Network-level Synchronization in Decentralized Social Ad-Hoc ...

Network-level Synchronization in Decentralized Social Ad-Hoc NetworksMatthew Dobson, Spyros Voulgaris, Maarten van SteenDept. of Computer ScienceVrije Universiteit Amsterdam, The Netherlands{mc.dobson, spyros, steen}@cs.vu.nlAbstractSocial ad-hoc networks can be composed of a largenumber of tiny, wireless, battery-powered nodes, and exhibitarbitrary mobility. We aim to enable the executionof distributed applications that depend on capturing networkdynamics across large-scale social ad-hoc networks.Conservation of the nodes’ fixed energy budget is thechief concern in all design decisions, and necessitatesthat wireless communication is kept to a minimum. Inthis paper we present GMAC 1 : a gossiping TDMA MAClayer created to address these problems. We analyze itsefficiency in achieving and maintaining a tight networklayersynchronization with respect to the active periods ofnode radios.Keywords: ad-hoc wireless networks, TDMA MAClayer, duty cycle synchronizationI. IntroductionAt the intersection of wearable computing, wirelessad-hoc networks, and social networks, lies an area wehave dubbed social ad-hoc networks. Such networks arecomposed of nodes carried or worn by people, use wirelesscommunication, do not rely on existing communicationinfrastructures, and are battery powered. Unlike typicalwireless ad-hoc networks, that focus on propagating collecteddata to a sink, social ad-hoc networks aim atcapturing network dynamics at regular intervals.A number of applications can benefit from capturingnetwork dynamics. Consider, for instance, a (large) groupof people at a conference or similar social event, eachwearing a small unobtrusive electronic badge with a limitedradio range. By simply measuring how often and for1 GMAC is protected by US Patent Application 12/215,040. GMAC isavailable free of charge for academic use.978-1-4244-9142-1/10/$26.00 c○2010 IEEEhow long two badges are within range of each other, wecan register social interaction and study the structure of thesocial network. Furthermore, by aggregating and disseminatingdata we can even stimulate social interaction, forinstance by a social game where groups of people (e.g.,students of the same department) increase their score bytalking to members of other groups, and lose points whensticking among themselves. Finally, a family or group offriends attending a large social event may be informedwhen they come in close proximity to each other, helpingthem to stay in contact. Other applications easily cometo mind, including peer-to-peer messaging, finding peoplewith specific profiles, and crowd management, to name afew.There are three main challenges at the MAC layer thatmust be overcome in order to implement large-scale socialad-hoc networking. The first, and primary concern foralmost any mobile device, is energy consumption. As thesemobile nodes are battery powered, careful and judicioususe of a node’s fixed energy budget is essential. Thesecond problem we face is scalability. We require that ourplatform operates efficiently on a wide variety of networksizes, from a few tens of nodes (e.g., a birthday partyor small restaurant) to thousands of nodes (e.g., a largesporting event or stadium rock concert). Our third majorchallenge is node mobility. In the social settings where ourapplications will be executing people are free to join andleave the network, and to move anywhere they please (i.e.,make arbitrary changes in the network topology). Thesedynamic changes to network topology can wreak havoc onmany algorithms (e.g., routing and leader election) whichassume stable and symmetric connections between nodes.As part of an overall solution, we propose GMAC: agossip-based MAC protocol, with energy-awareness beinga primary goal. GMAC is designed to run on highlyconstrained sensor nodes and it therefore has very low processingrequirements and a small memory footprint. Thisserves to reduce energy consumption, but GMAC’s chiefmechanism for energy conservation is duty cycling. Dutycycling is a common technique from the area of wireless

