Survey of buffer management policies for delay tolerant networks

Transcription

1 Survey of buffer management polces for delay tolerant networks Sweta Jan 1, Meenu Chawla 1 Department of CSE and IT, Maulana Azad Natonal Insttute of Technology, Bhopal, Inda E-mal: chawlam@mant.ac.n Publshed n The Journal of Engneerng; Receved on 5th March 2014; Accepted on 5th March 2014 Abstract: Delay tolerant networks DTN) are a class of networks that are a subset of the tradtonal moble ad-hoc networks. It dffers from moble ad hoc networks MANETs) n the sense that t can wthstand hgh delays n delverng data because of frequent network parttons, lmted bandwdth and storage constrants persstng n such networks. Owng to these nherent characterstcs of the delay tolerant networks mprovng delvery rato n such networks depends on two man factors-use of routng strategy and a good buffer management polcy. Many routng protocols have been proposed n the lterature for DTN. Buffer management s a very mportant factor n DTN because of the very lmted buffer space avalable n DTN nodes. Although a schedulng polcy n DTN determnes whch message has to be forwarded frst, the droppng polcy decdes whch messages are to be dropped n case of buffer overflow. Ths Letter presents a survey of the exstng buffer management polces proposed for DTN and dscusses the pros and cons of these approaches. The buffer management technques have been classfed on the bass of nformaton used by them whether they are based on local nformaton of messages avalable at the node or global nformaton of all the messages n the network. 1 Introducton Mllons of devces are connected to each other wth the help of Internet operatng on the transfer control protocol/nternet protocol TCP/IP) to provde a relable communcaton through a standard set of protocols. The usablty of Internet however reles on some mportant assumptons lke contnuous bdrectonal end-to-end path, short round trps, symmetrc data rates, low error rates etc. However wth the prolferaton of moble devces n each and every aspect of the current world, the Internet has lmtatons to provde a satsfactory performance n connectng the mllons of moble devces poppng up. In networks lke terrestral networks, mltary networks etc. where frequent network parttons are lkely to occur because of the frequent lnk dsruptons caused by obstacles or power shut downs usng a technology lke TCP/IP or tradtonal MANET routng protocols, whch requre an end-to-end connecton, cannot ensure a successful delvery. The envronments that are charactersed by ntermttent connectvty, long or varable delays, asymmetrc data rates and hgh error rates are known as delay tolerant networks [1 3]. Snce the chances for a destnaton to come nto contact wth a source node are very less n a DTN, these networks employ a store and forward paradgm to forward the messages to the destnaton. Ths means that whenever a node has a message for another node and ths node s not n ts vcnty, then the node wll forward the message to the next encountered node. The encountered node stores ths message n ts buffer n order to forward t to the destnaton. The store and forward approaches have two major problems that have evolved two separate research areas n the study of DTNs. The frst problem s that forwardng messages to any encountered node s not feasble. The man reason behnd ths s that forwardng to a large number of nodes puts a burden on lmted node resources lke buffer, energy and also wastes lmted network bandwdth. Ths also leads to network contenton because of the floodng of messages. Thus, nodes to whch message has to be forwarded should be chosen carefully. Ths problem s addressed n the desgn of routng protocol. Another problem arses because of lmted buffer space avalable at DTN nodes. Snce DTN nodes work on store carry and forward mechansm ther buffer space gets flled up quckly. A stuaton can arse where each node buffer s full wth messages from other nodes n the network. In such a stuaton, when a new encounter happens, ether the node has to drop a message from ts buffer or t may have to deny the sendng node of a new message transfer. Ths problem s taken care by the buffer management polces n a DTN. Buffer management s very mportant n DTN because the combnaton of long-term storage of messages and the message replcaton places a hgh bandwdth and storage overhead on nodes [4]. When the buffer s full, often n order to accommodate a new message a DTN node wll have to drop an mportant message. If an effcent droppng polcy s mplemented that can prortse message drop sequence, t wll have a huge mpact on the delvery rato of the network. Not only droppng polces defne buffer management n DTN, but also the schedulng polces. Snce DTN suffers from partal transmssons, sortng the messages n the order of ther relevance s very mportant for ther successful delvery. Partal transmssons occur when two nodes are transmttng the messages and ther transmsson has to be aborted because of some lnk falure or power shut down by ether of the nodes. Thus, both schedulng and droppng polces play a huge role n the delvery performance of a DTN network. The goal of an effcent buffer management polcy can be summarsed as below. The prmary and necessary goal must be to mprove the delvery rato of the network. Even though the DTNs are made to tolerate delays mnmsng the average delay of delverng messages s another factor to be taken nto consderaton whle developng an effcent buffer management polcy. If delay s too hgh, the messages can reach the destnaton when almost the relevance of the message s lost. To preserve the resource consumpton, a buffer management polcy must also try to reduce the overhead rato, that s, mnmses the number of relays to whch a message s forwarded. In ths paper, varous DTN buffer management polces that have been ntroduced n the lterature are surveyed and revewed. Buffer management polces have been manly classfed on the bass whether they use locally avalable nformaton or globally avalable nformaton about all messages. Each algorthm s descrbed n bref along wth ther advantages and dsadvantages wth respect to a DTN envronment. The rest of ths paper s organsed as follows. Secton 2 ntroduces the varous buffer management schemes and Secton 3 summarses the varous buffer management polces. Secton 4 provdes the concluson of ths paper. 1

