Modeling the Effect of Mobile Handoffs on TCP and TFRC Throughput

Transcription

1 Modeling the Effect of Mobile Handoff on TCP and TFRC Throughput Antonio Argyriou and Vijay Madietti School of Electrical and Computer Engineering Georgia Intitute of Technology Atlanta, Georgia , USA Abtract At the forefront of the recent advance in mobile network i the development of ophiticated mobility management mechanim that are uually baed on Mobile IP and it derivative. Baed on thee mobility management protocol, everal tudie that characterize tranport protocol performance have been preented. In thi work we move one tep further, and preent a joint performance evaluation model of TCP and TFRC, with the underlying IP-baed mobility protocol. We develop tochatic model that can characterize the protocol performance during handoff between heterogeneou wirele network like WLAN, cellular, or WMAN. We preent performance evaluation reult for validating the developed model under a et of different handoff cenario. The developed model can be utilized a a bai for further analytical evaluation of new mobility management protocol, allowing thu a fat and accurate comparion. I. INTRODUCTION The wide-pread ucce of IP-baed mobile and wirele device, ha created a need for new network protocol and architecture o that revenue-generating and eamle ervice can be provided to the end uer. One of the firt challenge that ha to be reolved i that of mobility management. The functionality that mobility management define conit of two eparate operation location management and handoff management. Currently, the protocol that i conidered to be a practical approach to the above problem i Mobile IP [1]. Depite however the elegant olution that Mobile IP offer, TCP performance uffer from everal problem due to handoff caued by hot mobility [2]. Neverthele, modeling the performance of TCP i uch a mobile environment ha received little attention. There are everal model for quantitatively analyzing TCP throughput [3], [4], pecifically in the context of the wired internet. On the other hand, coniderable amount of work on modeling mobility management protocol i available, and epecially for mobile IP and it derivative [5], [6]. The main focu of the thee mobility management modeling approache ha been the characterization of the ignalling and proceing load a a meaure of protocol performance. We are aware of only one recent tudy [7], where the author evaluate the performance of the TCP-friendly rate control protocol TFRC [8] during handoff between aymmetric network. However, they do not attempt to model or expre analytically thi behavior. II. SYSTEM MODEL AND ASSUMPTIONS The propoed ytem model aume an end-to-end connection with the data flowing from the correpondent hot CH to the mobile hot MH. According to the mobility cenario that we want to analyze, the MH i initially aociated with the firt acce network AN 1, and at ome time in the future it may move toward AN 2. The two acce network are modeled eparately, while they are both attached to the core network CN which i the Internet. We make the aumption that while the mobile hot i connected to AN 1, the tranport protocol i in teady tate. Subequently, after handoff i performed to AN 2, the protocol will reach the teady tate at ome time in the future. Many tudie have hown that the firt-order two-tate Markov chain i.e. the Gilbert path model can approximate fairly well the behavior of the Internet. However, in thi paper we adopt the Bernoulli path model which i a implification of the Gilbert model. The reaon why we elected a impler path model come from the need to model TCP on the packet level and obtain a cloed form reult. By following a more complex path model, and therefore TCP model, the overall ytem complexity will be unnecearily increaed. With the Bernoulli model the only quantity needed to model the channel i the average packet lo rate. The average packet lo rate i imply calculated by dividing the number of lot packet v. the total number of packet lot or received for each of the two path: #lot P 1,2 = 1 #lot + received However, in order to capture the effect of the handoff we define the handoff packet lo rate P H, which repreent packet loe caued only by handoff. III. TCP MODEL WITH MOBILE HANDOFFS Objective of thi performance model i to evaluate how the pecific parameter related to handoff affect the tranient packet lo, and ultimately the actual TCP throughput. More pecifically, we aume that TCP i in the congetion avoidance phae when a handoff take place. We alo model a TCP connection between two endpoint by conidering round, that have a duration of one RTT. We name the number of RTT