Fig. 1. GMAC’s TDMA structuresensor networks where individual nodes periodically turnon their wireless radio for only a short period of timeto allow for communication. GMAC’s communication isbased on a periodic gossip paradigm, in which one periodof gossip-based communication is known as a gossipround. During each gossip round, each node sends oneapplication message and one join message. Applicationmessages comprise a small GMAC header with synchronizationand strategy information, and the application data.Join messages carry debug and/or statistical informationfrom the GMAC core, and are also used to aid networksynchronization.Note that most existing MAC protocols, focusing onadapting radios’ duty cycles to variable data propagationneeds, are not appropriate for monitoring network dynamicsat regular intervals. GMAC’s periodic duty cyclingprovides an inherent solution to that, provided that nodesare properly synchronized with each other. The synchronizationmechanism, which is what we concentrate on inthis paper, is the glue that holds the entire network together.Properly synchronized, GMAC’s gossip-based datadissemination will epidemically spread application messagesthroughout the network. Improperly synchronized,network-level packets will collide or go unheard entirely.By extensive simulation, we show that it is feasible forGMAC to dynamically establish and adaptively maintaina synchronized duty cycle for networks of very largescale, even in the presence of drifting clocks and diversetopologies.II. System model / GMAC descriptionGMAC’s core design considerations were predictable(and low) energy consumption and self-adaptability. In orderto accomplish these goals, GMAC combines networklevelgossiping with a low duty-cycle Time-Division MultipleAccess protocol. TDMA protocols divide time into arepeated series of frames comprising a number of fixedlengthslots, each of duration T slot . They attempt toavoid message collisions by scheduling transmissions inunused slots. This scheduling mechanism is known asslot allocation, which GMAC provides in the form ofscheduling strategy modules. For the slot allocation tobe effective, nodes must be synchronized. That is, thenodes participating in this protocol must closely align theirslots, minimizing the offset between all pairs of nodes.To this end, GMAC also supports the use of varioussynchronization modules. GMAC’s TDMA structure isshown in Fig. 1.GMAC maintains a fixed duty-cycle when in a synchronizedstate, and thus constant and predictable energy consumption.At compile-time GMAC computes the largestinteger number of TDMA slots that fit within the desiredframe time, Slots frame = ⌊T frame /T slot ⌋. T slot is basedon the transmit time of a fixed-length GMAC packet anda guard time, which is padded to both ends of a singleTX time to compensate for a possible synchronizationoffset between nodes, as shown in Fig. 1. The TX andguard times are measured at the granularity of the node’sunderlying hardware clock tick, T 2 tick , which determinesthe minimum possible synchronization adjustment. Thetransmit time is based on the radio hardware present on anode, specifically the total time for the radio to transitionfrom standby mode to transmit mode (T StandbyT oT X ) andthe time to transmit a single packet ( P acketSizeDataRate). Thus,T icks T X = ⌊(T StandbyT oT X + P acketSizeDataRate )/T tick⌋ + 1.The guard time is a compile-time constant, currently setto 300µs, giving us T icks guard = ⌊300µs/T tick ⌋. Thisyields a final formula for the length of a TDMA slot:T slot = (T icks T X + 2 × T icks guard ) × T tick . Finally,the GMAC’s duty cycle is Slots active /Slots frame , whereSlots active (the number of active slots in a frame) is aGMAC parameter.A. GossipingGMAC addresses our issues of scalability and mobilityby using a simple gossip-based communication paradigm[1]. Gossiping has had much success in peer-to-peer protocolsfor wired networks, partly due to its scalabilityand robustness in the face of network churn (changes innetwork membership or topology). For wireless networks,our basic model is that nodes periodically broadcast messagesconsisting of various news items. Other nodes withintransmission range receive these messages and maintain acache of recently heard news items. During each gossipperiod, or round 3 , a node selects and sends a number of2 In the case of a 32,768Hz clock, T tick = 1s/32, 768 ≃ 30.5µs3 The term round and frame describe the same thing, a single periodof GMAC communication. Round is used to refer to gossip, while frameis used at the MAC layer