2 2 Buffer management polces n DTN In ths secton, we descrbe the varous buffer management polces for the delay tolerant networks that have appeared n the lterature. The methods are descrbed n the order n whch they were ntroduced n the past years along wth ther shortcomngs and advantages. 2.1 Drop-random, drop-least-recently receved, drop oldest and drop-least-encountered Davs et al. [5] ntroduced four droppng polces n The goal of the droppng polces was to ncrease the delvery performance n hghly parttoned ad-hoc networks. Ths paper stressed the mportance of these polces n terms of the ncreasng use of wearable computer for packet transport mechansms. Four droppng polces were ntroduced n ths paper: drop random DRA), dropleast-recently-receved DLR), drop oldest DOA) and drop-leastencountered DLE). Among the four strateges used, DRA uses the smplest strategy, that s, when the buffer s full drop randomly selected packets untl space for newly arrved packet s created. Ths strategy does not take nto account the network condtons and s smply blnd. The DLR drops the packets that have stayed n the agent s buffer for a long tme. The dea behnd such an approach s that DLR assumes the packets that have stayed for a long tme n the buffer must have been forwarded the most number of tmes n the past encounters and therefore can be dropped. It assumes that such packets have enough replcas n other nodes so that ther probablty of delvery s hgher than other packets n the buffer. The DOA polcy drops the messages that have stayed for the longest tme n the network. Such a polcy requres network nformaton and s therefore more adaptve. The messages that have stayed for a long tme n the network have hgher probablty to be already delvered and therefore are dropped. The major contrbuton of ths paper however s DLE. In ths polcy, messages are dropped based on the estmated lkelhood of delvery. It takes nto consderaton agent locaton and movement for ts mplementaton and hence s more adaptve than the other three approaches. The DLE sorts the packets accordng to the relatve ablty of two agents to forward a packet to the destnaton. To calculate the relatve ablty, each agent node) mantans a lst contanng the agent addresses of other agents n the network. For each tme step, agent A updates the meetng tme for another agent C, wth respect to co-located agent B, usng the followng rule M t+1 A, C) l M t A, C), f none co located = l M t A, C) + 1, f C = B l M t A, C) + am t B, C), for all C = B where M t A, C) s the meetng value at tme t, α = 0.1 s a parameter that decdes how much porton of B s meetng value s to be added wth A and λ = 0.95 s the decay rate of the meetng value. The value M t A, C) s ntally zero for all node pars. The meetng value s updated n three steps: ) update the meetng value whenever node A meets node B; ) update the transtvty n meetng values, for example, f node A encounters node B and node B has a hgh encounter value for C, then A s a good canddate for passng packets to C also; and ) wth tme the meetng values have to be aged f nodes A and B do not meet for a long tme. To sort the packets n the buffer accordng to the relatve ablty of two nodes to delver the packet to the destnaton, whenever node A meets B the packets are sorted accordng to M t A, C) M t B, C), where C s the destnaton of the message. Such a strategy ensures that the packets can always move from agents that have less probablty to delver the packets to the destnaton to agents that have a hgh delvery probablty wth the destnaton. Ths proves that DLE works best when compared wth other algorthms because of ts adaptveness wth the network condtons. 2.2 Buffer management n MaxProp MaxProp [6] s one of the earlest protocols that combned buffer management along wth a routng strategy. Routng protocols proposed earler to t lke Epdemc [7], Prophet [8], Spray and Wat [9] dd not specfy any specfc buffer management technque. For effectvely managng the buffer, the buffer s logcally dvded nto two parts based on whether the hop count of the packets have a hop count less than a threshold t hops. If the hop count s below t counts, the packets are then sorted n the ncreasng order of the hop count. For those packets whch are havng a hop count greater than t then they are sorted n the decreasng order of delvery lkelhood. Delvery lkelhood s a new metrc used n MaxProp to prortse the packets based on the cost that t takes to reach the destnaton. When a node encounter occurs, packets whch have hop count less than threshold t are forwarded frst. If all such packets are forwarded then only the packets havng hop count greater than threshold t are gven a chance for transmsson. By gvng prorty to the frst buffer secton, the MaxProp s gvng mportance to those packets whose hop count s very less. The assumpton s that the packets wth a very small hop count are relatvely new n the network and have not traversed far n the network. Therefore such packets are gven more mportance. For packets above the threshold t, they are sorted n decreasng order of delvery lkelhood. Those packets wth a hgh value for ths metrc are forwarded frst. Thus, the MaxProp uses an effcent schedulng polcy for forwardng the packets to the relay nodes to ensure better packet delvery. For a droppng polcy, when a packet s to be dropped t s dropped from the tal of the buffer. The tal of the buffer contans the packet wth the lowest delvery lkelhood and s not a new packet snce t s n the second secton of the buffer. Such a packet can be easly dropped snce t does not have enough prorty. MaxProp shows a hgh delvery rato often n comparson wth the other basc routng protocols n DTN thereby emphassng the mportance of buffer management. The communcaton cost for MaxProp s however too hgh because of the calculaton of shortest path to the destnaton for estmatng the delvery lkelhood. 