Data Dissemination in Opportunistic Networks using ... - IEEE Xplore

More documents

Recommendations

Info

Algorithm 1 Recognition algorithm1: Let i be an observed channel/item;2: Let H be a hashed index of removed channels/items3: Let R θ be the recognition threshold4: if Cache.contains( i ) then5: Increment i counter6: reset i.TTL7: else8: if Cache is full then9: Select the item o with the oldest TTL10: if o.counter ≥ R θ then11: Move o to H12: end if13: Drop o14: end if15: Put i in the Cache16: i.counter = 117: Set i.TTL18: end ifStartChannelRecognition# Recognized and old itemslower than SFigure 2.AvailabilityRecognitionStore theItemsMerge With theActual MemoryContentTake the first S objectsModified Take The Best AlgorithmAvailabilityRankwith respect to the cues. We propose to adapt this algorithmin the scenario we are considering. When a node meetsanother peer, it ranks the objects of the other node usingan adaptation of the Take The Best algorithm, depicted inFig. 2 and Algorithm 2. The first two cues we consideredconsist of the recognition of channels and items, using thealgorithm presented in Section IV. The first cue is thechannel recognition: items of recognized channels are rankedhigher than the others and selected for the next steps. Ifthe total size of remaining items (considering both the nodeand the peer shared storage spaces) is greater than S (thesize of the node’s shared storage space), items are furtherdiscriminated using the second cue, i.e. the recognitionof items. In this case the recognition assumes a negativemeaning, as recognized items (already very spread) areranked lower than the others and they are not consideredanymore. If further discrimination have to be carried outto fill the node’s shared storage space, the precise valueof estimated availability of items is considered, and lessavailable items are ranked higher. As for the original TakeThe Best Algorithm, not all the steps are required, and thelast (and more costly) one is run only on a subset of theitems.A. The Less–is–More EffectGoldstein and Gigerenzer show that the recognitionheuristic [5] and the Take The Best algorithm are subjectto the so-called less–is–more effect [6]: When increasingthe number of recognized items, the number of accurateinferences will increase up to a certain point, but thereafterdecrease. This is due to the diminished discrimination powerof the recognition heuristic, since too many items are recognizedat the same time. In order to take advantage of theAlgorithm 2 Modified Take The Best Algorithm1: Let R be a set of items received from another node;2: M be the actual storage content3: Let S be the storage capacity limit4: Let C be the channel chache and O be the item cache5: Let H C and H I be the hashes of old recognized channels and items6: Let RC θ be the recognition threshold for channels7: Let RI θ be the recognition threshold for items8: Let I = R − M9: Let recChannels = ∅10: for each i ∈ I do11: if C.contains(i.channel) & i.channel.counter ≥ RC θ ORH C .contains(r) then12: recChannles ∪ = i13: end if14: end for15: Let recItems = ∅16: if recChannels.size + M.size > S then17: for each r ∈ recChannels do18: if (O.contains(r) & NOT r.counter ≥ RI θ ) OR NOTH I .contains(r) then19: recItems ∪ = r20: end if21: end for22: if recItems.size + M.size > S then23: Let M ′ = M ∪ recItems24: Rank M ′ in ascending order according to the counters of itsitems25: Select and keep in M the first S objects of M ′26: else27: M ∪ = recItems28: end if29: else30: M ∪ = recChannels31: end ifless–is–more effect, the storage capacity of the recognitioncaches should be adapted in order to achieve the maximumaccuracy. Identifying autonomic algorithms to adjust thiscapacity is one of the key directions of future work.VI. EXPERIMENTAL RESULTSWe evaluate the proposed algorithm by simulating thefollowing scenario. We consider 45 nodes, divided into threedifferent groups. In order to simulate real user movementpatterns, nodes move in a 4 x 4 grid (1000 m wide),according to the HCMM model [3]. The HCMM modelis a mobility model that integrates temporal, social andspatial notions in order to obtain an accurate representationof real user movements. In the simulation scenario, groupsrepresent set of users that have social and spatial relationships.Groups are initially assigned to a home cell and anyphysical contact among groups is avoided. Thus, the onlyway to exchange and obtain data among groups is throughnode mobility. Nodes can move in the cell of their grouponly. A few nodes in each group (named travellers) bridgebetween communities by visiting more than one group.This model well represents social communities, in whichpeople typically stay, with a few people commuting betweendifferent communities due to different social relationships.In the simulation settings, each group has two travellers,one for each of the other groups. The objects available inthe network are assigned to channels and there are as manychannels (n c ) as groups. Each channel has a total of 99objects and each group originates 1/n c items per channel,i.e. there are 33 items of each channel per group. All theobjects are generated at the start of the simulation. Each
