An Infrastructureless End-to-End High Performance Mobility Protocol

Transcription

1 Proceeings of the 5 IEEE International Conference on Electro Information Technology, 5, Lincoln Nebraska, May 5 An Infrastructureless En-to-En High Performance Mobility Protocol Saneep Davu, Rai Y. Zaghal an Jave I. Khan {savo, rzaghal, jave }@cs.kent.eu Networking an Meia Communications Research Laboratories Computer Science Dept., Kent State University 33 MSB, Kent, OH 444 ABSTRACT Mobile IP is offers isconnection free hanoff by assuming availability of infrastructure. It requires intermeiate software agents in the network to be eploye ahea of time to circumvent IPs normal moe ientity base routing. This infrastructure base mobility management though offers connectivity but incurs significant hanoff an tunneling elays along with eployment costs. In this paper we emonstrate an alternate mobility scheme which oes not require any such infrastructure an uses only en-point technique an yet provies much faster loss-free hanoff. This En-to-En scheme name Interactive Protocol for Mobile Networks (IPMN) neither requires any functional changes to the network layers on the sening an receiving host machines nor an infrastructure in the network. It intelligently performs hanoff base on information provie by MAC Layer. The network aress change is hanle by renewing the existing connections by manipulating the TCP/IP stack at the en-points. Besies, the ifference in eployment scenarios, the IPMN offers blazingly fast event base hanoff an much faster an simplifie transport (no tunneling elay) than MIP. We provie a etail moel base performance comparison between the two. 1 INTRODUCTION Mobile IP (MIP) [1] is being use wiely currently to hanle mobility in TCP/IP network. In TCP/IP the ientity of a noe an all corresponing routing is inicate an manage by the same ientifier. At application level a connection is ientifie using the four tuple <source aress, source port, estination aress, estination port>. A L3 hanoff changes network aress resulting in stale non-routable connections. MIP aresses this problem by aing an inirection to the routing mechanism in the form of home agent (HA) an foreign agent (FA). In MIP a hanoff is etecte when a Mobile Noe (MN) leaves the present service area an enters a new agent s service area. The ifference between the service areas is ientifie by sensing a form of Agent Avertisements (AA). This requires FA to perioically an continually broacast a small signal. The next issue is to circumvent the static ientity attache routing. This is one by registering the MN to a new foreign agent an upating the same at the HA. To isable the MN s ientity base routing a tunnel is create from MN s HA to FA, where the actual IP packet ries insie another packet from HA to the FA. The return path from MN to Corresponing Host (CH) is irectly route through the FA without requiring any intervention of HA. The solution logically hanles the IP mobility, but also shows the following artifacts. The first is relate to inirection. This creates a triangular pattern an hence this system is also referre to as triangular routing. Resulting forwar path an return path also becomes asymmetric. The secon is performance. The mobility etection epens on beacon. The timer base avertisements place an upper boun on the hanoff elay. These elays egrae the transport an application level performance. Too fast a timer creates excessive network overhea, on the other han a slow beacon (which is use in most cases), elays the entire mobility management an in practice reacts aversely with the upper layer transport timers. Currently, it almost prohibits the use of connection oriente transport such as TCP [17,18]. Besies the above artifacts, MIP also requires infrastructure. It requires a significant mechanism to hie the aress change from higher layers. It also ictate the strong nee for an infrastructure implying greater eployment an management costs. Specifically it requires new IP layer in MN, HA & FA. In many scenarios however, the above moel of mobility hanling is inconvenient. Many are sensitive to triangulate long roun trip elay. For example, consier a remote surgery requiring highly reliable an time sensitive ata elivery. Many avance applications require connection oriente transport. All avance istribute applications funamentally require reliable connection. Even if they are force to use UDP- they have to anyway replicate the reliability an connection states at the upper level. Also in many scenarios the infrastructure might not be available at all. An example of this is a corporate business having branches sprea all over the country. It cannot change the alreay existing infrastructure to suit the mobility management rather they woul like something to work with the alreay existing infrastructure. In this paper we emonstrate a mobility solution which is base on only en-point technique. Further the change at the en-points is mostly restricte at application level. It is base on a new en-point network software architecture calle interactive transparent networking paraigm (InTRAN) [1, 11]. This new en-point network software architecture allows event base access to protocol states 1