2.3 Evaluaton of queung and schedulng polces n DTN Lndgren and Phanse proposed varous droppng and schedulng polces n [10]. These polces were based on the delvery probablty concept n Prophet routng protocol [8]. The authors proved that consderng delvery predctablty whle takng forwardng and droppng decson s far better than usng a smple random prunng as used n epdemc routng. The droppng polces compared here are: frst-n-frst-out FIFO), most forwarded frst MOFO), most favourably forwarded frst MOPR), shortest lfe frst SHLI) and least probable frst LEPR). Among the droppng polces, FIFO s the smplest polcy. The packets n the queue are sorted for droppng n the order those messages that were receved earler. The message that came to the queue frst wll be dropped, then the second and so on. MOFO forwards the messages that have been forwarded most number of tmes. For ths the node keeps a counter each tme the messages are forwarded. When a packet s to be dropped because of buffer overflow, the one whch was forwarded the most s deleted. Ths s because such packets have already reached many nodes and have hgh probablty to get delvered and can be easly dropped. MOPR uses a metrc FP to count the most favourably forwarded message. The value of FP s ntalsed to zero and each tme t s forwarded the value s updated usng the equaton FP = FP old + P, here P denotes the delvery probablty value wth whch the 2

3 message was forwarded. Although droppng the message wth hghest FP value s dropped frst. SHLI s based on the concept of message tme-to-lve TTL). The message havng lowest remanng TTL s dropped frst. Ths s based on the dea that a message havng a low remanng TTL does not have enough tme to reach the destnaton and wll get expred soon and therefore t can be easly dropped. LEPR drops the message for whch t has the lowest P-value. However, f the source has a small delvery probablty for the destnaton of a message, ths message wll never get forwarded and wll not be spread n the network further reducng ts chance of reachng the destnaton. The authors proved that MOFO s best n terms of delvery rato and SHLI for average delay. Ths s because MOFO ensures that each packet be forwarded at least once before t s dropped. In case of SHLI, the packets that were never forwarded can also be dropped. The forwardng strategy of Prophet was slghtly modfed n each schedulng mechansm proposed n [8]. In GRTR, the forwardng strategy s same as that of the Prophet. The message wth destnaton D s forwarded to an encountered node B by a current node A only f PB, D)>PA, D), where P, j) s the delvery probablty between two nodes and j. The next schedulng polcy GRTRSort selects the messages n the descendng order of PB, D) PA, D) and then forwards f PB, D)>PA, D). In GRTR, the messages were scanned n a lnear way; however, GRTRSort frst selects those messages whch have a hgh dfference between PB, D) and PA, D) and then performs forwardng. In GRTRMax, messages are sorted n the descendng order of PB, D) and then forwarded f PB, D)>PA, D). Ths s based on the same logc as that of GRTRSort, but t prortses based on the hghest delvery predctablty than by maxmsng the mprovements. The next strategy s not used n Prophet, but n Epdemc and s termed as COIN. Instead of smply forwardng that s used n Epdemc, the method frst performs a con toss to check whether the message s to be forwarded or not. For ths a message s forwarded only f a value X >0.5, where X U0, 1). The smulaton results presented n ths Letter showed that GRTRSort n combnaton wth MOFO has the hghest delvery performance. 2.4 Prortsed epdemc routng Ram Ramanathan et al. proposed a buffer management scheme that consdered the topology of the network n [11] known as Prortsed Epdemc routng PREP). The PREP orders the messages n the node for both transmsson and drop by assgnng prortes to them. It conssts of two mechansms: ) a method that estmates the routng cost from a gven node to the destnaton and ) a prorty mechansm for message processng deleton and transmsson). The costs for nter-node contacts are computed usng a metrc known as average avalablty AA). It measures the average fracton of tme n the near future for whch the lnk wll be avalable for use. When there s a threshold change n the value of AA for a node, then lnk state advertsements LSA) wth a lst of all current lnks and ther assocated metrcs s generated. The LSA s then crculated through whole network usng epdemc routng. The advantage of such dssemnaton s that when two parttons are joned, the node n each partton gets to learn the entre connected topology. Ths helps a node to keep a knowledge of the subgraph from whch an LSA was transportable to durng some recent tme perod. To calculate the routng costs, a cost s assgned to each lnk usng the equaton 1 AA) after whch Djkstra s shortest path algorthm s used to fnd the lowest cost route. To