SELF-ORGANISING QUORUM SYSTEMS FOR AD HOC NETWORKS

Transcription

1 SELF-ORGANISING QUORUM SYSTEMS FOR AD HOC NETWORKS Gareh Owen and Mo Adda School of Compuing, Universiy of Porsmouh Buckingham Building, Lion Terrace, Porsmouh England, PO1 3HE {gareh.owen, ABSTRACT There are many essenial applicaions for quorum sysems in ad-hoc neworks, such as ha of locaion servers in large-scale neworks. Exising research proposes many approaches o he problems, many of which are incomplee, cumbersome, or incur significan cos. We describe and analyse a self-organising quorum sysem ha creaes an emergen inelligence o minimise overhead and maximise survivabiliy. We hen examine he quorum s performance as a locaion server and sugges improvemens o he query mechanism and rouing algorihm using he informaion. KEY WORDS Ad-hoc neworks, quorum sysems, self-organising sysems. 1. Inroducion Ad-hoc neworks are a means of neworking compuing devices ogeher wihou requiring any seup or exising infrasrucure. We define an ad hoc nework as a graph of a se of nodes, V, and a se of communicaion pahs, E, ha vary hrough ime as nodes move and fail. G = V, E ) (1) ( Muli-hop ad-hoc neworks allow nodes o communicae wih each oher wihou being in ransmission range by relaying heir messages hrough inermediae nodes. Rouing in small neworks is usually achieved hrough an opimal broadcas mechanism; however, when he nework becomes large (>200 nodes), his form of discovery becomes exremely expensive. Rouing in large-scale ad hoc neworks can be achieved hrough using locaion informaion from a GPS device. Hou [1] describes Mos-Forward-wihin-Radius (MFR) whereby nodes are aware of heir geographical posiion (from GPS) and progressively roue packes closer o he locaion of he desinaion. Before a node can roue packes in his fashion, i will need o know an accurae locaion for he desinaion. Many rouing proocols propose a funcion ha maps a node s ID o is locaion, bu whils his is ideal when neworks are saic and such a funcion can be defined, i is no when nodes are mobile and neworks span many kilomeres. Locaion servers solve his problem by mainaining locaion informaion of a node ha sends i frequen updaes. Several ypes of locaion server have been proposed o provide his informaion. Li [2] divides he area ino gridsquares of differen orders. Saring wih he grid-square ha he node currenly occupies, an order-one square, an order-wo square is he grid-square conaining four of he order-one squares. Li proposes ha a number of nodes in each order-n square should hos he locaion informaion. Therefore, informaion abou a node s locaion becomes less densely available he furher one is from is locaion. Giordano proposed [3] he mos promising approach o locaion servers by allocaing nodes wihin a specific radius of a geographical poin o serve as a locaion server. Their work suggesed modifying he radius depending upon he node densiy so ha a minimum number of nodes paricipae; however, hey did no address issues such as node mobiliy, faul olerance and query success. Finally, Sojmenovic proposed [4] ha all nodes o he norh and souh in a fixed-widh column assume he role of a locaion server. Locaion searches are performed by rouing packes horizonally so ha hey will evenually inersec he verical column. This approach is no feasible for large-scale ad hoc neworks because of he large number of nodes. For a more complee review of approaches o locaion servers, he reader is referred o [5]. The approach shown in his paper is mos similar o Giordano s work on home-regions bu we furher develop he quorum sysem we proposed in [6] by looking a alernaive mehods of replicaion and analysing is performance as a locaion server. Mehods of querying he quorum and ways of miigaing he problem of incomplee quorum updaes, along wih a mehod o improve rouing success in he face of ou-daed informaion are also addressed. 2. Quorum sysems in ad hoc neworks

