Information Dissemination in Partitionable Mobile Ad Hoc Networks

Information Dissemination in Partitionable Mobile Ad Hoc NetworksGoutham Karumanchi Srinivasan Muralidharan Ravi PrakashDepartment of Computer ScienceUniversity of Texas at DallasRichardson, TX 75083-0688.E-mail: fgoutham,msrini,[email protected] wireless networks have no wired component,and may have unpredictable mobility pattern. Suchnetworks can get partitioned and reconnected severaltimes. One possible approach for information dissem-ination in such networks is to replicate informationat multiple nodes acting as repositories, and employquorum based strategies to update and query informa-tion. We propose three such strategies that also uselocal knowledge about the reachability of repositories tojudiciously select quorums. The primary goal is highavailability of information in the face of network par-titioning. We also consider four policies to determinethe appropriate time to perform updates. Experimentalresults indicate that a hybrid information managementstrategy and an absolute connectivity-based update trig-ger policy are most suited for partitionable ad-hoc net-works.1 IntroductionExisting solutions for dissemination of informationin static or cellular networks may not be applicable toad-hoc networks. First, these solutions usually do notconsider changing topology of the network backbonethat contains the information servers. Second, the pos-sibility of partitioning means that some nodes may notbe able to communicate updates to other nodes and/ormay be unable to retrieve the latest information onqueries. Third, earlier solutions that do consider net-work partitioning approach the problem from the di-rection of replica consistency in distributed databases.While the problems are similar, several instances of in-formation dissemination problem in ad-hoc networksare simpler in nature. Employing the sophisticatedreplica consistency solutions would result in reducedavailability of data, while also incurring unacceptablyhigh communication overheads. Fourth, unlike tradi-tional distributed database systems where the timingof updates is independent of network topology, in ad-hoc networks location sensitive information should beupdated as a function of network topology. So, it is

important to determine:1. When to update information?2. Where to send updates?3. Which nodes to query for information?1.1 Problem Description and System ModelLet us consider a building on �re. A large number,say N , of �re�ghters have been assigned to extinguishthe �re. Among the N �re�ghters are a small number,say n, of o�cers whose job is to collectively managethe entire operation. All the o�cers need to be able tocommunicate amongst themselves. Also, an o�cer canact as a proxy for all ordinary �re�ghters with whomit can communicate.To increase the e�ciency of the operation, and forthe safety of the �re�ghters, it is important to:1. track the location of each �re�ghter, and2. gather information about the surroundings of each�re�ghter.A �re�ghter is responsible for updating his/her locationand state information. Also every �re�ghter shouldbe able to retrieve the latest information about everyother �re�ghter. Let each �re�ghter have a small, light-weight wireless communication device with the samerange to send and receive information. The devices ofall the �re�ghters collectively form a connected multi-hop wireless network. The o�cers have additionalmemory in their devices to maintain the state infor-mation. Thus, the o�cers' devices collectively act asinformation servers for other �re�ghters. 1The multi-hop wireless network formed by the �re-�ghters may get partitioned into disjoint components.The duration for which the network stays partitioned,the identities of nodes in each partition, and the waysin which partitions merge are non-deterministic.1While it is possible to add more memory to every �re�ghter'sdevice without signi�cantly increasing the weight, broadcastingevery update to every node would incur high communicationcosts and drain the batteries quickly.

Thus, a �re�ghter should choose the o�cer(s) to re-ceive its updates without any prior knowledge of futureaccess patterns to its data. Similarly, a querying nodeshould send its query to some o�cer(s) without anyprior knowledge of the mobility pattern of the nodewhose information it is trying to retrieve. The goal isto maximize the availability of latest state informationabout nodes in a partitionable network while incurringreasonable communication overheads.The set of �re�ghters constitute an ad-hoc networkof N nodes. The n o�cers (n << N) correspond todesignated servers in the network. The remaining �re-�ghters are ordinary nodes or clients. There is no a pri-ori association between a client and a server, and thenetwork may get partitioned. To mitigate the impactof partitioning we propose to use a data replicationscheme for information management. Each update bya client is sent to a subset of servers. Similarly, eachquery is also sent to a subset of servers. Designingsubsets (quorums) that always intersect has been ex-tensively studied in the context of connected networks.The challenge is to construct the query and updatesubsets so as to maximize the probability of a hit (aquery returning latest state information about a node)even when the network is partitioned.The proposed solution has to perform e�cientlyboth when the network is connected and when it ispartitioned. Successive updates/queries by a nodeneed not be sent to the same subset of servers. Thisis because the set of reachable servers changes withtime. However, a node could employ recent informa-tion about server reachability to determine the subsetof servers to which queries and updates are sent.2 Related WorkDavidson, Garcia-Molina and Skeen [4] state thatdata replication results in high availability of informa-tion in the presence of failures and network partition-ing. However, if unconstrained updates by di�erentnodes are permitted in multiple partitions of a net-work, the replica values may diverge. Consequently,queries in di�erent partitions would return inconsis-tent values. To ensure correctness, one would need toprevent updates in all but one partition. Thus, avail-ability and correctness are mutually con icting goals.Basically, the causes of error are the write-write con- icts and read-write con icts among replicas in a par-titioned network. A detailed description can be foundin [4].In order to avoid con icts various mutual exclusionbased solutions are proposed. In the voting approachproposed by Gi�ord [6] each replica is assigned a num-ber of votes. A majority of votes is needed for updates.Also, the sum of votes needed for queries and updatesshould exceed the total number of votes. This ensuresthat two updates cannot happen concurrently. Also,if the network is partitioned, the same data item can-not be updated in two di�erent partitions. However,