mplement the prorty-based droppng and schedulng polces, each packet s assgned wth a drop prorty p d and a transmt prorty p t. For droppng those messages havng hop count greater than a threshold V hc are selected. Then, they are assgned prorty accordng to the cost path to destnaton from the current node. Those messages that are far from the current node to the destnaton are dropped frst. If the buffer occupancy s stll so hgh that new messages cannot be accommodated, then all the messages have a hop count less than a threshold V hc are dropped. For transmsson, the node calculates the cost to destnaton of the encountered node. If t has a smaller cost, then the message s places n downstream bn, otherwse n upstream bn. Messages n downstream bn are sorted based on remanng TTL and age of the message. It can be seen that ths method uses the hstory of contacts for transmsson and droppng of the messages. When n tradtonal floodng, the nodes have no dea of the topology of the peer; ths method can obtan an dea of the network to whch the message s beng forwarded. Thus t s a guded transmsson. The dsadvantage of ths method s the dssemnaton of LSAs. In a network wth hgh node densty, such dssemnaton leads to hgh overhead. It s useful n such networks where load s less relatve to the network resources. 2.5 Optmal jont schedulng and drop polcy for DTNs Krfa et al. proposed an optmal polcy for both schedulng and droppng n DTNs n [4]. The approach proposed n global knowledge based schedulng and drop GBSD) s based on optmsng two parameters: ) maxmsng delvery rato and ) mnmsng average delay. The GBSD calculates per-message utlty for a gven routng metrc. To maxmse the average delvery rate of all the messages, the message havng smallest value of below gven utlty functon s dropped 1 m )) T ) ) lr L 1 exp ln T R where t s assumed that there are K messages that are n the network wth L nodes and the elapsed tme s T for a message when a drop or forwardng decson s to be made. m T ) s the number of nodes that have seen the message snce ts creaton and n T ) are the nodes that have a copy of the message. The meetng tme between nodes s exponentally dstrbuted by the parameter λ. The GBSD polcy to mnmse the average delay drops the message havng smallest value of followng utlty functon 1 ) 2 n T l 1 m T ) ) L 1 Snce the calculaton of utlty functon n GBSD requres complete knowledge about every message n the network, that s, the number of nodes who have seen the copy of message and number of nodes who have a copy of the message, a hstory based schedulng and drop HBSD) polcy s used to employ a dstrbuted algorthm. Ths polcy learns the network statstcally usng the network hstory to obtan an estmate of the global status of the network. Wth HBSD, the per-message utlty for maxmsng the delvery rato s [ lr E 1 MT) ) ] ) exp lr L 1 NT) Here, m T ) and n T ) are supposed to be nstances of random varables MT ) and NT ). When global nformaton s not avalable, the HBSD reles on the method of calculatng the average delvery rate usng all possble values of MT ) and NT ) and then maxmse t. Smlarly, the HBSD polcy of mnmsng the average delvery 3

4 delay s as follows E[L 1 MT))/NT)] 2 ) ll 1) L 1 mt) The HBSD s ndependent of the actual global state as t s just a functon of the locally avalable hstory. In [4], the authors menton that for a large number of messages, ths polcy s expected to lead to the same average delay as n GBSD polcy. Even though ths polcy provdes an optmal framework, t s based on some assumptons lke the nodes movement path s known, the bandwdth s unlmted and all the messages have the same sze. Such a stuaton s very mpractcal n a DTN and therefore can only be consdered as an optmal soluton aganst whch other methods can be compared. 2.6 Adaptve optmal buffer management polces AOBMP) for realstc DTN Owng to the mpractcal assumptons made n [4], L and Qan proposed a new set of optmal polces for buffer management n [12]. Ths assumes that both the bandwdth and capablty for one communcaton contact s lmted and that the messages vary n ther sze. These match wth the real characterstcs of a DTN. The system model assumes that there for each node there are N t t) messages n ts buffer { at tme t. The set of messages } n at tme t, PK t) s denoted as P 1, P 2, P 3,..., P N t). Each message j n s denoted as a tuple Sou j t), Dstj t), T j t), sj t), nj t), Cj t) ) whch represent source, destnaton, remanng tme, sze, the number of copes and transmsson opportunty of message j, respectvely. The opportunty for packet j to be transmtted to destnaton { Dst j t by all the nodes ncludng s denoted as C j t). C j ) ) )} = u 1, l 1, u2, l 2,..., un j ), l n j ), where the kth tuple represents that there s copy of message j n the buffer of node u k and n the front of the buffer l k bytes of other messages are present. The AOBMP protocol conssts of three steps ncludng an ntalsaton, utlty computaton and message dscardng step. The ntalsaton step conssts of exchangng metadata between two nodes and updatng transmsson opportunty of messages and parameters of moblty model. In utlty computaton, the utlty functon for messages n node s calculated and the message dscardng step calculates a metrc f satsfed on whch the message shall be dscarded. When the buffer of node s full at tme t 0, the AOBMP polcy to maxmse the average delvery rate s to dscard the message j max satsfyng the followng condton 1 l j max = arg max 2 R j ) n 1 n=1 l 1 T j ) k ) je[ 1, N t 0 )] n 1)! k! n=1 k=0 e l 2R j l 1T j Here, T j and R j are the remanng tme and addtonal tme needed to transmt message j, respectvely, and λ 1 and λ 2 are the parameters of exponental dstrbuton of nter-meetng tme and contact tme, respectvely. To mnmse the average delay, the node should dscard a message whch maxmses the followng functon j max = arg max je[ 1, N t 0 )] 1 l n=1 1 l ) n 1 2 Rj n 1)! e l 2 R n 1 n l 1 T j ) k ) e l 1 T j k! j k=0 Moreover, the AOBMP adapts to each moblty model characterstc by adjustng the mean of exponental dstrbuton wth the ad of the nformaton exchanged through the control channel. Ths approach works well n DTN as the assumptons are more realstc. 