2 In his paper, we propose an approach ha is similar o ha of he home region bu ha addresses many of he problems associaed wih i. Insead of allocaing a circular region of nodes as he hoss of a locaion server, we say ha nodes direcly adjacen o a paricular geographic poin should be responsible. This permis a small number of nodes o be used wihou knowledge of node densiy. A node ses up a locaion server by sending an iniiaion packe oward is home locaion ha will be received by he node closes o he home locaion. This node will hen assume he role of maser of he server, and as i moves he role is ransferred o a node ha is closer o he home locaion. Migraion is he erm we call his process of quorum componens moving from node-o-node o remain close o a geographic poin. Presenly, wih only a maser, failure of he node hosing i could resul in failure of he server so i is imporan ha he daa is replicaed. Immediae neighbours of he maser are sen a duplicaion packe from he maser and assume he role of a slave; however, he server as a whole needs o make sure ha enough replicaed copies are kep in he even of node failures so ha he daa is no los. The maser is he only node permied o duplicae as i is mos likely o be in he cener of he quorum and so able o correcly coun he number of componens. When he number of neighbours hosing componens falls below a hreshold, he maser will begin a duplicaion phase o raise his number. We now have a se of quorum componens who do no direcly communicae wih each oher, bu who ac in a self-organising manner forming an emergen inelligence (like ans). Each aemps o ge as close as possible o he home locaion by moving from node o node (excep where a desinaion node is already in possession of one). Each componen assumes he role of maser and responsibiliy for duplicaing if i finds iself, hrough observing is environmen, o be on he closes node o he home locaion. If i realises i is no longer he closes hen i will auomaically remove is responsibiliy and assume he role of a slave. We illusrae he maser and slaves, along wih migraion in Figure 1, where he maser is held on he closes node o he poin and has duplicaed is daa o several slaves on adjacen nodes. Formally, we define our quorum or locaion server as a se Q where M and S are he ses of nodes ha hos maser and slave componens respecively: Q V = M S (2) S M S Such mainenance of wha is essenially a quorum [7] is very cosly process in wired neworks le alone wireless, and so a novel approach is needed. We draw our inspiraion from ans in a colony who find heir way o food by laying pheromones for each oher o follow [8]. These pheromones serve o modify he environmen so ha oher ans may deec hese modificaions avoiding he need for direced communicaion. Drawing upon his o develop a locaion server, we sugges ha insead of componens of a quorum communicaing wih each oher hey modify heir environmen. As every node has o beacon is locaion o is neighbours for rouing o occur, we add o his beacon packe (which we call he environmen) an iinerary of locaion server componens held a ha paricular node. A node s iinerary, I v,, is simply a lis of idenifiers for quorums which i hoss a componen of. Therefore, a beacon frame can be described as he n-uple: B = v, v, I } (3) v, { x y v, Each node mainains a lis of all received beacons so ha each componen of he quorum is able o examine a lis of which nodes hold which componens. Beacons sored in he lis expire afer wice he beacon inerval. Figure 1: Illusraion of quorum and componen migraion In addiion o he duplicaion ask, he componens perform one more asked based upon informaion gahered from he beacons. The componens self delee he quorum o sop a large number of replicaed componens consuming resources. They do his by each monioring he number of replicas, and if hey rise above a hreshold hen ha componen will delee iself. This process is performed on each node a regular inervals for each componen held, as described by he ManageComponen(c) funcion: MANAGECOMPONENT(c) 1 n closerneighbor(homelocaion) 2 if exiss(n) and hascomponen(n) 3 hen c role slave 4 else c role maser 5 if exiss(n) and!hascomponen(n) 6 hen c.migrae(n) S

3 7 else if c role =maser and hascomponen(all) < hreshold 8 hen replicae(c) 9 else if c role =slave and hascomponen(all) > hreshold 10 hen delee(c) The Replicae(c) funcion has several definiions represening several variaions of he algorihm. The unicas varian sends a replica o he firs node in is neighbour daabase wihou a componen already. I is expeced his approach will no survive well due o he period i would ake o replace a number of componens. REPLICATE(C) - UNICAST 1 for all n in neighbors 2 if no hascomponen(n) hen 3 unicas(c) 4 exi forloop 5 endif 6 end for The broadcas varian sends replicas o all neighbours in he viciniy and is expeced o survive beer due o i being able o recover from a number of los componens wih jus one packe. REPLICATE(C) - BROADCAST 1 broadcas(c) The broadcas algorihm