if the network is partitioned into more than two par-titions such that no partition has a majority of votesthe update operations are not possible. Thus, dataavailability is reduced.To address this problem, dynamic voting schemeshave been proposed by Paris and Long [13], Jajodia andMutchler [8], Barbara, Garcia-Molina and Spauster [3],among others. In [13], dynamic voting allows data ac-cess to proceed as long as strict majority of currentalive physical copies are accessible. If the number ofaccessible copies is equal to the number of inacces-sible copies data accesses are not possible. However,this problem can be resolved by lexicographic dynamicvoting wherein all nodes are totally ordered by nodeidentity which is used to break the ties [8]. In [3] twodynamic voting schemes have been described, namelythe group consensus approach and the autonomous re-assignment approach. In the group consensus approachthe nodes in the majority agree upon a new vote as-signment either in a distributed fashion or by electinga leader. Then, the two-phase commit protocol is em-ployed among the majority nodes to install the newvotes. In the autonomous reassignment approach eachnode independently picks a new vote value which canbe installed only if the node can obtain a majority ofthe votes. These dynamic voting schemes require ex-tensive communication between servers for vote reas-signment. This is a serious concern for ad-hoc wirelessnetworks with scarce communication bandwidth.El Abbadi, Skeen and Cristian [1] proposed theaccessible copies algorithm that employs the read-one/write-all protocol. A datum is accessible if a ma-jority of its replicas are present in the same partition.A query on accessible datum is performed by readingits nearest copy. An update writes to all copies of thedatum in the partition. Thus, queries and updates canbe performed in only one partition. Also, all copies ofthe datum remain consistent.Instead of relying on strict majorities, quorum basedsolutions create a set of subsets of the nodes in the sys-tem. Each subset is called a quorum. To perform anupdate or a query permission from all nodes in at leastone quorum is su�cient. The quorums need not bea majority of nodes. Herlihy [7] proposed a dynamicquorum adjustment scheme in which quorums can beadjusted on partitioning and mergers. If an operationis unable to progress using one quorum it may be ableto make progress by using another more favorable quo-rum. Such a selection is aided by hints provided by theunderlying system about node accessibility.Previous Research: Though the problem we setabout solving appears very similar to the replica con-sistency problem, there are a few subtle di�erences.The algorithms described above seek to avoid write-write con icts across partitions, and ensure serializ-ability among all the update transactions. For someweak consistency solutions it is acceptable if read-only

transactions are not serializable with respect to eachother.In the problem at hand, write-write con icts aresimply not possible. A datum is updated by only onenode. Let a node perform successive writes to two dif-ferent quorums of replicas. If we assume that eachnode has a monotonically increasing local clock, likeLamport's clock [10], the two writes can be easily seri-alized if each replica stores the data item as a <value,writer's timestamp> tuple. Let a query be sent to aset of replicas, returning di�erent values for the datum.The most recent of these values can be identi�ed usingthe timestamp received with the value.We employ a variation of the quorum based scheme.As in [1, 7] we, too, use previous knowledge aboutnode accessibility to select sets of replicas that receivequeries and updates. However, unlike [1], we do not as-sume an ideal network in which nodes detect networkpartitioning immediately.In most replica consistency solutions, on merger ofpartitions the replicas in the merging partitions aresynchronized so that they have the same value. Thisincurs very high communication overheads which arenot acceptable in low bandwidth wireless networks.The synchronization tra�c on mergers may precludeall other updates and queries in the network for sometime. So, we intend to adopt an optimistic approach.We do not perform synchronization between copies atthe time of mergers. This saves on communication costwithout signi�cantly degrading the accuracy of infor-mation returned by queries. Moreover, when new linksare established we do not have to run tests to deter-mine if hitherto disjoint partitions are merging. Also,the simplicity of the solution is important for its fastand error-free implementation in an actual network.Our approach is guided by similar principles as theBayou project [16]. However, there are also some im-portant di�erences. In Bayou servers engage in anti-entropy sessions to exchange information and reachconsensus about the order of updates. We do not usesuch a scheme for reasons described above. Also, inBayou each datum has a primary server responsiblefor committing updates. We do not have the notionof primary servers. Also, Bayou is not designed forreal-time applications and has tentative updates thatare later committed or rolled back. We are targetingreal-time applications where inaccuracy of informationis sometimes preferable to no information at all.3 Quorum Based SolutionGiven a set S of servers, a quorum system is a setof m subsets of S, namely S0, S1; : : : ; Sm�1, such that:1. [m�1i=0 Si = S.2. Si \ Sj 6= �, for 0 � i; j � m� 1.The sets Si are constructed a priori and every nodeknows the membership of these sets. Given n servers,it is possible to form quorums of size O(pn) [11].