2 Proceeings of the 5 IEEE International Conference on Electro Information Technology, 5, Lincoln Nebraska, May 5 by network layer processes or by L7 processes. The new scheme oes not require the battery of pre-eploye FAs in the Internet (thus the nee of an infrastructure!). Besies, offering avantage of being infrastructure-less, this also offer major performance avantage. It oes not require HA an thus any route inirection, can avoi triangulation, is loss-free, eliminates tunneling overhea (significantly reuce jitter or elay) an above all offer much faster hanoff. Before presenting this new approach, in the following section we first present a brief review of the current mobility hanling techniques in TCP/IP framework. Section 3 then explains the new interactive transparent networking paraigm. Section 4 then presents the propose scheme. In section 5 we show the etail performance moel of the scheme. For comparison we also present a moel of MIP. Finally in section 6 we present a comparative performance analysis of the propose scheme in managing selecte scenarios. RELATED WORK Mobility solutions can be categorize in two broa approaches- one which hanles mobility fully at the networking layers, hiing any changes in the network structure from the en systems. The others, those hanle mobility at the transport layers an involves only the en systems. Below we briefly present these two approaches..1 Network versus Higher Layer Solutions Mobile IP [1] provie a first effective crack of hanling mobility. It is now an IETF stanar an most common protocol in practice. Though this solve the problem of hanling stale connections, it introuce lots of communication overhea an longer triangular routing paths. Triangulation of routes can be reuce by enhancing Mobile IP with Route Optimization [] presente by Perkins an Johnson. The CH optimizes the route by maintaining a bining cache for each MN s care-of-aress. Base on this information a irect tunnel from CH to MN is create efying any further involvement of HA. MN nees to upate CH of its current care-of-aress explicitly uring each hanoff, accruing alreay complex MIP. Hint base hanoffs in [3] use triggers from L as hints to L3 about an impening hanoff. This eliminate the cost an complexity of avertisement base movement etection. [] also uses assistance form link layer to perform fast Mobile IP hanoff through MAC briging. The RAT (Reverse Aress Translation) architecture [4], base on the network aress translation (NAT) protocol, uses packet re-irection service between CH an MN to support IP mobility. However, both the techniques require functional enhancement of the layers involve. Higher layer solutions mostly involving TCP hanle mobility ifferently, using the split connection approach. Inirect-TCP [5, 9] splits the wire an wireless parts of the connection using the Base Station (BS) as a common en point. The state information of the connection is transferre between the ol BS an new BS transparently uring hanoffs. MSOCKS [8] achieves connection reirection using split connection proxy. TCP-R[1] is base on an iea- same as ours, renewing the connection to hanle the new IP aress. But their scheme bases its iea on the conventional timer base approach an hence lacke the robustness an scalability. Freeze-TCP[6], though oes not hanle mobility irectly, ais the TCP performance by freezing the connection uring perios of isconnection.. Our Approach In compare to the above approaches the solution we propose has several istinguishing characteristics. First the mobility reactions are event base as oppose to perioicity epenent. Seconly, the solution oes not require any significant overhauling of the existing protocols rather works in L7. The InTraN enables networking layer protocol events to be orerly an securely subscribe an respone at L7 without requiring functional moification of the existing protocols- with much simpler protocol meta-engineering. This is an interesting iversion from previous thinking that application layer solutions may slow own the overall performance. We show by moeling an analysis that the premium is much less than suspecte- almost negligible. On the other han- since mobility tracking an corresponing reconnection responses can be generate at application layer- its eployment is much easier an also the strategies can be mae much more sophisticate an powerful. Inee the scheme we show incorporates an extens the strengths of several previous approaches. It coul thus offer a ramatically improve hanoff performance- far out weighing the cost. 3 INTERACTIVE TRANSPARENT NETWORKING The framework assumes a transparent encasement of protocol functionality by making a subset of its internal protocol states an events visible to its service subscriber layer. It provies means to allow programs in upper layers to subscribe liste events in lower layers an be notifie when these (state transitions) occur. Upon notification, subscriber processes can further pull up the liste service states. Typically an application program can register an event hanler chil process calle Transientware to hanle a specific event. The transientware is an application level programmable process an thus can perform various aaptive intelligent actions at L7. Figure-1 shows the general architecture of an example interactive version of

3 Proceeings of the 5 IEEE International Conference on Electro Information Technology, 5, Lincoln Nebraska, May 5 user space system Kernel Application TCP Connectio Socket API 1 TCP kernel Subscription API 5 Event Monitor T-ware (1) Signal Hanler 3a 3b T-ware () 4a 4b Event Information Figure 1. TCP interactive extension an API. 7 Probing API T-ware (n) 6b Connection State 6a 4 INTERACTIVE PROTOCOL FOR MOBILE NETWORKS (IPMN) 4.1 Overview of Wireless Environment In a wireless scenario L hanoff is initiate by the MN when the Signal/Noise ratio an Signal strength of the current Access Point (AP) falls behin a certain threshol. To have a better unerstaning of the architecture let us first explain briefly the two wireless scenarios existing an how our architecture robustly hanles both while performing a L hanoff. When an MN receives service from more than one access point at any point of time then their areas of coverage are overlapping, which generally happens at the bounaries of the cells an this is calle Overlapping cell bounaries. Overlapping cell bounaries as shown in figure (a) have better performance in our approach as applications probing link layer yiels information about a b Fig Cell Bounaries a) Overlapping b) Non-Overlapping TCP. Upon opening the socket, an aaptive application may bin a Transientware moule to a esignate TCP event by subscribing with the kernel. This is represente by arrows 1 an in figure-1. The bining is optional; if the application chooses not to subscribe, the system efaults to the silent moe ientical to TCP classic. When the event occurs in TCP, the kernel sens a signal to the upper layers (3a) an at the same time it saves the event information (3b). A special hanler catches the signal an probes the kernel for the event type (4a, 4b), The hanler then invokes the appropriate Transientware moule to serve the event (5). A Transientware moule can also use the probing API to access the kernel state (6a, 6b) or to pass some information to the subscriber application itself (7). A freebsd version InTraN kernel an corresponing interactive protocols incluing itcp has been recently implemente [1, 11]. the next possible AP that woul serve MN after the hanoff. This information ais faster L3 hanoffs which woul be iscusse in etail later in the section. At times MN may encounter temporary layer isconnections, encountere ue to isjoint coverage area as shown in figure (b). This type of coverage is calle Non-Overlapping cell bounaries. Probing the link layer will not yiel any useful information. 4. Architecture In this section we will iscuss in etail the architecture of our approach which we have name Interactive Protocol for Mobile Networks (IPMN). This event-riven moel has network layer events (like congestion, retransmission, hanoff, etc.,) that the application can subscribe so that the application can be notifie upon occurrence of that event in the network layer. L hanoff consists of three steps a) probing, b) authentication an c) re-association explaine in etail in the next section. The hanoff proceure is 3

4 Proceeings of the 5 IEEE International Conference on Electro Information Technology, 5, Lincoln Nebraska, May 5 starte when the application receives a L trigger notifying an impening hanoff. Application woul then let the TCP layer avertise a zero winow temporarily interrupting the ata flow. Another signal triggers the application to notify the completion of the authentication process in L layer. Here the overlapping scenario an non-overlapping scenario follow ifferent approaches. In overlapping case probing the link layer now woul give information about the Future Access Point (FAP). This information is use to registers an IP aress for MN even before the completion of L hanoff. After getting the aress the same connection is renewe by manipulating the TCP/IP stack at both the en points. The application unfreezes the connection after manipulating the TCP stack. Since IP aress is attaine almost in parallel with L hanoff we reuce the hanoff latency by a great extent. In non-overlapping case the system waits until it gets an IP aress normally through DHCP. After obtaining the new IP the rest of the process follows as explaine earlier. Freezing the connection uring hanoff avois congestion control algorithms an increasing the performance of the transport layer. We are trapping 4 events, two state variable upate from three layers an corresponing 5 transientware processes in orer to provie mobility support. Figure-3 gives the architecture an event sequencing of IPMN. Probing in L is initiate when the Signal-to-Noise Ratio (SNR) of the current AP falls below a certain threshol. At the MN, the subscribing application (L7) is notifie of an impening hanoff after probing (event 1) is complete. When the event is receive at the L7, a Transientware moule (Hanler 1) is activate immeiately; this moule makes a system call Freeze which lets TCP avertise a zero winow temporarily interrupting the ata transmission. Freezing the connection uring hanoff avois TCP congestion control algorithms by invoking the persist timer. In overlapping case successful L authentication (event ) triggers L7 about the same. Hanler then probes L to ientify the cell bounary conition (if the probe yiels at least one AP then the cell bounaries are overlapping. Here we assume that the current connection is still vali) an extract information about the Future Access Point (FAP). MN uses this information to register with Future AP an attain an IP aress (L3 hanoff) in parallel with L hanoff. Hanler3 upon initiation transmits the IP aress to the CH through a system callrelayip an also manipulates the SourceIP fiel in the TCP substrate. A special TCP segment with option = SWITCH_IP is sent to the CH. On reception of this special segment at CH (event 5), Hanler5 is activate, which manipulates the DestinationIP fiel in the TCP substrate with the newly receive IP aress. MN on reception of an ACK for SWITCH_IP segment (event 4), invokes Wakeup Hanler allowing MN to resume the communication by avertising a non-zero winow to the CH. In non-overlapping case MN waits until L hanoff is complete an a DHCP IP aress (event 6) is assigne. Hanler3 is then invoke which probes IP layer to get the IP aress. The same process as explaine in the previous case follows when RelayIP Hanler is initiate. The timing iagrams of both overlapping an non-overlapping cases are epicte in the timing figures 4(a) an 4(b) respectively. E1, E an E5 are events that trigger the hanlers in overlapping case while E1, E5 an E6 trigger the respective hanlers in non-overlapping scenario. 5 FORMULATION OF HANDOFF LATENCY We ivie the hanoff process into two subsets an formulate the latency for each of the subset Layer hanoff an Layer3 hanoff. Depening on the cell 4

5 Proceeings of the 5 IEEE International Conference on Electro Information Technology, 5, Lincoln Nebraska, May 5 bounary conition the latency woul change yieling ifferent results. 5.1 Link Layer Hanoff Latency Link layer (8.11) hanoff can be classifie into 3 categories accoring to [13, 14, 18] a) Probing, b) Authentication an c) Association/Re-Association. Each of these categories woul constitute a elay parameter to L hanoff latency. Probing: Probing is one by a MN to etermine the availability of a wireless network an perform a L hanoff. A MN sens a probe request an waits for a reply governe by two timers. MinChanneltime(T e ) minimum time that a MN nees to wait after sening a probe request. If there is no activity in the channel for MinChanneltime then the channel is consiere empty. MaxChanneltime (T u ) time for MN to get a response from a use channel. In [14] Velayos an Karlsson et al., trie to estimate the various 8.11 hanoff elays an trie to optimize the hanoff latency. If u an e are the number of use an empty channels respectively, then the total time to probe the network calle the Scan Delay S can be compute as in Eq 1(a). S = u * Tu + e * T...1(a) e Furthermore both T u an T e sen probe requests twice so as to minimize the possibility of these probe packets being lost. Values of T u an T e are irectly epenent of the transmission elay T governe by the loa of the channel. Tu = * T + MaxChannelTime Te = * T + MinChannelTime...1(b) The transmission elay T can be moele as a function (Eq 1(b)) as the contention for the channel an retransmission (both being the critical parts of T ) always follow the ranomize binary exponential backoff. Equations 1(a), 1(b) an 1(c) will give the total probing elay. Authentication: This is the process that valiates whether the MN can use the services of an AP. T = ft ( t) t...1(c) If a mutually acceptable level of authentication has not been establishe between an AP an MN then, association/re-association woul not be establishe. The MN after probing the channels an obtaining the list of Access Points in range tries to prioritize them base on numerous criteria. The MN will sen an authentication request to the first entry in the list of prioritize APs an waits for an authentication reply. A = n ( * T + PTa ) i= 1 Where 1 n u 1() If the reply inicates a successful authentication then reassociation proceure is starte. If there is an unsuccessful authentication then the same proceure is repeate with the next entry until successful authentication. Re-Association: After successful authentication the MN s state information is transferre from the ol AP to new AP. Re-association is helpful in knowing the current attachment point of the MN as it is moving from AP to AP. Re-Association elay (RA ) is the request/response sequence RTT plus the time to exchange state information between the ol an the new AP(P Tr ). RA = * T + P...1(e) Tr 5. Hanoff Latency in Higher Layers. There are 5 system calls that are use to invoke hanlers. The latencies of these system calls range from 1µs to 1µs. Each system call will have latency Syc given by a ranom variable X between the aforementione intervals. Syc = Xµ s where 1 X 1 (a) In the same way the triggering latencies are between 5 an µs which is represente as Tg an governe by a ranom variable Y ranging form 5 to. Tg = Yµ s where 5 X...(b) MN woul irectly contact the Future AP to pro-actively register itself an get an IP aress. This elay terme as proactive registration elay(pr ) constitutes of the RTT between the MN an the PAP through the present AP T1. PR = * T1...(c) Where T1 is T1 = f 1 ( t) t t...() We have observe rountrip times for various emographic istances up to 4 miles an observe the fact that it is practically impossible to have access points whose emographic istance is more than 4 miles. This implies that RTT latency woul fall between 1µs to 1ms epening on the istance. Table 1 gives the hanoff elay in both overlapping an non-overlapping cases. From Eqs 1(), 1(e), (a) an (b) we have the elay of Overlapping cell bounary given by Equation (3). T = S + Tg + max max ( A, Syc ) ( RA, 3* Syc + PR + Ep ) + Syc + * Tg + Where max(a1, a) gives the maximum value of a1 an a, an represents the overlapping latencies. EP is the 3 5

6 Proceeings of the 5 IEEE International Conference on Electro Information Technology, 5, Lincoln Nebraska, May MIP Message Arrival Times IPMN Time (secs) 3 1 Roun Trip Time between en points of the connection governe by Relay Hanler. In Non-Overlapping cell bounaries the probe elay S no prolongs until it hears from at least one AP, also taking into account the time to traverse between the coverage areas. From Eqs 1(), 1(e), 1(f), (a) an (b) we have the total hanoff elay for non-overlapping cells T no is as given in Equation (4). Tn = S + Tg + max( A, Syc ) + * Tg RA, 4* Syc + PR + Ep + Syc Performance evaluation is then one in by comparing our IPMN hanoff moel with the Mobile IP hanoff moel. In orer to o this we also moele Mobile IP hanoff latencies. Similar efforts have been realize in moeling Mobile IP s hanoff latency. MIP hanoff elay H is given by Equation (5) H = L + M + R + IT...5 Where L is the total Link Layer elay moele earlier. M is the Movement Detection elay cause by MIP consisting of the outgoing an incoming agents beacon timers. R an IT are the registration an tunneling elays respectively. Detaile escription on moeling of Mobile IP can be foun in the technical report[16]. 6 PERFORMANCE EVALUATION We have simulate the moels for IPMN an MIP. Since most of the elay for IPMN is the signaling costs an system calls, it incurre only a very little hanoff elay. Figure 5 shows the latencies of non-overlapping (right sie) an overlapping (left sie) cell bounaries. MIP1 an MIP are versions of MIP that use ifferent agent avertisement life times. MIP1 uses a life time of 1ms an MIP uses 1 sec. Hanoff elay for MIP1 an IPMN is in the orer of 1ms. Though MIP1 promises better performance, it actually egraes the performance ue to banwith monopolization. As evient in figure, MIP fails to eliver real-time voice Message Number Figure 6: Message arrival times in an ieal channel without BER an congestion. an vieo traffic. MIP an IPMN have the hanoff elay ifference up to secons. In eveloping the simulation we assume that each message is of size 1KB an the total transfer size is 3MB. We also assume the expecte packet inter arrival time of 8 ms as banwith consieration. We also scheule a hanoff every one minute an observe the packet arrival times. Figure 6 gives the ifference between IPMN an MIP message arrival times. MIP experiences a elay of aroun 1 secon at each hanoff while IPMN only has a elay of 7 millisecons. A wireless environment is prone to higher Bit Error Rates (BER) which can be even upto1% of the transmission rates. BER can be moele as a Log-Normal istribution [1]. We have also consiere elay ue to congestion in our moel. This was simulate ifferently for ifferent congestion constraints. Many factors such as BER in the physical layer, or cross traffic in the network layer, causes congestion. Congestion incurre ue to BER is inuce as an exponential istribution. Congestion cause ue to cross traffic is explaine in etail in later sections separately for various traffic patterns. Various traffic patterns was generate using the moels given by NetSpec [19]. We subtracte each packet s arrival time from its expecte arrival time. This woul be the elay ue to either BER, congestion, hanoff or a combination of these factors. To give a better unerstaning we have also plotte the Normal case (where there is no hanoff) along with the Mobile IP an IPMN cases. The elay in this normal case is only ue to BER an/or congestion. 6.1 Ftp Cross Traffic Figure 7 shows the performance analysis of MIP IPMN to that of Normal case. Here the congestion is generate ue to ftp transfers suenly initiate while the ata is being transferre to the MN. Ftp traffic an many other traffic patterns follow a two step moel. The first is the Packet inter arrival times or packet size calle Session Level. The other is the Session uration or Session interarrival time calle Call Level. We have use the traffic moeling of NetSpec [19] a traffic generator evelope by 6

7 Proceeings of the 5 IEEE International Conference on Electro Information Technology, 5, Lincoln Nebraska, May 5 University of Kansas. In Netspec Ftp follows an exponential istribution for Session inter-arrival times an a log-normal istribution for the item sizes in Session Level. We use these ensity functions to generate cross traffic by incorporating these elays in RTT measurements. We use λ =.5 for exponential istribution mean = 6 an stanar eviation = for Log-Normal istribution. The burst inicate in figure 7 is the elay cause in packet arrival times ue to ftp cross traffic. This has been inuce as congestion into the packet arrival times in all the three cases. As evient, in MIP this elay compouns whenever there is a hanoff. The hanoff in MIP causes congestion in higher layers an cumulatively worsens the packet arrival time. This is evient from the figure 7. After every hanoff there is an exponential increase of packet arrival times. This tens to be linear as the transmission progresses. Our metric was packet arrival times with respect to expecte packet arrival. MIP triggers TCP Time(secs) Message Arrival Times NO HANDOFF MIP IPMN burst Message Number Figure 7: Message arrival times 1% BER an congestion ue to ftp cross traffic. congestion control every time a hanoff is initiate. In contrast, IPMN oes not have this elay because of the explicit event notification an the persist timer. Overall MIP takes aroun 5 secons more than that of IPMN to complete the ata transfer. 6. Voice Cross Traffic Voice has a Constant Bit Rate (CBR) traffic characteristic an its typical sampling rate is 8 khz an each sample consisting of is 8 bits. This gives the stanar bit rate of 64 Kb/sec for acceptable voice quality. Call inter-arrival times are moele to be exponentially istribute. Figure 8 shows the performance when the congestion is ue to backgroun voice call. We have use λ =.333 an incorporate the congestion factor in packet arrival times. As observe in figure 8 MIP has lot of performance egraation ue to cross traffic. It also suffers a ripple effect an the packet arrival times ten to increase after every hanoff. Hanoff also has elay inuce as Figure 8: Message arrival times at MN with 1% BER an congestion ue to voice cross traffic. congestion into packet arrival times. The ifference in the packet arrival times of MIP an IPMN is almost 1 secons in this case which gives a clear inication of how MIP woul fail elivering voice traffic. IPMN on the other han has no congestion builup ue to hanoff. Compare to IPMN is again almost linear an smooth. MIP cannot hanle real time traffic like voice an vieo while IPMN elivers the same efficiently with very little or no quality egraation. 6.3 WWW Cross Traffic Interactive traffic has the ocument size as a Power Law or Pareto istribution which is heavy taile. The probability ensity function an cumulative istribution function of a Pareto istribution are given by eq 18. Since this is a heavy taile istribution WWW possesses self similarity in network traffic. Call Level of the WWW traffic has a mean request time of 5.75 an can be moele as a homogeneous Poisson istribution over one hour perio implying that the inter-request time is an exponential moel. We have generate the istribution using the probability ensity function an the value of α=.45. This elay was also introuce as congestion into the packet arrival times in between the ata transfer shown in figure 9 as burst. MIP has longer hanoff jumps an as a result Time(secs) MESSAGE ARRIVAL TIMES NO HANDOFF MIP IPMN burst burst Message Number Figure 9: Message arrival times at MN with 1% BER an congestion ue to WWW cross traffic. 7

8 Proceeings of the 5 IEEE International Conference on Electro Information Technology, 5, Lincoln Nebraska, May 5 introuces its own congestion along with the heavy taile nature of cross traffic. This leas to longer arrival times of packets an the elay creeps up as every hanoff will a more an more latency in packet arrival times into WWW interactive traffic. MIP takes 7 secons more than IPMN for the same ata transfer. 7 CONCLUSION In this paper we have presente an infrastructure-less high performance mobility protocol which uses explicit En-to-En notification mechanism to hanle mobility. This eliminate the istribute nature of MIP s movement etection an hence the nee for a preeploye infrastructure. Thus MIP requires new software at one en-point (MN), one new entity at MN s traitional base station from which it gets ientity (HA) an also a battery of network eploye entities (FA). Copies of FA have to be available ahea of time in all parts of the Internet where the mobile noe might move. In comparison, the IPMN oes not require HA, oes not require the battery of pre-eploye FAs. Instea it requires the sening en-point to be mobility aware. MIP oes not nee this. Naturally, there will be mobility scenarios where one will be more appropriate than the other. When both options are feasible it is worthwhile to consier the traeoff. Traeoff for IPMN is that it offers much greater performance on several counts. Besies, the ifference in eployment scenarios, the IPMN offers blazingly fast hanoff (base on local event) compare to MIP (base remote timer/beacon). It also uses simplifie routing. A packet is irectly route in both irections (oes not have to go through HA). We have foun that just the hop count is reuce approximately by half. There is also no tunneling (packing/unpacking) elay/jitter ae to packets. With the elimination of infrastructure it also rastically reuces the eployment an maintenance costs. The performance avantage is likely to make IPMN a caniate technology for connection oriente mobility. This is currently very ifficult on MIP. REFERENCES [1] Perkins C., IP Mobility Support, RFC, IETF, Oct [] Charles Perkins, Davi B. Johns Route Optimization in Mobile IP, IETF, February 1999 [3] Fikouras N. an Görg C., Performance Comparison of Hinte an Avertisement Base Movement Detection Methos for Mobile IP Han-offs, In Proc. of the European Wireless, Dresen, Germany, September. [4] Singh R., Tay Y., Teo W., an Yeow S., RAT: A Quick (An Dirty?) Push for Mobility Support, n IEEE Workshop on Mobile Comp. Systems an Applications, pp. 3, Feb [5] A. Bakre., an B.R. Barinath., Hanoff an system support for Inirect TCP/IP, Proc. Secon Usenix Symp. On Mobile an Location Inepenent Computing [6] Goff T., Moronski J., Phatak D., Gupta V., "Freeze- TCP: A True En-to-En TCP Enhancement Mechanism for Mobile Environments," INFOCOM', Tel-Aviv, Israel, pp ,. [7] Gustafsson E., et al. Mobile IPv4 Regional Registration raft-ietf-mobileip-reg-tunnel-5, IETF, September 1. [8] MALTZ D., AND BHAGWAT P., MSOCKS: An architecture for transport layer mobility. In Proc. IEEE Infocom 98, March [9] Bakre A., an Barinath B.R., I-TCP; Inirect TCP for Mobile Hosts, In Proc. Of 15 th International conference on Distribute Computing Systems, May [1] Khan J., Zaghal R., an Gu Q., Symbiotic Streaming of Elastic Traffic on Interactive Transport, IEEE ISCC'3, Antalya, Turkey, July 3. [11] Khan J. an Zaghal R., "Protocol Moeling with Transparent Networking", CCCT'4, Austin, TX, August 4. [1] Funato D., Yasua K., an Tokua H., TCP-R: TCP Mobility Support for Continuous Operation, Proc. International Conference on Network Protocols, [13] Mishra A., Shin M., Arbaugh W., An Empirical Analysis of the IEEE 8.11 MAC Layer Hanoff Process, Dept. of Computer Science, University of Marylan, technical report number CS-TR [14] Velayos H., Carlson G., Techniques to Reuce 8.11b MAC Layer Hanover, Technical Report, April 3 [15] Perkins C., Mobile IP, Design Principles an Practices, Aison Wesley, [16] Davu S., Hanoff Moel of Mobile IP, Technical Report, Kent State University, 5. [17] Postel J.B., Transmission Control Protocol, RFC 1793, September [18] Carvalho M., Aceves J., Delay Analysis of IEEE 8.11 in Single-Hop Networks, ICNP, November 3. [19] Lee B., Frost V., Wie Area ATM Network Experiments Using Emulate Traffic Sources,Technical Report, Jan [] Yokota H, et al., Link Layer Assiste Hanoff Metho over Wireless LAN Networks, Proc. of MOBICOM, Sept.. [1] Carvalho M., Aceves J., Delay Analysis of IEEE 8.11 in Single-Hop Networks, ICNP, November 3. 8