Queue-Mngement Architecture for Dely Tolernt Networking Sotirios-Angelos Lens, Stylinos Dimitriou, Fni Tspeli, Vssilis Tsoussidis Spce Internetworking Center (SPICE), ECE Deprtment, Democritus University of Thrce, Xnthi, Greece {slens, sdimitr, ttspeli, vtsousi}@ee.duth.gr Astrct. During the lst yers, the interest in Dely/Disruption Tolernt Networks hs een significntly incresed, minly ecuse DTN covers vst spectrum of pplictions, such s deep-spce, stellite, sensor nd vehiculr networks. Even though the Bundle Protocol seems to e the prevlent cndidte rchitecture for dely-tolernt pplictions, some prcticl issues hinder its wide deployment. One of the functionlities tht require further reserch nd implementtion is DTN queue mngement. Indeed, queue mngement in DTN networks is complex issue: loss of connectivity or extended delys, render occsionlly meningless ny pre-scheduled priority for pcket forwrding. Our Queue-mngement pproch integrtes connectivity sttus into uffering nd forwrding policy, eliminting the possiility of stored dt to expire nd promoting pplictions tht show potentil to run smoothly. Therefore, our pproch does not rely solely on mrked priorities ut rther on ctive networking conditions. e present our model nlyticlly nd compre it with stndrd solutions. e then develop n evlution tool y extending ns- modules nd, sed on selective scenrios primrily from Spce Communictions, we demonstrte the suitility of our model for use in lowconnectivity/high-dely environments. Keywords: DTN, Queue mngement, Scheduling Introduction Queue mngement in trditionl networks is used minly to regulte trffic fluctutions, s well s to ssign priorities to specific trffic clsses. It often utilizes dropping mechnisms tht signl end-users, implicitly or explicitly, for impeding congestion events. Nevertheless, queue mngement in Dely/Disruptive Tolernt Networks (DTN) [] [], with long delys nd disruptions hs to ddress dditionl issues. hile fir resource lloction nd trffic clssifiction re still importnt, queue mngement needs lso to exploit every ville contct opportunity nd reschedule trffic prioritiztion nd dt storge to hndle communiction disruptions nd delys. e extend further the preliminry rchitecture proposed in [3] with mthemticl nlysis, rchitecturl enhncements nd systemtic evlution. Our rchitecture is composed now y three min components: i) n Admission Control unit, which determines the criteri tht DTN nodes use to ccept or reject incoming undles, ii)
Buffer nd Storge Mngement unit, which determines how ccepted undles should e stored, nd iii) Scheduling unit, which determines undle service priorities. Here, we refer to dt hndled y DTN nodes s undles, even though the rchitecture proposed is not confined y the stndrdized Bundle Protocol [4]. The chrcteristics of ech type of DTN network my vry nd the ojective ech time my e different. Here, we emphsize on spce pplictions: spce stisfies oth dimensions of DTN, tht is, disruptions nd long delys. e hve primrily two gols: i) to increse the DTN device throughput vi efficient link exploittion nd ii) to increse ppliction stisfction. From n engineer s viewpoint, DTN requires dditionl supportive functionlity primrily for resource mngement. In DTN, resource mngement incorportes storge cpcity s well, which in turn is ssocited with long delys for dt forwrding nd incresed complexity for scheduling. For exmple, unlike typicl IP network pckets, undles do not fce the dnger to e dropped, however they do fce the dnger to expire. Also, priority-mrked pckets need not prioritized service in cse connectivity disruptions hve dmged their scope lredy. A trditionl FIFO-Droptil queue policy my hve een n initil cndidte for such system. Aprt from its simplicity nd the fct tht it my perform decently in low-trffic networks, this pproch is flwed severely: we cnnot ssign different priorities to different trffic clsses nd, on top of tht, queuing delys punish uniformly nd cumultively ll users, even those tht my hve chnce to survive potentil short disruption. Alterntively, we could integrte prioritiztion lgorithm in our scheme, such s Priority Queuing (PQ), Fir Queuing (FQ), Clss-sed eighted Fir Queuing [5] (CBFQ) nd Low Ltency Queuing (LLQ) [6]. A common, undesirle chrcteristic, however, is tht priorities re typiclly predetermined nd/or sttic, in the sense tht do not incorporte connectivity feedck nd hence cnnot reflect scheduling policy to corresponding forwrding implementtion. This non-typicl requirement renders them only lind tools