2 round that pa until there i a packet lo a the NL round figure 1. During thi round, TCP end a burt of packet equal to the allowed window, and wait for acknowledgment. Thi approach, which i baed on renewal theory can lead to a cloed form olution for the average throughput. Sequence no. Fig. 1. RTT round FTT 1 A BTT 1 C: No more ack received, RTO tart RTO FTT 2 E L2 handoff diruption time X B packet ent packet received packet lot ack received BTT 2 time Packet-level TCP behavior at the ender during a handoff. A. Handoff packet lo rate P H Figure 1 will be ued for explaining the behavior of TCP during a handoff event. The variable FTT 1 /BT T 1, and FTT 2 /BT T 2 decribe the forward and backward trip time for the two acce network repectively. According to thi figure, at time intant t A, L2 connection i lot and at time t A +RTO, the TCP ender retranmit the firt lot packet. Even more, if t B >t C + FTT 1, then clearly no duplicate acknowledgment will be received at the ender, and the only way for TCP to reume the data flow i by expiration of the RTO of the firt lot packet. On the other hand, a T H hrink, and if t B t C + FTT 1 the probability to receive a number of the lat packet cloe to time intant t C i increaed. If thi happen, then the ender would fat retranmit the firt lot packet, reulting into a fater recovery. If we rewrite the previou equation we have and becaue t C = t A + BTT 1 : t B t C + FTT 1 T H RT T 1. Therefore, the probability of handoff induced packet lo i the probability of that the one way network latency L N, i maller than the handoff duration T H,or: P H = P [L T H ] 2 We define a f LN the ditribution of the end-to-end network induced latency. However, the election of f LN i a deciion orthogonal to our work in thi paper. In our cae we will model the one-way Internet latency a a Gamma ditribution that occur mainly due to buffering at the routing infratructure. On the other hand, the diruption time T H i a parameter that depend of the mobility management a protocol. B. Throughput Having calculated the packet lo probability due to handoff P H, we now have to calculate the actual number of packet ent and lot during the handoff event. Firt, we will ee how the congetion window i evolved when packet lo lead only to triple-duplicate event. Then we will calculate the probability that a packet lo event led to recovery of the lot packet with TO and we will alo calculate the number of packet ent during the TO event. Congetion window in congetion avoidance: Concerning the evolution of the congetion window, we next how the equation that decribe it evolution baed on the average endto-end packet lo rate. By re-uing the formula developed at [3], that decribe congetion window evolution a a function of the packet lo rate, we have: E[Wx h ]= 2 2+b 2b1 P x P H 2+b b 6 3P x + P H 6 3 where P x + P H i the aggregate end-to-end packet lo rate for either path 1 or 2 x =1, 2. Now if packet loe caued by handoff or other reaon in the end-to-end path, and thee loe create only triple duplicate event TD, the throughput formula would be of coure imilar to [3]: 1 P x P T x = x + E[Wx h ] RT T E[Wx h 4 ]b/2+1 However, there i the probability that packet loe caued by the handoff event or the end to end path lead to a timeout TO. Thi i what we calculate next. Probability of TO and TD event: Aume that the ender ha ent a window w worth of packet in the current RTT round. The probability that the firt i packet are acknowledged in thi round, given that the ret are lot becaue of a handoff or buffer overflow in the wireline path 1, i given by P HW = p H + p 1. Becaue packet lo in the wired path and handoff lo are independent, the previou equation become: P HW w, i = 1 P H i P H 1 1 P H w + 1 P 1 i P P 1 w 5 In the above equation the term 1 P i P expree the probability that i packet are received before one i lot on a channel with packet lo rate P. Alo 1 1 P w i the probability that a lo happen in a NLR where w packet were ent. We alo write the probability that m packet acknowledged from the n ent in the lat RTT round [3]: { P m Gn, m = H 1 P H if m n 1 PH m 6 if m = n Now, the probability that at mot two of them were acknowledged i gk = 2 i= Gk, i. The probability for a TO to happen, will be one of coure if w<3, ince not enough duplicate ACK can be received. Now let u ee what happen when w 3. In that cae a packet lo could lead to a TO in two cae: Firt, if TCP end

