advancing prediction is: ^ ^ R2 = s 1− s x2 + y 2 − 2 ( 2 1 1− s x1 + y1 + 2 2 ) ( ) X j ,k =Φ Xk 1 j (9) ⎛ sx 2 − x1 ⎞ ⎛ sy 2 − y1 ⎞ 2 2 ⎡ T ⎤ 2 ⎜ ⎟ +⎜ ⎟ ⎢1 T ⎥ ⎝ 1− s ⎠ ⎝ 1− s ⎠ ⎢ 2⎥ R2 − R1 − K Where Φ1 = ⎢0 1 T ⎥ s=e ε ⎢0 0 1 ⎥ ⎢ ⎥ f x ( x), f y ( y ) are probability density function of X and ⎢ ⎣ ⎥ ⎦ Y. And covariance matrix is j −1 Pj ,k = Φ 1j Pk Φ 1j ( ) + ∑ΦT 1 j −1−i ( Qk Φ 1j −1−i ) T i =k So we know ^ ^ ^ ^ ( jT ) 2 .. . x(t 0 + jT , t 0 ) = x 0 + jT x 0 + x0 2 ^ ^ ^ ^ ( jT ) 2 .. . y (t 0 + jT , t 0 ) = y 0 + jT y 0 + y0 2 And actual position in t 0 + jT can be expressed as Gauss distribution: ^ The probability described by duplicate-integral can x(t 0 + jT ) ~ N ( x(t 0 + jT , t 0 ), Pj ,k (1,1)) not be expressed by elementary functions. But we can ^ calculate it approximately by weighted sum of 4 point in y (t 0 + jT ) ~ N ( y (t 0 + jT , t 0 ), Pj ,k (4,4)) integral circle area: We suppose that L2 handoff adopts comparative πR 2 R R R R It can be denoted as Ph (t 0 + jT ) = ( f x ( ) f y ( ) + f x (− ) f y ( ) 4 2 2 2 2 y = k 0 + k1t + k 2 t 2 + k 3 t 3 + k 4 t 4 R R R R + f x ( ) f y (− ) + f x (− ) f y (− )) + O( R 6 ) 2 2 2 2 Suppose ϕ = [1, t , t 2 , t 3 , t 4 ]T , When R2 = R1 , especially in some networks (such as WLAN), K = 0 , then And θ = [k 0 , k1 , k 2 , k 3 , k 4 ]T doesn’t change in Ph (t 0 + jT ) = several sampling period, so it can be estimated it by ⎛ A+ B ⎞ Φ⎜ ⎟ RFM: ⎜ 4( x − x ) P (1,1) + 4( y − y ) P (4,4) ⎟ 2 2 ⎝ 2 1 j ,k 2 1 j ,k ⎠ Pk ϕ k +1ϕ k +1 Pk T ^ Qk +1 = Pk − A = ( x 2 − x1 )( x1 + x 2 − 2 x(t 0 + jT , t 0 )) 1 + ϕ k +1 Pk ϕ k +1 T ^ Qk +1ϕ1ϕ1T Qk +1 B = ( y 2 − y1 )( y1 + y 2 − 2 y (t 0 + jT , t 0 )) Pk +1 = Qk +1 − (12) 1 + ϕ 1T Qk +1ϕ1 Where Φ is the Guass probability distribution function. θ k +1 = θ k + Pk +1 (ϕ k +1 ∆y n + k +1 − ϕ1 ∆y n +1 ) It meanings that at time t 0 + jT , L2 connects with new ^ 2 subnet with probability Ph (t 0 + jT ) . Then d i (t 0 + jT , t 0 ) can be calculated And future handoff criterion is: 2.2 Handoff Prediction by RFM RSSI 2 (t 0 + jT , t 0 ) − RSSI 1(t 0 + jT , t 0 ) = R 2 − R1 When MN can’t get enough information to 2 d2 determine its position, we can use recursive forgetting ε 1 − lg d1 ≥ K memory linear square (RFM) to estimate future RSSI. 2 When handoff MN can observe 2 RSSI at least: Then L2 handoff probability at t 0 + jT : 2 ( Ri − RSSI i ) Ph (t 0 + jT ) di = e ε + η i = ( x(t ) − ai ) 2 + ( y (t ) − bi ) 2 2 ⎛ ^ 2 ^ 2 ⎞ ⎜ η d 1 (t 0 + jT , t 0 ) − d 2 (t 0 + jT , t 0 ) ⎟ Where ηi is the deviation of d i 2 .Then = Φ⎜ 2 ⎟ ⎜ η D(d 1 (t 0 + jT , t 0 )) + D(d 2 (t 0 + jT , t 0 )) ⎟ 2 2 Because ⎝ ⎠ 1 Where D(d i (t 0 + jT , t 0 )) 2 is deviation of x(t ) = s x 0 + vx t + a x t 2 2 ^ 2 1 d i (t 0 + jT , t 0 ) y (t ) = s y 0 + v y t + a y t 2 2 3. Handoff Trigger Scheme for Mobile IP So we have d i2 = (( s x 0 − a i ) 2 + ( s y 0 − bi ) 2 ) 3.1 Estimation of Ahead handoff time Tcos t + (2( s x 0 − a i )v x + 2( s y 0 − bi )v y )t In fast or smooth handoff Mobile IP to start handoff + (v x + v y + a x ( s x 0 − a i ) + a y ( s y 0 − bi ))t 2 2 2 ahead with L2 trigger, we must estimate Tcos t .It means + (a x v x + a y v y )t 3 L3 handoff latency such as pre-registration or time to prepare buffer(dispose). In common MN is belonging to 1 2 1 2 + ( a x + a y )t 4 individual who often move on particular path. So network 4 4 handoff will often implement between some certain subnets. Because of unchangeable of network topology, handoff latency is usually unchangeable when MN switches between two subnets. So we can estimate this 4. Experiment and Performance Analysis latency by MN’s handoff history. MN records handoff latency as follows: Handoff Process 1 2 … Last n Tcos t 4.1 Experiment of Predictive Handoff Accuracy FA1 dispose T1 T2 … (T1+T2 (T1+T2 To FA2 + … + … +Tn)/n +Tn)/n+ In the experiment, we create simulation scene as Transmit Current RTT RTT same as ZB [5]. We assume a rectangular are with cells of data size 2000m × 2000m . Each cell contains one base … … … … … … … station located at the center of the cell. The mobile Table 