for DTN mngement nd hence unsuitle, in their present form, for DTN networks. Additionl pproches presented in [7] tke ccount the low expirtion time left nd the numer of times pcket hs een forwrded, in order to drop it in cses of congestion. The most notle pproch is SHLI (drop shortest life time first) which drops the pcket with the lowest expirtion time. However, this might hve some undesirle side-effects when some pckets hve een delyed significntly, yet we cn still mnge to forwrd them efore they expire. Relevnt work on DTN queue mngement is limited, nd usully focuses on specific prolems, such s policies or scheduling, providing nrrow pproch to the queue mngement, occsionlly isolting joint prolems. However, some interesting work exists lredy. In [8], Amir Krif et l., focus on queue mngement policies, nd rech similr conclusions; they show tht trditionl uffer mngement policies such s drop-til or drop-front re su-optiml for use in DTN networks. Current implementtions of the DTN, such s ION [9] nd DTN [0], t this stge, dopt simple pproches to queue mngement considering the lck of corresponding stndrds. For exmple ION, deploys the undle protocol pproch, vi PQ scheme with three queues of outound undles, one queue for ech of the defined levels of priority ( clss of service ) supported y BP. Our pproch employs n enhnced clssifiction scheme tht integrtes oth network (i.e., connectivity) dynmics nd
trffic requirements. Therefore, scheduling is not product of pcket mrks nd hence, ppliction lone, ut rther joint decision of dt priority, ppliction potentil to survive disruptions nd the network disruptions per se. The rest of the pper is orgnized s follows. In section, we define our queue mngement model, including only the detils necessry for the stochstic nlysis, wheres in section 3 we present our stochstic nlysis nd the corresponding results. In section 4 we pply, vlidte nd compre prt of our model through simultions in ns-. Finlly, in section 5 we conclude nd set the frmework for future work. Dely Tolernt Queue Mngement Model In order to define our queue-mngement model, herefter referred s DTQM (Dely-Tolernt Queue Mngement) we consider the DTN network s network with low connectivity, high nd vrile delys nd sence of end-to-end pth. e define DTQM in wy tht llows us to include ll the necessry functionlity to stisfy generic set of requirements. Thus, we divide DTQM into three units; i) Admission Control, ii) Buffer nd Storge Mngement, nd iii) Scheduling. e discuss ll units functionlity, however, due to lck of spce we emphsize on the most novel nd sophisticted units, nmely the Buffer nd Storge Mngement nd Scheduling units. Admission Control: Admission Control determines how nd which dt my e ccepted from DTN node nd is minly relted to dt-custody requests. hen custody requests re ccepted y DTN node, tht node is oliged to mintin the undles in its memory, until it is le to forwrd them, or until they expire. Buffer nd Storge Mngement: Contrry to trditionl networks, where the routing nodes require uffers to implement store-nd-forwrd strtegy, DTN nodes need dditionl persistent storge to mintin those pckets tht cnnot immeditely e forwrded due to limited connectivity. In IP-sed routers, the min focus of reserchers is to increse chnnel utiliztion nd decrese dely through scheduling nd dropping. This pproch inherently ssumes tht end nodes respond to losses nd therefore recover in short time. However, when connectivity is scrce, the requirement for short-time recovery is lredy violted. Furthermore, DTN networks introduce n dditionl level of complexity, s result of comining oth voltile nd persistent storge. Clerly, the trivil pproch to store every incoming undle in persistent storge nd move it to uffer upon request increses the processing dely of ll the undles nd fils in cses of pplictions engged in low-dely trnsfers. In Fig. we depict grphiclly the Buffer nd Storge Mngement unit. Generlly, this model is composed y two units, the Policy unit nd the ctul Storge unit. The purpose of the Policy unit is to ccept ll the undles tht enter the node nd, depending on the conditions, move them to uffers or storge. Buffer nd Storge mngement is initilly differentited sed on whether there is connectivity etween the DTN node nd the next-hop. During periods of connectivity, pckets tht enter the node my e immeditely routed to the output without eing stored first. The totl sending rte is clculted y the sum of sending rtes of the Connectivity nd Non-Connectivity uffers.
Fig.. The Buffer nd Storge Mngement model. In more detil the purpose of ech storge unit cn e descried s follows: Connectivity uffer. The Policy unit moves undles to the Connectivity uffer only when there is connectivity nd therefore the corresponding undles cn e forwrded to the next node. After time-period which is determined y some threshold, when no connectivity exists, undles tht re stored temporrily in the Connectivity uffer move to Persistent storge. Persistent storge. The Policy unit moves undles to Persistent storge in three cses: i) when there is no connectivity, ii) when there is connectivity ut no Connectivity uffer spce ville, nd iii) when there is oth connectivity nd Connectivity uffer spce ville, however the contct grph, which is known priori, instructs tht time does not suffice to forwrd undles to the next hop. Non-Connectivity uffer. Bundles re moved from storge to the Non- Connectivity uffer in the following two cses: i) when undles re of high priority (re either urgent or scheduled contct is expected) nd there is no connectivity nd ii) when there is connectivity ut other undles re selected to e forwrded (opportunistic contct). The lgorithm tht determines which undles should e forwrded first t given communiction opportunity is descried riefly in the Scheduling section. Our proposed model dditionlly dels with the prolem of incresed processing dely. In the event of multiple nodes in row tht re ctively connected, trnsmission is rther strightforwrd, since pckets re trnsferred from uffer to uffer without the interference of storge. In the event of short connectivity no further dely due to storge retrievl is imposed; undles hve lredy moved to the Non- Connectivity uffer, nd hence ndwidth through short connectivity cn e fully exploited. Scheduling: Scheduling unit ressigns the priorities for ech undle nd determines which undles should e outputted from the DTN node when communiction opportunity occurs. A priority-oriented model should e inevitly considered; this model should incorporte ppliction requirements, dt requirements, Time-to-Live (TtL) for undles etc. Although we will not delve into more detils, our scheduling depends hevily on the rrivl timestmp nd on TtL. In order to enhnce ppliction service, we promote oth pckets tht hve recently rrived in the node nd pckets tht re ner their expirtion. This pproch decreses significntly witing delys nd promotes ppliction stisfction. In Fig. we depict the priority function used, where ToD (Type of Dt) is specific identifier tht denotes the pcket trffic clss nd TtL denotes the expirtion time.
ToD ToD ToD Fig. Scheduling priority function. 0.5 TtL TtL TtL 3 Stochstic Anlysis The purpose of this nlysis is to highlight the dvntges of our proposed model over trditionl scheduling pproches. Our model consists of primry FIFO queue (Connectivity uffer) nd secondry supportive queue (Non Connectivity uffer), which serves high-priority undles. e note tht, in the context of Spce, the queuing dely involved does not expect to contriute significntly to the totl ppliction dely, given the high propgtion dely, nd furthermore, the potentilly very high storge dely involved in typicl Spce pplictions. Therefore, priority queuing (PQ) or PQ-derived scheduling model for incoming pckets does not present conceptully tempting pproch, since it my fil with long-stored pckets. Clerly, our pproch deprts from FIFO scheme, nd therefore, clls for strightforwrd comprison with typicl FIFO scheme. However, for completeness, we extend our stochstic nlysis lso for PQ, which we consider s theoreticl upper ound (only) when connectivity is lwys present. It is pprent, from n engineer s perspective, tht our model is designed to hndle disconnectivity/disruption issues. As such, it is resonly expected to perform etter in environments with limited connectivity, especilly ginst trditionl queuing schemes, which do not include ntive mechnism to hndle intermittent connectivity. Nevertheless, to chieve firness towrds FIFO nd PQ, we use worstcse scenrio for DTQM, where connectivity is lwys ville on the system. The results shll indicte to which extent our model is le to perform stisfctorily. In such environments, where connectivity is lwys ville nd we do not exploit DTQM s full potentil, even performnce comprle to PQ nd etter thn FIFO is cceptle. e initite our nlysis y modeling network trffic. Pcket rrivl is modeled s s n exponentil process nd pcket deprture s generl distriution process. This is not, however, glolly vlid ssumption, since different types of trffic cn result in different distriutions concerning the pcket rrivl nd deprture. Nonetheless, s our knowledge on the possile DTN pplictions is limited, we ssume tht the environment nd the pplictions under investigtion mnifest ll the necessry chrcteristics of M/G/ system sed nlysis. The potentil existence of selfsimilr chrcteristics is not considered here. Furthermore, we ssume tht ll pckets hve the sme size, non-preemptive priority is enforced nd flows correspond to different trffic clsses.
One might rgue, from n nlyticl perspective, tht precision is rther duious when we ttempt to nlyticlly compre different queuing policies with potentilly distinct gols. However, there re severl occsions when these mechnisms re indeed equivlent nd directly comprle. For exmple, s throughput is decresing, the ehvior of these three systems converges. e egin our queuing nlysis y estimting the verge system dely for ech flow in PQ scheme. e consider single-server PQ system fed y three Poisson strems with rrivl rtes, nd 3. Ech strem cn e considered s flow of dt generted y vrious pplictions, in our cse Rel-time (RT), Telemetry (TM) nd Telecommnd (TC) pplictions. The uffers corresponding to different flows re infinite nd pckets in ech uffer re served in the order they rrive. Thus, we use three queues, one per flow, nd three priority clsses. e limit the overll system utiliztion y setting ρ i < nd ρ +ρ +ρ 3 <. This keeps the system from eing overloded nd cncels the possiility of flow strvtion. Tle, presents the nottion used throughout the present mthemticl nlysis. TABLE I. NOTATION N i Averge numer of pckets in ech queue i, Pcket rrivl rte t clss-i queue / Totl pcket rrivl rte μ i, μ Pcket service rte t clss-i queue / Totl pcket service rte ρ i, ρ System utiliztion fctor per clss-i / System utiliztion fctor R Men residul service time i Averge queuing dely of clss-i pcket T i Averge system dely of clss-i pcket X Second service moment Tht sid, we consider the i th dt pcket rrivl t the first queue of the PQ system. Since clss- pckets hve the highest priority, the i th pcket tht hs just rrived must wit in queue for men residul time R until the end of the current pcket trnsmission, plus the trnsmission time required for men numer of pckets N currently in the first queue, preceding the i th pcket. N () R + e clculte the men residul time, for M/G/ systems, y the formul ([]): R n i i X i, X i Now, ccording to Little s lw [], the verge numer of pckets witing in the system is equl to the verge dely multiplied y the verge rrivl rte of the system. e pply Little s lw to the clss- queue. As the verge queuing dely for clss- pckets is nd the verge queue occupncy is N with rrivl rte, we hve N (3) R (),(),(3), ρ ρ (4) Similrly, y enforcing non-preemptive priority queuing nd ccording to [] the verge queuing dely for clss nd clss-3 pckets is ccordingly: ()
3 R ρ )( ρ ρ ) ( R ρ ρ )( ρ ρ ρ ) ( 3 (5) (6) Finlly, y dding the service time /μ of the i th pcket in the equtions (4), (5) nd (6), we cn clculte the verge system dely for clss-k pckets: T k k + Next, we estimte the verge dely for ech flow in FIFO scheme. In this scheduling scheme, ech flow hs the sme verge queuing dely, which depends on the verge numer of pckets in queue N, nd the men residul time, R. According to [] the verge system dely in FIFO scheme is: T + T ( ρ) FIFO FIFO However, in order to gurntee tht the service distriution of the FIFO queue corresponds to n ppropritely weighted sum of the service distriutions for the different clsses in the priority queue scheme, we set: T NFIFO 3T R FIFO e continue our nlysis y estimting the verge dely for ech flow in DTQM scheme. e divide the nlysis in two prts. Since DTQM uses two outgoing queues (plus the permnent storge see Fig. ) one for connectivity nd the other for non-connectivity dt, we cn sfely ssume tht the first one emultes FIFO queue while the second one pproches the ehvior of PQ scheme. The ltter ssumption holds since the prioritiztion function tht we pply (see Fig. ), requires pcket sorting. Tht sid, the first queue nlysis in terms of verge system dely is directly comprle with FIFO scheme, while the second queue clls for PQ-sed nlysis. In order to firly evlute our proposed scheme, we omit permnent storge from the nlysis. This llows the three systems to present similr properties nd hence e comprle. e consider n verge system rrivl rte equl to the other systems nd set rrivl rtes, without loss of generlity, for the two queues 0.4 nd 0.6. The selection of coefficients 0.4 nd 0.6 s the preferred vlues for our stochstic nlysis ws sed on some initil empiricl clcultions nd will e clirted further ccording to emultion results. Finlly, the totl verge system service rte will e μ where μ 0.4μ nd μ 0.6μ for first nd second queues, respectively. To strt off, the verge queuing dely for the first queue does not differentite per flow nd cn e clculted sed on the verge numer of pckets in queue- nd the men residul time, using eqution () nd replcing,, μ nd μ with,, μ nd μ respectively. (7) (8) (9)
N (0) R CON + By pplying Little s lw for queue- we get: N () CON From equtions (0) nd () we get: CON R 0.4, where ρ ρ ρ The verge system dely for ll the flows in queue- is: T CON CON + () (3) e now pproximte the ehvior of the second queue s follows. The queue cn split into three su-queues or suclsses. Ech clss will e served using PQ-sed scheme. Furthermore, the proportion of pckets for ech clss on the totl ville cpcity of the queue- uffer is / for clss- pckets, / for clss- pckets nd 3 / for clss-3 pckets. Therefore, the verge queuing dely for clss- pckets is: N + NONCON By pplying Little's lw [], we otin: N N NONCON 0. 6 NONCON From equtions (4) nd (5) we get: R, where ρ NONCON ρ R ρ 0.6 The verge queuing dely for the second suclss depends on N pckets, which re uffered in the first suclss, plus N pckets, which re uffered in the second suclss, plus the residul time R. In our cse, unlike the ordinry properties of PQ, our design ssumptions do not permit the possiility of higher priority pckets to rerrnge the queue t ny given stge. Therefore, we hve: N N R + + + NONCON NONCON NONCON N N NONCON 0. 6 NONCON From equtions (7) nd (8) we otin: N (4) (5) (6) (7) (8)
NONCON NONCON ρ By the sme token, verge queuing dely for third suclss is: 3NONCON NONCON ρ 3 Finlly, system's verge dely for ech suclss is: T k NONCON k NONCON + (9) (0) () Hving clculted the verge delys for the two queues seprtely, we will comine these vlues to cquire the totl dely. A sttisticlly cceptle method for doing this is y using weights. Considering the wy tht we hve defined the prolem, it is logicl to expect tht ech queue will contriute with different percentge to the overll system dely. Hence, we will consider tht queue- nd queue- will contriute to the totl verge system dely y 40% nd 60%, respectively. Considering the vlues of nd μ in ech queue, we hve. T 0.4T + 0. 6T DTQM CON T 0.4T + 0. 6 NONCON DTQM CON T T 0.4T + 0. 6 3 T DTQM CON 3NONCON,, NONCON As for the numericl results presented elow, the vlue of system service rte is constnt t 0 pckets/sec, wheres the vlues of vry in order to otin the possile rnge of system utiliztion, 0% - 90%. Τhe vlue of service rte, considering pcket sizes in the order of KB, provides n cceptle rte of dt trnsmission, especilly in spce environments nd mong low energy sensors. () Fig 3. Numericl results The numericl results of our nlysis re presented in Fig. 3. e compre the forementioned queuing schemes sed on the verge dely for ech queue in ech system nd the verge queue occupncy. The results show tht our pproch
chieves performnce clerly etter thn FIFO nd in some cses etter thn PQ, especilly when the system utiliztion fctor is high. 4 Experimentl Evlution Evluting such queue mngement policy requires extensive experiments using n ctul spce network, since DTQM ffects oth lower-level (ttery lifespn) nd higher-level (throughput, ltency) performnce metrics. e implement nd evlute our proposed solution using the ns- softwre simultor [3]. This implementtion ws not trivil nd required significnt time considering the fct tht ns- does not support DTN. In this work, we focus on the second nd third components of DTQM, nmely the Buffer nd Storge Mngement nd Scheduling prts, leving the rest of the rchitecture evlution s future work. e pply DTQM to the network topology of Fig. 4. e ssume three sending nodes, S, S, S 3, which re locted in spce nd send trffic generted y vrious pplictions (Rel-time, Telemetry nd Telecommnd; ll constnt it rte with different sending intervls) nd receiving node R, which is locted on erth. All trffic generted y the sending nodes is routed through the routing node Q in which we deploy DTQM. Fig 4. Simultion topology Since we ssumed tht the sending nodes re locted in spce, the connectivity of the wireless links should e intermittent. In order to emulte DTN environment where connectivity is not predetermined, thnks to lterntive routes nd connections, we select rndom disconnectivity pttern with uniformly distriuted connectivity disruptions spred cross the entire durtion of the experiment, s the most pproprite for evluting the proposed rchitecture. In this context, link S -Q is uniformly unville in totl % of the time of the experiment. Similrly link S -Q is 5% unville, link S 3 -Q is 0% unville, nd finlly, link Q-R is 0.5% unville. e set the totl time of the experiments to one hour, nd mesure the totl numer of received pckets for ll three sending nodes, s well s the Appliction Stisfction Index (ASI) [4] of the network, metric tht highlights the contriution of the queuing dely to the totl dely. In order to dd reliility to our results nd enforce rndomness to tke effect we repet the experiment severl times. In prticulr, the performnce of DTQM ws evluted in five connectivity scenrios, ech one utilizing different (rndomly generted) connectivity schedule. e compre the otined results with the corresponding results of FIFO policy. In line with the priority function used (see Fig. ) we ssign ToD for ech ppliction s follows: for the RT ppliction, for the TM ppliction nd 3 for the TC ppliction.
Furthermore, we set the pcket intervls for ech ppliction s we would expect in rel-life, tht is, Rel-Time pckets re generted with the shortest intervl nd Telecommnd pckets re generted with the longest intervl; hence the difference in pcket numers. In Fig. 5 elow, we present the results from the comprison of DTQM ginst Droptil, using ASI nd verge dely s our performnce metrics. The first oservtion tht we cn mke is tht DTQM outperforms Droptil in ny cse. In prticulr, we experience 60% dely decrese on verge nd in some cses it cn rech up to 90% reduction of the corresponding undle dely using typicl FIFO scheme. This dely decrese is lso reflected on the system ASI, which is incresed 0%, on verge. Fig 5. Experimentl results TABLE II. EXPERIMENTAL RESULTS Received pckets Averge Dely System ASI Best Cse orst Cse RT TM TC RT TM TC FIFO 5680 696 5 573 645 4 DTQM 568 754 5 5500 869 7 FIFO 308 96 477 306 3 9 DTQM 357 67 50 565.9 789.7 877 FIFO 0.49 0.53 DTQM 0.69 0.55 Tle demonstrtes in detil the est nd worst cse results for DTQM. By viewing Tle, we notice tht the received pckets re lmost the sme in ny cse (with the exception of the TM pckets of the worst cse experiment) regrdless of queuing policy. Nevertheless, we my experience up to 40% increse in received Telemetry nd Telecommnd pckets when DTQM is deployed, since we ssign them with higher ToD. Moreover, the most interesting oservtion is the undoutle improvement of the verge dely regrdless of the ppliction. However, since we trnsfer the sme numer of pckets in oth cses, how cn we justify the dely decrese? DTQM uses sophisticted scheduling lgorithm tht ssigns higher priority to the pckets most recently rrived in the node nd promotes them in the queue (see Fig. ). Thus, pckets re reordered in the uffer sed on the time they entered the routing node. Clssic scheduling uses rigid FIFO pproch, which lthough it seems to promote firness, in fct it increses the communiction time, with the risk of dropping pcket due to TtL expirtion.
5 Conclusions In this pper we proposed novel rchitecture for queue mngement in DTN nodes. Although the ville spce confines us from presenting more detiled version nd evlution of our model, we sketched severl of its chrcteristics. One of the most interesting results ws initilly introduced y the stochstic nlysis, which yielded positive results for the opertion of our model tht exhiits ehvior fr superior to FIFO nd comprle to PQ, even in network conditions tht re unfvorle for our scheme. e lso otined supportive results from the conducted experiments tht llevite worries from dopting numerous ssumptions on the stochstic nlysis section. Therefore we cn sfely clim tht DTQM hs the potentil to chieve smller queuing delys nd higher ppliction stisfction when connectivity is scrce. Our next step is to enhnce our evlution towrds two directions: i) to present more detiled nlysis, which incorportes totl cpcity nd storge cpcity s well, in order to highlight one mjor property of DTN nd ii) to extend the experiments y using the spce-oriented tested [5] tht we hve developed in our l. Acknowledgments. The reserch leding to these results hs received funding from the Europen Community's Seventh Frmework Progrmme ([FP7/007-03_FP7- REGPOT-00-, SP4 Cpcities, Coordintion nd Support Actions) under grnt greement n 646 (project title: Spce Internetworking Center-SPICE) References. Fll, K.: A dely-tolernt network rchitecture for chllenged internets. ACM SIGCOMM 003, Krlsruhe, Germny, 5-9 August (003).. Cerf, V., Burleigh, S., Hooke, A., Torgerson, L., Durst, R., Scott, K., Fll, K., eiss, H.: Dely Tolernt Networking Architecture. RFC 4838, April (007). 3. Dimitriou, S., Tsoussidis, V.: Effective Buffer nd Storge Mngement in DTN Nodes. E-DTN 009, St. Petersurg, Russi, 4 Octoer (009). 4. Scott, K., Burleigh, S.: Bundle protocol specifiction. RFC 5050, Novemer (007). 5. Clss-sed eighted Fir Queuing, Cisco IOS Softwre Releses.0 T. 6. Low Ltency Queuing, Cisco IOS Softwre Releses.0 T. 7. Lindgren A. nd Phnse K. S.: Evlution of queuing policies nd forwrding strtegies for routing in intermittently connected networks. Proc. of IEEE COMSARE, Jnury 006. 8. Krif, A., Brkt, C., Spyropoulos, T.: Optiml uffer mngement policies for dely tolernt networks. IEEE SECON 008, Sn Frncisco, Cliforni, June 6-0 (008). 9. Jet Propulsion Lortory: ION: Interplnetry Overly Network. https://ion.ocp.ohiou.edu. 0. The Dely-Tolernt Networking Reserch Group (DTNRG): http://www.dtnrg.org/.. Bertseks, D. P., Gllger, R.: Dt Networks. Prentice Hll (99).. Little, J. D. C.: A Proof of the Queueing Formul L. Opertions Reserch, 9, (96). 3. Network Simultor, http://www.isi.edu/nsnm/ns/ (997). 4. Mmts, L, Tsoussidis, V: Differentiting services with Non-Congestive Queuing (NCQ). IEEE Trnsctions on Computers, 58(5):59 604 (009). 5. Smrs, C. V., Komnios, I., Dimntopoulos, S., Koutsoginnis, E., Tsoussidis, V., Ppstergiou, G., Pecci, N.: Extending Internet Into Spce - ESA DTN Tested Implementtion nd Evlution. Moilight 0, Athens, Greece, 8-0 My (009).