3 uccefully le than three packet from a round of w packet end, thi would lead to a TO becaue not enough duplicate acknowledgement will be received at the ender. Second, there i alo the probability that the number of acknowledged packet i more than three in one RTT round, while a TO happen becaue in the next round le than three packet are acknowledged. The contribution of the aforementioned two cae in the probability that a packet lo i a TO for path x will give the final: P TOx w = 2 w P HW w, k+ P HW w, kgk 7 k= k=3 Note alo that P TDx w =1 P TOx w. TO duration: Beide the probability for a TO to happen, we have to calculate the average duration of a timeout, baed on the probability that t B > t C + FTT 1 figure 1. The number of the retranmitted packet that will be lot leading to further RTO increae determine when the data flow will be reumed on the new link. Given that the duration of a handoff i T H = t B t A, the number of retranmitted packet depend on the duration of the timeout period, and ubequently on the number of exponential growth the TO timer experienced. We know that for TCP, j conecutive RTO event will have a duration: { 2 = j 1RT O if j< j 6RT O if j 7 where RT O i the initial value of the retranmiion timer. By inverting thi expreion, we obtain the number of RTO expiration: { log 2 RT O j = +1 if 2 6 1RT O 9 RT O +5 otherwie Therefore, the number of experienced TO will be obtained L by taking the of h RT O. Thi value will give the number of RTO expiration and the number of retranmitted packet. The previou equation finally give the expected number of packet ent during t C <t<t B : { E[S h ]= log 2 TH +1 if T H 2 6 1RT O TH +5 otherwie 1 and the duration of the induced TO i E[ ]+T H. Final throughput formula: Finally, by combining all the previou equation, the complete TCP throughput model that conider handoff between aymmetric link i given by: 1 P x Tx h P = x + E[Wx h ]+P TO E[S h ] RT T x E[Wx h ]b/2+1+p TO T H + E[ ] With thi formula we are able to etimate the throughput of TCP connection that i characterized by packet lo rate P x and RT T x after a diruption of T H econd happen. For the new path the P H will be zero which mean that the throughput etimate will only depend on the characteritic of the new path. It i intereting to note that we have decribed the TCP throughput a a function of the diruption time T H, which i eentially controlled by the mobility management cheme in ue, and the handoff packet lo rate P H. St RTT 1 link 1 link 2 B : A: L2 lot with link1 Convergence time t_inter t_inter 2 1 t_interx Fig. 2. B: L2 with link 2 etablihed time TFRC behavior during IP layer handoff. IV. TFRC THROUGHPUT After dealing with TCP, we proceed with the characterization of the rate control protocol TFRC [8]. Our goal i to quantify the effect of mobile handoff on a rate control protocol which ue a completely different congetion control algorithm. In thi paper we conider TFRC to be implemented on top of UDP. TFRC ue the cloed form equation for TCP throughput in order to regulate the ender output rate: T TFRC = 11 2p RT T 3 + t 3p RT O3 8 p1 + 32p2 where i the packet ize, RT T i the RTT etimate, and t RT O i the value of the retranmiion timer. In addition TFRC follow a packet pacing algorithm at the ender [8]. Note that, equation 11 doe not repreent the actual TFRC ending rate but only an upper bound for it. The actual output rate of TFRC i calculated uing the algorithm 1. In thi algorithm, repreent the packet ize, tld i time when the rate wa lat doubled, tmbi i the maximum back-off time 64 econd by default, and X recv i the average receive rate. If p i zero, no packet lo ha yet been een by the flow and in thi phae, the TFRC ender emulate low tart of TCP by doubling the tranmiion rate every RTT. However, the main feature of TFRC i that it react lowly and cautiouly to RTT change and thi mean that the throughput etimate change lowly when compared with TCP. Therefore, a udden change in the RTT and bandwidth of a link, a i the cae with handoff, will lead to coniderable packet loe. We will now attempt to model thi behavior. We tart by identifying the RTT etimation procedure that TFRC ue. Given a decay factor df,then-th RTT etimate i calculated a follow: RT T n = df RT T n 1 +1 df RT Tcur, with RT T cur be the lat real RTT meaurement.