2.7 N-drop congeston control strategy under epdemc routng n DTN L et al. proposed a buffer management strategy specfc to Epdemc routng n [13]. In ths strategy, the congeston s prevented by desgnng an effcent droppng polcy when the buffer s full. Each node keeps track of the number of tmes a message has been forwarded snce the tme of ts creaton. A threshold N s calculated by each node as a functon of ts buffer sze. As the buffer sze ncreases, the threshold N also ncreases and can be shown as N = fbuffer sze). When the buffer s full, the node checks the number of forwardng of each packet n the buffer. The packet whose number of forwardng s greater than or equal to the threshold N s dropped. If all the packets n the buffer have a value less than N, then the packet at the end of the buffer s dropped. Ths method s proved to be better than the drop-last DL) packet strategy n whch the packets at the tal of the buffer are dropped on the occurrence of congeston. However when all packets have ther forwardng count less than N, then the strategy degrades to DL tself. One soluton s to rearrange the packets n descendng order of ther forwardng count and then drop the packet wth the hghest count. Ths method moreover does not control the blnd forwardng of epdemc whch s also another mportant factor that leads to congeston. 2.8 Performance analyss of schedulng and droppng polces n VDTN In [14], Vasco et al. has shown a performance analyss of varous schedulng and droppng polces along wth a combnaton of both methods to show the effect of buffer management n effectvely delverng the packets n a vehcular delay tolerant network VDTN). The schedulng polces consdered n ths paper were FIFO, random, remanng lfetme RL) and replcated copes RCs). The FIFO as known schedules the packets based on ther arrval tme. The basc concept n FIFO s that the packets that arrved frst has to be gven the top prorty. In random, ths dea s challenged because n a typcal DTN envronment where short encounter opportuntes and fnte bandwdth s n commonplace gvng bundles that arrved frst the only chance can lead to starvaton of other bundles. Therefore packets are randomly forwarded to avod the starvaton of other packets. Ths method s shown to have better delvery rato than FIFO. The next schedulng polcy s based on the RL or TTL RTTL) of the packets. The authors have studed performance of both schedulng polces one n whch packets wth smallest RTTL are scheduled frst [RL ascendng order RL ASC)] and the other method n whch the packets wth largest RTTL are [RL descendng order RL DSC)] scheduled n the begnnng. The dea behnd forwardng packets wth smallest RTTL s that these packets have small tme before ther expry and therefore should be forwarded quckly. However what f such packets gets expred soon before reachng ther destnatons? So forwardng those packets that can reman untl they reach the destnaton seems good n a DTN where destnaton nodes may be qute far from the nodes. Another method s schedulng the packets based on the RCs. In ths method, each node has to keep track of the number of tmes t got replcated. Based on ths value, we can have two approaches: RC ASC and RC DSC. Ether the bundle that has been replcated more s forwarded or the one replcated less s forwarded. These approaches are followed n droppng polces also lke FIFO drop, random drop, RL ASC/RL DSC drop and RC ASC/RC DSC drop. The smulaton results n [14] showed that best performances are acheved usng a polcy that combnes both schedulng and 4

5 droppng by gvng more preferences to the packets wth less number of RCs. Ths analyss proved that n a DTN the best consderaton should be gven for newly created copes snce they have not been dffused much nto the network. 2.9 Novel buffer management polcy for DTN In [15], Ma and Nengha Lu have proposed a method whch calculates the usefulness of a packet n a node s buffer n terms of how well t can be delvered to the destnaton node. To acheve ths, each packet n the node s assgned a weght W. The message weght s composed of two parts: one part takes nto account the possblty that the packet reaches the destnaton drectly and second part consders the possblty that t may be relayed to other nodes and then reach the destnaton. The frst part Wd s calculated as Wd = 1 e lt where λ s the parameter assocated wth the nter contact tme between the nodes whch s exponentally dstrbuted and T s the remanng TTL of the message. The second part of the weght s calculated usng the nformaton about how many copes of ths packet have been dstrbuted to the network. For ths each node records the number of copes of a packet t has created as NC and the total copy number of packet as TC. Durng an encounter opportunty, each node updates ts TC usng the encountered node s NC. W = NC TC NC The total weght for a packet s then calculated as: W = αwd + βw. Whenever the buffer s full, the packet wth the lowest W s dropped. Ths approach can work effectvely n both sparse and dense networks. In dense networks, the nodes tend to relay more than drectly delver the packets and therefore the parameter W s mportant and n sparse networks where nodes can drectly delver the packets Wd are more mportant Congeston avodance n a data