news items from its cache, possibly including a freshlygenerated news item of its own. Similarly, a node thatreceives messages selects a number of news items that itheard during the round to store in its local cache. Thissimple protocol provides all nodes with an approximationof the true global view of all the news items in thenetwork based purely on local data. The accuracy of thisapproximation depends on several factors, including theconnectivity of the network, the size of a node’s cache,and the algorithm a node uses to select which news itemsto send or store.Network-level gossiping does not rely on routing, nodeIDs, or neighborhood maintenance. Still, it is highly faulttolerant, an ideal property in networks where both wirelesslinks and individual nodes are unreliable. Such a simplegossiping protocol is sufficient for many ad-hoc networkingoperations, such as dissemination, aggregation, andeven maintenance of routing tables, as in [2], in case theyare needed.B. Duty-Cycle SynchronizationGMAC’s use of duty cycling makes proper node synchronizationessential. If a node’s frame is not synchronizedwith those of its neighbors, it is very likely that thenode’s broadcasts will go unheard. For example, with atypical duty-cycle of 1.5%, the odds of having the activeperiods of unsynchronized nodes overlap is quite low,meaning these nodes will not hear each other’s messages.On the other hand, the better synchronized the nodes are,the more energy we can save. This is because GMACis designed to be pessimistic, and currently uses a guardtime that is quite high, approximately equal to the radio’sTX time. A guard time that is too large will result inwasted energy as receivers wait with their radios on fortransmitters to begin. On the other hand, a guard time thatis too short will result in increased message collisions,as transmissions from adjacent slots begin to overlap.Shrinking this guard time as much as possible is one ofthe direct benefits of sharp synchronization.Groups of nodes whose active periods overlap are saidto form a synchronized subnetwork, or simply subnetwork.Synchronization of all nodes participating in thenetwork then takes two forms: maintaining synchronizedsubnetworks and joining separate subnetworks. Because notwo clocks are exactly identical, GMAC’s synchronizationmodule must maintain existing synchronized subnetworksby compensating for the inherent clock drift between anytwo nodes. GMAC operates in a completely decentralizedmanner, so there is a chance that nodes will form independentlysynchronized subnetworks or remain isolated.GMAC’s join messages are responsible for joining theseisolated nodes and subnetworks together.The goal of the maintenance portion of GMAC’s synchronizationis to minimize the offset (see Fig. 1), ∆T i,j =T i − T j between all pairs i, j of communicating nodes.GMAC divides the offset between nodes’ notions of timeinto two parts: slot offset and phase offset. We define slotoffset as the number of whole TDMA slots by which thetwo nodes differ, i.e. ⌊∆T i,j /T slot ⌋. We define phase offsetas the number of clock ticks (intervals of T tick ) the twonodes are offset within one TDMA slot. So, the total offsetbetween the nodes is the sum of the slot offset and phaseoffset between them.Currently GMAC provides one synchronization moduleimplementing a method known as the median algorithm,described in Alg. 1. The median algorithm uses the numberof packets received during the current frame and the arrayof packets as input. The algorithm sorts the received packetentries by their offset from the local time, selects themedian entry, and increases (or decreases) the number ofslots for the next frame based on the timing data pertainingto that entry.Algorithm 1 Median SynchronizationRequire: NumEntries > 0, array RxEntriesSortByOffset(RxEntries)MedianEntry ← RxEntries[ NumEntries2]T ickCorrection ← MedianEntry.Offset/2SlotCorrection ← T ickCorrection div T icks slotP haseCorrection ← T ickCorrection mod T icks slotNextF rameSlots ← Slots frame − SlotCorrectionNextSlotT icks ← T icks slot − P haseCorrectionAs GMAC nodes execute their synchronization algorithmin a completely decentralized manner, it is possiblethat separately synchronized subnetworks form (i.e.,subnetworks partitioned in time, rather than space). In anattempt to join these separate subnetworks together to forma single network, GMAC nodes send out join messages inevery TDMA frame. Each node randomly selects a slotduring the inactive portion of the frame and transmits amessage filled with synchronization and debugging information.Occasionally, a join message from Subnetwork Awill be received during the active period of Subnetwork B .Nodes receiving a join message will modify their localclocks in an attempt to synchronize to the sending node.Note that this join behavior is purely probabilistic, and itis not uncommon for disjoint subnetworks to exist, despitethe nodes being in close physical proximity.C. Slot Scheduling StrategyGMAC’s slot allocation algorithms are based on SlottedAloha [3], and are used to share access to the wirelessmedium. At a high level, the main difference between