Information Update: When a node x wishes to up-date some information it timestamps the datum withits local clock value. We assume loosely synchronizedclocks such that the time between successive updatesis greater than the maximum clock skew. Then, thefollowing actions are performed:1. Node x randomly selects a quorum Si from theset of quorums and sends UPDATE message,timestamped with its local clock value, to allservers in the quorum.2. The servers, on receiving the UPDATE message,overwrite their old copy of the data item with thenew copy. If they do not have an old copy of thatdata item, they simply add the information re-ceived in the message to their database. Option-ally, the servers may also send a positive acknowl-edgment to x.Let an UPDATE message be �rst sent to quorum Si.Later, let another UPDATE, for the same data item,be sent to quorum Sj . Then, all servers in the setSi �Sj have outdated versions of the data item, whileall servers in Sj have the latest version of the data item.An alternative would have been to send a DELETEmessage to servers in the set Si � Sj . While this al-ternative would reduce the memory requirements, itwould increase the communication overheads.Information Query: When node y wishes to makea query the following actions are performed:1. Node y randomly selects a quorum Sj and sendsa QUERY message to all servers in the quorum.2. When a server receives a QUERY for a datum andhas a copy of it, the server sends a REPLY con-taining the information along with the timestampassociated with the datum. Otherwise, the serversends a NULL reply.3. Receiving all the REPLY messages,y selects thevalue of the datum with the greatest timestamp.As two quorums always intersect in the �xed net-work, the set of queried servers is bound to containat least one server that belonged to the quorum thatreceived the latest update. Hence, each query returnsthe latest value of the queried data item. Such a queryand update strategy has been previously employed forlocation management in cellular networks [9, 14, 15].Why random selection of quorums: Let a quo-rum be initially selected by a mobile node such that allservers in the quorum are reachable from that node.Subsequently, if the same quorum is selected wheneverthe node has to make an update or a query, there isno guarantee that all serves in that quorum would bereachable. So, deterministic quorum selection is nota good idea. Moreover, random selection gives a bet-ter chance of load balancing among the servers evenif some node suddenly starts generating updates andqueries at a higher rate than other nodes.

4 Proposed SolutionWe need to address the issues of when to triggerupdates, and where to send updates and queries so asto mitigate the impact of network partitioning.4.1 When to Update?Our aim is to propagate information about nodessuch that other nodes, on querying for information, getas recent a version as possible. Had there been a meansto predict network partitioning, updates could be per-formed just before such partitions occur. However, inthe absence of such an oracle, a node may have to ob-serve changes in network topology and make guessesabout reachability to/from other nodes. We proposefour di�erent policies to trigger updates in increasingorder of sophistication:1. In the time-based strategy the time between suc-cessive location update attempts by a node is ex-ponentially distributed, with a mean of t units.For our simulations we set t to 1:0 unit.2. The time and location-based strategy is similar tothe time based strategy. However, in this strategya node remembers its location when it last sentan update. If the node's location unchanged sincethe last successful update, no update requests aresent.3. In the absolute connectivity-based strategy a nodesends an update when a certain pre-speci�ed num-ber of links incident on it have been established orbroken since the last update.4. In the percentage connectivity-based strategy anupdate is triggered when a pre-speci�ed percentof the links incident on it have changed since thelast update.The last two update strategies are motivated by thefeeling that in ad-hoc networks the frequency of up-dates should be a function of the dynamism exhib-ited by the network. In [2], Bar-Noy, Kessler andSidi showed that movement-based and distance-basedupdates performed better than time-based updates ina cellular network. As there is no notion of cells inad-hoc networks the movement-based strategy is notapplicable. Also, the relative position of nodes (andthe resultant topology) is a more important piece ofinformation than their absolute location or absolutedistance moved. Hence, distance-based updates mayhave limited usefulness in ad-hoc networks. The ab-solute connectivity- and the percentage connectivity-based approaches are an attempt to capture the rela-tive change in topology.4.2 Where to Update and Query?There is an implicit assumption in the quorum basedscheme that all the servers in the selected quorum