centrc opportunstc network In [16], Bjurefors et al. have proposed droppng polcy based on the nterests of nodes n topcs and the degree of replcaton of packets. The scheme was evaluated on Haggle project whch tself s a data centrc network archtecture based on a publsh/subscrbe model. The strateges evaluated are least nterested LI), most nterested MI), max max copes, most forwarded MF), least forwarded LF), random, nfnte buffer and no buffer. In LI, when the buffer s full the packet n whch least number of neghbours are nterested n s deleted. Such a method ncreases the overall delvery rate n a data centrc network because the overall nterest of the network keeps ncreasng. Ths s sensble because the data n whch no nodes are nterested wll eventually expre and may be dropped. MI performs just the opposte of LI by droppng the packets that most neghbours are nterested n. Ths s judcous n the sense that t can reduce the number of copes of the packets whch most nodes keep lookng for. In max max approach, the data objects are removed after they are coped a max number of tmes. Ths reduces the data objects beng coped for a lot number of tmes. MF strategy drops the packet that has made the hghest number of replcatons. The packet whch has replcated many tmes must have reached many nodes and has a hgher chance of reachng the destnaton and therefore s sensble to drop t frst. LF operates just the opposte of MF by droppng the packet whch has been least replcated. Infnte buffer polcy assumes that each node has an nfnte relay buffer and therefore no droppng polcy s requred, whereas no buffer polcy assumes that nodes do not have a relay node and the data are accepted only f node s nterested n t. The smulaton results show that MF strategy s the best n terms of delvery rato, delay and overhead Buffer management polcy for DTN based on message sze In [17 19], message sze s used as the crtera for buffer management. Drop largest [17] drops the message wth the largest sze. Thus, a large packet s sacrfced for preventng the drop of many small packets. In the T-drop [18], the packet whch les n the threshold range of the buffer s dropped. Equal drop E-drop) [19] s a polcy where a packet of the same sze s dropped when the node s congested. Ths does not take care of the network condton to drop the packet. When a message has to be put nto the buffer, E-drop chooses that another message of equal length be dropped to accommodate the new message. The advantage s that n some cases when a packet wth a smaller sze s deleted to make more space some other packets wll have to be dropped addtonally. Thus, for accommodatng one packet we have to drop more packets. In E-drop such a stuaton s saved. The mean-drop polcy [20] calculates the mean of the sze of the messages n the congested node and then drops only those messages whose sze are greater than or equal to the mean. These polces gave a new look nto the buffer management polcy by ntroducng a new parameter apart from just consderng how old or how much replcas the message to be dropped has Buffer management for preferental delvery n opportunstc networks The approach proposed n [21] consders the message prortes for the schedulng and droppng of messages. The messages are assgned three prortes: bulk, normal and expedted. The bulk messages have the lowest prorty and the expedted have the hghest wth normal n a medum prorty. The schedulng polcy gves prorty to the expedted messages whenever an encounter opportunty arses. When a packet arrves n the node, the bundle classfer stores them n ther approprate queue n terms of the class to whch the message belongs to. When an encounter opportunty occurs the bundle scheduler s nvoked to schedule the packets for forwardng. The bundle dropper drops the message by gvng more prorty to bulk and normal messages. The polcy takes care that a node never drops a message that t has created. The scheduler transmts the bundles from hgh prorty to low prorty n a round robn fashon Buffer management scheme based on message transmsson status Lu et al. proposed a buffer management scheme nspred by the law of dmnshng margnal utlty n [22]. The law states that as a user ncreases consumpton of a product, there s a declne n the margnal utlty that user derves from consumng each addtonal unt of that product. More RCs of a message n the network s therefore not good as they consume the buffer space of other nodes preventng other messages to have a place n them. The buffer management scheme conssts of two parts: a buffer replacement scheme and a buffer schedulng scheme. The buffer management scheme works on two man propertes of the message: the number of RCs of a message n a network and the dssemnaton speed of the message. The replcaton number of message known to node m s denoted as R m. Whenever node meets node j that does not have message m, the replcaton number of m s calculated as: 1. Let node j be selected as a relay node by the routng protocol. Then f node j has no nformaton about m, then replcaton 5