GMAC and Slotted Aloha is GMAC’s use of duty cycling.Slotted Aloha is an always-on protocol, whereas GMACturns the radio on only for a small percentage of the totalslots in order to save energy. Slots in which the radiois powered up are known as active slots, in contrast toinactive, or idle slots, where the radio is powered down.At the moment, GMAC provides three different TDMAstrategy algorithms. For the purposes of this paper, werestrict our investigation to the most basic strategy, knownas Simple TDMA.The Simple TDMA strategy takes as a parameter thenumber of active slots in a frame, denoted Slots active .In each frame, the algorithm selects a random TX slotin the range [1..Slots active ]. This algorithm works wellin low- to moderate-traffic networks. However, highlycongested neighborhoods (i.e., where #Neighbors >>#ActiveSlots) will result in a large number of collisions,and thus significantly reduced throughput. The DistributedTDMA strategy attempts to remedy this problem by adaptingthe number of active slots based on the perceivednumber of neighbors. We leave the investigation of thisand other strategies to our future work.III. Experimental setupWe implemented our GMAC simulator using theMiXiM (http://mixim.sourceforge.net) extensions overOMNET++ (http://www.omnetpp.org). The OMNET++platform is expressive, efficient, modular, and increasinglythe de-facto simulation environment for mobile ad-hoc andsensor networks [4].A. Simulated NodesAs we are interested only in synchronization, we canignore the difficulties of modeling sensors and actuators.The behavior of the GMAC layer is then driven solelyby the interrupts from the internal clock and the radio.We chose to model our simulated clocks and radios inaccordance with the hardware GMAC was designed torun on in the real world, MyriaNed nodes. These nodeshave an Atmel ATXMega128 CPU with 8k of RAM, 128kof Flash memory, a Nordic nRF24L01+ wireless radio,and four colored LEDs, and several sensors. The internaltimer is an oscillator designed to operate at 32,768Hz,and are specified to have a frequency offset ≤ ±20ppm(parts per million). This may seem small, but it meansthat after thirty seconds 4 , two initially synchronized clocksmay have drifted as much as 1.2ms (40 ticks) apart. Thatis, one of the nodes may have counted as many as onemillion and twenty ticks in this time interval, while another4 one million time intervals of T tick = 1F req ≃ 30µsTABLE I. Networks InvestigatedNodes Dimensions Spacing16 320m × 320m64 640m × 640m256 1280m × 1280m1024 2560m × 2560m80m Matrix, Randommay have counted only nine-hundred ninety-nine thousand,nine-hundred and eighty ticks in that same interval. WithSlots active = 8 active slots of T icks slot = 28 ticks each,two nodes can become completely isolated from each other(i.e. their active periods do not overlap) in as little as3 minutes. Compensating for this inherent divergence ofseparate clocks is the goal of any clock synchronizationalgorithm.We designed our own OMNET++ modules to representthe type of internal clocks described above. OMNET keepstrack of the global simulation time, T sim , while an individualnode computes its own local time. A node bases this onits own clock’s frequency offset (F offset ) and phase offset(P offset ) from the global simulation clock (provided asOMNET parameters): T i = (T sim ×F offseti )+P offseti . Anode’s phase offset determines the length of time betweenthe global start of the simulation and the start of thatparticular node. The frequency offset determines howmuch faster or slower than simulation time the node’s clockruns.For the purposes of this paper, we have restrictedour investigation to GMAC’s most basic synchronizationprimitives: median synchronization maintenance and theSimple TDMA strategy, which were discussed in Section II.B. Input parameters to investigateAs our primary interests lie in GMAC’s synchronizationand stability in large mobile networks, our investigationsfocus on the following parameters: clock drift, transmitpower and starting topology. Due to space constraints, ourinvestigation into GMAC’s handling of mobility will bedeferred to a future publication.1) Clock Drift: Even clocks of the same specificationsexhibit slight differences, as previously discussed.We investigate a range of maximum clock drift settingsin our experiments. By selecting a maximum offset ofO max , the simulator will assign each simulated node’sclock a random clock frequency multiplier in the range[1 − O max , 1 + O max ]. This results in each node havinga similar but slightly varied value for T tick , causing oursimulated clocks to slowly drift apart just as real clocksdo.2) Starting Topology: In this investigation, we examinetwo different network topologies: matrix topologies and