node subscribes to one channel only. Within each group,node interests follow a Zipf law with parameter 1. Interestsare rotated, so that the most popular channel in a groupis the second in another and the third one in the other,and so on. Note that in this scenario data items can reachinterested users in communities other than those where theyare generated only through nodes mobility. Therefore, itallows us to highlight the effectiveness of the data disseminationalgorithm. The simulation runs for 50,000 seconds.At the start of the simulation each node subscribes to a givenchannel. The performance figure is the hit rate, computed atvarious time instants after the simulation starts. It is definedas the ratio between the number of retrieved objects of thesubscribed channel and the total amount of objects of thechannel. By default, on each node, the channel recognitioncache size is 3 (i.e. all channels can be recognized), the itemsrecognition cache is 10, and the shared storage space has 10slots (items are assumed to be of equal size). Simulationhave been replicated in independent conditions 10 times,and confidence intervals (with 95% confidence level) havebeen computed. As they are very narrow, in the plot we onlyshow the central values of the confidence intervals.Simulations were run changing the values of the thresholdsfor both the channel and the items recognitions and byvarying size of the shared storage space, S. The first set ofexperiments (Figs. 3–4) shows, first of all, that by using therecognition heuristic the system is able, after some time, toreach 100% hit rate. This is a very important result. Usingthe proposed recognition heuristic result in an extremelylightweight data dissemination system, which neverthelessproves to be very effective. Another aspect highlighted bythis set of results is the effect of varying the item recognitionthreshold with different values of the channel recognitionthreshold. The main result is that for very low value ofthe item threshold (RI θ = 2) the information diffusionalgorithm is not able to obtain a 100% hit rate. For theother values, changes in the item threshold implies little oralmost no differences in the algorithm performance. Thisresults (particularly evident in Fig. 4) are a sign of theless−is−more effect. In order to have a in-depth analysesof this case, Fig. 5 shows the results obtained by fixingRI θ = 2 and varying the channel threshold. Such a lowvalue of the item threshold implies a higher probability ofrecognizing items as widespread. As a consequence, at agiven time all the items are recognized as too diffused byall the nodes. In this case, items are not exchanged anymore,since all of them are not recognized as useful. In the othercases in Fig. 3 and 4, not all the items were recognized as toodiffused, thus allowing to reach a 100% hit rate. RI θ = 2,as in Fig. 5, always leads to lower performances. Moreover,this figure highlights that different values of the channelthreshold imply different information diffusion speeds. Thelower RC θ , the more rapid the convergence of the algorithm.The longer it takes to converge, the higher is the probabilityfor a node to see more copies of a given data before startingto take it. Thus, with a slower convergence rate, is moreprobable to have all the items recognized as widespreadat increasingly further points from 100% of hit rate. TheHit Rate10.80.60.40.20Item Thr. = 2Item Thr. = 5Item Thr. = 101 10 100 1000 10000 100000TimeFigure 3. Hit Ratio with a Channel threshold = 3Hit Rate10.80.60.40.20Item Thr. = 2Item Thr. = 5Item Thr. = 101 10 100 1000 10000 100000TimeFigure 4. Hit Ratio with a Channel threshold = 10effect of RC θ values on the convergence speed is visiblealso with other values of RI θ . Fig. 6 reports the resultsachieved with RI θ = 10. In this case, the hit rate alwaysconverges to 100%, but , as in the previous figure , theconvergence velocity is strongly determined by the differentvalues of RC θ . Fig. 7 highlights the effects of the sharedstorage space to the algorithm effectiveness. The results areobtained with RC θ = 2, RI θ = 5. The number of slots ofthe shared memory is fixed to 2 and 50, respectively. A largeramount of shared memory allows a greater replication factor,since is more probable that multiple copies of the same datacan be stored by different nodes at the same time. This facteasy the diffusion of items, thus increasing the convergespeed. Nonetheless, even a very low number of slots doesnot affect the final performance (100% hit rate). On the otherhand, they result in a slower convergence rate, due to theincreased difficulty to share and spread items in the network.VII. CONCLUSIONSCognitive heuristics are models of how the human brainassess the relevance of information using only partial knowl-Hit Rate10.80.60.40.20Ch. Thr. = 3Ch. Thr. = 5Ch, Thr. = 10Ch. Thr. = 151 10 100 1000 10000TimeFigure 5. Hit Ratio with an Item threshold = 2
Page 3: decision-making processes [11], for

Data Dissemination in Opportunistic Networks using ... - IEEE Xplore

Create successful ePaper yourself

Delete template?

Save as template?