should perform beer in he face of high failure raes; however, if jus one componen is los from he quorum hen a broadcas could increase he number of componens significanly over he hreshold, which would hen require pruning causing an expand and collapse scenario o occur. Therefore, we also produce a hybrid approach ha uses broadcas when he number of componens is low and unicas when i is nearer he hreshold. REPLICATE(C) - HYBRID 1 if hascomponen(all) > 50% of hreshold 2 hen unicas(c) 3 else broadcas(c) Finally, we also examine using mulicas for replicaion of he daa. When a node senses ha here is less han he hreshold of slaves, hen i will creae a mulicas packe o a number of neighbours defined by he hreshold minus he number of slaves. The neighbours chosen will be hose who do no all ready have a componen. REPLICATE(C) - MULTICAST 1 needed hreshold hascomponen(all) 2 descoun 0 3 while needed > 0 and neigh no equal null 4 if no hascomponen(neigh) hen 5 c dess [descoun++] = neigh 6 needed-- 7 endif 8 neigh neigh nex 9 endwhile We can define he seup cos, C Seup, as he number of hops o he home locaion, h, plus one for a broadcas duplicaion packe (eq. 4) or hreshold packes for he unicas (eq. 5). The mainenance cos of he quorum, C Mainenance, is defined (eq. 6) as he number of migraions of componens due o node mobiliy (m) plus he number of replicaion packes (r) o mainain he hreshold. C Seup = h + 1 (4) C Seup = h + hreshold (5) C = m r Ma in enance + (6) Keeping wih he simpliciy of he curren algorihm, we follow he same philosophy here o develop a quorum query mechanism. When a packe s desinaion locaion is no known, i is roued owards he home locaion and upon encounering a node conaining a quorum componen for he desinaion, i updaes he packe (p) wih he node locaion and roues i owards. We describe his in he RouePacke(p) funcion below. ROUTEPACKET(p) 1 c componen(p des ) 2 if exiss(c) 3 hen p locaion c locaion 4 if exiss(p locaion ) 5 hen roue o closesneighbor(p locaion ) 6 if no exiss(p locaion ) 7 hen roue o closesneighbor(homelocaion(p des )) 2.1 Query and updae mechanisms The mechanism o updae he quorum is equally simple and based upon a zero knowledge approach o minimise overhead. When receiving a new updae packe, a node simply updaes iself if i holds a componen and hen rebroadcass i. This way, each elemen of a quorum should receive a copy and he number of overhead packes will be equal o he number of componens presen (q componen ) as shown in equaion 7. C = h + (7) updae q componens This approach does no guaranee a complee quorum updae however due o is self-organising approach. Therefore, we examine he age of he daa reurned and propose a echnique ha we call Muli-query (MQuery) o ry o reduce he likelihood of rerieving old daa. MQuery requires every quorum query o aemp o query a leas wo quorum componens and ake he mos up-odae resul. Simply, if a query is received firs by a componen, hen he packe will be roued o he nex

4 closer node o he home locaion. Hopefully, his node will also be a paricipan in he quorum and herefore could have more correc daa. In our resuls, we examine he effec his has on he age of daa. RECVUPDATEPACKET(p) 1 c Componen(p nodeid ) 2 if no exis(c) or seenbefore(p) hen ignore 3 if p revision c revision hen ignore 4 c daa p daa 5 broadcas(p) Quorum sysems are radiionally analysed based upon heir performance in erms of load [9], faul-olerance [10] and failure probabiliy [11]. Therefore, we measure he survivabiliy or failure probabiliy of he quorum over 30-minues and he overhead in packes per minue of one quorum. In addiion o analysing he performance as a quorum, we need o look a he abiliy of his approach o ac as a locaion server for large-scale ad hoc neworks. The merics we will measure are updae cos, query success, age of resuls and rouing success. 