are reachable from the node that sends the update orquery. This assumption is valid for cellular networkswhere base stations and servers maintaining locationinformation belong to the wired backbone network andare reachable from each other virtually all the time.However, in an ad-hoc network there is a possibilitythat the network gets partitioned as described earlierin the context of �re�ghters in a building.Let node x select quorum Si = fsx1; sx2; : : : ; sxngto update the value of a data item. If some elementsof Si are in a di�erent partition from x at the time ofthe update. They do not receive the update. Only theservers in S0i � Si receive the update. Subsequently, letanother node y select a quorum Sj = fsy1; sy2; : : : ; syngto query the value of the data item.Some elements of Sj may be in a di�erent partitionfrom y at the time of the query. So, only servers in S0j �Sj receive the query. Thus, there is a possibility thatthe query may not return any information, or returnstale information. This is due to the fact that eventhough Si \Sj 6= �, it is possible that S0i \S0j is empty.One possible solution would be to send the queryonce again to another quorum Sk hoping that S0i \ S0kis non-empty. However, this solution is not acceptablefor two reasons:1. It increases the communication overheads in abandwidth poor network, without any guaranteesof returning the latest information.2. The querying node may not be able to realize thatit has stale information. If at least one serverreturns a non-NULL reply the node accepts thevalue with the highest timestamp. There may beno way for the querying node to know that thereexists at least one copy of the data item with ahigher timestamp, somewhere in the network.To alleviate the problem of query failures we needto make informed selection of quorums at the time ofupdates and queries. The idea is to select quorums sothat the sets Si�S0i and Sj �S0j are as small as possible.The smaller these sets, the greater the probability ofthe set of reachable queried servers intersecting withthe set that received the latest update.Disquali�ed List: We propose to use a heuristic toselect servers for updates and queries. Node x main-tains a disquali�ed list, DQLx, containing servers thatx believes to be unreachable. Note that DQLx repre-sents x's perception of unreachability. There may beservers that are not reachable from x, yet not in DQLx.Similarly, there may be servers that are reachable fromx, but also in DQLx.The instance of the heuristic running at node x per-forms the following tasks:1. Maintenance of DQLx.2. Selection of servers for updates/queries on the ba-sis of DQLx's composition.

3. Communication with selected servers for updatesand queries.One of the goals of the heuristic is to be able toupdate the disquali�ed list as part of the query andupdate operations, without incurring any extra com-munication overheads and delays.4.3 HeuristicInitially, disquali�ed list, DQLxis empty. We pro-pose three strategies to select servers for updatesand queries using information in the disquali�ed list,namely: (i) Select then Eliminate (STE), (ii) Elimi-nate then Select (ETS), and (iii) Hybrid strategies.Select then Eliminate (STE) Strategy: In thisstrategy, node x randomly selects a quorum Si. Fromthe set of servers Si, x eliminates those in DQLx. LetS0i = Si �DQLx. Update/query messages are sent toall servers in S0i.Node x expects to receive a reply from every serverto which a message is sent within time Ttimeout. Anupdate is considered to be successful if at least oneserver sends an acknowledgment of receipt of the up-date message. In the case of queries, if one or moreservers send non-NULL timestamped information, thereceived copy with the greatest timestamp is selected,and the operation is considered to be successful. Thisprocess is repeated until there is success.If a reply is not received from a server s withinTtimeout, node x concludes that s is not reachable andadds s to DQLx. Once a server s has been addedto DQLx it stays in it for a disquali�cation dura-tion, �DQL. At the end of �DQL, s is removed fromDQLx.The value of �DQL is a parameter of the heuris-tic and should be determined on the basis of the con-nectivity characteristics of the network.If all the nodes are concentrated in a small area, orif the nodes have a large wireless range, the probabilityof network partitioning is low. Even if the network getspartitioned, the mean time before the partitions mergewith each other is also expected to be small. In suchcases having a small value of �DQL would be advisable.If a much smaller value of �DQL is selected, a servermay be removed from DQLx when it is still unreach-able from x. So, x will end up trying to probe serversthat are still not reachable, increasing the communi-cation overheads. On the other hand, if a very largevalue of �DQL is selected, some previously unreachableservers will not be probed for a long time even afterthey are once again reachable. This may result in un-availability of information, reduced utilization of someservers, and excessive load on other servers.Eliminate then Select (ETS) Strategy: In thisstrategy, node x �rst eliminates all quorums that haveat least one node in DQLx. One of the remaining quo-rums is randomly selected and messages are sent to

servers in this quorum. If at least one server sendsan acknowledgment in the case of an update, or anon-NULL value in the case of a query, the opera-tion is said to succeed. Let some servers not respondwithin Ttimeout, while all others send a NULL reply.Then the servers that did not respond are added toDQLx for �DQL time. Also, quorums containing atleast one server in the expanded DQLx are eliminated,and one of the remaining quorums is selected for theupdate/query operation. This process is repeated untilthe update/query succeeds. If all quorums get elimi-nated before there is success, the update/query opera-tion is said to fail.Hybrid Strategy: The hybrid strategy uses ETS forupdates and STE for queries. The idea is to maximizethe number of servers receiving the updates with thehope that this will result in more accurate informationretrieval by queries. The motivation for the hybridstrategy will become clear in the following paragraphs.Comparison of STE, ETS and Hybrid Strate-gies: STE tries to maximize availability of informa-tion, while ETS tries to maximize accuracy of queriesat the expense of availability. In the ETS strategy anattempt is made to identify a quorum with no serverin the disquali�ed list. This reduces the number ofquorums available for update or query while maximiz-ing the number of servers receiving the query/update.In the STE strategy all the quorums are available forupdate or query thus increasing availability.However, the increased availability of STE may ac-tually reduce the accuracy of information. One of theconcerns of a transaction processing system for parti-tioned networks is correctness within partition. Cor-rectness within partition means that the queries in apartition should return the values stored by the latestsuccessful update in that partition to the same dataitem. When updates are sent to fewer servers, theprobability of queries hitting at least one server withthe latest information is also lower. This probability isfurther reduced as the queries may also be sent to fewerservers. Even though the query and update sets inter-sect, their intersection may belong to another networkpartition. STE seems to run counter to the desired goalof minimizing the size of the set Si � S0i, where S0i isthe set of reachable servers of quorum Si. Hence, theprobability of obtaining stale information is greater.Let us consider the situation when DQLx accuratelyrepresents the servers that are not reachable from nodex. Then, due to the reason stated above, queries inthe STE strategy may return stale information or noinformation at all. However, in the ETS strategy, dueto the way the quorums are eliminated a priori, allservers in the quorums selected for an update and asubsequent query belong to the same partition, and areindeed reachable. Hence, correctness within partitionis guaranteed for ETS in this situation.