4 Algorithm 1 TFRC rate etimation. 1: if p> then 2: X calc = T p, RT T, T 3: tfrc x = max[minx calc, 2 X recv, tmbi ] 4: ele 5: if tnow tld RT T then 6: tfrc x = maxmin2 tfrc x, 2 X recv,/rtt 7: end if 8: tld = tnow 9: end if When the TFRC ender doe not receive feedback during an entire RTT, it cut the output rate in half. If we aume that at intant t B, where handoff i over the ender end at leat on packet and receive feedback, then in a period of RT T 2 the ender will receive the firt feedback report. So for the handoff duration T H, the ender will reduce the rate in half every RT T 1, ince thi the lat RTT etimate. Therefore the total number of packet ent during T H will be gradually reduced with a total number: E[S h ]= T H RT Tcur i= 1 2 i T TFRC RT T 12 Therefore, for two link with different characteritic, p 1,RTT 1 and p 2,RTT 2, for link 1,2 or when t < t A or t>t C, the actual rate at the ender will be given by the baic TFRC formula: T A B + TFRC = max when t A <t<t B then: T AB TFRC = 1 T H 2 mint TFRC, 21 P 1,2 T TFRC, T H RT Tcur i= 1 2 i T TFRC RT T tmbi If t B <t<t C, the ender gradually tart to catch up with the new link characteritic: TTFRC BC = 1 ncv T TFRC RT T n cv RT T n 15 t cv n=1 We define a convergence time, t cv, the time needed for TFRC to obtain it fair hare of the bandwidth on the new link. Then we can rewrite the equation for RT T n : RT T n = df n RT T + df n 1 1 df RT T cur df RT T cur 16 which if we olve for n give: RT T n 1 df RT T cur n cv = log df RT T 1 df 1 df 17 RT T cur Therefore, n cv will give the number of RTT round needed in order for the RT T n etimate to converge to the RT T cur. In the handoff cae, RT T cur repreent the RT T 2 of the new link while RT T i the lat etimate we had and it i RT T 1. Practically we would like for RT T n to be cloe to 9% of the RT T cur. So the total convergence time will be: t cv = T H + n cv RT T n 18 The throughput i calculated from equation 13, 14, 15. V. EXPERIMENTS FOR MODEL VALIDATION The topology that form the bai of the imulation aume the handoff between two WLAN acce point. Each AP i connected to a LAN which i configured a follow: The value for bandwidth and delay of the LAN link are 1Mbp/1m and 5Kbp/1m for the GFA link, and two FA link, repectively. The elected parameter for TCP and TFRC in the imulation are: W max i 6MB, W i one egment, MSS i 146 byte, the RT O i 2 mec, and the i 15 byte. The cae tudy imulated i now decribed Initially, the MH initiate an FTP data flow from the correpondent hot CH. According to the cenario, at the time intant ec, the MH tart moving away from the firt acce point AP, at a peed of 1m/ec, and i heading toward the other AP. The FA follow the Mobile IP procedure in order to notify the HA after the MH regiter with new FA. In the cae of Hierarchical MIP, the GFA i the one that handle the handoff from the two FA that correpond to the two acce point: AP 1, and AP 2. Finally, for the MIP-RO cae we configured the MH to end a binding update directly to the CH. For all thee experiment we ued the n-2 [9] network imulator. HMIP and MIP-RO: With thi experiment, we want to evaluate the effect of the handoff diruption time T H,onthe throughput of a eion between the erver and the client. Throughput reult are hown in figure 3. A expected, even with MIP-RO, TCP throughput uffer coniderably when diruption time i increaed. The propoed model, predict a logarithmic decreae in throughput a the packet lo rate i increaed, which of coure depend on the duration of the diruption. Concerning the throughput for the combination of HMIP/buffering, we can oberve in figure 3 that TCP throughput remain coniderably table until the point where the diruption time come cloe to the average RTT of the end-to-end eion. Packet are buffered in the old AP and forwarded to the new AP, but after the buffer overflow, packet are dropped and the throughput decreae. Thi mean that after the point where the forwarding buffer i full, the throughput will continue to decreae very fat ince any new packet will be dropped. In figure 4 we preent reult for TFRC throughput. TFRC ue a rate control algorithm that react lowly to packet lo and RTT fluctuation. In addition TFRC make ue of a packet pacing algorithm that arrange the packet in time. A hown in figure 4, the combined MIP-RO/TFRC uffer from minimal packet loe and throughput degradation. With the addition of a buffer in the previou AP, thee loe are reduced further. Alo it i important to note that ince TFRC pace the packet

