Proactive selective neighbor caching for enhancing ... · Proactive Selective Neighbor Caching for Enhancing Mobility Support in Information-Centric Networks Xenofon Vasilakos, Vasilios

Proactive Selective Neighbor Caching for EnhancingMobility Support in Information-Centric Networks

Xenofon Vasilakos, Vasilios A. Siris, George C. Polyzos, and Marios PomonisMobile Multimedia Laboratory

Department of InformaticsAthens University of Economics and Business76 Patission Str., Athens, Greece GR10434

{xvas, vsiris, polyzos}@aueb.gr, [email protected]

ABSTRACTWe present a Selective Neighbor Caching (SNC) approachfor enhancing seamless mobility in ICN architectures. Theapproach is based on proactively caching information re-quests and the corresponding items to a subset of proxiesthat are one hop away from the proxy a mobile is currentlyconnected to. A key contribution of this paper is the def-inition of a target cost function that captures the tradeoffbetween delay and cache cost, and a simple procedure for se-lecting the appropriate subset of neighbors which considersthe mobility behavior of users. We present investigations forthe steady-state and transient performance of the proposedscheme which identify and quantify its gains compared toproactively caching in all neighbor proxies and to the casewhere no caching is performed. Moreover, our investiga-tions show how these gains are affected by the delay andcache cost, and the mobility behavior.

Categories and Subject DescriptorsC.2.1 [Computer-Communication Networks]: NetworkArchitecture and Design—distributed networks, network com-munications.

General TermsDesign, Performance

KeywordsMobility, Information-Centric Networking, Publish-Subscribe

1. INTRODUCTIONInformation-Centric Networks (ICNs) have two key dif-

ferences compared to IP networks that have implications tomobility: First, they employ a receiver-driven model wherereceivers request content by its name, asynchronously, from

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.ICN’12, August 17, 2012, Helsinki, Finland.Copyright 2012 ACM 978-1-4503-1479-4/12/08 ...$15.00.

the publishers. Second, the transport of content from pub-lishers to receivers is performed in a connectionless (state-less) manner, in contrast to TCP’s connection-oriented (state-full) end-to-end control involving location-dependent IP ad-dresses. Both features allow mobile receivers to re-send re-quests for content they did not receive while they were attheir previous attachment point or could not receive whilethey were disconnected, without requiring re-establishing aconnection or cumbersome and costly overlay solutions suchas mobile IP.

Despite the aforementioned support for mobility, increaseddelay for receiving data can be incurred when a mobile sendsa request for content but disconnects or moves to a new at-tachment point before receiving the requested data. Thisdelay can be significant for applications with strict delayrequirements. Examples of such applications include real-time emergency notification services, teleconferencing, andonline gaming, which are sensitive to the end-to-end delay.Another example is streaming multimedia services which aresensitive to delay jitter.

The key premise of this paper is that this delay can be re-duced by exploiting knowledge from information requests (orsubscriptions), which is available due to the receiver-drivenmodel of ICN architectures, and knowledge from the users’mobility behavior; this knowledge can be utilized by prox-ies that enhance mobility support. Proxies can be viewedas special caches with additional functionality to handle in-formation requests on behalf of mobiles, and pre-fetch andcache items that match a mobile’s requests while the mobileis in a handover phase or is disconnected from the network.Such an application of caching is different from its typicalapplication of serving requests for the same content fromdifferent users. Understanding the advantages of cachingfor enhancing mobility support, we believe, is important forunderstanding the overall benefits from in-network cachingin ICN architectures.

There are three types of caching solutions identified in lit-erature based upon when and where subscriptions and infor-mation items are cached: i) reactive approaches [2, 5, 12, 13],ii) durable subscriptions [4], and iii) proactive approaches [1,7]. In reactive approaches, when a mobile disconnects fromits current proxy, the latter keeps caching items that matchthe mobile’s subscriptions. When it reconnects, the mobileinforms the new proxy of the old proxy’s identity, and thenew proxy requests from the old proxy all items that havebeen cached during the disconnection period. This reactiveprocedure has the disadvantage of increased delay for the

61

new proxy to start forwarding items to the mobile, since theproxy must receive all the items cached in the old proxy be-fore it can start forwarding. This delay can be avoided withboth durable subscriptions and proactive caching solutions.

With durable subscriptions, proxies maintain a mobile’ssubscriptions and cache items matching these subscriptions,independent of whether the mobile is connected to the proxyor not. Without any additional mechanisms, in order toavoid loosing information, all proxies within a domain thata mobile can possibly connect to would need to maintainsubscriptions and cache matching items; this incurs signifi-cant memory costs.

Finally, proactive caching solutions enhance mobility byproactively caching items matching a mobile’s subscriptionsin single hop distance “neighboring” proxies, before the mo-bile disconnects. The selection of proxies to proactivelycache items can be based on prediction [1], or knowledgeof all neighboring proxies lying one hop ahead [7]. Thework of [1] does not propose a specific prediction algorithm,whereas with the approach of [7], when a mobile discon-nects all neighboring proxies start caching items matchingthe mobile’s subscriptions. Hence, when the mobile connectsto one of these proxies, it can quickly receive items that weretransmitted during its disconnection. Proactive approachestrade-off buffer space for reduced delay in forwarding itemsto subscribers.

A selective neighbor caching approach has been proposedfor reducing the handover delay in WLANs [9]. The moti-vation is that, even when the number of neighboring APs islarge, at most 3 or 4 access points are targets of the hand-offs [9]; it is likely that a similar motivation applies for ICNs,since the movement of users is the cause for mobility in bothcases. However, the application of such an idea for support-ing mobility in ICNs is different due to the different natureof the problem: in ICNs the objective is to forward to mo-biles items that match their subscriptions, whereas in [9] theobjective is to proactively send a mobile’s context to neigh-boring access points in order to reduce association delay. Asa result, the model and corresponding procedure for select-ing the subset of neighbors is fundamentally different. Wealso note that mobility prediction has been used to improveQoS support during handovers in cellular networks, e.g., see[3, 11].

In this paper we propose a Selective Neighbor Caching(SNC) approach which selects an appropriate subset of neigh-boring proxies that minimizes the mobility costs in termsof expected average delay and caching costs. The prelim-inary ideas of the target cost function used in SNC werepresented in an extended abstract [10]; in the current paperwe present the procedure for selecting neighbor proxies toproactively cache information items and discuss its imple-mentation. Moreover, we present and discuss performanceresults that illustrate the tradeoff between the delay andcache cost, and we identify and quantify SNC’s gains com-pared to full proactive caching, i.e., caching at all neighborproxies, and when caching is not used.

The remainder of the paper is structured as follows: InSection 2 we present the target cost function and decisionprocedure for the proposed SNC approach, and in Section 3we discuss its application over two representative ICN ar-chitectures: Publish-Subscribe Internetworking (PSI) andContent-Centric Networking (CCN). In Section 4 we evalu-ate the steady-state and the transient performance of SNC,

comparing it with the case of full proactive caching and whenno caching is performed. Finally, in Section 5 we concludethe paper identifying ongoing research directions.

2. SELECTIVE NEIGHBOR CACHINGConsider a mobile that is initially connected to proxy i,

Fig. 1. Assume that the probabilities pij of the mobile con-necting to proxy j ∈ J (set of i’s neighbor proxies) after itdisconnects from i are known. Note that the term neighbordoes not correspond to geographic proximity, but is relatedto the mobile’s sequence of network attachment points; forexample, when a mobile is connected to a WLAN in a sub-way station, the neighboring proxies can be those located inother subway stations.

The key idea of the SNC approach proposed in this paperis to select an optimum subset of neighbor proxies to whichi will send the mobile’s subscriptions, and which will proac-tively cache information items matching these subscriptions.If a mobile connects to one of the proxies in this subset, thenthe mobile can immediately receive information items it didnot receive due to its disconnection from the previous proxy.Let the probability that the mobile connects to a proxy inthe subset S ⊆ J be Phit(S) =

∑j∈S pij . The optimum

subset of neighbor proxies S∗ is the set that minimizes thefollowing target cost function:

Phit(S) · Chit + (1− Phit(S)) · Cmiss +N(S) · Ccache , (1)

where N(S) is the number of proxies in set S. Chit is thedelay experienced by the mobile in order to receive its re-quested information items from a proxy j that has proac-tively cached items matching its subscriptions; the proba-bility for the mobile to move to such a proxy is Phit(S).Cmiss > Chit is the delay cost for a mobile to receive itemsfrom their publishers, which can occur with probability 1−Phit(S). Finally, Ccache is the cost for a proxy to cache a mo-bile’s subscriptions and the matching items. If we assumethat the memory requirements for storing subscriptions issmall relative to information items, then Ccache depends lin-early on the memory requirements for storing a single item.

The proposed target cost function captures the tradeoffbetween the average delay for a mobile to receive an infor-mation item, which is given by the first and second termsof (1), and the corresponding cache cost, which is given bythe third term of (1). Increasing the set S of neighborsthat proactively cache items increases Phit(S), hence on onehand reduces the average delay for obtaining informationitems since Cmiss > Chit, but on the other hand increasesthe buffers that are required to cache the items. A key issueis how to quantify the cache cost in order to add it to thedelay. In Section 2.2 we discuss how the cache cost can bequantified in the practical case where proxies have a fixedcache size.

When a mobile disconnects1 from its current proxy i, theni notifies its neighbor proxies in the set S∗ to start cachingitems that match the mobile’s subscriptions, starting fromthe item after the last item forwarded to the mobile. Whenthe mobile connects to a new proxy, then the latter informsthe old proxy that the mobile has reconnected, and subse-

1Depending on the disconnection period’s duration, an alterna-tive is that neighbor proxies start to proactively cache prior to amobile’s disconnection. The target cost function (1) still remainsthe same.

62

Proxy: Cache with mobility support mechanisms

i

j

Mobile

Current proxy:S: Subset of neighbor proxies

pij

Figure 1: A mobile is currently connected to proxy i. Based onthe probabilities for the mobile to connect to proxy i’s one hopneighbors, SNC selects the subset of neighbor proxies to proac-tively transmit the mobile’s subscriptions in order to minimizethe target cost function (1).

quently the old proxy informs all neighbor proxies in S∗ tostop caching items for that specific mobile. In Section 3 wediscuss how the above communication between proxies canbe implemented in two representative ICN architectures.

2.1 Finding S∗

Next we discuss how to find the set S∗ that minimizesthe target cost function (1). Consider a subset S′ ⊆ J . Ifitems matching a mobile’s subscriptions are pre-fetched byall proxies in S′, the total cost would be

Phit(S′) ·Chit+

(1− Phit(S

′)) ·Cmiss+N(S′) ·Ccache . (2)

Assume now that we want to decide whether to include aproxy j in the subset of proxies that pre-fetch informationitems. The total cost when items are pre-fetched by proxiesin the set S′ ∪ {j} is

Phit(S′∪{j})·Chit+

(1 − Phit(S

′ ∪ {j}))·Cmiss+(N(S)+1)·Ccache .(3)

Pre-fetching items in proxy j would be beneficial if and onlyif cost (3) ≤ (2), which is equivalent to

pi,j ≥ Ccache

Cmiss − Chit=

CcacheChit

CmissChit

− 1. (4)

The numerator in (4) denotes the cost for caching, whilethe denominator denotes the gain with caching, in termsof reduced delay. Hence, finding the set S∗ that minimizesthe cost function (1) involves deciding individually for eachneighboring proxy whether it should proactively cache infor-mation items matching a mobile’s subscriptions, using (4).In addition to involving a simple comparison, (4) can be ap-plied centrally at the original proxy i, or decentralized atevery neighbor j ∈ J to decide whether to proactively cacheitems or not.

In equations (1) and (4) we have assumed that the de-lay cost Cmiss for obtaining items from their publishers isthe same for all publishers and for all items; the latter isequivalent to assuming that all items have the same size.In practice, there can be a different delay cost for receiv-ing items from different publishers and for different itemsizes. The delay costs Chit and Cmiss can be estimated fromactual measurements of the time for the mobile to receiveitems from a local proxy or from the item’s publisher, re-spectively, after issuing the request for that item. Note thatthese delay costs refer to the neighbor proxies J , which themobile can connect to. Hence, the estimation of the costs

Chit and Cmiss should be performed at each neighbor proxyj ∈ J . Moreover, the cache cost Ccache can depend on thebuffer utilization, as we discuss in Section 2.3, hence differ-ent neighbor proxies can have different cache costs. Fromthe above, (4) can be written more generally as

pi,j ≥C

jcache

Cjhit

Cjmiss

Cjhit

− 1. (5)

An advantage of implementing (5) in the neighbor proxiesj ∈ J , rather than the original proxy i, is that this wouldavoid the need for communicating the ratios Cj

cache/Cj

hit and

Cjmiss/C

j

hit from all the neighbor proxies to proxy i. On theother hand, the transition probability pi,j can be estimatedby the ratio of the number of transitions from proxy i toproxy j, which is known to j, over the total number of tran-sitions from proxy i; the latter needs to be communicatedfrom i to all its neighbor proxies j ∈ J . Estimation of thetransition probabilities can be enhanced with informationsuch as location, orientation, and road/path topology, whichhave been used for mobility prediction in cellular networks[3, 11].

2.2 Cache costThe target cost function (1) and (5) involves the delay cost

for obtaining an information item and a cost for proactivelycaching an information item. In a practical application ofthe proposed SNC approach to a system where proxies havea fixed cache size, a goal would be to achieve a high cache uti-lization. This suggests that when the cache utilization is low,then the cache cost should also be low to allow more proac-tive caching in neighbor proxies. On the other hand, whenthe cache utilization is high then the cache cost should alsobe high in order to reduce the amount of proactive caching.Based on the above, the cache cost can be defined as

Cj

cache =a

1− ρjutil, (6)

where ρjutil is the cache utilization at proxy j. The cachecost coefficient a can be adjusted to achieve a minimumcache utilization.

2.3 ExtensionsThe transition probability on the left-hand side of (5) cap-

tures the average behavior of all mobiles that connect toproxy i. If there are different types of mobiles, with mobilesof the same type having a similar mobility behavior, whereasmobiles of a different type have a distinct mobility behavior,then one can apply (5) considering the specific proxy tran-sition probability for each type of mobile. A mobile’s typecan be identified through some characteristic or informationmade available by the mobile.

Equation (5) assumes that the cost for proactively cachingan information item is incurred by one mobile, that re-quested the specific information item. However, when morethan one mobile request the same information item, thenthe caching cost should be shared by the mobiles requestingthat item. If the average number of requests for an infor-mation item k is nk, then the decision inequality (5) can be

63

written as:

pi,j ≥C

jcache

Cjhit

nk ·(

Cjmiss

Cjhit

− 1

) .

3. SNC OVER PSI AND CCNSNC can be implemented in software modules running in

the proxies, without modifying the underlying ICN archi-tecture. The implementation of SNC involves the following:i) estimation of the cost ratios Cj

miss/Cj

hit and Cj

cache/Cj

hitand the transition probabilities pij that appear in (5), ii)selection of the neighbor proxies that will perform proac-tive caching, iii) transmission, from the original proxy to itsneighbor proxies, of a mobile’s requests and notifications tostart/stop proactive caching, and iv) notification contain-ing the mobile’s new attachment point which is sent by themobile’s new proxy to its old proxy.

Recall from Section 2.1 that costs and transition probabil-ities can be estimated locally at each neighbor proxy j ∈ J .The only information that needs to be communicated in or-der to estimate the transition probabilities is the total num-ber of transitions from the old proxy i to any of its neighbors;this communication involves a one-to-many dissemination ofthe same information, from proxy i to all its neighbors. Sim-ilarly, the transmission of a mobile’s requests and notifica-tions to start/stop proactive caching from the original proxyto neighbor proxies also involves a one-to-many dissemina-tion of the same information. If the decision of whether toproactively cache information items is taken at the neighborproxies, then the aforementioned one-to-many communica-tions can be performed in a receiver-driven (pull-based) fash-ion. Hence, they can be appropriately implemented usingthe receiver-driven (pull-based) communication primitive inthe Content-Centric Networking (CCN) architecture [8], ex-ploiting CCN’s ability to effectively disseminate that sameinformation to multiple interested receivers. Similarly, itcan be implemented in the Publish-Subscribe Internetwork-ing (PSI) [6] architecture, which supports a receiver-drivenmodel at the rendezvous (or resolution) layer. Moreover,one-to-many dissemination can exploit both CCN and PSI’snative multicast capabilities.

Finally, the last communication mentioned above (in iv))involves the old proxy being notified of the mobile’s newproxy. This can be implemented in both PSI and CCN byhaving the old proxy issue a subscription message (in PSI)or an Interest (in CCN) that corresponds to an informationitem containing the mobile’s connection status. Once themobile connects to a new proxy, the latter issues a pub-lication announcement (PSI) or a Data packet (CCN) thatmatches the aforementioned subscription or Interest, respec-tively. Note that this communication exploits the ability toasynchronously issue subscriptions or Interests prior to pub-lication announcements or Data packets in the PSI or CCNarchitecture, respectively.

4. PERFORMANCE EVALUATIONIn this section we present analytical investigations for the

steady-state and simulation investigations for the transientperformance of SNC. The analytical investigations illustratethe tradeoff between the average delay, the cache cost, theirinfluence on the total cost, and how the tradeoff and gains

of SNC depend on the cost ratios Cmiss/Chit, Ccache/Chitand on the mobility pattern. The simulation investigationsshow the transient delay gains of SNC when the cache size isfixed and when the proxy transition probabilities, based onwhich the selection of neighbors to proactively cache itemsis made, are estimated on-line rather than a priori known asin the analytical investigations.

For the analytical investigations we consider scenarios withone proxy and 4 neighboring proxies, whereas in the tran-sient investigations we consider scenarios with 5 and 8 neigh-bor proxies. The analytical investigations show the averagedelay, the cache cost, and the total cost, which are definedbelow:

Total cost = Phit · Chit + (1− Phit) · Cmiss︸ ︷︷ ︸

Average delay

+N(S) · Ccache︸ ︷︷ ︸

Cache cost

.

(7)

The analytical results show SNC’s gain, defined as the re-duction of the total cost relative to full proactive caching,for which S = J and Phit = 1, and when no caching is used,for which S = ∅ and Phit = 0. The simulation results showSNC’s gains in terms of average delay, compared to full andno caching.

4.1 Analytical investigationsIn this subsection we present analytical investigations that

illustrate the tradeoff between the average delay and cachingcost, and how the cost ratios Cmiss/Chit and Ccache/Chit,and the proxy transition probabilities influence SNC’s gainscompared to the case where no proactive caching is per-formed and when all neighbors proactively cache a mobile’ssubscriptions; the latter corresponds to the proactive ap-proach of [7].

Average delay and cache cost tradeoff: Fig. 2 shows thetradeoff between the average delay and the caching cost.As expected, the average delay decreases when more neigh-bor proxies proactively cache a mobile’s subscriptions. Onthe other hand, the caching cost increases when more neigh-bor proxies proactively cache subscriptions. SNC capturesthis tradeoff and selects the subset of neighbors that givesthe lowest total cost. Different cost ratios Cmiss/Chit andCcache/Chit result in a different selection of neighbors thatproactively cache information items: SNC selects 4 neigh-bor proxies (N∗ = 4) in Fig. 2(a), 2 neighbors in Figs. 2(b)and 2(d), and zero neighbors in Fig. 2(c). Note that the twoneighbors SNC selects in Figs. 2(b) and 2(d) are the twowith the highest transition probability (0.5 and 0.3).

Influence of Ccache, Cmiss, and proxy transition probabili-ties: Fig. 3 shows that as the ratio Cmiss/Chit increasesSNC’s gains compared to full proactive caching are lower;this occurs because the optimal number of proxies for proac-tive caching increases, Fig. 3(b). On the other hand, SNC’sgains compared to the case where no caching is performedincrease with Cmiss/Chit.

Fig. 4 shows the influence of the ratio Ccache/Chit onthe gains and the optimal number of neighbors that per-form proactive caching, for two proxy transition probabili-ties. Observe that increasing the ratio Ccache/Chit reducesSNC’s gains compared to the case where proactive caching isnot performed, whereas it increases SNC’s gains comparedto the case of proactive caching in all neighbors. Fig. 4(a)shows that when mobiles move to neighboring proxies with

64

(a) Cmiss/Chit=12 ,

Ccache/Chit=0.5,N∗=4

(b) Cmiss/Chit=12 ,

Ccache/Chit=2,N∗=2

(c) Cmiss/Chit=4,

Ccache/Chit=2,N∗=0

(d) Cmiss/Chit=4 ,

Ccache/Chit=0.5,N∗=2

Figure 2: Tradeoff between average delay and caching cost. SNCconsiders this tradeoff and selects the subset of neighbors thatgives the lowest total cost.

(a) Gain (b) N∗

Figure 3: Influence of Cmiss/Chit on gains and optimal numberof proxies that perform pre-fetchingN∗ = N(S∗). Ccache/Chit =1, {pij} = {0.5, 0.3, 0.1, 0.1}.

a uniform probability, then SNC’s gain compared to fullcaching and no caching are smaller in comparison to thecase of non-uniform probabilities for intermediate cost ra-tios Cmiss/Chit (2 and 4 in Fig. 4). Fig. 4(b) shows thatwhen the proxy transition probabilities are uniform, the op-timal number of proxies to perform pre-fetching exhibits abimodal behavior: either it is optimal to not pre-fetch in anyproxy or it is optimal to pre-fetch in all neighboring proxies.

4.2 Simulation investigationsNext we evaluate SNC using the OMNeT++ simulation

framework. Due to space limitations, we present the resultsfrom a small set of simulations. We consider 100 mobileswhich are uniformly distributed to 5 proxies. From its ini-tial proxy, a mobile can move to one of 5 other proxies withprobabilities 80%, 10%, 5%, 2.5%, 2.5%. We assume that thedestination proxy with probability 80% is different for differ-ent initial proxies, hence the number of mobiles that connect

(a) Gain (b) N∗

Figure 4: Influence of Ccache/Chit on gains and optimal num-ber of proxies that perform pre-fetching. Cmiss/Chit = 12,{pij} = {0.5, 0.3, 0.1, 0.1} (non-uniform), {0.25, 0.25, 0.25, 0.25}(uniform).

to the destination proxies from any initial proxy is uniform.After a mobile moves to a new proxy, a new mobile con-nects to one of the initial 5 proxies, hence the total numberof mobiles in the system remains 100, and each proxy has onaverage 20 mobiles. We also consider a scenario with 8+8proxies, where the total number of mobiles is 160, uniformlydistributed to the different proxies. Information items re-quested by mobiles have the same size, and each proxy cancache up to 20 items. For the above scenario, we have foundthat an appropriate value for the cache cost coefficient in(6) is a = 2. If we consider a minimum target cache utiliza-tion (e.g., 80%), the cache cost coefficient can be estimatedfrom equations (5) & (6), and the cache size, distributionof mobiles, and proxy transition probabilities. Finally, notethat the transition probabilities are estimated online, ratherthan a priori known as in the analytical investigations.

Fig. 5(a) shows that SNC can achieve lower delay com-pared to both full proactive caching and no caching. Asexpected, the delay improvement is higher for a higher ratioCmiss/Chit. The convergence time is necessary for the cacheutilization and the estimated proxy transitions probabilitiesto converge. Fig. 5(b) shows SNC’s delay gain for a differentnumber of proxies. The above results show that the intelli-gent selection of neighbor proxies to cache information canefficiently utilize caches, significantly improving delay.

5. CONCLUSIONS AND FUTURE WORKWe have presented a Selective Neighbor Caching (SNC)

approach for enhancing seamless mobility support in ICNarchitectures. SNC selects the set of neighbor proxies toproactively cache information requests based on the tradeoffbetween cache cost and delay. We have presented analyti-cal and simulation investigations that quantify SNC’s gainscompared to full proactive caching and when no caching isperformed.

Ongoing work includes investigating how the network topol-ogy, traffic demand, information item size, disconnection pe-riod duration, and in-network caching influence SNC’s per-formance, and how the cache cost can be adapted on linefor scenarios with a different number of mobiles and prox-ies, proxy transition probabilities, and cache sizes. Also, weare investigating the case of multiple levels of proxies. Whenthere are proxies at different levels, the delay cost for a proxyat a higher level is higher. On the other hand, a proxy at ahigher level can potentially serve a larger number of mobileattachment points, hence using proxies at a higher level can

65

70%

80%

60%

70%

40%

50%

n

30%

ay gain

Full Caching, N=8, Cmiss = 4 x ChitNo Caching N=8 Cmiss = 4 x Chit

10%

20%

Dela No Caching, N=8, Cmiss = 4 x Chit

Full Caching, N=8, Cmiss = 10 x ChitNo Caching, N=8, Cmiss = 10 x Chit

10%

0%0 250 500 750 1000 1250 1500 1750 2000

‐20%

‐10%

(a) Cmiss/Chit = 4, 10

70%

80%

60%

70%

40%

50%

30%

ay gain

Full Caching, N=8, Cmiss = variableNo Caching N=8 Cmiss = variable

10%

20%

Dela No Caching, N=8, Cmiss = variable

Full Caching, N=5, Cmiss = variableNo Caching, N=5, Cmiss = variable

10%

0%0 250 500 750 1000 1250 1500 1750 2000

‐20%

‐10%

(b) 5 and 8 destination proxies

Figure 5: Delay gains based on simulation. In the left figure,Cmiss/Chit is randomly selected in the range [4, 5].

reduce the number of proxies that need to proactively cacheinformation, hence the total cache cost.

6. ACKNOWLEDGMENTThe work in this paper was supported by the FP7 ICT

project PURSUIT, under contract ICT-2010-257217.

7. REFERENCES[1] I. Burcea, H. Jacobsen, E. Lara, V. Muthusamy, and

M. Petrovic. Disconnected Operation inPublish/Subscribe Middleware. In Proc. of IEEE Int’lConference on Mobile Data Management (MDM),2004.

[2] M. Caporuscio, A. Carzaniga, and A. L. Wolf. Designand Evaluation of a Support Service for MobilePublish/Subscribe Applications. IEEE Transactionson Software Engineering, 29(12):1059–1071, December2003.

[3] M.-H. Chiu and M. A. Bassiouni. Predictive Schemesfor Handoff Prioritization in Cellular Networks Basedon Mobile Positioning. IEEE Journal on SelectedAreas in Communications, 18(3):510–522, March 2000.

[4] U. Farooq, E. Parsons, and S. Majumdar. Performanceof Publish/Subscribe Middleware in Mobile WirelessNetworks. In Proc. of International Workshop onSoftware and Performance (WOSP), 2004.

[5] L. Fiege, F. C. Gartner, O. Kasten, and A. Zeidler.Supporting mobility in content-basedpublish/subscribe middleware. In Proc. of Middleware2003, LNCS, volume 2672, 2003.

[6] N. Fotiou, P. Nikander, D. Trossen, and G. Polyzos.Developing information networking further: Frompsirp to pursuit. In Proc. 7th ICST Int’l Conf. onBroadband Commun., Networks, and Systems, 2010.

[7] A. Gaddah and T. Kunz. Extending mobility topublish/subscribe systems using a pro-active cachingapproach. Mobile Information Systems, 6(4):293–324,2010.

[8] V. Jacobson, D. K. Smetters, J. D. Thornton,M. Plass, N. Briggs, and R. Braynard. NetworkingNamed Content. In Proc. of CoNEXT, 2009.

[9] S. Pack, H. Jung, T. Kwon, and Y. Choi. SNC: ASelective Neighbor Caching Scheme for Fast Handoffin IEEE 802.11 Wireless Networks. ACM MobileComputing and Communications Review, 9(4):39–49,October 2005.

[10] V. Siris, X. Vasilakos, and G. Polyzos. A SelectiveNeighbor Caching Approach for Supporting Mobilityin Publish/Subscribe Networks. In Proc. of ERCIMWorkshop on eMobility. Held in conjunction withWWIC, 2011.

[11] W.-S. Soh and H. S. Kim. QoS provisioning in cellularnetworks based on mobility prediction techniques.IEEE Communications Magazine, 41(1):86–92,January 2003.

[12] V. Sourlas, G. S. Paschos, P. Flegkas, andL. Tassiulas. Mobility support through caching incontent-based publish/subscribe networks. In Proc. ofIEEE/ACM Int’l Conf. on Cluster, Cloud and GridComputing, 2010.

[13] J. Wang, J. Cao, J. Li, and J. Wu. MHH: A NovelProtocol for Mobility Management inPublish/Subscribe Systems. In Proc. of IEEE Int’lConference on Parallel Processing (ICPP), 2007.

66