publish-subscribe in mobile information centric networks ...€¦ · keywords: internet of things,...

Publish-subscribe in mobile information centric

networks: modeling and performance evaluation

Agnese V. Ventrella, Giuseppe Piro, L. Alfredo Grieco

Dep. of Electrical and Information Engineering (DEI), Politecnico di Bari, v. Orabona4, 70125, Bari, Italy; e-mail: {name.surname}@poliba.it.

Abstract

Emerging Internet of Things (IoT) services require efficient content dissemi-nation mechanisms based on the publish-subscribe model in static and mobilescenarios. The Information-Centric Networking (ICN) architecture can suc-cessfully satisfy these requirements. In its native formulation, ICN can fulfillpublish-subscribe data dissemination and natively supports mobile applica-tions. At the time of this writing, several ICN-based solutions have beenproposed to implement the publish-subscribe model, but none of them isexplicitly tailored to mobile scenarios. To bridge this gap, the present con-tribution: (i) formulates new pull-based and push-based publish-subscribecommunication schema, able to support user mobility in ICN networks; (ii)provides analytical models describing the communication overhead they in-cur, (iii) investigates their accuracy through computer simulations. Theconducted study considers well-known benchmark network topologies, realIoT monitoring services, and standardized settings for urban and rural en-vironments. From one side, obtained results validate the conceived analyt-ical models. From another side, they highlight pros and cons of pull-basedand push-based approaches by emphasizing the conditions under which onescheme should be preferred to the other one.

Keywords: Internet of Things, Information-Centric Networking,publish-subscribe, analytical models, computer simulations

1. Introduction

Latest reports from key stakeholders of the telecommunication marketforecast that up to 100 billion of connected things are expected to join the

Preprint submitted to Computer Networks (Elsevier) August 29, 2017

Internet of Things (IoT) by 2020 [1] [2] and several novel applications areemerging in smart health, smart transportation, smart metering, smart city,and smart building domains [3]-[6].

Most of IoT applications share three main characteristics: first, servicesare inherently information-centric [7]-[13]. In fact, consumers are more in-terested in retrieving contents generated by IoT devices rather than commu-nicating with a specific node [7]. Second, IoT contents may be, at the sametime, real-time and event-triggered. Therefore, to avoid to continuously polla given data producer, contents’ updates should be disseminated by using thepublish-subscribe communication paradigm. Accordingly, producers publishinformation objects under a specific topic; consumers express their interestby issuing subscriptions; when a new information object is generated, it issent to the list of subscribers [14]-[16]. In between producers and consumers,a scalable data dissemination architecture is required to match publishedtopics with subscriptions and send notifications. In addition, consumers andproducers may be mobile [17]-[19].

At the time of this writing, the Information-Centric Networking (ICN)paradigm, widely considered as a key building block for the Future Inter-net [20]-[22], has all the potentials to enable IoT services by encompassinginformation-centric, publish-subscribe, and mobility aspects. ICN, in fact,natively provides an explicit separation between contents and locators, iden-tifies data with names (that do not contain any details about the location),handles data request and dissemination through routing-by-name strategies,and allows a simplified management of user mobility [23][24]. Each of theseimportant features were already applied and investigated in IoT scenarios[7]-[26].

Several publish-subscribe approaches for the current Internet architecturewere already proposed in the literature. Some of them, like [27]-[30], also fo-cus on mobile wireless networks. However, built on top of the host-centriccommunication paradigm, these mechanisms cannot be directly applied toICN deployments. Technical solutions to implement the publish-subscribemodel in ICN, instead, are presented in [31]-[38]. In general, they lever-age two baseline approaches (see section 2 for more details): pull-based andpush-based [39]. In the first case, the consumer continuously polls the pro-ducer to retrieve the new version of the content of interest. In the secondcase, the consumer establishes a stable communication with the producer andcontents’ updates are delivered without requiring continuous solicits. Unfor-tunately, except for a preliminary assessment discussed in [38], none of these

2

contributions consider mobility.This goal could be accomplished by introducing dedicated control mes-

sages that maintain the (multi-hop) communication path between consumersand producers. Anyway, in mobile scenarios, when either the consumer orthe producer switch the access network, it is necessary to activate a newnetwork path to sustain data exchange. The older paths (herein referred toas stale paths) are de-activated after some transient according to the logicof network architecture. Stale paths could be partially overlapped with re-spect to the new one. Therefore, when a new content is generated, it will bedelivered to the consumer through both new and stale paths, thus incurringa non-negligible communication overhead.

These considerations have been discussed (but only in a very preliminaryformulation) in [40] by the same authors of this work. In [40], however, onlythe mobility of the data consumer is investigated and no analytical modelsare provided. This paper, instead, significantly extends that previous contri-bution by deepening the understanding of ICN publish-subscribe protocolsand their interplay with node mobility. In summary, it offers the followingcontributions:

1. baseline pull-based and push-based approaches available in literatureare extended to support the mobility of consumers and producers;

2. the communication overhead incurred by mobility support is evaluatedthrough analytical models;

3. the analytical models are validated through computer simulations inrealistic scenarios;

4. pros and cons of pull-based and push-based approaches are discussedin all the investigated scenarios.

In particular, the average communication overhead is investigated in ur-ban and rural scenarios, where node density and user speed are set accordingto [41], the ICN network topology is generated by Boston university Rep-resentative Internet Topology gEnerator (BRITE) [42], and many real IoTservices (including home security system, health sensors, smart meter, trafficsensors) are taken into account.

Simulation results, obtained trough an ad-hoc simulator written in C++,clearly demonstrate that the average communication overhead increases with

3

the number of nodes belonging to the network topology (specifically, urbanscenarios reach the higher communication overheads). Moreover, in a sce-nario with mobile consumer and stationary producer, both pull-based andpush-based approaches experience higher communication overheads when theuser speed increases. Same considerations can be done in scenarios whereboth consumer and producer are mobile. Differently from the previous case,however, a lower communication overhead is registered. On the contrary,in a scenario with stationary consumer and mobile producer, such kind ofbehavior is only registered for the pull-based approach. With respect to theinvestigated IoT services, the analysis demonstrates that the communicationoverhead always increases with the content generation rate, except for thecase where the push-based approach is adopted in a scenario with stationaryconsumer and mobile producer (in which all the IoT applications reach thesame communication overhead). Finally, the study highlights that the pull-based approach always guarantees the lowest communication overhead. Toprovide a further insight, the accuracy of the conceived analytical models isevaluated by calculating the absolute relative error. Resulting values, alwaysless than 10%, demonstrate the good fit of conceived models.

The rest of the paper is organized as follows: section 2 presents anoverview of the state of the art, with particular reference to ICN and publish-subscribe communication schema. section 3 provides related analytical mod-els that catch the overhead incurred by publish-subscribe scheme in mobileICN deployments. Simulations and theoretical results are presented andcompared in section 4. Finally, section 5 provides closing remarks and drawsfuture research activities.

2. State of the art

Nowadays, Internet has to support information dissemination and re-trieval, devices mobility, a scalable management of storage and bandwidth,and security issues [7]. To meet these requirements, the scientific communitystarted working on the definition of the so-called Future Internet, based on adata-centric architecture [20][21]. In this context, the Information-CentricNetworking (ICN) paradigm [20] emerged as one of the most promisingapproach to concretely accomplish this revolution. Several ICN architec-tures have been already designed. They are: Data-Oriented Network Ar-chitecture (DONA) [43], Publish Subscribe Internet Technology (PURSUIT)[36], MobilityFirst [44], CONVERGENCE [45] and Named-Data Network-

4

ing (NDN) [46]. Without lack of generality, this contribution is grounded onthe NDN architecture, but its findings are quite general to be extended alsoto different architectures. To face IoT traffic, two primitives can be used:request-response and publish-subscribe. NDN natively supports only therequest-response communication scheme. Thus, extensions to the publish-subscribe are required.

In the following, NDN architecture and publish-subscribe approaches pro-posed in the literature are presented.

2.1. NDN

NDN [46] [20] natively supports a receiver-driven communication model.Only two types of messages are admitted: Interest and Data. A user mayask for a content by issuing an Interest which is routed towards the nodes inposses of the required information. Then, these nodes are triggered to replywith Data packets that will follow the way back to the user.

Moreover, each network node implements networking functionalities bymeans of three main data structures: Content Store (CS), Forwarding In-formation Base (FIB), and Pending Interest Table (PIT). CS represents acache memory where received/forwarded contents can be stored. FIB is sim-ilar to the one used for IP, except for the fact that its entries do not containIP addresses with a fixed length, but hierarchical content names with vari-able lengths. Finally, the PIT table is used to keep track of Interest packetspreviously forwarded in upstream, and that still remain unsatisfied: in thisway, Data messages can be sent back to the users by following a breadcrumbrouting that traces the same path opened by Interest messages but in theopposite direction. It is worth to note that Interest packets for the samecontent can be aggregated at each node into one PIT entry, thus enablingnative support to multicast communications [46]. Moreover, a PIT entry isremoved in two cases: when the corresponding Data is received or when thetimeout expires.