andom topologies. In matrix topologies, N nodes aredeployed in a √ N × √ N grid topology, where rows (andcolumns) and placed 80m apart. This will show us thebehavior of the GMAC in a more ‘friendly’ setting, wherethe network is fully connected and there exist many pathsbetween all nodes. In the random topologies, N nodesare deployed at random locations in the same area as theequivalent matrix topology (i.e., with the same numberof nodes). Random topologies present more difficulty insynchronization as there will generally be some veryhighly connected areas and some less connected areas, andpotentially even physically isolated nodes. These randomtopologies, however, are more representative of the typeof topologies that will be encountered in the real world,and hence of more interest to our investigation. Sincescalability is one of our primary interests, we investigatenetworks of various combinations of size and deploymentpattern. A full list is given in Table I.C. Observable metricsWhen a simulated node begins a new round, it recordsthe round number and global simulation time (T sim ). Bylater comparing the times that the nodes began each round,we can analyze the synchronization behavior of a simulatedGMAC network.IV. Simulation resultsIn this section, we present the results of our experimentssimulating the GMAC layer in the OMNET++ environment.We elected to analyze the GMAC’s synchronization performancein two ways. First, we ran a series of experimentswhere all nodes were started in a synchronized state at thevery beginning of the run (P offset = 0). Second, we ranexperiments where each node i was started at a randomlychosen time in the first 15 seconds (0.0 ≤ P offset < 15.0).The first set of experiments is designed to show howGMAC’s synchronization maintenance, i.e., the medianalgorithm, performs in an already synchronized network.The second set will give us insight into the performanceof GMAC’s subnetwork joining mechanism, i.e., the joinmessages.A. Synchronous StartFor this series of experiments, our primary metric formeasuring the level of synchronization in the networkis the standard deviation of the node start times. Asmentioned earlier, each node reports the time it beginseach round. For each round, we compute the standarddeviation of all node’s reported time for the start ofthat round. This tells us how tightly synchronized thenetwork is at every round. This is a somewhat broadmeasurement, as it pertains to the whole network. Thatis, it cannot give us insight into which sections of thenetwork are suffering from poor synchronization. Note thatthe best synchronization level we can realistically hope foris T tick2≃ 15µs, or half the smallest adjustment that theGMAC can make in its effort to compensate for the driftingclocks. Consistent and stable synchronization to the levelof a single clock-tick is quite good, though, and indicatesthat we could reduce T guard significantly below its currentvalue of 300µs in order to save energy.We begin by looking at Fig. 2a, depicting a set ofindividual simulated runs. Each line is the result of a singlerun on a 256-node matrix topology, with the standard deviationplotted per round. These simulations used the defaulttransmission power of 20mW and MaxClockDrift=25ppm.Here we can see much similarity between individual runs.The median algorithm maintains a steady-state where itcompensates for the clock drift between the nodes.In Fig. 2b, each line shows the average standard deviationfor each simulated round across the 10 experimentalruns for each MaxClockDrift setting. We explore the effectsof different clock drifts on the median algorithm, using thesame topology and transmit power settings as above. Thus,the 10 lines from Fig. 2a comprise the single line in Fig. 2bat 25ppm. Here we can see how the median algorithmcopes with various clock settings. More clock drift betweennodes leads to progressively looser synchronization. Thatis, as clocks drift apart faster and faster, the median cannotmaintain the same level of synchronization between nodes.This can be seen by the increased standard deviation athigher drift settings.Finally, in Fig. 2c we see the aggregation of a largeamount of experimental data. Each data point representsthe average standard deviation for the last half of the10 runs of a particular parameter setting. We plot theclock drift settings along the x-axis and average convergedstandard deviation for the 10 runs of each MaxClockDriftvalue. Thus, the single points along the line {Matrix, 256}show the average value over the last 30 minutes of the10 runs represented by the individual lines from Fig. 2b.The solid lines show matrix topologies, while the dashedlines show random topologies and, as expected, the randomtopologies show much more variability.B. Asynchronous StartFor our second series of experiments, our metric foranalyzing the joining of the nodes into a single synchronizednetwork is the fraction of nodes not belonging tothe largest synchronized subnetwork. When analyzing ourlogs, we group nodes into subnetworks as follows: we