However, the ETS strategy provides a lower avail-ability of data than STE. Once again, let DQLx accu-rately represent reachability information, and let up-dates or queries succeed in one partition. This meansthat all servers of at least one quorum are within thatpartition. Due to the non-empty intersection propertyof quorum pairs, all quorums will be eliminated in otherpartition(s) by the ETS strategy. So, queries and up-dates cannot be performed in those partition(s).If availability of stale information is preferable to noinformation at all, STE may be preferred over ETS.Of course, if �DQL is di�erent from the elapsed timebetween partition detection and merger, DQLx will beinaccurate, and even ETS may return stale informationwhile STE may have reduced availability.The hybrid strategy attempts to increase the accu-racy of future queries by performing updates using theETS strategy. At the same time it strives to maximizethe availability of information during queries by usingthe STE strategy. It is hoped that performing updatesat a greater number of sites will actually increase theprobability of a query returning fresh information, eventhough the queries are sent to fewer servers. This maybe especially useful in systems where there are manymore queries than updates. Savings in communicationoverheads during the more frequent queries may morethan o�set for the extra communication overheads in-curred by the less frequent updates.4.4 Communication and Storage OverheadsA copy of every data item, fresh or stale, may beresident at each server. So, every data item requiresO(n) storage. This can be reduced if, at the time ofupdate, the node also sends delete messages to serversthat were in its previous update quorum, and are not inits present update quorum. However, this will requireadditional communication.Given n servers, there exist quorum formationschemes that give quorums of size O(pn) [11]. So, evenif an update or query requires a constant number of re-tries before success, the communication complexity isO(pn). Note that n is usually much smaller than thetotal number of nodes N . Hence, the communicationand storage overheads are reasonable.5 Simulation ExperimentsWe conducted simulation experiments using theCSIM18 simulation engine [12]. We compared the per-formance of the STE,ETS and Hybrid strategies withthe simple quorum based strategy that does not ac-count for partitioning and does not use the disqual-i�ed list. In the simple quorum based scheme, anode continues to randomly select a quorum and sendqueries/updates until either the operation is performedsuccessfully, or all quorums have been tried withoutsuccess. In the latter case, the operation is said to fail.We assumed a system composed of 100 mobilenodes. Of these, 25 nodes were assumed to act as

servers. Initially, the 100 nodes are randomly distrib-uted in a square region with edges of 1000 length units.During the simulation experiment all the nodes are al-lowed to move only within this square.In order to model mobility of nodes we cycle throughthe following two steps: (i) the duration for which anode stays at a location is exponentially distributedwith mean 0:5 time units, (ii) at the end of this dura-tion, the node randomly selects another location withinthe square region that is at most 75 length units awayfrom its current location, and instantaneously movesto the new location.A wireless link is assumed to be present betweena pair of nodes if they are within 150 length units ofeach other. Each time a new link is established, oran old link is broken, the simulator executes Floyd-Warshall's algorithm [5, 17] to determine the reacha-bility of nodes, and the presence of partitions in thenetwork. Note that this computation of reachabilitydoes not have to be performed by nodes as part of theirupdate and query function. In subsequent experimentswe vary the wireless range to values other than 150 andmeasure the impact of partitioning on the informationdissemination strategies.The 25 servers are logically arranged in a 5�5 squaregrid, and 25 quorums are formed: each quorum is com-posed of the union of a row and a column of serversin this logical grid. Thus, each quorum consists of 9servers. The quorums are assigned distinct sequencenumbers in the range 1 � 25. Every pair of quorumshas two servers in common.The update and query operations work on locationdependent information.2 As each node has its uniquelocation, there are 100 distinct pieces of information.The time between successive queries by a node is ex-ponentially distributed, with a mean of 0:5 time units.A querying node randomly selects one of the 100 nodesas the node whose location it wishes to query. Querymessages are sent to servers based on the strategy inuse: ETS, STE, or Hybrid. If a node discovers that aserver is unreachable, the node places the server in itsdisquali�ed list for a period �DQL = 0:25 time units.We also varied the value of �DQL to measure its impacton server availability, utilization and load balancing.In each simulation run, no data collection was per-formed until the �rst 10; 000 queries were �nished (suc-cessfully or unsuccessfully). This was done to elimi-nate the impact of any initial transient e�ects. Subse-quently, data collection was performed until the next100; 000 queries were completed. Each data point inthe graphs represents the mean value obtained from�ve simulation runs with identical input parameters,but di�erent random number generation seeds.2Note that each �re�ghter sends its location and the statusof the �re, etc. in his/her location as part of the update.