2.2 DSR-aided MFR o improve rouing success I is possible ha he locaion informaion received could be ou-of-dae and so we analyse he effec his has on a rouing and ways o improve i. For comparison, we use he MFR approach used in RouePacke(p) and compare i wih he echnique proposed here. I is worh noing ha several oher echniques have been examined o improve rouing success (recursive search [12], GRA [13], and a planar sub-graph [14]); however, hese do no examine success in he face of incorrec locaion informaion. Here, we propose a echnique o handle no only small voids bu also ou-of-dae locaion informaion. We also propose ha if a packe reaches a node in which he MFR algorihm is unable o recover, hen he node iniiaes a DSR [15] search (wih limied TTL) o find a node ha is closer. Upon finding one, he packe is source roued here and hen resumes using MFR o coninue o he desinaion. If he DSR discovered pah becomes invalid before he packe is roued compleely, hen he node will reiniiae discovery. This is described in he DsrMfrRouePacke funcion below. A densiy of 200 nodes per square kilomere wih 140.5m TX range almos guaranees conneciviy. Therefore, we se he TTL of he DSR search o hree hops as we expec ha rouing failures will be due o a locally poor node arrangemen raher han large voids. The TTL could be varied depending on he node densiy bu we do no examine his here. 3. Mehodology Simulaion is widely acceped as a means for analysing ad hoc neworking proocols due o he mahemaical complexiy of he scenario. Here we use he GloMoSim (v2.03) simulaor [16] ha provides models of all he layers experienced in a real experimen. We configure he simulaor wih he parameers shown in Figure 2 o represen ha of a kilomere-squaed secion of a largescale ad-hoc nework. Figure 2: Simulaion parameers Parameer Value Parameer Value Terrain Propagaion Model Tx/Rx Range Mobiliy Model Simulaionime 4. Resuls 1000x1000 Two-ray 140.5m Random Waypoin v min = 0.1m/s v max = 3.0m/s 30 minues Number of nodes Number of quorums hreshold (number of slaves) Beacon rae Manage Componen call rae Every 5 seconds + jier Every 6 seconds + jier Firsly, we examine (in Figure 3) he abiliy of he algorihm o survive over a 30-minue simulaion given various node failure raes DSRMFRROUTEPACKET(P) 1 if no rouepacke(p) hen 2 if roues.exis(p locaion ) hen 3 use exising roue 4 else 5 iniiaedsrdiscovery(p locaion, p des ) 6 queuepacke(p) endif Loss percenage Node power failures per minue Unicas Broadcas Hybrid Mulicas Poly. (Unicas) Poly. (Broadcas) Poly. (Hybrid) Poly. (Mulicas) Figure 3: Survivabiliy

5 All four versions of he algorihm survive excepionally well even given he excepionally high failure raes, wih up o 20% of he quorums surviving. Packes per minue Unicas Broadcas Hybrid Mulicas Linear (Unicas) Linear (Broadcas) Linear (Hybrid) Linear (Mulicas) Percenage of queries 30.0% 25.0% 20.0% 15.0% 10.0% 5.0% 0.0% Pause ime Figure 6: Ou-of-dae resuls reurned UC BC UC+MQuery BC+MQuery Linear (UC) Linear (BC) Linear (UC+MQuery) Linear (BC+MQuery) Node pow er failures per minue Figure 4: Overhead (packes/min) The overhead (Figure 4) is less han 10 packes per minue despie node mobiliy and node failures. Ineresingly, he hybrid version of he algorihm incurs less overhead and we explain his as due o he reducion of several duplicaion packes ino one when near he hreshold, and he avoidance of he compacion and expansion problem explained earlier. Perhaps even more ineresing is ha he mulicas version has higher overhead han he hybrid given ha one would expec his o provide he ideal soluion. This is explained as due o he hybrid algorihm mainaining an opimum beween replicaion and quorum size. Figure 5 shows he overhead in erms of memory used per node. Fify quorums are disribued across he nework and each node incurs a memory overhead of approximaely 6.5 o 9.0 componens per node given zero node failures The overhead incurred o fully updae he quorum depends on he number of componens presen. Figure 7 clearly shows he unicas algorihm mainains he lower number of componens while he broadcas version mainains almos wice as many. Packes per updae Pause ime UC BC Linear (BC) Linear (UC) Figure 7: Updae cos (packes) Figure 8 shows he resuls obained when he informaion from he quorum is ou-of-dae by varying degrees. Wih updae inervals of wo minues, almos a hird of packes do no reach heir desinaion. However, when we use he DSR-aided MFR rouing algorihm, less han 3% of packes fail o reach heir desinaion % Average iems per node Unicas Broadcas Hybrid Mulicas Linear (Unicas) Linear (Broadcas) Linear (Hybrid) Linear (Mulicas) Rouing failure % 25.00% 20.00% 15.00% 10.00% Wihou DSR Wih DSR Poly. (Wihou DSR) Log. (Wih DSR) % Node pow er failures per minue Figure 5: Memory overhead (componens/node) Figure 6 shows he number of queries reurned by he quorum ha were ou-of-dae due o an incomplee updae of all componens. We hen also show how his can be improved by querying a leas wo componens and aking he mos up-o-dae resul. This simple echnique halves he number of old-resuls reurned. 0.00% Updae inerval Figure 8: DSR-aided geodesic rouing 5. Conclusion Exising algorihms are no comparable wih he approaches we have presened here, as hey do no address node mobiliy and failure ha are he key moivaion o our work. Therefore, we presen our resuls wihou comparison wih hem. The resuls show ha he broadcas algorihm performs beer in erms of survivabiliy while he hybrid provides