6 number of m s set as R m + 1 n both the nodes. Otherwse t s set as max R m, ) Rj m + 1 n both the sendng and recevng nodes. 2. If node j s not selected as relay node, then f node j has no nformaton about m the replcaton number s set as R m on both sdes. Otherwse t s kept as max R m, ) Rj m. The messages wth lowest replcaton numbers are gven hgh prorty. When a message s to be dropped then the messages wth lowest prorty, that s, the ones havng hghest replcaton numbers are dropped. If all the messages have the same replcaton number then the dssemnaton speed of messages s used to break the te. The dssemnaton rate of a message s calculated as K m rate = TTL nt TTL where K m s the number of hop counts experenced by message m. The messages wth hgher dssemnaton speeds are dropped frst on a buffer overflow. For schedulng purposes, all the messages selected by the routng protocol to be forwarded are sorted accordng to ther prortes. Let ths set of messages be called the ready set RS). If the lowest prorty of messages n RS s greater than the hghest prorty of messages n the peerng node, then all the messages are forwarded. If the hghest prorty of messages s lower than the lowest prorty n the peerng node, then only those messages are forwarded that can be accommodated n the free space. In all other cases, the RS and the peerng node queue are merged. Ths method s unque n the sense that t ensures farness among the messages and thus t provdes overall good performance n terms of delvery rato and delay. Ths further underlnes the fact that the number of RCs n the network s a good consderaton for buffer management Enhanced buffer management polcy that utlses message propertes for DTN Shn and Km have proposed a buffer management polcy that s also based on the message propertes n [23]. Ths technque utlses three message propertes namely the estmated number of replcas, the age and the remanng TTL of the message. For calculatng the numbers of replcas of a message n the network, two parameters are used: the estmated total number of replcas ETRs) and my forward MF). Whenever node A meets node B and sends a message to node B, then the MF value of node B s set to 0 and then B sets ETR value from node A ETR B ETR A, MFB 0 ETR A ETR A + 1, MF A MF A MF A ) + 1 Now when node B meets node D whch has the same message, then the MF values of nodes B and D are exchanged so that both nodes can update the ETR value ETR B ETR B + MF D, ETR D ETR D + MF B The ETR value of a message depcts how well the message s dstrbuted n the network. Therefore a hgher ETR value says that the message has more replcas n the network. The remanng TTL and age of a message show how much tme the message has been n the network. The more the tme spent n the network, the more s the chance to obtan ts replcas dstrbuted. In [23], two utlty functons are defned. EBMP delvery = ETR logage ) EBMPdelay = 1 ETR + logrttl ) The frst utlty functon gves more emphass on droppng the messages that has been replcated many number of tmes and the second functon tends to delete an old message whch has a shorter remanng TTL. Ths strategy uses a better way to calculate the number of messages replcas by exchangng MF. In addton, t not only consders a sngle parameter, but addtonally consders the age and remanng TTL to drop a message. However f two nodes meet frequently before meetng any other new nodes, there wll be false updaton of ETR values Buffer management polces n opportunstc networks A novel buffer schedulng and droppng polcy s proposed n [24] whch uses the concept of average contact frequency among nodes for takng buffer management decsons. The message wth the hghest average contact frequency ACF) value s duplcated to the peer and the message wth the most replcas s dropped when the buffer s full. TheACF f j s defned as the number of encounters between node and node j durng a unt tme perod. ACF can be approxmated wth f j = N j /T where T s the pre-defned fxed length of tme and N j s the number of contacts between and j durng T tme. When two nodes meet each other, they exchange those messages that are not n common n the followng prncple. The messages destned to the encountered node are frstly replcated to the peer n descendng order of messages creaton tme. Other messages that meet forwardng condton are replcated to the encountered node n the descendng sequence of the ACF value. A droppng polcy of messages based on the copes of messages s used. The copes of each message s calculated usng a smple hop-count method. 3 Summary From the above study of varous buffer management polces proposed for DTNs, t can be concluded that varous factors whch affect the buffer droppng and schedulng decson are: Remanng TTL of message the amount of tme left before the message des). Hop count of message number of hops a message has travelled from the source node to the current node). Number of replcas of a message number of nodes n the network havng a message copy). Sze of the message. Age of a message tme snce the message has been created). Delvery cost e.g. delvery predctablty of message). Servce count number of transmsson events for a message). Dstance to the destnaton dstance between the current buffer node and destnaton node of a buffer). Out of these factors hop count, number of replcas of a message and forward count propertes of a message sgnfy how well a message has been dstrbuted n the network. A message whch has hgh value for these parameters means t has been spread well. Remanng TTL of message and age sgnfy the amount of 6