1 History Handoff Delay station moving along the test trajectory receives signals According to table 1, when MN switches between subnets, it records disposing time and transmitting time. from the base stations. The parameter R0 is assumed to Before the next handoff, it will search handoff between be -30db for all base stations and ε is set to be 40. The the same FAs. For disposing time, average time of front N shadowing variance is taken as 6 dB. The hysteresis times is used as current predictive time (N can be margin is 3 dB. configure).Transmitting time could be measured by Ping. Test trajectories are generated using the linear dynamic system model of mobility [6]. The parameters of 3.2 Ahead Handoff criterion the Linear dynamic system-based model of mobility are follows: α = 1000s −1 set as , T = 1s and According to (11), L3 should get handoff probability σ = 1dB .The discrete command process u x (t ) and function from L2, the probability of L2 handoff after time Tcos t is u y (t ) are independent semi-Markov processes, each Ph (t 0 + Tcos t ) . Assume that the false alarm cost taking on five possible levels of acceleration comprising is C1 and the miss detection cost is C2, then handoff the set (-0.5,-0.25, 0, 0.25, 0.5) m / s . This set of 2 evaluation expression is C1 (1 − Ph (t 0 + Tcos t ) ) < C 2 Ph (t 0 + Tcos t ) acceleration levels produces an average speed of approximately 15 m / s . The initial probability vector That is π for the semi-Markov model (SMM) governing C1 u x (t ) and u y (t ) is initialized to a uniform distribution. Ph (t 0 + Tcos t ) > =β C1 + C 2 The elements of the transition probability matrix for the Where β is the probability threshold for Mobile SMM are initialized to a common value of 1/5. The initial position of the mobile is randomly selected within IP handoff. a 200 m × 200 m area centered at the origin of the coverage area. 3.3 Handoff procedure with L2 trigger Figure 2 shows a snippet simulation result. We record the predictive handoff time together with actual Above we have illustrate the estimate method of L2 handoff. We can see there is 2 false alarm after 1400s Ph (t0 + Tcos t ) and Tcos t , with which we can define and we can calculate the predictive difference time handoff procedure with L2 trigger: In L2: 1) Detects signal strength of surrounding BS. If there is at least 2 BS, goto 2), else goto 1). 2) Calculates Ph (t 0 + Tcos t ) in period and pass it to L3. In L3: 1) Estimates Tcos t and β 2) Receive Ph (t 0 + Tcos t ) from L2 in period. 3) Calculate Ph (t 0 + Tcos t ) , if Ph (t 0 + Tcos t ) > β , goto 4), else goto 1). 4) Starts L3 handoff. Figure 2 Predictive and Actual handoff For each simulation run, we calculate the ratio of missed detection, the ratio of false alarm, and the mean prediction interval of handoff time. The mean prediction interval is the average length of the interval from the predictive handoff time until the occurrence the actual handoff event. Simulation experiments are performed for j ={1s,…，15s}and β ={0.2, 0.3…，0.8}. The results are averaged over 300 simulation runs for each parameter setting. We compared experiment results with trigger scheme proposed by ZB[5] when ahead time is 11s and 15s in figure 3, 4 and 5. Because of Perform mobility prediction on Zhou’s “current” adaptive model instead of AR-I, and nonlinear least square to estimate position instead of EFK, the performance of our handoff trigger scheme is higher than ZB[5]. In figure 4 we can see when Tcos t = 2s , mean prediction interval is less than 0.3s, Figure 5 Mean Prediction Interval which is reliable to ensure the seamless handoff of Mobile IP. 4.2 Evaluation of Mobile IP Performance In the mobile IP protocol, the handoff latency may directly lead to the interruption of communication, which affects both TCP and UDP traffic, especially TCP. Because of exponential backoff of TCP timeout timer and MTU discover of tunnel on HA, TCP traffic can not recover immediately after handoff. So we test TCP performance when MN leaves home subnet on dual-link Mobile IP base on