5.1 Simulation ResultsWe measured various performance characteristics ofthe ETS, STE and Hybrid strategies versus the simplequorum based strategy that does not use the disquali-�ed list. For updates the time and location-based policywas employed.Query success and freshness: The impact of wire-less range of nodes on the number of successful queriesand the number of queries returning fresh information(information stored by the latest update for that da-tum) is shown in Figure 1. When the wireless range isequal to 250, or 150 length units almost all the queriessucceed regardless of the strategy employed, as seen inthe graph on the left hand side of Figure 1. This is be-cause for these values of wireless range there are veryfew partitions in the network. Even when partitionsdo occur, the partitions get merged quickly. However,there is a signi�cant drop in the number of successesfor wireless range of 100 length units. This is due to in-creased probability of network partitioning. The ETSstrategy su�ers the most due to increased partitioningprobability. This is due to the fact that when the net-work is partitioned the ETS strategy may not be ableto �nd any quorum because at least one server of everyquorum is in the querying node's disquali�ed list.When the wireless range is equal to 250 or 150 mostof the queries return latest information. But, whenthe range is reduced to 100, the number of queries re-turning the latest information is signi�cantly reduced.Once again, this is due to increased likelihood of net-work partitioning. Moreover, the fraction of successfulqueries that yield the latest information is much higherfor the ETS strategy than the STE and simple quorumbased strategies, as observed by looking at the graphson the left as well as on the right in Figure 1. Thisis consistent with our prior analysis of the STE strat-egy emphasizing availability over accuracy, and ETSemphasizing accuracy over availability.In Figure 1, the value of �DQL is set to 0:25. Similarresults were obtained for other values of �DQL as well.Among the four strategies, ETS has fewest queries re-turning the latest information. Also, ETS is signif-icantly poorer than the other strategies in terms ofquery success. This leads us to believe that ETS isnot a good strategy for information dissemination.Communication Overheads: Figure 2 shows thatthe ETS, STE, Hybrid, and simple quorum basedstrategies incur comparable communication costs forlarger values of wireless range. However, for wirelessrange of 100 length units when the probability of net-work partitioning is more, the simple quorum basedstrategy incurs very high communication cost. For thesame wireless range the communication costs are com-paratively low for the STE, Hybrid, and ETS strate-gies, as was expected. The ETS strategy has the lowestcost because quite often it cannot even �nd a quorum to

send the queries to. The simple quorum based strat-egy has the highest overheads because the nodes tryto send their queries to even those servers that are notreachable. Also, when there is a failure, the nodes keeptrying other quorums until either some information isretrieved or all quorums have been tried.

0

500000

1e+06

1.5e+06

2e+06

2.5e+06

3e+06

3.5e+06

4e+06

4.5e+06

5e+06

250 150 125 100

Total

no.of

query

mes

sage

s sen

t Wireless Range

Impact of wireless range and DQL on communication overheads (delta = 0.25)

Without DQLSTEETSHybrid

Figure 2. Impact of wireless range and quo-rum selection policy on communication over-heads.Load Distribution: We measured the distributionof load (total number of messages received by a server)on all servers when �DQL was set to 0:25 and the wire-less range was 100. In all the four strategies di�erentservers receive a di�erent number of query messages,with the simple quorum based scheme (without DQL)experiencing the maximum tra�c and also the max-imum variance in tra�c. The reason for the load tobe unevenly distributed is that the network gets par-titioned. Due to this every server is no longer equallylikely to receive the query messages. However, whenthe wireless range was increased to 150 and 250 dis-tance units network partitioning was a rare occurrenceand the load was evenly distributed across all servers.Number of attempts: Figure 3 shows the cumula-tive number of successful queries versus the number ofquery attempts, where a query attempt corresponds toselection of a quorum and sending messages to reach-able servers in that quorum. The STE, Hybrid, and thesimple quorum based strategies require almost samenumber of attempts for a success. The ETS strategy re-quires fewer attempts because the number of quorumsavailable after eliminating those that contain serversin DQL is less. Similar results were obtained for othervalues of wireless range and �DQL.Most of the successful queries require only one at-tempt. As the curves plateau after about four to �veattempts it can be safely said that if a query has not

0

20000

40000

60000

80000

100000

120000

250 150 125 100

No.of

succ

esse

s

Wireless Range

Impact of wireless range and DQL on successful queries (delta = 0.25)

Without DQLSTEETSHybrid

0

5000

10000

15000

20000

25000

30000

35000

40000

45000

50000

250 150 125 100

No.of

queri

es re

turnin

g late

st inf

ormati

on

Wireless Range

Impact of wireless range and DQL on retrieval of latest data (delta = 0.25)

Without DQLSTEETSHybrid

Figure 1. Impact of wireless range and quorum selection poli cy on the number of successful andaccurate queries.succeeded after four to �ve attempts, there is no pointin trying further. The small increase in the number ofsuccesses may not be worth the signi�cant increase inthe communication cost incurred in doing so.

0

20000

40000

60000

80000

100000

0 5 10 15 20 25

No.o

f que

ries

No.of attempts

Cumulative no. of successful queries vs. no. of attempts (delta=0.25,range=100)

Without DQLSTEETSHybrid

Figure 3. Number of attempts required by asuccessful query.Impact of Update Policy: Figure 4 shows that theupdate policy can have a signi�cant impact on thenumber of information updates performed by nodesduring the lifetime of the simulation. As the Hybridstrategy performs signi�cantly better than ETS, andhas lower communication overheads than simple quo-rum based strategy, we only studied the impact of theupdate policy on the Hybrid strategy. For the absoluteconnectivity-based policy, an update is triggered when�ve edges incident on the node have changed sincethe last update. For the percentage connectivity-based

strategy, an update is triggered when 20% of the linkshave changed since the last update.Maximum number of updates were performed whenthe percentage connectivity-based policy was em-ployed. This is because with 100 nodes scattered overthe entire region, the network is very sparsely con-nected. So, only a few links need to change to exceedthe 20% threshold. For the same reason, it takes along time for �ve links to change. Hence, the absoluteconnectivity based policy triggers very few updates.However, inspite of the signi�cantly fewer updates, thenumber of successful, fresh and stale queries for the ab-solute connectivity-based update policy is comparableto that of the other policies.However, before we rush to judgment, let us alsoconsider the error in the location information returnedby all queries (fresh as well as stale). Note that a queryreturns the location of the queried node at the timethat node updated its location and state information.Error indicates the distance between the actual loca-tion of the queried node at the time of the query, andthe location information returned to the querying node.Recalling our example, a large error in location deter-mination will mislead the �re�ghters about the posi-tion of their colleagues, and seriously jeopardize boththeir safety and the success of their operation. Figure 4shows the number of queries on the y-axis with loca-tion error no more than the value on the x-axis. Thepercentage connectivity-based policy has the best per-formance with about 20,000 queries returning locationinformation that is within 25 distance units of the ac-tual location of the queried node. About 40,000 queriesreturn location information that is no more than 100distance units away from the actual location of thequeried node, and so on. The graph also indicates thatinspite of signi�cantly fewer updates, the error in loca-tion determination for the absolute connectivity-based

policy is comparable to other policies.This seems to indicate that, at least for sparse net-works, the absolute connectivity-based policy is mostsuitable. Even though the other policies perform moreupdates, due to the highly partitioned nature of thenetwork most of the updates are visible to a small sub-set of querying nodes. A majority of nodes in the otherpartitions do not see the updates and continue to re-turn outdated location information.Impact of Network Density: We increased thenumber of nodes in the network from 100 to 200, whilekeeping the number of servers unchanged at 25. Ascan be seen in Figure 5, with increased density (due tothe presence of 200 nodes in the same area), networkconnectivity improved and there were fewer partitions.As a result, the error in location information returnedby queries was reduced. This is because the serversreceiving the latest update were reachable from morenodes in the network.Also, with increased connectivity the absoluteconnectivity-based update policy triggered more loca-tion updates. Greater the number of links incident ona node, the less time it takes for �ve links to changesince the last update. Still the absolute connectivity-based update policy performs fewer updates than thetime-based and percentage connectivity-based updatepolicies, and only slightly more updates than the timeand location-based update policy. Just as in the caseof sparse networks, the number of successful, fresh andstale queries are comparable for all the four policies.Thus, our simulation experiments seem to indicatethat the absolute connectivity-based update policy isthe most suitable among the four policies considered.When the network is sparse and updates are going tohave less of an impact it triggers fewer updates. Whenthe network is dense, the topology is changing morerapidly, and updates can be seen by more nodes it trig-gers a greater number of updates.Impact of disquali�cation period: For values of�DQL greater than 0:25 (we tried values of 0:50 and1:00) the simple quorum based strategy (without DQL)exhibited a higher number of successes than the ETS,STE and Hybrid strategies. This may be due to bedue to the fact that with a larger value of �DQL, if aserver is found to be unreachable, a node does not sendupdates and queries to that server for quite some timeeven after the server has once again become reachable.A �DQL value lower than 0.25 yielded the sameperformance in terms of query success, freshness, andstaleness. However, the communication overheads in-creased. This seems to be due to the possibility thatnodes may resume sending updates and queries toservers prematurely: the servers may still be in a dif-ferent partition.