1000100100 1000 2000 3000 4000 5000 6000 7000Round NumberStandard Deviation of Round Start Times (us)100010010MaxClockDrift1 255 5010 10010 1000 2000 3000 4000 5000 6000 7000Round Number1 5 10 25 50 100MaxClockDrift(a) 10 individual runs of same settings. Topology: (b) Average over 10 runs for each MaxClockDrift. (c) Converged offset standard deviations for combinations256-node Matrix, MaxClockDrift: 25ppm. Topology: 256-node Matrix.of MaxClockDrift and Topology. From topto bottom: {1024, 256, 64, 16} nodes, {Random,Matrix} topologies.Fig. 2. Variation in round start times, synchronous startStandard Deviation of Round Start Times (us)10 610 510 410 310 210 1100%10%1%0.1%% of Nodes OUTSIDE Largest Subnetwork100%10%1%0.1%MaxClockDrift15102550100% of Nodes OUTSIDE Largest Subnet100%10%1%0.1%0%0 1000 2000 3000 4000 5000 6000 7000 8000Round Number0%0 1000 2000 3000 4000 5000 6000 7000 8000Round Number(a) 10 individual runs of same settings. Topology: (b) Average over 10 runs for each MaxClockDrift.256-node Matrix, MaxClockDrift: 25ppm. Topology: 256-node Matrix.0%1 5 10 25 50 100MaxClockDrift(c) Per-Parameter Per-Topology AverageFig. 3. Percentage of unsynchronized nodes, asynchronous startsort all the reported times for the start of a particularround. We iterate through the times, computing the offsetbetween the previous and current time. When there is anoffset between consecutive reported times greater than thesubnetwork threshold, we decide that the current reportedtime marks the beginning of a new subnetwork. By separatingthe nodes into disjoint sets representing synchronizedsubnetworks at various thresholds, we can analyze howwell GMAC’s join mechanism is working. We use thresholdsof 7000µs, 850µs, and 30µs which correspond toapproximately the length of an active period, a TDMA slot,and a simulated clock tick, respectively. At each round, foreach threshold, we calculate the percentage of all simulatednodes that are outside (i.e., not synchronized with) thelargest subnetwork. As such, we should expect this metricto begin at or near 100% and fall towards 0%. In Fig. 3,we show the results of our asynchronous start experiments.In general, most runs behave as expected and rapidlyconverge to a single network. However, in some runs, wefind that separate subnetworks can exist for a long time, tothe point where they never form a single network duringthe entire simulated hour. We can see just such behaviorin Fig. 3a, which depicts the results of a set of twentyfive256-node matrix runs with a maximum clock drift of±25ppm. In Fig. 3b, we show the average of 25 such runsfor each MaxClockDrift setting. As previously explained,the behavior of the join mechanism is probablistic, andthis can be clearly seen in the variability of the simulatedresults.Finally, in Fig. 3c we plot (analogously to Fig. 2c) theaverage fraction of nodes outside the largest subnetworkover 25 runs for each setting, considering only the lasthalf of the rounds of each run. The settings correspond toall combinations of clock drifts and topologies we experimentedwith. Clearly, in all of our simulations GMACeventually achieved synchronization of at least 99% ofthe participating nodes, irrespectively of the topology andmaximum clock drift.V. Related workIn [5], the authors discuss why traditional synchronizationmechanisms, like NTP, are unsuitable for ad-hocsensor networks. In [6] the authors give an overview of the

problem of time synchronization in sensor networks andevaluate several important protocols. [7] provides a morerecent survey of time synchronization protocols designedfor sensor networks, however the most scalable protocolanalyzed was run on networks up to 300 nodes, far belowwhat is required here.Many popular energy-aware MAC protocols, e.g. S-MAC [8], use carrier-sensing or ready-to-send/clear-tosendhandshakes, which necessarily use more energy thana properly synchronized TDMA protocol that avoids collisionswithout these mechanisms. Both Timing-sync Protocolfor Sensor Networks (TPSN [9]) and Flooding TimeSynchronization Protocol (FTSP [10]) rely on creating aspanning tree over the whole network, stemming from aglobally elected root node. The cost of leader electionand tree building make such solutions unsuitable for highdiameter and/or mobile networks. Reference BroadcastSynchronization (RBS [11]) provides only post-facto synchronizationand is thus infeasible in our setting. Post-factosynchronization is used to retroactively