7 tme a message has spent n the network. A message wth smaller RTTL or large age means t has been n the network for a long tme and should have been dstrbuted well. Sze of the message sgnfes the amount of buffer space occuped by the message or the amount of bandwdth the message would occupy durng a transmsson opportunty. Delvery cost or dstance to destnaton parameters sgnfy how far a message s from ts destnaton. Bascally buffer management polces can be dvded nto three types: Global buffer management polcy whch utlse network-wde nformaton regardng all messages. Local buffer management polcy whch use partal network knowledge lke number of copes of message n the network, nstead of all network-wde nformaton correlated wth messages and addtonal message propertes lke remanng TTL, sze etc. Tradtonal buffer management polces lke drop head, drop tal drop random. Tradtonal buffer management polces perform poorly n DTN envronments. Although global buffer management schemes utlsng network-wde nformaton all outperform the ones based on local nformaton n respects of optmsng specfc network performance; polces based on partal network nformaton can compensate ther lmtatons and they can mprove the routng metrcs to near optmal levels. 4 Concluson In ths paper, a survey of varous buffer management schemes proposed n the lterature for delay tolerant networks has been conducted. As DTNs use store and forward method for forwardng the messages to the destnaton, buffer schedulng and droppng polces play a very mportant role n the effcent delvery of messages. It may be seen that a number of message parameters have been used n dfferent works for effcent buffer management; lke remanng TTL of the message, hop count of the message, number of replcas of a message, sze of a message, age of a message, delvery cost, forward count, dstance to the destnaton etc. Most of the works have gven emphass to the number of replcas a message has. However, determnng the exact number of replcas of a message s not easy snce the network s qute dstrbuted. Each method has used a unque way to determne the number of replcas. It may also be observed that a combnaton of multple message propertes can result n better buffer management decsons. Buffer management schemes n a DTN should be desgned consderng the lmted storage of nodes and the short contact duraton between nodes. In ths paper we have tred to present a varety of buffer management schemes that are generc for any routng protocol and specfc for some routng protocols. It would be nterestng to combne varous message schedulng and message droppng polces and study ther effects on varous routng protocols desgned for DTN. Use of soft computng technques for takng buffer management decsons would be another nterestng future drecton to work. 5 References [1] Cao Y., Sun Z.: Routng n delay/dsrupton tolerant networks: a taxonomy, survey and challenges, IEEE Commun. Tutor., 2012 [2] Delay Tolerant Networkng Research Group. Avalable at dtnrg.org [3] Delay Tolerant Network: Challenges and Applcaton by Fard Farahmand AITISl) [4] Krfa A., Barakat C., Spyropoulos T.: Optmal buffer management polces for delay tolerant networks. IEEE Conf. Sensor, Mesh and Ad Hoc Communcatons and Networks SECON 2008), June 2008, pp. 1 9 [5] Davs J.A., Fagg A.H., Levne B.N.: Wearable computers as packet transport mechansms n hghly parttoned ad-hoc networks. Int. Symp. Wearable Computng, 2001, pp [6] Burgess J., Gallagher B., Jensen D.: MaxProp: routng for vehclebased dsrupton-tolerant networks. INFOCOM 2006 The 25th IEEE Int. Conf. Computer Communcatons, Barcelona, Catalunya, Span, Aprl 2006, pp [7] Vahdat A., Becker D.: Epdemc routng for partally connected ad hoc networks. Techncal Report, CS , Duke Unversty, 2000 [8] Lndgren A., Dora A., Schelen O.: Probablstc routng n ntermttently connected networks, SIGMOBILE Mob. Comput. Commun. Rev., 2003, 7, 3), pp [9] Spyropoulos T., Psouns K., Raghavendra C.S.: Spray and wat: an effcent routng scheme for ntermttently connected moble networks. ACM SIGCOMM 2005 Workshop on Delay Tolerant Networkng and Related Networks WDTN-05), Phladelpha, PA, USA, August 2005, pp [10] Lndgren A., Phanse K.S.: Evaluaton of queung polces and forwardng strateges for routng n ntermttently connected networks. Proc. IEEE COMSWARE, January 2006, pp [11] Ramanathan R., Hansen R., Basu P.: Prortzed epdemc routng for opportunstc networks. ACM MobOpp 07, PR, USA, June 2007, pp [12] L Y., Qan M.: Adaptve optmal buffer management polces for realstc DTN. IEEE Global Telecommuncaton Conf., GLOBECOM 2009, USA, 30 November 4 December, 2009, pp [13] L Y., Zhao L., Lu Z., Lu Q.: N-drop congeston control strategy under epdemc routng n DTN. Ffth ACM Conf. Moble Computng Conf. IWCMC 2009), Lepzg, Germany, June 2009, pp [14] Vasco N.G.J., Farahmand S.F., Rodrgues J.J.P.C.: Performance analyss of schedulng and droppng polces n vehcular delay-tolerant networks, Int. J. Adv. Internet Technol., 2010, 3, 1), pp , ISSN: , [15] Ma K.Y., Nengha Lu B.: A new packet droppng polcy n delay tolerant network. USTC, Informaton Process Center 4th malbox Hefe cty, Anhu provnce, Chna, pp [16] Bjurefors F., Gunnngberg P., Rohner C., Tavakol S.: Congeston avodance n a data-centrc opportunstc network. ACM ICN 11, ON, Canada, 2011, pp [17] Rashd S., Ayub Q.: Effectve buffer management polcy DLA for DTN routng protocols under congeston, Int. J. Comput. Netw. Secur., 2010, 2, 9), pp [18] Ayub Q., Rashd S.: T-drop: an optmal buffer management polcy to mprove QOS n DTN routng protocols, J. Comput., 2010, 2, 10), pp [19] Rashd S., Ayub Q., Soper Mohd Zahd M., Abdullah A.H.: E-Drop: an effectve drop buffer management polcy for DTN routng protocols, Int. J. Comput. Appl., 2011, 13, 7), pp [20] Rashd S., Abdullah A.H., Soper Mohd Zahd M., Ayub Q.: Mean drop an effectual buffer management polcy for delay tolerant network, Eur. J. Sc. Res., 2012, 70, 3), pp [21] Fathma G., Wahdabanu R.S.D.: Buffer management for preferental delvery n opportunstc delay tolerant networks, Int. J. Wrel. Mob. Netw., AIRCC Publ., 2011, 3, 5), pp [22] Lu Y., Wang J., Zhang S., Zhou H.: A buffer management scheme based on message transmsson status n delay tolerant networks, IEEE Globecom, 2011, pp. 1 5 [23] Shn K., Km S.: Enhanced buffer management polcy that utlses message propertes for delay-tolerant networks, IET Commun., 2011, 5, 6), pp [24] Tang L., Cha Y., L Y., Weng B.: Buffer management polces n opportunstc networks, J. Comput. Inf. Syst. Bnary Inf. Press, 2012, 8, 12), pp