2.2. Publish-Subscribe in NDN

Publish-subscribe scheme is based on the following principle: the con-sumer issues a subscription request for a given topic. Then, every time anew content is generated under that topic, the producer (or, more in general,the network) delivers that Data to all the subscribed users [14].

In NDN, the implementation of publish-subscribe communication is stillan open issue. Nevertheless, some solutions were proposed and implemented

5

in the literature. In summary, they can be grouped in two main categories,pull-based and push-based approaches [39]. In the first case, the consumerpolls the producer to retrieve the next update. In the second case, the con-sumer establishes a semi-persistent communication path with the producerthat sends update without being solicited.

For instance, [31] presents a pull-based approach. It is assumed that theconsumer issues a window of W Interest packets, asking for W consecutiveversions of contents identified with the same topic-name. These requests aredelivered towards one or more potential sources according to FIBs entries,configured in intermediate routers. When the publisher produces the con-tents, they are sent back to the consumer. Moreover, as soon as a Datapacket is received, the consumer releases a new Interest packet. It is im-portant to note that this approach perfectly matches baseline networkingprimitives already designed for NDN. As a consequence, no further modifi-cations are required at the network level, because the subscription processis intrinsically managed by the Interest generation process at the consumerside.

Some push-based approaches are summarized in [8] and deeply investi-gated in [34], [33] and [32]. Differently from the previous approach, all ofthese solutions require some enhancements to the baseline NDN communi-cation primitives. [33] suggests to include contents within Interest messagessent by the producer towards subscribed consumers. This mechanism is alsoapplied in [32], but a dummy Data packet for acknowledgement is intro-duced in order to preserve the 1-to-1 matching between Interest and Data.An interesting approach comes from [34][32], where semi-permanent Interestpackets are used. According to NDN specifications, these kind of Interestpackets are stored in PIT entries of intermediate routers. However, they arenot deleted if one or more contents are sent back to the consumer. Instead,they remain in PIT tables till the expiration of a timeout. Subscriptionsare periodically refreshed by issuing new semi-permanent Interest packetsrelated to the considered topic.

Push-based publish-subscribe schema can also be implemented by adopt-ing a new set of messages and architectural elements [37]. For instance,[35] presents the COPSS (Content-Oriented Pub/Sub System) architecture,which introduces a new network element (i.e., the Rendezvous Point), twonew messages (i.e., subscribe and publish), and a new table in intermedi-ate routers (i.e., Subscription Table). To make a subscription, the consumersends a subscribe message to a given Rendezvous Point. Multiple Rendezvous

6

Points can be deployed in the network, for handling subscriptions related todifferent groups of topics. Subscribe messages are routed to the most suitableRendezvous Point, according to information stored in FIBs of intermediaterouters. During the delivery of subscribe messages, intermediate routers alsopopulates their Subscription Table. This table is used to track this new con-trol message in the same way the Pending Interest Table handles Interestmessages. When a new content is generated, the producer sends it within apublish message to the reference Rendezvous Point. The received messageis forwarded to the subscribers, based on information stored in SubscriptionTables. The COPSS architecture was applied to disaster scenarios [47] andvehicular networks [38].

Among all the contributions mentioned before, only [38] addresses mobil-ity. To this end, to support the consumer’s mobility, the subscription requestis periodically renewed; to support the producer’s mobility, the subscriptionrequest is sent again if no Data are received for a specific amount of time.

To summarize, a big picture of works investigated in this subsection ispresented in Table 1, highlighting design principle (pull or push), the need ofan extension of NDN communication primitives, the support for the mobility,as well as analytical models provided. From this table, it clearly emergesthat a comprehensive study of publish-subscribe approaches for NDN, whichdeeply investigates the impact of the mobility, is still missing in the currentliterature. Indeed, as remarked in the latest column of Table 1, this lack isgoing to be addressed in this paper.

3. On designing and modeling of publish-subscribe schema in ICN

The reference architecture considered in this contribution is depicted inFigure 1. An overlay NDN network1 is configured on top of network at-tachment points [7][48][49]. Without loss of generality, it is assumed thatproducer and consumer are not connected to the same network attachmentpoint. Therefore, Data exchange requires the set up of a multi-hop commu-nication path. Also, producer and consumer could be mobile: while theychange the network attachment point, the multi-hop communication pathsshould be frequently re-configured. The present contribution assumes that

1The same considerations apply for ICN realms that embed NDN functionalities at thenetwork layer.

7

Table 1: State of the art solutions for publish-subscribe.

[31] [8] [34] [33] [32] [35] [38] Present work

Pull-based yes no no no no no no yesPush-based no yes yes yes yes yes yes yesRequiresextensionsto NDN

no yes yes yes yes yes yes yes (push-based)

Supportsmobility

no no no no no no yes yes

Providesanalyticalmodel

no no no no yes no no yes

such a multi-hop communication path represents the shortest path connect-ing consumer and producer available within a given network topology. With-out loss of generality, the way this path is configured is neglected. Specificmechanisms based on the Software-Defined Networking paradigm could beused for this purpose [50]. But, any related technical detail remains out ofscope for this work.

Contents are disseminated according to the publish-subscribe scheme.The producer generates contents for a given topic, namely ndn://[topic-name]. Let TD be the average time interval between the generation of twoconsecutive contents, belonging to the same topic. Then, each content isuniquely identified with an incremental identifier. The resulting name ap-pears in the form:

ndn://[topic-name]/#id.

Based on these premises, this section (i) provides a formal definition ofthe communication overhead, (ii) describes some enhancements to baselinepull-based and push-based approaches required to support the mobility inNDN networks, and (iii) proposes analytical models able to quantify com-munication overheads due to mobility management.

To ease the comprehension of the notions presented in the following, asummary of main symbols is reported in Table 2.

8

Table 2: List of mathematical symbols

Symbol Description

TD Time interval between the generation of two consecutive con-tents

ti Time instant in which a communicating entity (i.e., consumeror producer) attaches to the generic ith network node

∆tC Average cell residence time of the consumer

∆tP Average cell residence time of the producer

Oi,i+1I Communication overhead due to control messages during

∆ti,i+1

Oi,i+1D Communication overhead due to data dissemination during

∆ti,i+1

O Average communication overheadW Window size, adopted in the pull-based approachTRTT Round Trip TimeTIL Interest lifetime

T pullO Pending request timeout, adopted in the pull-based approachdi Shortest path between consumer and producer, established at

tiA(t) Stale disjoint links in a specific time instantSI Interest message sizeSD Data message size

T pushO Pending request timeout, adopted in the push-based approachRD Data packets generation ratev Average speed of the mobile consumer or mobile producerr Cell radiusN Total number of nodes in the topology

9

DataConsumer

Multi-hop communication pathData

Producer

Figure 1: Reference scenario.

3.1. Formal definition of the communication overhead

As anticipated in the Introduction, the implementation of publish-subscribemechanisms in mobile NDN deployments can generate a non negligible over-head. NDN natively relies on a receiver-driven communication paradigm:a content can be disseminated only in response to an Interest. Note thatInterest packets cover a key role also in the publish-subscribe mechanism. Infact, in the current literature (see section 2), they act as control messagesneeded to establish and maintain the multi-hop communication path betweenconsumers and producers. Therefore, they represent the first contribution tothe overhead.

A given communication path may remain active until entries stored inPIT tables of intermediate NDN routers expire. When a consumer switchesthe network attachment point, a new path is created. Path established in thepast and not yet expired are simply referred to as stale paths. Specifically,routers belonging to stale paths store, for a given transient and in theirPIT tables, the request previously sent by the user. Therefore, as soon asthe producer generates a new content, it will be disseminated along all theactive paths, including the stale ones. Note that stale paths could be partiallyoverlapped with respect to the latest generated one. Hereby, the concept ofstale disjoint links is introduced to identify the set of links that belong to

10

stale paths and that are not present in the new one. As a consequence, Datapackets transmitted through stale disjoint links inflates the overhead.

To provide a further insight, Figure 2 qualitatively explains the contribu-tion to the communication overhead due to both Interest and Data packets.The first example refers to the scenario where the consumer is mobile andthe producer is stationary. In this case, the consumer sends Interest packetsto the producer while it is moving across network attachment points. All ofthese messages provide the first contribution to the communication overhead.When the producer generates a new content, it is delivered across the latestpath established between consumer and producer and towards the two stalepaths, that in the depicted example are still active. Thus, Data packets sentthrough stale disjoint links give the second contribution to the communica-tion overhead. In the second example, the consumer is stationary and theproducer is mobile. Differently from the previous case, now the producermay deliver Data packet only through the latest path established with theconsumer. As a result, the communication overhead can only be generatedby the exchange of Interest packets. The scenario considering both consumerand producer mobility embraces all the events described for the first two ex-amples. Here, the communication overhead is generated by the exchangeof Interest packets (sent from the consumer to the producer) and by Datapackets (sent through stale disjoint links).