our handoff trigger mechanism (DLT)[3] comparative with standard mobile IP. For the test, the TCP traffic is generated by ttcp benchmark. The results are shown in Figures 5 and 6 (two vertical lines present the beginning and ending of handoff respectively). In figure 5, we can see standard Mobile IP Protocol will bring on packet lost during the handoff process, so Figure 3 Probability of Miss Detection the sender will use exponential backoff algorithm to resend packet. Because MN can’t receive this retransferred packet during handoff, TCP layer of receiver has to retransfer this packet again after timeout. The last time retransfer is caused by MTU discover mechanism. The MTU of the tunnel on HA is 1440 bytes while the size of the packets generated by ttcp is 1460 bytes. So when HA receives this packet with DF (Don’t fragment) bit set, it rejects this packet and sends an ICMP “Datagram Too Big” message to indicate the tunnel’s MTU. After this retransfer failure, due to the congestion control mechanism, TCP will enter slow restart process. In comparison with the Mobile IP, our system does not affect the TCP traffic. From figure 6, we can see that the TCP traffic doesn’t interrupt during handoff and there Figure 4 Probability of False Alarm is no packet lost or packet retransfer during the handoff process. Support Optimized Handover for IP Mobility, work in progress, draft-manyfolks-l2-mobilereq-02.txt, June 2002. [5] Z. R. Zaidi and B. L. Mark, A Mobility-aware Handoff 16000 Trigger Scheme for Seamless Connectivity in Cellular sequence 14000 Networks, IEEE Vehicular Technology Conference, v 60, n 12000 5, 2004 IEEE 60th Vehicular Technology Conference, 10000 VTC2004-Fall: Wireless Technologies for Global Security, 80000 2004, p 3471-3475 [6] Hong-Ren Zhou, Zhong-Liang Jing, Pei-De Wang, 6000 Tracking of Maneuvering Targets, (Beijing National 40000 Defence Publishing House,1991, 178-181) 20000 [7] H. H. Xia, An analytical model for predicting path loss in 0 urban and suburban environments, in Proc. PIRMC, 1996 0 0.5 1.0 1.5 2.0 2.5 [8] Tong Liu, Paramvir Bahl, and Imrich Chlamtac, Mobility time(s) Modeling, Location Tracking, and Trajectory Prediction in Figure 5 TCP Performance of Mobile IP Wireless ATM Networks,IEEE journal on selected areas in communications,vol. 16, No. 6, Aug 1998 500000 [9] R.A.Singer, Estimating optimal tracking filter performance 450000 for manned maneuvering targets, IEEE Trans. Aerospace 400000 and Electronic Systems,July,1970 sequence 350000 [10] Pubudu N.Pathirana. Mobility modelling and trajectory 300000 prediction for cellular networks with mobile base stations. 250000 International Symposium on Mobile Ad Hoc Networking & Computing Proceedings of the 4th ACM international 200000 symposium on Mobile ad hoc networking & computing. 150000 Annapolis, Maryland, USA, 2003. 213-221 100000 50000 0 0.5 1.0 time(s) Figure 6 TCP Performance of DLT 5. Conclusion We have proposed a predictive handoff approach for Mobile IP to support fast or smooth handoff. It integrates handoff of L2 and L3. Using a real-time mobility estimation to predict a future handoff, based on L2 handoff criterion in cellular networks --the difference of RSSI between current and target base stations. From the probability distribution function L3 of MN can start handoff ahead by comparing missed detection cost and false alarm cost as a threshold. This paper evaluates handoff prediction performance. Our simulation results show that the proposed handoff trigger scheme can reliably predict future L2 handoff and improve the Mobile IP performance. Reference [1] Tan C. L., Pink S and Lye K M. A Fast Handoff Scheme for Wireless Networks[A]. Proc. of ACM/IEEE WoW-MoM[C], 1999. [2] Ericsson, Perkins C. Low Latency Handoffs in Mobile IPv4 . IETF Internet Draft, Oct.2003. [3] Da-Wei Zhang, Yi Yao, Lin Tian, "A predictive Handoff Approach of Mobile IP on Dual-Link", International Conference on Communications, Internet, and Information Technology (IASTED CIIT 2005),2005.10.31~11.2 [4] James Kempf, Requirements for Layer 2 Protocols to