6 lower overhead. Maximising survivabiliy is he main prioriy and all he algorihms cope wih failure raes significanly higher han expeced. Therefore, we recommend he use of he hybrid algorihm due o he avoidance of he expand-collapse problem and he abiliy o recover quicker han he unicas approach o a number of node failures. In erms of abiliy o query he quorum, he approach used o mainain he quorum had lile effec of he resuls. We found ha approximaely 4% of queries reurned ou of dae resuls, bu ha his is halved by querying a leas wo componens. Furher improvemens could be made if elemens of he quorum communicae wih each oher, or revision informaion is added o he beacon packe. We showed ha by aiding he rouing algorihm wih DSR when he packe was unable o make any furher progress forward, wih a search radius of jus hree hops, increased he rouing success by up o a facor of 10 even when he locaion informaion is up o wo minues old. This work demonsraes ha i is possible o implemen reliable quorum sysems in ad-hoc neworks wih negligible overhead. The echnique used is a selforganising sysem relying on an inelligen behaviour emerging from he individual acions of each componen. We demonsraed he performance of he quorum for obaining locaion informaion and measured he percenage of ou-of-dae daa reurned. Locaion informaion does no necessarily have o be up-o-dae as long as he node has no moved oo far and he rouing algorihm is able o recover. One way o improve geodesic rouing algorihms o cope wih incorrec informaion would be o aid he algorihm wih a limied DSR search when he node is no a he given locaion. Underaking his simple improvemen increased he rouing success by up o a facor of en; however, his will of course incur a higher rouing overhead. References [1] T.-C. Hou & V. Li, Transmission range conrol i mulihop packe radio neworks, IEEE Transacions on Communicaions, 34 (1), 1986, [2] J. Li, J. Jannoi, D. S. J. De Couo, D. R. Karger & R. Morris, A Scalable Locaion Service for Geographic Ad Hoc Rouing. Proceedings of 6h ACM Inernaional Conference on Mobile Compuing and Neworking (MobiCom), [3] S. Giordano & M. Hamdi Mobiliy Managemen: The Virual Home Region, EPFL, Lausanne, Swizerland(1999). [4] I. Sojmenovic A scalable quorum based locaion updae scheme for rouing in ad hoc wireless neworks, Universiy of Oawa(1999). [5] T. Camp, J. Boleng & L. Wilcox, Locaion Informaion Services in Mobile Ad Hoc Neworks. Proceedings of IEEE Inernaional Conference on Communicaions, 2002, [6] G. H. Owen & M. Adda, Quorum based geographically saic daa sorage in ad-hoc neworks. Proceedings of Inernaional Nework Conference, Plymouh, UK, [7] D. Malkhi, M. K. Reier, A. Wool & N. Wrigh, Probabilisic Quorum Sysems. Proceedings of Symposium on Principles of Disribued Compuing, 1997, [8] S. Johnson, Emergence: The Conneced Lives of Ans, Brains, Ciies and Sofware (Scribner 2001). [9] M. Naor & A. Wool, The load, capaciy and availabiliy of quorum sysems., SIAM Journal of Compuing, 27 (2), 1998, [10] D. Barbara & H. Garcia-Molina, The vulnerabiliy of voe assignmens, ACM Transacions on Compuer Sysems, 4 (3), 1986, [11] D. Peleg & A. Wool, The availabiliy of quorum sysems, Informaion and Compuaion, 123 (2), 1995, [12] G. Finn Rouing and addressing problems in large meropolian-scale inerneworks, USC/Informaion Sciences Insiue(1987). [13] R. Jain, A. Puri & R. Sengupa, Geographical Rouing Using Parial Informaion for Wireless Ad Hoc Neworks, Ieee Personal Communicaions, 8 (1), 2001, [14] B. Karp & H. T. Kung, GPSR: Greedy Perimeer Saeless Rouing for Wireless Neworks. Proceedings of Mobile compuing and neworking; MobiCom 2000, Boson, MA, 2000, [15] D. Johnson & D. Malz, Dynamic source rouing in wireless ad-hoc neworks.in. Mobile Compuing. T. Imelin-sky&H. Korh, Kluwer Academic Publishers, [16] X. Zeng, R. Bagrodia & M. Gerla, GloMoSim: a library for parallel simulaion of large-scale wireless neworks. Proceedings of 12h Workshop of Parallell and Disribued Simulaions - PADS '98, 1998.