Now, to provide a formal definition of the communication overhead, it isimportant to introduce the following details. Let ti be the time instant inwhich a communicating entity (i.e., a consumer or a producer) attaches to thegeneric ith network attachment point. Then, the cell residence time, ∆ti,i+1,represents the amount of time in which the consumer or the producer remainsconnected to the ith network attachment point before jumping to the nextone. Moreover, let Oi,i+1

I be the total amount of bits due to Interest packetstransmitted, during ∆ti,i+1, across the links belonging to multi-hop pathsconnecting consumer and producer. Also, let Oi,i+1

D be the total amount ofbits due to Data packet transmitted across the stale disjoint links, during∆ti,i+1. The average communication overhead that can be produced by apublish-subscribe communication scheme during a unit of time, i.e., O, canbe formally defined as:

O =E[Oi,i+1

I ] + E[Oi,i+1D ]

E[∆ti,i+1], (1)

where the E[x] operator returns the expectation of the random variable x.

11

Fixed Producer Mobile

Consumer

Data packetData packet (overhead)

Interest packet (overhead) Router belonging to the stale pathsRouter belonging to the current path

(a)

Mobile Producer

FixedConsumer

(b)

Figure 2: Examples showing the communication overhead in a scenario with (a) stationaryproducer and mobile consumer and with (b) mobile producer and stationary consumer.

12

Closed-form expressions for the average communication overhead related toboth pull-based and push-based approaches in different mobility scenariosare formulated in the following subsections. Nevertheless, a summary ofconceived models is reported in Table 3.

Table 3: Summary of analytical models

Stationary producer, mobile consumerPull-based Push-based

E[Oi,i+1I ] dSI

¯∆tC

(1 + ∆t

TD

)dSI

1

1−e1¯

∆tCTpushO

E[Oi,i+1D ] ASD ARD

(T pushO

1−e1¯

∆tCTpushO

− ∆tC)

O dSI¯∆tC

(1 +

¯∆tC

TD

)+ ASD

¯∆tC

dSI1

1−e

1¯

∆tCTpushO

+ARD

TpushO

1−e

1¯

∆tCTpushO

− ¯∆tC

¯∆tC

Mobile producer, stationary consumerPull-based Push-based

E[Oi,i+1I ] dSI

¯∆tP

TDdSI

¯∆tP

T pushO

E[Oi,i+1D ] 0 0

O dSI1TD

dSI1

T pushO

Mobile producer, mobile consumerPull-based Push-based

E[Oi,i+1I ] dSI

¯∆tC

(1 +

¯∆tC

TD

)dSI

1

1−e1¯

∆tCTpushO

E[Oi,i+1D ] ASD ARD

(T pushO

1−e1¯

∆tCTpushO

− ∆tC)

O dSI¯∆tC

(1 +

¯∆tC

TD

)+ ASD

¯∆tC

dSI1

1−e

1¯

∆tCTpushO

+ARD

TpushO

1−e

1¯

∆tCTpushO

− ¯∆tC

¯∆tC

3.2. Pull-based approach

The pull-based publish-subscribe communication scheme was initially de-scribed in [31]. To support the mobility, it should be extended as describedbelow:

13

• the consumer issues contents’ requests (i.e., Interest packets) accordingto a sliding window mechanism: a burst of W requests is sent at thebeginning; then, every time a new content is received, a new request isreleased;

• to optimize the flow control, the window size should be properly con-figured as a function of both Round Trip Time (RTT) and TD:

W =

⌈TRTT

TD

⌉, (2)

where the dxe operators returns the nearest integer greater than orequal to x. In general, the flow control acts as a receiver-driven Go-Back-N automatic repeat-request (ARQ). Nevertheless, the most of

IoT applications generate data sporadically and TRTT

TD< 1 (see section

4 for more details). For this reason, we can simplify Eq. (2) by settingW = 1. Therefore, the flow control mechanism becomes a receiver-driven stop-and-wait ARQ;

• in NDN, the Interest lifetime represents the timeout used by networknodes to remove stale Interests from the PITs. In order to avoid thatPIT entries are deleted before the reception of a Data packet, the In-terest lifetime (TIL) is set as in the following:

TIL ≥ TRTT + (TDW − TRTT ). (3)

Now, considering that in our case W = 1, Eq. (3) can be simplyrewritten as TIL = TD.

• the consumer assigns a timeout T pullO to each pending request. When

the timeout expires, the same request is sent (again) to the producer.In order to reduce the polling frequency, the pending request timeoutT pullO is set equal to TIL. Indeed, is it possible to set T pull

O = TIL = TD.The polling strategy natively permits the establishment of a new multi-hop communication path in the case of producer mobility. For the sakeof simplicity, we assume that the update of FIBs tables in intermediateNDN routers is a task of the network [24][51], thus becoming out-of-scope of the publish-subscribe approach;

14

• if the consumer changes the network attachment point, the latest Wrequests not yet satisfied are sent again to the producer. In this way,a new path between consumer and producer is created. According to[52] and [53], this strategy supports consumer mobility.

3.2.1. First use case: stationary producer, mobile consumer.

Theorem 1. Let d, SI , ∆tC, TD, A, SD be the average shortest path, theInterest packet size, the average cell residence time of the consumer, thetime interval between the generation of two consecutive contents, the averagenumber of stale disjoint links, and the Data packet size, respectively. Theaverage communication overhead generated by the pull-based approach in thecase the producer is stationary and the consumer is mobile is equal to:

O =dSI

∆tC

(1 +

∆tC

TD

)+ASD

∆tC. (4)

Proof. At the beginning of a given cell residence time ∆tCi,i+1, the mobileconsumer polls the stationary producer by sending W = 1 Interest packet.Then, a new Interest packet is issued when a new content is received or thetimeout, TD, assigned to the pending request, expires. This is done till ti+1.During ∆tCi,i+1, the average number of Interest packets sent by the consumer

is approximately equal to∆tCi,i+1

TD. Therefore, the resulting communication

overhead due to control messages is equal to:

Oi,i+1I = diSI +

∆tCi,i+1

TDdiSI , (5)

where di and SI are the shortest path between consumer and producer, es-tablished at ti, and the Interest packet size, respectively.

Let A(t) be the number of stale disjoint links, active when the producergenerates a new content. Accordingly, the communication overhead due toextra data dissemination (i.e., Data packets sent across stale disjoint links)can be expressed as:

Oi,i+1D = A(t)SD, (6)

where SD represents the size of the Data packet.

15

Now, by substituting Eq. (5) and Eq. (6) in Eq. (1), the average overheadgenerated by the pull-based approach in the case the consumer is mobile andthe producer is stationary becomes:

O =E[diSI +

∆tCi,i+1

TDdiSI

]+ E [A(t)SD]

E[∆tCi,i+1]

=E[di]SI

E[∆tCi,i+1]+ E[di]

SI

TD+E[A(t)]SD

E[∆tCi,i+1]

(7)

By setting E[di] = d, E[∆tCi,i+1] = ∆tC , and E[A(t)] = A, Eq. (7) can berewritten as:

O =dSI

∆tC

+ dSI

TD+ASD

∆tC

=dSI

∆tC

(1 +

∆tC

TD

)+ASD

∆tC. (8)

3.2.2. Second use case: mobile producer, stationary consumer.

Theorem 2. Let d, SI , and TD be the average shortest path, the Interestpacket size, and the time interval between the generation of two consecutivecontents, respectively. The average communication overhead generated by thepull-based approach in the case the consumer is stationary and the produceris mobile is equal to:

O = dSI

TD. (9)

Proof. Since the consumer does not modify its network attachment point, anInterest packet is periodically sent every TD. Indeed:

Oi,i+1I =

∆tPi,i+1

TDdiSI . (10)

Moreover, stale paths are never generated, because the producer maydeliver contents it generates only through the latest path established withthe consumer. Therefore, the following Equation holds:

Oi,i+1D = 0. (11)

16

By substituting Eq. (10) and Eq. (11) in Eq. (1), the average overheadgenerated by the pull-based approach in the case the producer is mobile andthe consumer is stationary becomes:

O =E[

∆tPi,i+1

TDdiSI

]E[∆tPi,i+1]

=

E[∆tPi,i+1]TD

E[di]SI

E[∆tPi,i+1]= E[di]

SI

TD(12)

By setting E[di] = d, it can be rewritten as:

O = dSI

TD. (13)

3.2.3. Third use case: mobile producer, mobile consumer.

Theorem 3. Let d, SI , ∆tC, and TD be the average shortest path, the Interestpacket size, the average cell residence time of the consumer and the timeinterval between the generation of two consecutive contents, respectively. Theaverage communication overhead generated by the pull-based approach in thecase both consumer and producer are mobile is equal to:

O =dSI

∆tC

(1 +

∆tC

TD

)+ASD

∆tC. (14)

Proof. At the beginning of a given cell residence time ∆tCi,i+1, the mobileconsumer polls the stationary producer by sending W = 1 Interest packet.Then, it sends a new Interest packet after receiving a new content or thetimeout, TD, assigned to the pending request expires.

During ∆tCi,i+1, the average number of Interest packets sent by the con-

sumer is approximately equal to∆tCi,i+1

TD. Therefore, the resulting communi-

cation overhead due to control messages is equal to:

Oi,i+1I = diSI +

∆tCi,i+1

TDdiSI , (15)

where di and SI are the shortest path between consumer and producer, es-tablished at ti, and the Interest packet size, respectively.

17

Let A(t) be the number of stale disjoint links, active when the producergenerates a new content. Accordingly, the communication overhead due toData packets sent across stale disjoint links can be expressed as:

Oi,i+1D = A(t)SD, (16)

where SD is the size of the Data packet.It is important to note that, differently from the first use case, the vari-

able A(t) is influenced by both consumer mobility and producer mobility.Specifically, when the consumer changes the network attachment point, anew multi-hop communication path is established and some of the links be-longing to the old one are included in A(t). On the contrary, as soon theproducer changes its network attachment point, A(t) becomes an empty set.In this case, in fact, all the stale disjoint links will do not produce commu-nication overhead anymore: since the producer is not attached to previousstale paths, next Data packets will be sent only through the new establishedpath.

Now, by substituting Eq. (15) and Eq. (16) in Eq. (1), the averageoverhead generated by the pull-based approach in the case both consumerand producer are mobile becomes:

O =E[diSI +

∆tCi,i+1

TDdiSI

]+ E [A(t)SD]

E[∆tCi,i+1]

=E[di]SI

E[∆tCi,i+1]+ E[di]

SI

TD+E[A(t)]SD

E[∆tCi,i+1]

(17)

By setting E[di] = d, E[∆tCi,i+1] = ∆tC , and E[A(t)] = A, Eq. (17) canbe rewritten as:

O =dSI

∆tC+ d

SI

TD+ASD

∆tC=dSI

∆tC

(1 +

∆tC

TD

)+ASD

∆tC. (18)

3.3. Push-based approach

According to the standard NDN implementation, the PIT entry associ-ated to a request is deleted as soon as a new Data packet is sent back to

18

the consumer. Then, any information related to the communication path be-tween consumer and producer is definitively lost. The push-based approachextends the normal behavior of NDN communication primitives by intro-ducing semi-persistent Interest packets [32]. Specifically, a semi-persistentInterest packet is not erased from the PIT table even if one or more cor-responding Data packets were already delivered to the consumer. But, itremains in the PIT table until its timeout expires.

To support the mobility, the push-based approach should be implementedas follows:

• the consumer releases a single request through a semi-permanent In-terest packet;

• the consumer assigns a timeout, T pushO , to the pending request. When

that timeout expires, the same request is sent (again) to the producer.T pushO has been set in order to have only one stale path during ∆ti,i+1, as

discussed below. Note that the polling strategy permits the establish-ment of a new multi-hop communication path in the case of producermobility;

• if the consumer changes the network attachment point, a semi-permanentInterest is sent to the producer. In this way, a new path between con-sumer and producer is created. Thus, the consumer mobility is sup-ported too.

3.3.1. First use case: stationary producer, mobile consumer.

Theorem 4. Let d, SI , ∆tC, RD, and A be the average shortest path, theInterest packet size, the average cell residence time of the consumer, the con-tents’ generation rate, and the average set of stale disjoint links, respectively.Moreover, let the pending request timeout, T push

O , be set in order to have onlyone stale path during ∆tCi,i+1. The average communication overhead gener-ated by the push-based approach in the case the producer is stationary andthe consumer is mobile is:

O =

dSI1

1−e1¯

∆tCTpushO

+ ARD

(T pushO

1−e1¯

∆tCTpushO

− ∆tC)

∆tC. (19)

Proof. At the beginning of a given cell residence time ∆tCi,i+1, the mobile con-sumer polls the stationary producer by sending an Interest packet. When

19

the timeout, T pushO , expires, the same request is sent (again) to the producer.

This is done till ti+1. Besides the first request, the additional number of In-

terest packets sent by the consumer in ∆tCi,i+1 is equal to⌊

∆tCi,i+1

T pushO

⌋. Therefore,

the communication overhead due to control messages, evaluated during thetime interval ∆tCi,i+1, can be expressed as:

Oi,i+1I = diSI + diSI

⌊∆tCi,i+1

T pushO

⌋. (20)

The adoption of semi-permanent Interest implies that a stale path mayexist for a long period of time. Also, the number of stale disjoint links maychange during the cell residence time. The communication overhead dueto transmission of Data packets across stale disjoint links could be formallyexpressed as a function of ∆tCi,i+1, A(t), the size of each Data packet SD, andcontents’ generation rate TD:

Oi,i+1D = F

(∆tCi,i+1,A(t), SD, TD

). (21)

From Eq. (21) it emerges that the estimation of the contribution to theoverhead due to extra data dissemination is difficult to achieve. Nevertheless,under some specific assumptions on T push

O , a valid approximation exists.In this contribution, it is assumed that T push

O is set in order to have onlyone stale path during ∆tCi,i+1. According to [54] and [55], the cell residencetime can be modeled through an exponential distribution with parameter1/∆tC > 0:

f∆tC (t) =1

∆tCe− 1

¯∆tC

t. (22)

Thus, in order to have only one stale path (which represents our designassumption), the relation:

T pushO < ∆tCi,i+1 + ∆tCi+1,i+2, (23)

should be valid with a high probability (i.e., > 99%)Let ε be a very small number (i.e., 10−2), Eq. (23) is satisfied if:{P (T push

O < ∆tCi,i+1 + ∆tCi+1,i+2) = 1− P (∆tCi,i+1 + ∆tCi+1,i+2 < T pushO )

P (∆tCi,i+1 + ∆tCi+1,i+2 < T pushO ) < ε.

(24)

20

Since the sum of two random variables, characterized by exponential dis-tribution and same average value, is a Gamma distribution2, the secondequation of (24) can be rewritten as:

P (∆tCi,i+1 + ∆tCi+1,i+2 < T pushO ) =

∫ T pushO

0

(1

∆tC

)2

te− 1

¯∆tC

tdt < ε, (25)

that, by integrating by part, becomes:

1− e−1¯

∆tCT pushO

(1

∆tCT pushO + 1

)< ε. (26)

Indeed, the values of T pushO that satisfy the initial assumption can be

calculated by solving Eq. (26), as deeply commented in section 4.2.Under these assumptions, the communication overhead due to the Data

packets sent through the set of stale disjoint links, A(t), during the timeinterval they are still active, that is T push

O − ∆tCi,i+1modTpushO = T push

O −∆tCi,i+1 + T push

O

⌊∆tCi,i+1

T pushO

⌋, is:

Oi,i+1D =

(T pushO −∆ti,i+1 + T push

O

⌊∆tCi,i+1

T pushO

⌋)A(t)RD, (27)

where RD is the contents’ generation rate.By substituting Eq. (20) and Eq. (27) in Eq. (1), the average com-

munication overhead generated by the push-based approach in the case the

2Gamma(

2, 1¯∆tC

)=(

1¯∆tC

)2

te− 1

¯∆tC

t

21

producer is stationary and the consumer is mobile becomes:

O =E[diSI + diSI

⌊∆tCi,i+1

T pushO

⌋]E[∆tCi,i+1]

+E[A(t)RD

(T pushO −∆ti,i+1 + T push

O

⌊∆tCi,i+1

T pushO

⌋)]E[∆tCi,i+1]

=

=dSI

(1 + E

[⌊∆tCi,i+1

T pushO

⌋])∆tC

+ARD

(T pushO − ∆tC + T push

O E[⌊

∆tCi,i+1

T pushO

⌋])∆tC

. (28)

Now, to solve the expected value of a floor function, let a =⌊

∆tCi,i+1

T pushO

⌋with

a ∈ Z and let it solve:

E[a] =inf∑a=0

ap∆tC (a). (29)

From the floor function definition, it results that a ≤ ∆tCi,i+1

T pushO

< a + 1, or

aT pushO ≤ ∆tCi,i+1 < (a+ 1)T push

O . The probability associated to a is:

p∆tC (a) = Pr[aT pushO ≤ ∆tCi,i+1 < (a+ 1)T push

O ]

=

∫ (a+1)T pushO

aT pushO

f∆t(t)dt =

∫ (a+1)T pushO

aT pushO

1

∆te− 1

¯∆tC

tdt

= e1¯

∆tCT pushO a

(1− e1¯

∆tCT pushO ) (30)

By substituting Eq. (30) in Eq. (29), it results that:

E[a] = (1− e1¯

∆tCT pushO )

inf∑a=0

ae1¯

∆tCT pushO a

= (1− e1¯

∆tCT pushO )

e1¯

∆tCT pushO

(e1¯

∆tCT pushO − 1)2

= − e1¯

∆tCT pushO

e1¯

∆tCT pushO − 1

. (31)

22

To conclude, by substituting Eq.(31) in Eq. (28), the average communi-cation overhead becomes:

O =

dSI

(1− e

1¯

∆tCTpushO

e1¯

∆tCTpushO −1

)+ ARD

(T pushO − ∆tC − T push

Oe

1¯

∆tCTpushO

e1¯

∆tCTpushO −1

)∆tC

=

dSI1

1−e1¯

∆tCTpushO

+ ARD

(T pushO

1−e1

∆tCTpushO

− ∆tC)

∆tC(32)

3.3.2. Second use case: mobile producer, stationary consumer.

Theorem 5. Let d, SI , and T pushO , be the average shortest path, the Interest

packet size, and the pending request timeout, respectively. The average com-munication overhead generated by the push-based approach in the case theproducer is mobile and the consumer is stationary is:

O =dSI

T pushO

. (33)

Proof. Since the consumer does not modify its network attachment point, anInterest packet is periodically sent every T push

O . Therefore, the contributionto the communication overhead due to Interest packets can be expressed as:

Oi,i+1I = diSI

∆tPi,i+1

T pushO

. (34)

During the cell residence time ∆tPi,i+1, the producer may deliver contentsonly through the latest multi-hop path established with the consumer. Hence:

Oi,i+1D = 0. (35)

Therefore, by substituting Eq. (34) and Eq. (35) in Eq. (1), the averagecommunication overhead becomes:

O =E[diSI

∆tPi,i+1

T pushO

]E[∆tPi,i+1]

=E[di]SI

E[∆tPi,i+1]T pushO

E[∆tPi,i+1]= E[di]

SI

T pushO

. (36)

By setting E[di] = d, Eq. (36) can be rewritten as:

O = dSI

T pushO

. (37)

23

3.3.3. Third use case: mobile producer, mobile consumer.

Theorem 6. Let d, SI , ∆tC, RD, and A be the average shortest path, theInterest packet size, the average cell residence time of the consumer, the con-tents’ generation rate, and the average set of stale disjoint links, respectively.Moreover, let the pending request timeout, T push

O , be set in order to have up toone stale path during ∆tCi,i+1. The average communication overhead gener-ated by the push-based approach when both producer and consumer are mobileis:

O =

dSI1

1−e1¯

∆tCTpushO

+ ARD

(T pushO

1−e1¯

∆tCTpushO

− ∆tC)

∆tC. (38)

Proof. At the beginning of a given cell residence time ∆tCi,i+1, the mobileconsumer polls the stationary producer by sending an Interest packet. Afterthe expiration of the timeout, T push

O , the same request is sent (again) to theproducer. This is done till ti+1. Besides the first request, the additional num-

ber of Interest packets sent by the consumer in ∆tCi,i+1 is equal to⌊

∆tCi,i+1

T pushO

⌋.

Therefore, the communication overhead due to control messages, evaluatedduring the time interval ∆tCi,i+1, can be formulated as:

Oi,i+1I = diSI + diSI

⌊∆tCi,i+1

T pushO

⌋. (39)

Similarly to the first use case, also in this context the number of stale dis-joint links, A(t), may produce overhead during the time interval, whose upper

bound is set to T pushO −∆tCi,i+1modT

pushO = T push

O −∆tCi,i+1 + T pushO

⌊∆tCi,i+1

T pushO

⌋.

Of course, T pushO is properly set in order to have up to one stale path during

∆tCi,i+1. Thus, the communication overhead generated by Data packet sentthrough stale disjoint links is equal to:

Oi,i+1D =

(T pushO −∆tCi,i+1 + T push

O

⌊∆tCi,i+1

T pushO

⌋)A(t)RD, (40)

where RD is the contents’ generation rate.As expected, the variable A(t) is influenced by both consumer mobility

and producer mobility. Specifically, when the consumer changes the networkattachment point, a new multi-hop communication path is established and

24

some of the links belonging to the old one are included in A(t). On thecontrary, as soon the producer changes its network attachment point, A(t)becomes an empty set. In this case, in fact, all the stale disjoint links willdo not produce communication overhead anymore: since the producer is notattached to previous stale paths, next Data packets will be only sent throughthe new established path.

Indeed, by following the same mathematical procedure already presentedfor the first use case, it is possible to conclude that the average communica-tion overhead is equal to:

O =

dSI

(1− e

1¯

∆tCTpushO

e1¯

∆tCTpushO −1

)+ ARD

(T pushO − ∆tC − T push

Oe

1¯

∆tCTpushO

e1¯

∆tCTpushO −1

)∆tC

=

dSI1

1−e1¯

∆tCTpushO

+ ARD

(T pushO

1−e1¯

∆tCTpushO

− ∆tC)

∆tC(41)

25

4. Results

This section investigates the pull-based and push-based mechanisms de-scribed in section 3 and validates related analytical models through com-puter simulations. To this end, the conducted study considers well-knownbenchmark network topologies, real IoT monitoring services, and standard-ized settings for urban and rural environments.

4.1. Network attachment points and user speed

A number of network attachment points are distributed on a geographicalarea of 106 m2. They offer wireless connectivity to a mobile user in a coveragearea having an average cell radius equal to r. Each attachment point is aNDN node and all of them form an overlay NDN network [7][48][49].

To jointly consider both urban and rural environments, cell radius anduser speed are set as reported in Table 4 [41].

Table 4: Cell radius and user speed for both rural and urban scenarios.

Urban Rural

Cell radius, r [m] 1000 650 300 150 100 50Resulting aver-age node density[nodes/km2]

10000 2500 1111 227 61 25

Average user speed,v [km/h]

3, 30,50

3, 30,50

3, 30,50

30, 50,120

30, 50,120

30, 50,120

4.2. Setting of timers

Many time-related parameters should be properly set.With reference to the average communication overhead related to the

push-based approach, section 3.3.1 already anticipated that the values ofT pushO that satisfy the assumption to have only one stale path during a cell

residence time can be calculated by solving Eq. (26). Now, according to [56],the average cell residence time can be expressed as:

∆tC = ∆tP =πr

2v, (42)

where r and v are the cell radius and the speed of the mobile user (i.e, theconsumer or the producer), respectively.

26

Therefore, by setting ε = 10−2, the resulting T pushO values are shown in

Figure 3. These values are taken into account during computer simulations.

0 200 400 600 800 1000r, [m]

0

10

20

30

40

50

T opush

, [s]

7v = 30 km/h7v = 50 km/h7v = 120 km/h

(a)

0 50 100 150 200r, [m]

0

10

20

30

40

50

T opush

, [s]

7v = 3 km/h7v = 30 km/h7v = 50 km/h

(b)

Figure 3: T pushO values evaluated for the (a) rural scenario and (b) urban scenario.

Finally, typical TRTT values are of the order of millisecond [57]. Therefore,since TRTT � TD, the sliding window size W has been set to 1 in Eq. (2).

4.3. IoT services and packet size

The conducted study considers realistic IoT applications, as described in[58]:

• Smart Meter (SM): electronic device records energy consumption, al-lowing the monitoring of power consumption and billing;

• Home Security System (HSS): outdoor and indoor sensors provide mo-tion detection, alarm system, but also gas, water, heating measurementto monitor home status;

• Traffic Sensor (TS): sensors located on vehicles or along routes monitortraffic conditions, travel speed, traffic anomalies, toll highway informa-tion or pollution.

• Health Sensor (HS): sensors, such as wearable devices, provide biomed-ical parameters (heart rate, blood pressure, and so on) to monitor userwellness;

27

At the application layer, the main simulation settings include (i) theaverage time interval between the generation of two consecutive contents,TD, (ii) the application payload, (iii) the Data packet size, SD, (iv) theaverage content generation rate, RD = SD

TD, (v) and the Interest packet size,

SI . In line with [58], these parameters are set as reported in Table 5. Notethat both SI and SD are computed by considering the application payloadsuggested in [58], the average content name of 40 byte [59] and the typicalstructure of Interest, and Data packets depicted in Figure 4 [60][61].

Table 5: Parameters related to the considered IoT services.Application TD [s] App payload

[Byte]SD [bit] RD [bit/s] SI [bit]

SM 9090 2017 18552 2.04 432HSS 600 20 2576 4.29 432TS 60 1 2424 40.40 432HS 60 128 3440 57.33 432

FIELD BYTENoce 4Scope 1NackType 1InterestLifetime 2

Name 2+namesizeSelectors 2Options 2

FIELD BYTEName 2 +namesizeContent 2 +Application

payloadSignature 1+256

Datapacketformat Interestpacketformat

Figure 4: Data and Interest packet formats for NDN.

4.4. Simulation methodology

As already said, analytical models conceived in section 3 are validatedthrough computer simulations. To this end, two different tools are used:Boston university Representative Internet Topology gEnerator (BRITE) andan ad-hoc simulator.

28

4.4.1. About BRITE

BRITE [42] is used to generate network topologies that, in line withcurrent literature [62], follow the scale-free model. Indeed, it is adoptedto generate the NDN overlay network, as well as to define the distributionof nodes in a pre-defined geographical area and the connection links. Inparticular, the topology is created step-by-step. At each step, only one nodeis added to the network and attached to an existing node, properly chosenthrough a power-law formula [62]. The resulting topology has an averageshortest path, d, approximately equal to [62]: d ≈ logN , where N is the totalnumber of available nodes. Note that d is one of the parameters introducedin section 3 for evaluating the average communication overhead of both pull-based and push-based mechanism. The resulting network topology is thenprocessed by the ad-hoc simulator, as discussed below.

4.4.2. Description of the ad-hoc simulator

The ad-hoc simulator is written in C++ and provides a system-levelmodel of the publish-subscribe communication scheme based on ICN net-working primitives. Conceived as a synchronous simulator, it tracks thestatus of the network and executes some key operations during the time,useful to calculate (at the end) the communication overhead based on a listof parameter settings.

The tool receives in input the network topology generated with BRITE,the specific approach used for implementing the publish-subscribe communi-cation scheme, the speed of both producer and consumer, and the parametersrelated to the application layer.

At the beginning of each simulation, the initial position of both producerand consumer is randomly chosen within the geographical area of 106m2.Indeed, the tool identifies the network attachment point for both producerand consumer according to the position-based handover algorithm: the userattaches to the nearest node belonging to the modeled network topology.

From this moment on, five parallel processes are executed over the time.They include: Mobility Manager, Handover Manager, Paths Manager, Inter-est Generation Process, and Data Generation Process. The algorithm usedfor the pull-based and push-based approach is shown in Figure 5 and Figure6, respectively.

The Mobility Manager periodically updates the position of both producerand consumer. This is done according to the well-known random-walk mo-bility model. Every time a new position is chosen, the tool triggers the exe-

29

cution of functionalities offered by Handover Manager and Paths Manager.The Handover Manager implements the position-based handover algorithm.Thus, it verifies if the current user position requires a new network attach-ment point. In the affermative case, the Paths Manager establishes a newshortest path between consumer and producer and updates the list of disjointlinks still available in the network. Note that disjoint links will be deletedform the list handled by the Paths Manager when they are crossed by aData packet (for the pull-based approach) or theit assigned lifetime expires(for the push-based approach).

The Interest Generation Process runs at the consumer side. It is in chargeof sending the control messages (i.e., the Interest packets) related to subscrip-tion requests. As deeply illustrated in section 3, pull-based and push-basedapproaches implement their own mechanisms. In the first case, a new con-trol message is sent when a new content is received by the consumer, whena new attachment point is selected for the consumer, or when the timeoutT pullO expires (see section 3.2 for more details). In the second case, instead,

a new control message is sent when a new attachment point is selected forthe consumer, or when the timeout T push

O expires (see section 3.3 for moredetails). When the Interest Generation Process issues a new control mes-sage, the Paths Manager provides the list of links through which sendingthe packet and reports these information in a log file. These details will beprocessed at the end of the simulation for calculating the first contributionof the communication overhead.

Finally, the Data Generation Process creates a new content every anaverage amount of time equal to TD. Every time a new content is generated,the Paths Manager provides the list of links through which sending the packetand reports these information in a log file. These links include also the staleones. Also in this case, these details will be processed at the end of thesimulation for calculating the secondo contribution of the communicationoverhead.

The code of the developed ad-hoc simulator is publicly released as anopen source project, under the GNU General Public License v03. It can befreely downloaded from the link telematics.poliba.it/pubsubmodel.

4.4.3. Towards the validation of analytical models

The analytical average communication overhead is provided by analyticalmodels, as described in section 3. To this end, the Equations reported inTable 3 are solved by considering input parameters, the average shortest

30

Figure 5: Algorithm used for implementing the pull-based approach in the ad-hoc simu-lator.

31

Figure 6: Algorithm used for implementing the push-based approach in the ad-hoc simu-lator.

32

path given by d = logN , and the average number of stale disjoint linksobtained as described before.

A bunch of computer simulations are executed to estimate the simulatedaverage communication overhead. Specifically, for each combination of pa-rameters settings, 300 different network realizations are simulated. Moreover,for each realization, BRITE is used to generate a new network topology andthe ad-hoc simulator is used to calculate the communication overhead. Then,all the obtained results are processed for estimating the simulated averagecommunication overhead. Note also that the average number of stale dis-joint links A, used in Eq. (4), Eq. (19), Eq. (14) and Eq. (38), is estimatedfrom simulations. This value is used for estimating the analytical averagecommunication overhead, as described below.

Finally, in order to evaluate the accuracy of analytical models, simulationand analytical average communication overheads are compared through theabsolute relative error (see section 4.12 for more details).

4.5. Communication overhead generated by the pull-based approach, when theproducer is stationary and the consumer is mobile.

Figure 7 shows the average communication overhead generated by thepull-based approach, obtained through analytical models and computer sim-ulations in scenarios where the producer is stationary and the consumer ismobile.

First of all, the higher the user speed, the higher the communication over-head. This result is due to the fact that when the consumer moves at a highervelocity, it may change the network attachment point more frequently. In-deed, the average cell residence time of the consumer, ∆tC , reduces as well.Now, when ∆tC decreases, the number of Interest packets sent for estab-lishing the multi-hop communication path between consumer and producerincreases. At the same time, frequent handover processes provoke the gener-ation of a higher number of stale paths. Therefore, the higher the user speed,the higher the contribution to the overhead due to extra data dissemination.

The content generation rate, and in more detail the average time intervalbetween the generation of two consecutive contents (i.e., TD) and Data packetsize (i.e., SD), also influences the communication overhead. IoT applicationswith lower TD register a higher number of Interest packets transmitted by themobile consumer. Moreover, the higher the application payload, the higherthe contribution to the overhead due to the transmission of Data packetacross stale disjoint links. In general, the communication overhead increases

33

0 20 40 60RD, [bps]

10-1

100

101

102

103

104Av

erag

e ov

erhe

ad, [

bps]

7Osim, 7v = 3 km/h7Oth, 7v = 3 km/h7Osim, 7v = 30 km/h7Oth, 7v = 30 km/h7Osim, 7v = 50 km/h7Oth, 7v = 50 km/h

(a) Urban, r = 50 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(b) Urban, r = 100 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(c) Urban, r = 150 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(d) Rural, r = 300 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(e) Rural, r = 650 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(f) Rural, r = 1000 m

Figure 7: Communication overheads related to the pull-based approach, when the produceris stationary and the consumer is mobile.

34

with the content generation rate. A counterintuitive result, instead, is givenby the Smart Meter application. In this case, even if the application presentsthe lowest RD, it does not reach the lowest communication overhead. FromTable 5 it emerges that the Smart Meter application has a very high SD value.Thus, the transmission of Data packet across stale disjoint links inflates theoverall communication overhead.

According to Eq. (4), the distance between consumer and producerstrictly influences the communication overhead. Results demonstrate thatthe communication overhead increases with cell radius. In fact, the higherthe cell radius, the lower the number of NDN nodes in the considered networktopology. Therefore, while higher cell radius values bring to lower shortestpath lengths, the communication overhead decreases as well. Urban andrural environments present different cell radius values. Specifically, urbanenvironments have lower cell radius than urban ones. Thus, in line with theprevious comments, urban scenarios always register higher communicationoverheads.

4.6. Communication overhead generated by the pull-based approach, when theproducer is mobile and the consumer is stationary.

Figure 8 shows the average communication overhead generated by thepull-based approach, obtained through analytical models and computer sim-ulations in scenarios where the producer is mobile and the consumer is sta-tionary. As described in Eq. (9), the communication overhead is only dueto the transmission of control messages. Interest packets are sent everyT pullO = TD. Therefore, the resulting overhead increases as the average time

interval between the generation of two consecutive contents (i.e., TD) de-creases. Since the lower TD, the higher the average content generation rate(i.e., RD), Figure 7 reports the average communication overheads that in-creases with RD.

As already discussed, the distance between consumer and producer strictlyinfluences the communication overhead: higher values of the the cell radiusimply lower numbers of NDN nodes and lower shortest path lengths. Thus,the average communication overhead decreases when the cell radius grows.Furthermore, as expected, urban scenarios always register higher communi-cation overheads because of their lower cell radius values.

It is also important to highlight that the communication overhead doesnot depend on the producer speed. In fact, the producer speed does notinfluence the Interest generation process at the consumer side. Independently

35

0 20 40 60RD, [bps]

10-1

100

101

102

103

104Av

erag

e ov

erhe

ad, [

bps]


(a) Urban, r = 50 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(f) Rural, r = 1000 m

Figure 8: Communication overheads related to the pull-based approach, when the produceris mobile and the consumer is stationary.

36

from the producer mobility, the consumer sends Interest packets when thetimeout expires or when a new Data packet is retrieved. For this reason, thecurves reported in Figure 8 overlap.

4.7. Communication overhead generated by the pull-based approach, whenboth producer and consumer are mobile.

Figure 9 shows the average communication overhead generated by thepull-based approach, obtained through analytical models and computer sim-ulations in scenarios where both producer and consumer move at the samespeed. Instead, Figure 10 reports the result related to the case where con-sumer and producer move at different speeds.

In general, all the considerations already argued for the previous two casesare still valid. First, the higher the consumer speed, the higher the commu-nication overhead. When the consumer moves at a higher velocity, in fact, itmay change the network attachment point more frequently and its averagecell residence time, ∆tC , reduces as well. This leads to a higher numberof Interest packets sent for establishing the multi-hop communication pathbetween consumer and producer. At the same time, frequent handover pro-cesses provoke the generation of a higher number of stale paths. Therefore,the higher the user speed, the higher the contribution to the overhead dueto extra data dissemination. The content generation rate, and in more detailthe average time interval between the generation of two consecutive contents(i.e., TD) and Data packet size (i.e., SD), also influences the communicationoverhead. IoT applications with lower TD register a higher number of In-terest packets transmitted by the mobile consumer. The distance betweenconsumer and producer influences the communication overhead: higher val-ues of the the cell radius imply lower numbers of NDN nodes and lowershortest path lengths.

In addition, by focusing the attention on Figure 10, it is possible to ob-serve that lower values of the communication overhead are experienced whenthe producer moves faster than the consumer. As already anticipated in sec-tion 3, every time the producer changes its network attachment point, allthe stale disjoint links do not produce overhead anymore. As a consequence(and differently from other scenarios) the average set of stale disjoint links,i.e., A, decreases and the average communication overhead reduces as well.

37

0 20 40 60RD, [bps]

10-1

100

101

102

103

104Av

erag

e ov

erhe

ad, [

bps]


(a) Urban, r = 50 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(f) Rural, r = 1000 m

Figure 9: Communication overheads related to the pull-based approach, when both pro-ducer and consumer are mobile and move with the same speed.

38

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]

7Osim, 7vp = 3 km/h, 7vc = 50 km/h7Oth, 7vp = 3 km/h, 7vc = 50 km/h7Osim, 7vp = 50 km/h, 7vc = 3 km/h7Oth, 7vp = 50 km/h, 7vc = 3 km/h

(a) Urban, r = 50 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(c) Rural, r = 300 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(d) Rural, r = 1000 m

Figure 10: Communication overheads related to the pull-based approach, when both pro-ducer and consumer are mobile and move with different speed.

39

4.8. Communication overhead generated by the push-based approach, whenthe producer is stationary and the consumer is mobile.

Figure 11 shows the average communication overhead generated by thepush-based approach, obtained through analytical models and computer sim-ulations in scenarios where the producer is stationary and the consumer ismobile. From results it clearly emerges that the average communication over-head increases with the user speed. When the consumer moves at a highervelocity, it may change the network attachment point more frequently. Thus,the number of Interest packets sent for re-establishing a multi-hop commu-nication path between consumer and producer increases as well. Moreover,the higher the user speed, the lower the average cell residence time of theconsumer ∆tC . In this context, the publish-subscribe approach is configuredin order to ensure, with a probability close to 1, the presence of only onestale path every ∆tC . Hence, while Interest life time decreases with ∆tC , thecorresponding generation rate increases with ∆tC . As a result, the higher theuser speed, the higher the impact of control messages to the communicationoverhead.

Of course, the transmission of Data packets across stale disjoint links in-flates the communication overhead. But, its contribution is very low withrespect to the one produced by control messages. For this reason, the com-munication overhead registers a slight increment with the average contentgeneration rate.

As already discussed before, the communication overhead decreases withthe cell radius and registers higher values in urban scenarios.

4.9. Communication overhead generated by the push-based approach, whenthe producer is mobile and the consumer is stationary.

Figure 12 shows the average communication overhead generated by thepush-based approach, obtained through analytical models and computer sim-ulations in scenarios where the producer is mobile and the consumer is sta-tionary. As in the pull-based approach, also in this case the communicationoverhead is only due to subscription requests. However, when the push-basedapproach is used, the Interest generation process does not depend on the av-erage content generation rate, but it is configured according to the user speed(see section 4.2). Indeed, the higher the user speed, the higher the Interestgeneration rate, the higher the resulting communication overhead.

40

0 20 40 60RD, [bps]

10-1

100

101

102

103

104Av

erag

e ov

erhe

ad, [

bps]


(a) Urban, r = 50 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(f) Rural, r = 1000 m

Figure 11: Communication overheads related to the push-based approach, when the pro-ducer is stationary and the consumer is mobile.

41

0 20 40 60RD, [bps]

10-1

100

101

102

103

104Av

erag

e ov

erhe

ad, [

bps]


(a) Urban, r = 50 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(f) Rural, r = 1000 m

Figure 12: Communication overheads related to the push-based approach, when the pro-ducer is mobile and the consumer is stationary.

42

Again, the communication overhead decreases with the cell radius andregisters higher values in urban scenarios and reasons are the same of thosewidely explained in the previous paragraphs.

4.10. Communication overhead generated by the push-based approach, whenboth producer and consumer are mobile.

Figure 13 shows the average communication overhead generated by thepush-based approach, obtained through analytical models and computer sim-ulations in scenarios where both producer and consumer move at the samespeed. Instead, Figure 14 reports the result related to the case where con-sumer and producer moves at speed.

As expected, the average communication overhead increases with the userspeed. In fact, a higher velocity of the consumer leads to more frequentchanges of network attachment point. Therefore, the number of Interestpackets sent for re-establishing a multi-hop communication path betweenconsumer and producer increases as well. Additionally, the speed of the con-sumer, the lower its average cell residence time ∆tC . The publish-subscribeapproach is configured in order to ensure, with a probability close to 1, thepresence of only one stale path every ∆tC . Hence, while the Interest lifetime decreases with ∆tC , the corresponding generation rate increases with∆tC . As a result, the higher the user speed, the higher the impact of controlmessages to the communication overhead.

The contribution of the transmission of Data packets across stale disjointlinks is lower than the one produced by control messages. Moreover, such acontribution decreases with the producer speed because of the reduction ofthe average set of stale disjoint links.

As already discussed before, the communication overhead decreases withthe cell radius and registers higher values in urban scenarios.

4.11. Cross comparison

With reference to the set of results presented above, it is possible to ob-serve that push-based and pull-based approaches share one common charac-teristic: the communication overhead they produce increases with the densityof nodes belonging to the NDN network topologies.

When the consumer is mobile and the producer is stationary, both ap-proaches register an increment of the average communication overhead withthe user speed. Same considerations can be done for scenarios where both

43

0 20 40 60RD, [bps]

10-1

100

101

102

103

104Av

erag

e ov

erhe

ad, [

bps]


(a) Urban, r = 50 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(f) Rural, r = 1000 m

Figure 13: Communication overheads related to the push-based approach, when bothproducer and consumer are mobile and move with the same speed.

44

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(a) Urban, r = 50 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]



0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(c) Rural, r = 300 m

0 20 40 60RD, [bps]

10-1

100

101

102

103

104

Aver

age

over

head

, [bp

s]


(d) Rural, r = 1000 m

Figure 14: Communication overheads related to the push-based approach, when bothproducer and consumer are mobile and move with different speed.

45

consumer and producer are mobile. In this case, however, the average com-munication overhead reduces because of the decrement of the average set ofstale disjoint links. This behavior is mainly registered when the producermoves faster than the consumer. On the contrary, different behaviors areobserved in a scenario with stationary consumer and mobile producer. Here,in fact, the user speed only impacts on the communication overhead relatedto the pull-based approach.

By observing the absolute values of both estimated and measured com-munication overheads, it is possible to conclude that the pull-based approachalways guarantees the lowest bandwidth consumption, thus promising poten-tial performance gains of the overall network performance.