5 Throughput Kbyte/ Analytical for MIP-RO Simulation for MIP-RO Analytical for HMIP+buffering Simulation for HMIP+buffering Stabilization time ec Analytical for TCP+MIP Simulation for TCP+MIP Analytical for TFRC+MIP Simulation for TFRC+MIP Average diruption time T H ec Fig. 3. Throughput reult for a TCP eion of 15 econd and handoff between two different WLAN ubnet. Fig. 5. TFRC. Required recovery time veru diruption time for both TCP and Throughput Kbyte/ Analytical for MIP-RO Simulation for MIP-RO Analytical for HMIP+buffering Simulation for HMIP+buffering Average diruption time T H ec Fig. 4. Throughput reult for a TFRC eion of 15 econd and handoff between two different WLAN ubnet. in time, the buffer at the previou AP hould not overflow frequently ince it doe not receive packet burt a with TCP. Recovery period: We define a recovery period the time duration that the MH ha to be out of the handoff tate, o that the achieved throughput in thi time period reache the nominal for thi link. Eentially we want to know how fat TCP will recover from a handoff with a pecific duration T H. The anwer to thi quetion can alo give u ueful information concerning the effect of the handoff rate on the TCP throughput. Figure 5, preent reult for thi experiment when baeline Mobile IP wa ued for both TCP and TFRC protocol. We ee that for our model the required recovery period, i increaed exponentially diverging from the real meaurement which are not o peimitic. We believe that thi behavior i due the interpretation of more loe a a TO indication intead of TD. However, the TFRC model doe not have to claify packet lo type, allowing thu a more accurate etimation a figure 5 alo indicate. VI. CONCLUSIONS In thi paper we preented a model for tudying the effect of mobile handoff on two tranport protocol, namely TCP and TFRC. The model wa found to be accurate for TCP in both the cae where HMIP and MIP-RO were ued a the underlying mobility management protocol. However, the TFRC model predict the expected throughput with even better accuracy, due to the impler protocol algorithm. For example the wort cae error for the TCP model wa nearly 22% while for the TFRC model it wa 13%. We alo introduced in thi paper the notion of the recovery period. The low-reponive rate control algorithm of TFRC, require le time in order to recover when compared with TCP. However, we found that a the diruption time i increaed, TFRC uffer from more packet loe than TCP, due to the low-reponive algorithm. REFERENCES [1] C. Perkin, IP Mobility Support for IPv4, RFC 322, January 22. [2] H. Balakrihnan, Challenge to Reliable Data Tranport over Heterogeneou Wirele Network, Ph.D. diertation, UC Berkeley, [3] J. Padhye, V. Firoiu, and D. Towley, Modeling TCP Reno Performance: A Simple Model and It Empirical Validation, IEEE/ACM Tranaction on Networking, vol. 8, no. 2, pp , April 2. [4] B. Sikdar, S. Kalyanaraman, and K. Vatola, Analytic model for the latency and teady-tate throughput of TCP Tahoe, Reno and SACK, in INFOCOM, 21. [5] C. Blondia et al., Performance Analyi of Optimized Smooth Handoff in Mobile IP, in MSWiM, 22. [6], Performance Comparion of Low Latency Mobile IP cheme, in WiOpt, 23. [7] A. Gurtov, Effect of Vertical Handover on Performance of TCP-Friendly Rate Control, in ACM Computer Communication Review, 24. [8] M. Handley, S. Floyd, J. Pahdye, and J. Widmer, TCP Friendly Rate Control TFRC: Protocol Specification, RFC 3448, January 23. [9] Network imulator. [Online]. Available: