DUE to the rapid progress of wireless communication

Transcription

1 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 2, NO. 3, JULY-SEPTEMBER An Effcent Fault-Tolerant Approach for Moble IP n Wreless Systems Jenn-We Ln and Joseph Arul, Member, IEEE Abstract Ths paper presents the fault tolerance of Moble IP n wreless systems. Moble IP can support wreless users wth contnuous network connectons whle changng locatons. It s acheved by allocatng a number of moblty agents (foregn agents and home agents) n the archtecture of a wreless system. If a falure occurs n a moblty agent, the wreless users located n the coverage area of the faulty moblty agent wll lose ther network connectons.to tolerate the falures of moblty agents, ths paper proposes an effcent approach to mantanng the network connectons of wreless users wthout beng affected by the falures. Once detectng a falure n a moblty agent, falure-free moblty agents are dynamcally selected to be organzed as a backup set to take over the faulty moblty agent. Compared to the prevous approaches, the proposed approach does not take any actons aganst falures durng the falure-free perod. Besdes, the hardware redundancy technque s also not used n the proposed approach. The overhead of the proposed approach s analyzed usng the M/G/c/c queung model. The results show that the proposed approach can effectvely resolve the fault-tolerant problem of Moble IP n wreless systems. Index Terms Fault tolerance, Moble IP, wreless systems, M/G/c/c queung model. æ 1 INTRODUCTION DUE to the rapd progress of wreless communcaton technology, there s a growng demand for accessng data by wreless systems (wreless data accesses). Moblty s one mportant characterstc of wreless systems. Each wreless user (moble node) may change ts locaton several tmes durng the executon of ts data servce. To avod nterruptng the ongong data sesson, the Internet Engneerng Task Force (IETF) has defned Moble IP [1] as one whch enables wreless users to mantan ongong data sessons wthout nterrupton whle changng locatons. To provde the functonalty of Moble IP n a wreless system, a number of moblty agents are requred n the archtecture of the wreless system. The moblty agents are classfed nto two types: foregn agent () and home agent (). The logcally connects wth several rado access networks (RANs) to form a wreless data servng area. The mantans the current addresses of moble nodes (MNs). When an MN starts a data sesson, a data request s frst sent to the located RAN of the MN, then to a servng and, fnally, to a correspondng applcaton server. The applcaton server processes the data request and sends back the packets of the handlng results to the MN. The packets sent to the MN wll be frst ntercepted by the MN s correspondng. The looks up the current address of the MN and tunnels the packets to the servng the MN. Then, the detunnels the packets and forwards them to the MN. From ths scenaro of a wreless data sesson, we can know that, f a falure occurs n a, all MNs located n ts data servng. The authors are wth the Department of Computer Scence and Informaton Engneerng, Fu Jen Catholc Unversty, 510 Chung Cheng Road, Hsnchuang, Tape 242 Tawan, ROC. E-mal: {jwln, arul}@cse.fju.edu.tw. Manuscrpt receved 18 Aug. 2002; revsed 5 June 2003; accepted 25 June For nformaton on obtanng reprnts of ths artcle, please send e-mal to: tmc@computer.org, and reference IEEECS Log Number area cannot perform wreless data sessons agan. Lkewse, f an crashes, the response packets of the data sesson cannot be sent back to the correspondng MN. Ths paper presents an effcent approach to provdng fault-tolerant capablty n the wreless system wth Moble IP functonalty. The fault-tolerant mportance s descrbed as follows: For a telecom system, fve nnes ( percent) relablty s a usual requrement for the network desgn [2]. To acheve such hgh relablty, the falure rate of equpment n the telecom system may need to be very low. However, the falure rate vares as tme, where the relatonshp can be modeled as a bathtub curve [3]. In the ncreasng part of the bathtub curve, the falure rate of telecom equpment ncreases as tme ncreases. Ths means that telecom equpment wll be prone to ncur falures after t has been functonal for a long perod of tme. The wreless system s also one telecom system. For supportng the fve nnes relablty, t s necessary to provde fault-tolerant solutons for a wreless system. The proposed approach s based on the concept of resource sharng to redrect the workloads of a faulty () to other falure-free s (s). Unlke the prevous approach [4], [5], the fault-tolerant method of an n the proposed approach s dfferent from that of an. Once detectng a falure n a, one or more falure-free s are dynamcally selected to be organzed as a backup set for the faulty. Then, a system-ntated handoff s ssued to vrtually move the MNs now served by the faulty to the servng areas of the s n the backup set. The data executable capabltes for the MNs can be contnuously supported by the falure-free s n the backup set. For the fault tolerance of an, f a falure s detected n an, one or more falure-free s are also dynamcally selected to be the backup members of the faulty. The selected falure-free s, nstead of the faulty, ntercept the /03/$17.00 ß 2003 IEEE Publshed by the IEEE CS, CASS, ComSoc, IES, & SPS

2 208 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 2, NO. 3, JULY-SEPTEMBER 2003 Fg. 1. Wreless data network model. packets movng toward the faulty, and then send the packets to correspondng MNs. The remander of ths paper s organzed as follows: Secton 2 presents the background of ths paper. Secton 3 proposes our approach. Secton 4 descrbes the detals of mplementng the proposed approach. Secton 5 evaluates the overhead of the proposed approach. Secton 6 makes the comparson between the proposed approach and prevous approaches. Fnally, conclusons are made n Secton 7. 2 BACKGROUND Ths secton descrbes the background materals of ths paper. Frst, we gve an overvew of the Moble IP. Then, the system model s descrbed. Next, the fault assumpton s made. Fnally, related work s revewed. 2.1 Moble IP Overvew Moble IP s desgned to support node moblty by usng two types of moblty agents: home agent () and foregn agent (). Intally, each moble node (MN) has a unque address to be managed by an on ts home network. If an MN moves from ts home network to a foregn network, a care-of-address (COA) s allocated to the MN to reflect the MN s current pont of attachment. The on the foregn network also adds an entry correspondng to the MN n ts vstor lst. Then, a regster message s sent to the MN s to create or modfy the MN s moblty bndng whch assocates the MN s home address, the located, and the COA. Packets from a correspondent host (CH) to an MN are addressed to the home address of the MN. If the MN s away from ts home network, the on the home network can ntercept the packets and tunnel them to the located of the MN. Then, the decapsulates the packets and forwards them to the MN. 2.2 System Model The system model consdered n ths paper refers to the archtecture of a thrd generaton (3G) wreless system [6], [7], as shown n Fg. 1. The system model contans three major components: moble node (MN), rado access network (RAN), and core network. The MN nteracts wth an RAN to obtan rado resources for performng wreless data sessons. The RAN provdes the transmsson across the ar nterface. Wth the core network, t s an IP-based network and contans the followng equpment: foregn agent (), home agent (), and ntermedate routers. The and provde the wreless data sessons wth Moble IP functonalty. The ntermedate routers assst s and s to forward packets. In addton, between RANs and s, there s an nterconnecton network to delver the wreless data requests from MNs to the core network and to send back the response packets from the core network to MNs. Based on the specfcatons of [8], [9], [10], the nterconnecton network can be a Frame Relay network, ATM network, or IP network, respectvely. To manage all equpment n the wreless system, operatons, admnstraton, and mantenance (OA&M) functons are necessary. The OA&M functons are classfed nto confguraton management, fault management, performance management, and securty management [11]. The confguraton management confgures equpment (e.g., RAN,,, AAA server, router, etc.) wth sutable resource parameters. The performance management measures the resource utlzaton, loadng status, and other concerned values n the equpment. The fault management s capable of detectng and reportng falures n the equpment. The securty management montors the access rghts to the equpment. 2.3 Fault Assumpton In ths paper, falures are only assumed to occur n moblty agents. To announce the presence of a moblty agent ( or

3 LIN AND ARUL: AN EFFICIENT ULT-TOLERANT APPROACH FOR MOBILE IP IN WIRELESS SYSTEMS 209 ), the moblty agent perodcally transmts an agent advertsement message on ts located subnet [1]. When a falure occurs n a moblty agent, the falure can be detected by not recevng an agent advertsement message wthn a perod of tme. The fal-stop assumpton s also made so that the faulty moblty agent cannot send any agent advertsement message agan. As an fals, ts n-processng data requests wll be lost. For a faulty, ts n-processng response packets are also lost. However, these lost data requests and response packets can be retransmtted by an end-to-end relable transport layer such as Transmsson Control Protocol (TCP). Several modfcatons have been proposed to mprove TCP n wreless systems [12], [13]. In ths paper, we do not dscuss how to resend the lost data requests and response packets. 2.4 Related Work For Moble IP, many fault-tolerant approaches have been proposed [4], [5], [14], [15], [16], [17]. However, only [4] and [5] deal wth the ssue of toleratng falures n moblty agents. Accordng to [4], a moblty agent s statcally equpped wth one or more redundant moblty agents as ts backup set. The moblty agent can cooperate wth ts backup set to work n the standby or load-sharng model. If a moblty agent fals, one member n ts backup set wll be selected to act as the prmary moblty agent. Here, the Address Resoluton Protocol (ARP) [18] s used to map the IP address of the faulty moblty agent onto the network lnk-layer address of the selected backup member. Other backup members stll cowork wth the new prmary moblty agent, but the total processng capablty s less one. However, n the approach of [4], there s nontrval overhead for preservng the moblty bndng nformaton. When an MN regsters wth a moblty agent, the regstraton s requred to be addtonally done on all backup members. Ths possble long regstraton delay wll ntroduce sgnfcant performance degradaton. To avod the long regstraton delay, the approach of [5] suggests checkpontng and loggng technques to store the moblty bndngs n stable storage. After a moblty agent fals, ts backup s also selected from one of ts equpped redundances. Unlke the approach of [4], the backup does not have the moblty bndngs of the faulty moblty agent. It needs to restore the moblty bndngs from stable storage. The stable storage s assumed to be nvulnerable. Bascally, the approach of [5] s smlar to the approach of [4]. 3 THE PROPOSED APPROACH Ths secton presents a new approach to toleratng the falures of moblty agents. To avod sgnfcant performance degradaton, the approach dynamcally selects multple falure-free moblty agents to form a backup set for the faulty moblty agent. The workloads of the faulty moblty agent are redrected to the falure-free moblty agents n the backup set. 3.1 Fault Tolerance of Foregn Upon detectng a falure n an, the MNs located n ts servng area wll not be able to execute wreless data Fg. 2. Remappng the relatonshp between RANs and s for fault tolerance. sessons. The new arrvng MNs are also affected. These MNs are called as _falure-affected MNs. To resume the data executve abltes of the _falure-affected MNs, a system-ntated handoff s performed to dynamcally select multple falure-free s to organze a backup set for the faulty. Then, the _falure-affected MNs are vrtually moved to the servng areas of the selected falure-free s. Each of the selected falure-free s adds a number of vstor entres for the _falure-affected MNs that have moved to t and nforms those MNs correspondng s of the new servng s and the care-of-addresses to update ther moblty bndngs. After performng the system-ntated handoff, the network connectons of the _falure-affected MNs can be contnuously supported by ther new servng s. The workloads of the faulty are redrected to other falure-free s. Orgnally, a handoff functons to swtch an MN s wreless communcaton and mantans ts network connecton when the MN moves from one rado coverage area to another. In the proposed approach, the system-ntated handoff s to redrect the workloads of the faulty, but t does not ntroduce locaton changes of the _falureaffected MNs. Ths objectve s attaned by modfyng the relatonshp between RANs and s, descrbed as follows: As shown n Fg. 1, there s an nterconnecton network to connect RANs wth s. From the logc pont of vew, there exsts at least one delvery path from an RAN to each. Therefore, f an RAN receves a data request from a MN, t can delver the data request to any to process t. The RAN determnes whch to process ts receved data requests based on ts nternal -servng record. Intally, the -servng record of an RAN s set to be the dentfer of a fxed. After detectng a falure n an, one or more falure-free s are selected to be the backup members of the faulty. Then, the -servng records of the falureaffeced RANs (the RANs whch have an ntal servng relatonshp wth the faulty ) are reset as the dentfers of the backup members. By modfyng the servng records, the servng s of the falure-affected RANs become the backup members. Therefore, the _falure-affected MNs are now served by the backup members, but ther locatons are not changed (they stll locate n ther respectve rado coverage areas), as shown n Fg. 2.

4 210 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 2, NO. 3, JULY-SEPTEMBER 2003 Fg. 3. Packets destned to an _falure-affected MN. 3.2 Fault Tolerance of Home An asssts a correspondent host (CH) to send packets to an MN based on the followng three functons: moblty bndng mantenance, packet ntercepton, and packet tunnelng. If an fals, the _falure-affected MNs (the MNs managed by the faulty ) wll not be able to receve packets from CHs. To resume the packet recevng abltes of the _falure-affected MNs, one or more falure-free s are also dynamcally selected to be the backup members of the faulty. Then, the above three functons of the faulty are restored on the backup members, descrbed as follows: As mentoned n Secton 2.1, when an MN enters nto a data servng area, the correspondng adds a vstor entry n ts vstor lst to record the followng nformaton: the MN s data lnk-layer address, IP address, and home agent address. When the MN leaves the data servng area, ts correspondng vstor entry s made nvald. Therefore, f an MN has a vald entry n an s vstor lst, the MN s now located n the servng area of the. From the s vstor lsts, the up-to-date locatons of all MNs can be known. Ths ponts out that the moblty bndngs of the faulty can be restored by searchng all the s vstor lsts. The restored moblty bndngs are then dstrbuted to the backup members. Wth regard to the packet ntercepton of the faulty, t s restored on the backup members usng the tunnelng technque. Usually, one or more routers are collocated wth an on the same network segment to connect the wth the core network and the Internet. When an CH sends packets to an MN, the packet s frst receved by the collocated routers, and s then ntercepted by the MN s. After detectng a falure n an, ts collected routers do not forward ther receved packets to the faulty. Instead, the collected routers tunnel the packets to the backup members, as shown n Fg. 3. In such a case, a packet from a CH to an _falure-affected MN s sent by twce tunnelng. One s to let one backup member ntercept the packet. Another s to tunnel the packet to the located of the _falure-affected MN. As for the restoraton of the packet tunnelng, the falurefree s selected as the backup members, already possess the tunnelng functon. 4 IMPLEMENTATION The mplementaton of the proposed approach can be ntegrated nto the OA&M. As descrbed n Secton 2.2, there are four man functons n the OA&M. When a falure n an () s detected, the falure event s sent to the fault management. The fault management ntates the proposed fault-tolerant approach for the (). At frst, t nteracts wth the performance management to acqure the loadng status of the falure-free s (s). Based on ths loadng nformaton, multple falure-free s (s) wth low traffc are selected to be the backup members of the faulty (). Then, the confguraton management s nformed to confgure the backup members of the faulty () by resettng approprate parameters to some equpment n the core network. Thereafter, the ncomng workloads of the faulty () are redrected to ts backup members. 4.1 Implementaton of Foregn To understand the mplementaton n more detal, the fault-tolerant procedure for the s gven n Fg. 4. There are three man tasks n the procedure. The frst task Fg. 4. The fault-tolerant procedure for the. (a) Collectng the loadng status of falure-free s and the number of _falure-affected MNs.

5 LIN AND ARUL: AN EFFICIENT ULT-TOLERANT APPROACH FOR MOBILE IP IN WIRELESS SYSTEMS 211 Fg. 4. (Contnued). (b) Selectng the backup members. s to collect the loadng status of each falure-free and calculate the number of _falure-affected MNs, as shown n Fg 4a. The loadng status of each falure-free Fg. 4. (Contnued). (c) Performang the system-nated handoff. can be collected from the performance management of OA&M. For the number of _falure-affected MNs, t can be calculated by countng the number of assocaton entres n the falure-affected RANs. In an RAN s coverage area, f an MN ssues a wreless data sesson, a correspondng entry s generated n the RAN to record the assocaton between the RAN and the servng the MN. The second task selects multple falure-free s to be the backup members of the faulty, as shown n Fg. 4b. Frst, the falure-free s wth lower traffc are selected. If the total resources donated by the selected falure-free s are nsuffcent to serve all falure-affected MNs, some falure-free s wth hgher traffc wll be selected to be the backup members. In addton, each selected falure-free also estmates the number of _falure-affected MNs to be vrtually moved to ts servng area based on the number of the avalable resources n t. The thrd task, as shown n Fg. 4c, performs the systemntated handoff both to vrtually move _falure-affected MNs to the servng areas of the backup members and to update the moblty bndngs of such MNs. From Secton 3.1, we know that the vrtual moves of the _falure-affected MNs are acheved by modfyng the -servng records of the falure-affected RANs. However, the resources of a backup member are derved from a porton of the avalable resources n a falure-free. It s possble that the MNs n the falure-affected RAN cannot be fully served by a sngle backup member. Each falure-affected RAN frst needs to specfy enough backup members to serve the MNs n ts coverage area. Then, t resets ts -servng record to the dentfers of the specfed backup members. Next, each specfed backup member adds the vstor entres for the _falure-affected MNs served by t. Meanwhle, the s of the _falure-affected MNs are notfed to update ther moblty bndngs. Thereafter, the falure-affected RANs can delver ther receved data requests to the backup members. In addton, the packets destned to the _falure-affected MNs can be tunneled to the backup members. 4.2 Implementaton of Home There are also three man tasks n the procedure for mplementng the fault tolerance of the. Based on the s loadng status gven by the OA&M, the frst task selects multple falure-free s wth low traffc to be the backup members of the faulty. One of the backup members s specally desgnated as the s backup manager. The second task restores the moblty bndngs of the faulty by sendng a moblty-reconstructon message to each. Upon recevng the message, each searches ts vstor lst to fnd the vstors entry wth the dentfer of the faulty. Each qualfed vstor entres s then reorganzed as the form of a moblty bndng entry and sends back to the s backup manager. The s backup manager dvdes ts receved moblty bndng entres nto several groups based on ther correspondng MNs IP addresses and then sends each group to a backup member, as shown n Fg. 5a. If a backup member s desgnated to mantan the

6 212 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 2, NO. 3, JULY-SEPTEMBER 2003 Fg. 5. An example of toleratng the falure. (a) Reconstructng the moblty bndngs. (b) Changng the packet nterceptor. (c) Redrectng the packet ntercepton. moblty bndngs of MNs wth IP addresses n the range ½IP 1 ;IP 2 Š, t wll take the responsblty for nterceptng the packets whose destnaton addresses belong to ½IP 1 ;IP 2 Š, and then tunnelng such packets. The thrd task s to change the packet nterceptors of the _falure-affected MNs from the faulty to the backup members. To change the packet nterceptors, the collocated routers of the faulty must remove ther routng entres correspondng to the faulty and add the routng entres correspondng to the backup members, as shown n Fg. 5b. However, the backup members may not have drect physcal nterfaces wth the collocated routers. In each routng entry correspondng to a backup member, ts nterface feld s set to a vrtual nterface. The vrtual nterface ponts to a software program to perform the packet tunnelng. By tunnelng, the packets destned to a _falure-affected MN are ntercepted by backup members (1-4), not the faulty (0), as shown n Fg. 5c. 4.3 Falure Recovery When a faulty s recovered from falure, the _falureaffected_mns can be served back by the recovered. The recovery procedure s shown n Fg. 6a. Frst, the recovered determnes the nformaton whch RANs are preserved by t (the nformaton about the falure-affected RANs). Then, the falure-affected RANs reset ther -servng records to the dentfer of the recovered. The recovered gets back to be the common servng of all _falure-affected MNs. Next, the recovered creates a lot of vstor entres for the _falure-affected MNs. In addton, the respectve s of the _falure-affected MNs are also notfed to update the MNs moblty bndngs. The recovery procedure for an s shown n Fg. 6b. When a faulty s recovered, the responsblty of packet ntercepton can be returned to t. By modfyng the routng tables of the collocated routers of the recovered, the packet nterceptors of _falure-affected MNs are changed back from the backup members to the recovered. In addton, the moblty bndngs of the recovered are also requred to be reconstructed, whch can be done by searchng all s vstor lsts. Ths reconstructon task s smlar to the task of restorng the moblty bndngs of the faulty (see Secton 4.2).

7 LIN AND ARUL: AN EFFICIENT ULT-TOLERANT APPROACH FOR MOBILE IP IN WIRELESS SYSTEMS 213 Fg. 6. The falure recovery procedure. (a) and (b). 5 EVALUATION The proposed fault-tolerant approach redrects the workloads of a faulty () to falure-free s (s). To perform the workload redrecton, some control messages are ssued, whch also ntroduce a certan overhead on the system performance. Ths secton analyzes the performance degradaton of a falure-free () and the control message overhead. To obtan the two desred overheads, the parameter notatons are frst defned n Table 1. The data requests sent to an and the response packets ntercepted by an are assumed to follow Posson processes, but the servce tme of a data request and the processng tme of a packet tunnelng are not assumed to obey any specfc dstrbuton. Based on these traffc dstrbutons, the traffc behavor of an and that of an can both be modeled as the M/G/c/c queung model [19]. 5.1 Performance Degradaton of a Moblty As descrbed n Secton 3, the performance degradaton of a falure-free s due to the system-ntated handoff whch vrtually moves some _falure-affected MNs to ts servng area. Therefore, the resources of a falure-free are now contended by the MNs vrtually moved and the MNs orgnally located. The performance of a falure-free ncurs a certan degradaton, whch can be represented n terms of the followng metrc:. Increasng blockng probablty P Blockng that causes a new data request to be more possbly blocked at a falure-free n comparson to prefalure. The probablty P Blockng s derved as follows: The arrval rate of data requests to a falure-free s. Usng the well-known Erlang s loss formula from the M/G/c/c queung model [19], the blockng probablty of a data request to a falure-free s: c c! Pc : ð1þ

8 214 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 2, NO. 3, JULY-SEPTEMBER 2003 TABLE 1 Parameter Notatons After some s fal, the data requests to the faulty s are redrected to falure-free s. Compared to prefalure, the arrval rate of data requests to the falure-free k s larger and becomes as þ F w K f there are F faulty s. The new blockng probablty s as follows: Pc c þf w k c! þf w k : ð2þ From (1) and (2), P Blockng can be derved as follows: c c þf w k P Blockng ¼ Pc c! þf w k c! Pc : ð3þ Lkewse, the performance degradaton of a falure-free k s represented n terms of the followng metrc:. Increasng blockng probablty (P Blockng ) that causes an ntercepted packet to be more possbly blocked at a falure-free n comparson wth prefalure. The dervaton of P Blockng s smlar to P Blockng, whch s represented as: c c þf w k P Blockng ¼ Pc c! þf w k c! Pc : ð4þ 5.2 Control Message Overhead In the proposed approach, the followng control messages are ssued from the OA&M for assstng the fault tolerance of the ():. _Loadng: Collectng the loadng status of each falure-free for reducng the performance degradaton of a falure-free.. RAN_Mappng: Remappng the relatonshp between the RANs and the s for vrtually movng the _falure-affected MNs.. Bndng_Update: Updatng the moblty bndngs of _faure-affected MNs for regsterng ther new servng s.. _Loadng: Collectng the loadng status of each falure-free for reducng the performance degradaton of a falure-free.. Interceptor_Change: Modfyng the routng tables for changng the packet nterceptors of _falureaffected MNs.. Bndng_Restoraton: Searchng the vstor lst of each for restorng the moblty bndngs of a faulty. The frst three control messages are ssued for toleratng the falure. The frst control message nqures of each falure-free ts loadng status. The ntroduced cost of the message ncludes the transmsson tme (T Loadng ) of sendng the message and the transmsson tme (T Re sponse ) of respondng the s loadng status. The second control message notfes each falure-affected RAN to modfy ts -servng record. Snce the modfcaton of the -servng record s a sngle memory access operaton, the ntroduced cost of the message can be only determned by the transmsson tme (T RAN Mappng ) of sendng the message to an RAN. The thrd control message ndcates the backup members of a faulty to send the regstraton messages for updatng the moblty bndngs of _falure-affected MNs. The ntroduced cost of the message s estmated as follows: T Moblty Update þ T Re g ð5þ ¼ T Moblty Bndng þðn MN t Re g Þf MN ; where T Moblty Update s the transmsson tme of sendng a command of the moblty bndng update to each (note that the command can be smultaneously sent to each from the OA&M), and T Re g s the total requred tme for the

9 LIN AND ARUL: AN EFFICIENT ULT-TOLERANT APPROACH FOR MOBILE IP IN WIRELESS SYSTEMS 215 falure-free s to send the moblty bndng updates about all _falure-affected MNs. N MN s the total number of the _falure-affected MNs, and t Re g s the average transmsson tme of sendng a regstraton message from an to an. Snce each can smultaneously send the regstraton messages to the correspondng s, T Re g s only requred a fracton (f MN ) of the tme N MN t Re g (the tme for updatng the moblty bndngs of all _falure-affected MNs usng the seral transmsson). The last three control messages are ssued for toleratng the falure. The fourth control message nqures of each falure-free ts loadng status. The ntroduced cost of the message ncludes the transmsson tme (T Loadng ) of sendng the message and the transmsson tme (T Re sponse ) of respondng to the s loadng status. The ffth control message notfes the collocated routers of a faulty to modfy ther routng tables to add and delete routng entres. Snce the routng table of a router s stored n memory, the modfcaton of the routng tables nvolves several trval memory access operatons. The ntroduced cost of ths control message s only dependent on the transmsson tme (T Inteceptor Change ) of sendng the message to a collocated router of the faulty. The last control message ndcates each to search ts vstor lst to select the vstor entres correspondng to the _falure-affected MNs. Snce the s vstor lst s also stored n memory, the cost for searchng the vstor lst can be gnored. For example, the searchng tme of a vstor lst wth 10,000 entres s only sec. The ntroduced cost of the last control message s: T Moblty Bndng þ T Entry ð6þ ¼ T Moblty Bndng þðn MN t Entry Þf MN ; where T Moblty Bndng s the transmsson tme of sendng a command of the moblty bndng restoraton to each, and T Entry s the total requred tme for restorng the lost moblty bndng table of the faulty. N MN s the total number of the _falure-affected MNs, t Entry s the average tme of sendng a qualfed vstor entry, and f MN s a fracton. The dervaton of (6) s smlar to that of (5). The tme metrcs: T Loadng, T Re sponse, T RAN Mappng, T Moblty Update, T Loadng, T Re sponse, T Inteceptor Change, and T Moblty Bndng are manly dependent on the transmsson delays between the OA&M and the s, RANs, s, and routers n the core network. The OA&M and the equpment n the core network are connected through an OA&M network, as shown n Fg. 1. The physcal nterfaces of the OA&M network n a commercal wreless system are usually equpped wth hgh-speed lnes (T1 or above). In addton, the szes of the messages correspondng to the above tme metrcs are also small. The costs of the above tme metrcs can be neglected. To conclude, the control message overhead for toleratng the falure and that for the falure (Control Over and Control Over ) are manly dependent on the moblty bndng updates of the _falureaffected MNs and the moblty bndng restoraton of the _falure-affected MNs, respectvely. Both the overheads have been estmated as ðn MN t Re g Þf MN and ðn MN t Entry Þf MN. Wth respect to N MN and N MN, they can be represented as the followng equatons: N MN ¼ F Xc n P n n¼0 N MN ¼ F Xc n P n ; where P n ðp n Þ s the probablty that there are n nprocessng data requests (response packets) n a faulty (). If each MN s not allowed to smultaneously ssue more than one data sesson, P n ðp n Þ can be represented by the probablty that there are n _falure-affected MNs (_falure-affected MNs) n a faulty (). If the condton s not supported, the N MN and N MN are less than (7) and (8), respectvely. From [19], the formulas of P n and P n has been also gven. Control Over and Control Over can be further deduced as: 0 n¼0 Control Over ¼ F Xc n n¼0 0 f MN Control Over ¼ F Xc n n¼0 6 COMPARISON f MN : n Pc n! n Pc n! 1 t Reg C A 1 t Entry C A ð7þ ð8þ ð9þ ð10þ Table 2 makes a comparson between the proposed approach and the approaches of [4] and [5] wth respect to hardware redundancy, fault-tolerant range, fault-tolerant overhead, falure-free overhead, and falure-recovery overhead. The approaches taken n [4] and [5] are based on the hardware redundancy to equp a prmary moblty agent wth several redundances n a network segment. There are two alternatves: standby and load-sharng to confgure the coworkng mode between the prmary moblty agent and ts redundances. After a prmary moblty agent fals, the two coworkng modes n the above approaches take the ARP to select one redundancy as the new prmary moblty agent, but they have dfferent degrees of performance degradaton on the selected redundancy. In the standby mode, the blockng probablty of the selected redundancy before a falure s 0, snce t does not handle any workload. After a falure, the selected redundancy alone handles all the workloads n the network segment. Based on (3), the performance degradaton mposed on the selected redundancy can be represented as:

10 216 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 2, NO. 3, JULY-SEPTEMBER 2003 TABLE 2 Comparson between the Proposed Approach and the Approaches of [4] and [5] c c! P Blockng ¼ c P : ð11þ In the load-sharng mode, the selected redundancy s responsble for more workloads than other redundances n the same network segment snce t needs to addtonally handle the workloads of the orgnal prmary moblty agent. The performance degradaton mposed on the selected redundancy s: P Blockng ¼! c 2 ð1þr Þ c P c!! 2 ð1þr Þ! c ð1þr Þ c P c!! ð1þr Þ ; ð12þ where ð1 þ R Þ s the number of moblty agents (prmary moblty agent and ts redundances) n a network segment, and ð1þr Þ s the arrval rate of data to a moblty agent based on the load-sharng mode. After the falure, the arrval rate of data to the selected moblty agent s 2 ð1þr Þ. For the fault-tolerant range of the two approaches, f the prmary moblty agent and ts equpped redundances fal smultaneously, the redundances n other network segments (subnets or LANs) cannot be used to tolerate such falure stuaton snce the ARP s only avalable n the LAN envronment [19]. The fault-tolerant range s restrcted n a network segment. In addton, for restorng the moblty bndngs, the approach of [4] ncurs a long regstraton delay durng the falure-free perod. The approach of [5] nvolves a lot of tme-consumng operatons for savng moblty bndngs n stable storage. In the aspect of the falure-recovery overhead, both the approaches use the ARP to actvate the recovered moblty agent. The reconstructon of the moblty bndngs for the approach of [4] and that for the approach of [5] are done by retrevng the lost moblty bndngs from one redundancy and the stable storage, respectvely. Bascally, the falure-recovery overhead of each of the two approaches s smlar to ts respectve faulttolerant overhead but the performance degradaton. In contrast to the prevous approaches, the proposed approach does not ncur the hardware and falure-free overheads. Also, unlke the approaches of [4] and [5], the backup members of a faulty () are not necessary to be located wth the faulty () on the same network segment. In theory, the proposed approach can smultaneously tolerate N 1 faulty s (s) f there are N s (s) n the system. However, for toleratng a faulty (), the falure-free s (s) selected to be the backup members ncur a certan degree of performance degradaton due to handlng a porton of workloads of the faulty (). Two addtonal fault-tolerant overheads are also ncurred due to the ssues of control messages (see Secton 5.2). As for the falure recovery, ts procedure s smlar to the fault-tolerant procedure (see Secton 4.3). The ncurred falure-recovery overhead s nearly same as the fault-tolerant overhead except the performancedegradaton. From Table 2, we can further observe that the approaches of [4] and [5] and the proposed approach gve a dfferent trade off among the above gven comparson metrcs. Among all the approaches, the approach of [4] takes the least falure-recovery overhead. The approach of [5] mproves the approach of [4] to avod the long regstraton delay, but t has to take tme-consumng operatons to save moblty bndngs n stable storage. Comparatvely, the proposed approach has obvous advantages on the hardware overhead, fault-tolerant range, and falure-free overhead. In the aspect of the fault-tolerant overhead, t s dffcult to make a comparson. Intutvely, the cost of the ARP executon s trval. However, all three approaches

11 LIN AND ARUL: AN EFFICIENT ULT-TOLERANT APPROACH FOR MOBILE IP IN WIRELESS SYSTEMS 217 TABLE 3 Analytcal Parameters have the performance degradaton. Next, we perform the analytcal comparson to quantfy the dfferences of the three approaches on the performance degradaton. 6.1 Analytcal Comparson The analytcal parameters used are specfed n Table 3. denotes the expected number of arrvals per mean servce tme at a moblty agent ( or ); t s used to represent the traffc ntensty of a moblty agent. The values of the traffc ntensty are, respectvely, set to be 10, 50, 100, and 200. The four chosen values of the traffc ntensty all supply large workloads to a moblty agent ( or ) [20]. As mentoned before, the performance degradaton s n terms of the ncreasng blockng probablty of a falure-free moblty agent. Fg. 7 llustrates the comparson of the ncreasng blockng probabltes n the approaches of [4] and [5] and the proposed approach as a functon of F, when ¼ 10, 50, 100, and 200. For the proposed approach, the workloads of faulty moblty agents (faulty s or s) are evenly redrected to all the falure-free moblty agents snce the same traffc dstrbuton s made n each moblty agent. The rato w k ðw k Þ of redrectng 1 workloads to a falure-free moblty agent s N F. For the approaches of [4] and [5], there are two coworkng modes for the prmary moblty agent and ts redundances. From (11) and (12), we know that f the load-sharng mode s adopted, the number of redundances n a network segment affects the ncreasng blockng probablty. To observe the effects, the number of redundances equpped n a network segment s respectvely set to 1, 2, and 3. As shown n Fg. 7, t s clear that the ncreasng blockng probablty for the proposed approach ncreases as F ncreases. If F ncreases, the workloads of the faulty s (s) redrected to the falure-free s (s) become more. However, the ncreasng blockng probablty for the approaches of [4] and [5] s ndependent of the varance of F snce each redundancy s responsble for takng over only one faulty moblty agent regardless of how many faulty moblty agents n the same network segment. In the aspect of, the ncreasng blockng probablty for the approaches of [4] and [5] ncreases as ncreases snce only one redundancy s selected to take over one faulty moblty agent. When s larger, Fg. 7. Comparson of ncreasng blockng probablty under varous numbers of faulty moblty agents. (a) ¼ 10. (b) (d) ¼ 200. ¼ 50. (c) ¼ 100.

12 218 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 2, NO. 3, JULY-SEPTEMBER 2003 the addtonal workloads to the selected redundancy become more. For the proposed approach, the ncreasng blockng probablty does not always ncrease as ncreases. For example, the ncreasng blockng probablty for the proposed approach at ¼ 200 s less than that at ¼ 50 and 100. The stuaton s explaned as follows: When ¼ 200, the blockng probablty (0.7516) of a falure-free () before a falure s obvously larger than that (0.1048) at ¼ 50 and that (0.5093) at ¼ 100. After some s (s) fal (e.g., F ¼ 8), the new blockng probablty of a falure-free ðþ at ¼ 200 becomes , whch s not much larger than that (0.8010) at ¼ 50 and that (0.9002) at ¼ 100. Therefore, the ncreasng blockng probablty ( = ) of a falure-free () at ¼ 200 s smaller than that ( = ) at ¼ 50 and that ( =0.3909) at ¼ 100. For the comparson results of the ncreasng blockng probabltes n all the three approaches, when ¼ 10, the ncreasng probablty for the approaches of [4] and [5] s close to 0 regardless of the varance of F. In ths traffc ntensty, the resource unts n a redundancy are suffcent to handle the workloads to t snce c s set to 50. The ncreasng blockng probablty of a redundancy s very small. For the proposed approach n such traffc ntensty, f F < 8, the ncreasng blockng probablty s also very small. If F ¼ 8, the ncreasng blockng probablty of each falure-free () s slghtly larger snce only two falure-free s (s) handle all the workloads of the whole system, but t s stll less than When ¼ 100, f the ncreasng blockng probablty for the approaches of [4] and [5] s requred to be less than that of the proposed approach, the two approaches must use the load-sharng mode and equp two or more redundances n each network segment. When ¼ 200, even f the number of the redundances n a network segment s set to 3, the ncreasng blockng probablty for the approaches n [4] and [5] s stll larger than that of the proposed approach. In the proposed approach, there are two addtonal faulttolerant overhead for the fault tolerance of the and, respectvely. These two addtonal overheads are manly dependent on the average number N MN of the moblty bndng updates and the average number N MN of moblty bndng restoraton (see Secton 5.2). Fg. 8 plots N MN ðn MN Þ under one faulty (). Intally, N MN ðn MN Þ s approxmately equal to the value of.as ncreases up to a threshold value, N MN ðn MN Þ ncreases slowly as ncreases agan. Snce the total number of resource unts n an () s 50, f s already larger than 50, the new Fg. 8. Average number of moblty bndng updates (moblty bndng restoraton). data request (response packets) to an () s very possble to be blocked. Therefore, at any tme, the average number of n-processng data requests (response packets) n an () cannot be greater than 50. We have also made the assumpton that each MN s not allowed to smultaneously ssue more than one data sesson (see Secton 5.2). Ths also means that the maxmum value of the number of _falure-affected MNs (_falure-affected MNs) s 50 f only one () fals. Correspondngly, the maxmum number of N MN ðn MN Þ s 50. We can make the concluson that the overhead of the moblty bndng update (the overhead of moblty bndng restoraton) s restrcted by the total number of resources n an (). Furthermore, the analytcal results plotted n Fg. 8 can be also used to represent the falure-recovery overhead of the proposed approach snce the falure-recovery overhead s also dependent on N MN and N MN (see Table 2). 6.2 Smulaton Valdaton To valdate the analytcal results of the ncreasng blockng probablty for the proposed approach, smulaton experments are performed by extendng the gven Moble IP smulaton model n ns-2 [21]. In the extended smulaton model, there are 10 s, 10 s, 1000 MNs, and 5 applcaton servers. The sze of the packet queue n each () s set to 50 unts. Each MN randomly moves between the servng areas of the s and s. The capactes of the communcaton lnks n the smulaton model are randomly set n the nterval 1 to 10 Mpbs. The unform traffc s generated from fve applcaton servers to the 1,000 MNs, and vce versa, where the traffc ntensty s, respectvely, set to 10, 50, 100, and 200. The falures randomly occur to generate one, two, four, and eght faulty moblty agents durng the smulaton tme. The total smulaton tme s set to 5,000 mnutes. The smulaton results are shown n Table 4 n comparson to the analytcal results. The analytcal results match closely wth the smulaton results. In the tables, the

13 LIN AND ARUL: AN EFFICIENT ULT-TOLERANT APPROACH FOR MOBILE IP IN WIRELESS SYSTEMS 219 TABLE 4 The Valdaton of the Increasng Blockng Probablty for the Proposed Approach (a) ¼ 10. (b) ¼ 500. (c) ¼ 100. (d) ¼ 200. dfference rates between the smulaton results and the janalyss Smulatonj analytcal results Analyss are all below 10 percent. Note that the same workloads to an and, P Blockng and P Blockng, have the same value n the numercal analyss based on (3) and (4), but the two probabltes have dfferent values n the smulaton experments. 7 CONCLUSION Ths paper has presented an effcent approach to toleratng the falures of moblty agents n a wreless system. The proposed approach utlzes the avalable resources n other falure-free moblty agents to dynamcally generate a backup set for each faulty moblty agent. Compared to the prevous approaches, the proposed approach possesses the followng advantages:. Not requrng hardware support.. Not ncurrng falure-free overhead.. Dstrbutng the fault-tolerant overhead to avod the sgnfcant performance degradaton on a sngle falure-free moblty agent. However, the proposed approach ssues a more complcated fault-tolerant procedure to redrect the workloads of faulty moblty agents to multple falure-free moblty agents. Wth the fault-tolerant overhead, the M/G/c/c queung model has been used to make the analytcal comparson of the

14 220 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 2, NO. 3, JULY-SEPTEMBER 2003 proposed approach and the approaches of [4] amd [5]. The smulaton experments have been also performed to valdate the analytcal comparson. The comparson results show that when s enlarged to 200, the ncreasng blockng probablty for the proposed approach s obvously less than that n the approaches of [4] and [5]. Ths comparson result s also true n most cases of small faulty moblty agents. The approaches of [4] and [5] gve the smaller ncreasng blockng probablty when s not too large (e.g., ¼ 10 or 50), but they need to equp several redundances n each network segment. ACKNOWLEDGMENT The authors would lke to thank the anonymous revewers. Ther comments have sgnfcantly mproved the qualty of ths paper. Ths research was supported by the Natonal Scence Councl, Tawan, R.O.C., under Grant NSC E and the Socety of the Dvne Word under Grant SVD9021. REFERENCES [1] C. Perkns, IP Moblty Support, Techncal Report IETF RFC 3220, Jan [2] Goahead Software Inc., Buldng Hghly Relable Systems: The Role of Embedded Mddleware, techncal report, avalable at [3] B.W. Johnson, Desgn and Analyss of Fault Tolerant Dgtal Systems. Addson-Wesley, [4] R. Ghosh and G. Varghese, Fault-Tolerant Moble IP, Techncal Report WUCS-98-11, Washngton Unv., Apr [5] J.H Ahn and C.S. Hwang, Effcent Fault-Tolerant Protocol for Moblty s n Moble IP, Proc. 15th Int l Parallel and Dstrbuted Processng Symp., pp , [6] B. Sarkaya, Packet Mode n Wreless Networks: Overvew of Transton to Thrd Generaton, IEEE Comm. Magazne, vol. 38, no. 9, pp , Sept [7] P.J. McCann and T. Hller, An Internet Infrastructure for Cellular CDMA Networks Uusng Moble IP, IEEE Personal Comm., vol. 7, no. 4, pp , Aug [8] ETSI GSM 08.16: Dgtal Cellular Telecommuncatons System (Phase 2+); General Packet Rado Servce (GPRS); Base Staton System (BSS) Servng GPRS Support Node (SGSN) Interface; Network Servce, [9] 3GPP TS : UTRAN Iu Interface: General Aspects and Prncples, [10] 3GPP TR : Archtecture for an All IP Network, [11] R. Mstry, P. Savll, and A. Tofanell, OA&M for Full Servces Access Networks, IEEE Comm. Magazne, pp , Mar [12] R. Caceres and L.R. Iftode, Improvng the Performance of Relable Transport Protocols n Moble Computng Envronments, IEEE J. Selected Areas n Comm., vol. 13, no. 5, pp , June [13] H. Balakrshnan, V.N. Padmanabhan, S. Seshan, and R.H. Katz, A Comparson of Mechansms for Improvng TCP Performance over Wreless Lnks, IEEE/ACM Trans. Networkng, vol. 5, no. 6, pp , Dec [14] C. Graff, M. Bereschnsky, M. Patel, and L.F. Chang, Applcaton of Moble IP to Tactcal Moble Internetworkng, Proc. Mltary Comm. Conf., vol. 2, pp , [15] Y. Mun, Y. Km, Y.J. Km, and G. Hwang, IP Moblty Support over Wreless ATM, Proc. IEEE Int l Conf. Comm., pp , [16] S. Khurana, A. Kahol, S.K.S. Gupta, and P.K. Srman, An Effcent Cache Mantenance Scheme for Moble Envronment, Proc. 20th Int l Conf. Dstrbuted Computng Systems, pp , [17] H. Ahn and C.S. Hwang, Low-Cost Fault-Tolerance for Moble Nodes n Moble IP Based Systems, Proc. 15th Int l Parallel and Dstrbuted Processng Symp., pp , [18] D.C. Plummer, Ethernet Address Resoluton Protocol: Or Convertng Network Protocol Addresses to 48.bt Ethernet Address for Transmsson on Ethernet Hardware, Techncal Report IETF RFC 826, Nov [19] D. Gross and C.M. Harrs, Fundamentals of Queueng Theory. John Wley & Sons, Inc., [20] P. Ln and Y.B. Ln, Channel Allocaton for GPRS, IEEE Trans. Vehcular Technology, vol. 50, no. 2, pp , Mar [21] Ns-2, Creatng Wred-cum-Wreless and Moble IP Smulatons n ns, Jenn-We Ln receved the MS degree n computer and nformaton scence from Natonal Chao Tung Unversty, Hsnchu, Tawan, n 1993, and the PhD degree n electrcal engneerng from Natonal Tawan Unversty, Tape, Tawan, n He s currently an assstant professor n the Department of Computer Scence and Informaton Engneerng at Fu Jen Catholc Unversty, Tawan. He was a researcher at Chunghwa Telecom Co., Ltd., Taoyuan, Tawan, from 1993 to Hs current research nterests are faulttolerant computng, moble computng and networks, dstrbuted systems, and broadband networks. Joseph Arul receved the BSc degree n mathematces n 1981 from Indore Unversty, Inda, and the MS degree n computer scence n 1994 from De Paul Unversty, Chcago. From , he was a lecturer at Fu Jen Catholc Unversty, Tawan. He completed hs PhD degree n computer engneerng at the Unversty of Alabama n Huntsvlle (UAH) n He was also a research assocate n computer engneerng at UAH. He s currently an assstant professor n the Department of Computer Scence and Informaton Engneerng, Fu Jen Catholc Unversty, Tawan. Hs current research nterests are computer archtecture, parallel and dstrbuted computng, multthreaded programs, and complers. He s a member of the IEEE and the IEEE Computer Socety.. For more nformaton on ths or any computng topc, please vst our Dgtal Lbrary at