determine the timea past event occurred at a neighbor node (i.e., to come toconsensus on what time an event sensed by several nodetook place), but is not designed to synchronize the currentclock state of the nodes.VI. Conclusions and Future WorkOur results indicate that the simple median synchronizationmechanism provided by GMAC is capable ofadaptively maintaining good clock synchronization, even atvery large network sizes. Simulations show that the medianalgorithm is capable of compensating for clock frequencyoffsets far greater than can be expected from real clockcomponents. Furthermore, simulations show the medianalgorithm can maintain synchronization in networks aslarge as 4096 nodes, and possibly beyond. This level ofscalability is an absolute necessity in the setting of largesocial ad-hoc networks.Though this synchronization maintenance algorithm isquite simple, in practice it turns out to be quite effective.One problem with the median algorithm is that it has no‘memory’, or state. This means that the median algorithmcan maintain synchronization only while it is in regularcontact with other nodes. If a node becomes isolated(without neighbors) for an extended period of time, thenode’s clock will tend to drift away from the other nodesit was previously communicating with. It also seems thatfor smaller networks, GMAC keeps nodes tightly coupled.This indicates that there may be much energy to be savedby reducing the TDMA guard time.Though GMAC’s synchronization maintenance mechanismwas shown to be scalable, its join mechanism didnot prove as effective. As a node can only receive joinmessages during its (short) active period, separately synchronizedsubnetworks can persist for a long time. This isundesirable as subnetworks will be unable to communicatewith each other, leading to significantly reduced utility ofthe network as a whole. We plan to investigate using amore active join mechanism, which periodically listens forother subnetworks and actively tries to deterministicallyjoin them. The frequency of this extended listening periodcould be adaptively based on the size of the network,neighborhood density, or other metrics.If we can modify GMAC to efficiently join all reachablenodes into a single synchronized network, we will havea firm foundation on which to build scalable distributedapplications running on massive, dynamic social ad-hocnetworks.References[1] A. Demers, D. Greene, C. Hauser, W. Irish, J. Larson, S. Shenker,H. Sturgis, D. Swinehart, and D. Terry, “Epidemic algorithms forreplicated database maintenance,” in Proceedings of the sixth annualACM Symposium on Principles of distributed computing. ACMPress New York, NY, USA, 1987, pp. 1–12.[2] K. Iwanicki and M. van Steen, “Multi-hop cluster hierarchymaintenance in wireless sensor networks: A case for gossip-basedprotocols,” in Proceedings of the Sixth European Conference onWireless Sensor Networks (EWSN 2009). Cork, Ireland: Springer-Verlag LNCS 5432, February 2009, pp. 102–117. [Online].Available: http://www.few.vu.nl/ iwanicki/publications/2009-02-EWSN/[3] N. Abramson, “The throughput of packet broadcasting channels,”IEEE Transactions on Communications, vol. 25, no. 1, pp. 117–128,1977.[4] E. Weingartner, H. vom Lehn, and K. Wehrle, “A performancecomparison of recent network simulators,” in IEEE InternationalConference on Communications, 2009. ICC’09, 2009, pp. 1–5.[5] J. Elson and K. R”omer, “Wireless sensor networks: A new regime for time synchronization,”ACM SIGCOMM Computer Communication Review,vol. 33, no. 1, p. 154, 2003.[6] F. Sivrikaya and B. Yener, “Time synchronization in sensor networks:A survey,” IEEE network, vol. 18, no. 4, pp. 45–50, 2004.[7] S. Rahamatkar, A. Agarwal, and N. Kumar, “Analysis and ComparativeStudy of Clock Synchronization Schemes in Wireless SensorNetworks,” Analysis, vol. 2, no. 03, pp. 536–541, 2010.[8] W. Ye, J. Heidemann, and D. Estrin, “An energy-efficient MACprotocol for wireless sensor networks,” in IEEE INFOCOM 2002.Twenty-First Annual Joint Conference of the IEEE Computer andCommunications Societies. Proceedings, 2002, pp. 1567–1576.[9] S. Ganeriwal, R. Kumar, and M. Srivastava, “Timing-sync protocolfor sensor networks,” in Proceedings of the 1st international conferenceon Embedded networked sensor systems. ACM New York,NY, USA, 2003, pp. 138–149.[10] M. Maróti, B. Kusy, G. Simon, and Á. Lédeczi, “The flooding timesynchronization protocol,” in Proceedings of the 2nd internationalconference on Embedded networked sensor systems. ACM, 2004,pp. 39–49.[11] J. Elson and D. Estrin, “Time synchronization for wireless sensornetworks,” Parallel and Distributed Processing Symposium, International,vol. 3, p. 30186b, 2001.