6 ConclusionA scenario for the use of an ad-hoc network of mo-bile nodes was presented, where the network can getpartitioned and reconnected several times. The prob-lem of information dissemination in such situations wasformalized. In order to increase the availability ofinformation, data replication was considered. Threequorum based strategies STE, ETS and Hybrid (ETSfor updates, STE for queries) have been proposed.The STE strategy optimizes availability of information,whereas ETS tries to maximize accuracy of informationat the expense of availability. This is also indicated bythe simulation results. The Hybrid strategy exploitsthe advantages of both the strategies. Four policies totrigger updates were considered. Of these the absoluteconnectivity-based policy was found to have the bestperformance. The absolute connectivity-based policytriggers more updates in dense networks where the im-pact of updates will be visible to a greater number ofnodes. In sparse networks this policy triggers fewerupdates as the updates have a lower potential of beingvisible to other nodes.References[1] A. El Abbadi, D. Skeen, and F. Cristian. An E�cientFault-Tolerant Algorithm for Replicated Data Man-agement. In Proceedings of the 5th ACM SIGACT-SIGMOD Symposium on the Principles of DatabaseSystems, pages 215{229, 1985.[2] A. Bar-Noy, I. Kessler, and M. Sidi. Mobile Users: ToUpdate or not to Update? In Proceedings of IEEEINFOCOM, pages 570{576, 1994.[3] D. Barbara, H. Garcia-Molina, and A. Spauster. In-creasing Availability Under Mutual Exclusion Con-straints with Dynamic Vote Reassignment. ACMTransactions on Computer Systems, 7(4):394{426, No-vember 1989.[4] S.B. Davidson, H. Garcia-Molina, and D. Skeen. Con-sistency in Partitioned Networks. ACM ComputingSurveys, 17(3):341{370, September 1985.[5] R.W. Floyd. Algorithm 97 (SHORTEST PATH).Communications of the ACM, 5(6):345, 1962.[6] D.K. Gi�ord. Weighted Voting for Replicated Data. InProceedings of the 7th Symposium on Operating Sys-tems Principles, pages 150{162. ACM, 1979.[7] M. Herlihy. Dynamic Quorum Adjustment for Parti-tioned Data. ACM Transactions on Database Systems,12(2):170{194, June 1987.[8] S. Jajodia and D. Mutchler. Integrating Static andDynamic Voting Protocols to Enhance File Availabil-ity. In Proceedings of the 4th International Conferenceon Data Engineering, pages 144{153. IEEE, 1988.[9] G. Krishnamurthi, M. Azizo glu, and A.K. Somani.Optimal Location Management Algorithms for Mo-bile Networks. In Proceedings of The Fourth AnnualACM/IEEE International Conference on Mobile Com-puting and Networking (MobiCom'98), pages 223{232,October 1998.

0

20000

40000

60000

80000

100000

Time&Loc Time Absolute Percentage

Num

ber

Update policy

Impact of update policy on performance for a sparse network

Successes Fresh Stale Updates

0

20000

40000

60000

80000

100000

0 200 400 600 800 1000 1200 1400

Cum

ulativ

e no.o

f que

ries

Error range

Location error distribution for different update policies using hybrid query strategy

Time & LocationTime BasedAbsolute ConnectivityPercentage Connectivity

Figure 4. Impact of update policy on freshness and accuracy o f information returned by queries.

0

20000

40000

60000

80000

100000

0 200 400 600 800 1000 1200 1400

Cum

ulativ

e no.o

f que

ries

Error range

Location error distribution for sparse & dense networks using percentage connectivity policy for updates

Sparse NetworkDense Network

0

20000

40000

60000

80000

100000

Time&Loc Time Absolute Percentage

Num

ber

Update Strategy

Impact of update strategy on performance for a dense network

Successes Fresh Stale Updates

Figure 5. Impact of network density on freshness and accurac y of information returned by queries.[10] L. Lamport. Time, Clocks and the Ordering of Eventsin a Distributed System. Communications of the ACM,21(7):558{565, July 1978.[11] M. Maekawa. A pN Algorithm for Mutual Exclusionin Decentralized Systems. ACM Transactions on Com-puter Systems, pages 145{159, May 1985.[12] Mesquite Software, Inc., 3925 West Braker Lane,Austin, TX 78759. CSIM18 Simulation Engine, 1998.[13] J.-F. Paris and D.D.E. Long. E�cient Dynamic Vot-ing Algorithms. In Proceedings of the 4th Interna-tional Conference on Data Engineering, pages 268{275. IEEE, 1988.[14] R. Prakash, Z. J. Haas, and M. Singhal. Load Bal-anced Location Management for Mobile Systems us-ing Dynamic Hashing and Quorums. Technical ReportUTDCS-05-97, The University of Texas at Dallas, Oc-tober 1997.[15] R. Prakash and M. Singhal. A Dynamic Approach toLocation Management in Mobile Computing Systems.In Proceedings of the 8th International Conferenceon Software Engineering and Knowledge Engineering

(SEKE'96), pages 488{495, Lake Tahoe, Nevada, June1996.[16] D. B. Terry, M. M. Theimer, K. Peterson, A. J. De-mers, M. J. Spreitzer, and C. H. Hauser. ManagingUpdate Con icts in Bayou, a Weakly Connected Repli-cated Storage System. In Proceedings of SIGOPS'95,pages 172{183, 1995.[17] S. Warshall. A Theorem on Boolean Matrices. Journalof the ACM, 9(1):11{12, 1962.