4.12. Validation of analytical modelsTo provide a further insight, the accuracy of analytical models formulated

in section 3 is evaluated by means of the absolute relative error, calculatedbetween the theoretical communication overhead and the one measured fromsimulations, as defined in Eq. (43):

Error =

∣∣∣∣ ¯Osim − Oth

¯Osim

∣∣∣∣ 100, (43)

where ¯Osim and Oth are the average communication overhead evaluated bysimulations and the one estimated through analytical models, respectively.

The average and the maximum absolute relative errors related to resultsreported in sections 4.5-4.9, are shown in Table 6. Obtained values clearlydemonstrate that the proposed analytical models are accurate enough tocapture the behavior of the publish-subscribe mechanism based on the pull-based and push-based approach, in a wide range of scenarios.

5. Conclusion

This work devised publish-subscribe communication mechanisms for emerg-ing Internet of Things applications, in different mobile conditions. Built ontop of the Information-Centric Networking paradigm, developed solutionsleverage both pull-based and push-based approaches. In addition, analyti-cal models, describing the communication overhead they can introduce, wereformulated and validated through computer simulations. To this end, bench-mark network topologies, real Internet of Things applications, and standard-ized settings for urban and rural environments were taken into account. Ob-tained results highlighted how the communication overhead is influenced by

46

Table 6: Average and maximum values for the absolute relative errors, calculated for bothpull-based and push-based approaches.

Stationary producer, mobile consumerPull-based Push-based

Average error 4.80% 2.77%Maximum error 9.53% 4.97%

Mobile producer, stationary consumerPull-based Push-based

Average error 2.07% 2.03 %Maximum error 3.95% 6.53 %

Mobile producer, mobile consumerPull-based Push-based

Average error 4.75% 4.11%Maximum error 10.13% 9.10 %

the parameters settings, which include cell radius, user speed, average con-tent generation rate, and communication environment. Also, they allowed todetermine pros and cons of pull-based and push-based approaches by under-ling the conditions under which one scheme should be preferred to the otherone. Furthermore, the absolute relative error, calculated between the theo-retical communication overhead and the measured one, clearly demonstratedthe good level of accuracy offered by the proposed analytical models. Futureresearch activities will further investigate the behavior of conceived publish-subscribe communication mechanisms through experimental results, whilealso considering scenarios with different kind of services, mobility patterns,and routing algorithms. The impact of concrete Software Defined Networkingsolution supporting networking operations will be considered too.

Acknowledgments

This work was performed in the context of the project BonVoyage, whichreceived funding from the European Union’s Horizon 2020 research and in-novation programme under grant agreement No 635867.

47

References

[1] Cisco, “The Zettabyte Era: Trends and Analysis,” Cisco Systems, Inc.,Tech. Rep., 2016.

[2] Huawei, “Global Connectivity Index 2016,” Huawei Technologies Co.,Ltd., Tech. Rep., 2016.

[3] K. H. Yeh, “A Secure IoT-Based Healthcare System With Body SensorNetworks,” IEEE Access, vol. 4, pp. 10 288–10 299, December 2016.

[4] M. Zhu, X. Y. Liu, F. Tang, M. Qiu, R. Shen, W. Shu, and M. Y. Wu,“Public Vehicles for Future Urban Transportation,” IEEE Transactionson Intelligent Transportation Systems, vol. 17, no. 12, pp. 3344–3353,December 2016.

[5] J. Lloret, J. Tomas, A. Canovas, and L. Parra, “An Integrated IoTArchitecture for Smart Metering,” IEEE Communications Magazine,vol. 54, no. 12, pp. 50–57, December 2016.

[6] G. Piro, I. Cianci, L. Alfredo Grieco, G. Boggia, and P. Camarda, “In-formation Centric Services in Smart Cities,” Elsevier Journal of Systemsand Software, vol. 88, pp. 169–188, February 2014.

[7] Y. Zhang, D. Raychadhuri, L. A. Grieco, E. Baccelli, J. Burke, R. Ravin-dran, G. Wang, A. Lindren, B. Ahlgren, and O. Schelen, “Design Con-siderations for Applying ICN to IoT,” Internet Engineering Task Force,Internet-Draft, March 2017.

[8] W. Shang, A. Bannis, T. Liang, Z. Wang, Y. Yu, A. Afanasyev,J. Thompson, J. Burke, B. Zhang, and L. Zhang, “Named data net-working of things,” in proc. of IEEE First International Conference onInternet-of-Things Design and Implementation (IoTDI), April 2016, pp.117–128.

[9] M. Amadeo, C. Campolo, J. Quevedo, D. Corujo, A. Molinaro, A. Iera,R. L. Aguiar, and A. V. Vasilakos, “Information-centric networking forthe internet of things: Challenges and opportunities,” IEEE Network,vol. 30, no. 2, pp. 92–100, March-April 2016.

48

[10] N. Fotiou and G. C. Polyzos, “Realizing the Internet of Things usinginformation-centric networking,” in proc. of IEEE 10th InternationalConference on Heterogeneous Networking for Quality, Reliability, Secu-rity and Robustness (QShine), August 2014, pp. 193–194.

[11] L. A. Grieco, M. B. Alaya, T. Monteil, and K. Drira, “Architectinginformation centric ETSI-M2M systems,” in proc. of IEEE InternationalConference on Pervasive Computing and Communications Workshops(PERCOM Workshops), March 2014, pp. 211–214.

[12] J. Chen, S. Li, H. Yu, Y. Zhang, D. Raychaudhuri, R. Ravindran,H. Gao, L. Dong, G. Wang, and H. Liu, “Exploiting ICN for Realiz-ing Service-Oriented Communication in IoT,” IEEE CommunicationsMagazine, vol. 54, no. 12, pp. 24–30, December 2016.

[13] S. Li, Y. Zhang, D. Raychaudhuri, R. Ravindran, Q. Zheng, L. Dong,and G. Wang, “IoT Middleware Architecture over Information-CentricNetwork,” in IEEE Globecom Workshops (GC Wkshps), February 2015,pp. 1–7.

[14] P. T. Eugster, P. A. Felber, R. Guerraoui, and A.-M. Kermarrec, “Themany faces of publish/subscribe,” ACM Computing Surveys (CSUR),vol. 35, no. 2, pp. 114–131, June 2003.

[15] D. Guinard, V. Trifa, F. Mattern, and E. Wilde, “From the internetof things to the web of things: Resource-oriented architecture and bestpractices,” in Architecting the Internet of things. Springer, March 2011,pp. 97–129.

[16] A. Hakiri, P. Berthou, A. Gokhale, and S. Abdellatif,“Publish/subscribe-enabled software defined networking for effi-cient and scalable IoT communications,” IEEE CommunicationsMagazine, vol. 53, no. 9, pp. 48–54, September 2015.

[17] L. Atzori, A. Iera, and G. Morabito, “The internet of things: A survey,”Computer networks, vol. 54, no. 15, pp. 2787–2805, June 2010.

[18] M. Zorzi, A. Gluhak, S. Lange, and A. Bassi, “From today’s intranetof things to a future internet of things: a wireless-and mobility-relatedview,” IEEE Wireless Communications, vol. 17, no. 6, December 2010.

49

[19] G. Piro, M. Amadeo, G. Boggia, C. Campolo, L. A. Grieco, A. Molinaro,and G. Ruggeri, “Gazing into the crystal ball: when the future inter-net meets the mobile clouds,” IEEE Transactions on Cloud Computing,vol. PP, no. 99, pp. 1–1, May 2016.

[20] G. Xylomenos, C. N. Ververidis, V. Siris, N. Fotiou, C. Tsilopou-los, X. Vasilakos, K. V. Katsaros, G. C. Polyzos et al., “A survey ofinformation-centric networking research,” IEEE Communications Sur-veys & Tutorials, vol. 16, no. 2, pp. 1024–1049, July 2014.

[21] G. Piro, L. A. Grieco, G. Boggia, and P. Chatzimisios, “Information-centric networking and multimedia services: present and future chal-lenges,” Transactions on Emerging Telecommunications Technologies,vol. 25, no. 4, pp. 392–406, October 2014.

[22] B. Ahlgren, C. Dannewitz, C. Imbrenda, D. Kutscher, and B. Ohlman,“A survey of information-centric networking,” IEEE CommunicationsMagazine, vol. 50, no. 7, July 2012.

[23] G. Xylomenos, X. Vasilakos, C. Tsilopoulos, V. A. Siris, and G. C.Polyzos, “Caching and mobility support in a publish-subscribe inter-net architecture,” IEEE Communications Magazine, vol. 50, no. 7, July2012.

[24] Y. Zhang, A. Afanasyev, J. Burke, and L. Zhang, “A survey of mo-bility support in named data networking,” in proc. of IEEE Confer-ence on Computer Communications Workshops (INFOCOM WKSHPS),September 2016, pp. 83–88.

[25] O. Waltari and J. Kangasharju, “Content-Centric Networking in theInternet of Things,” in proc. of 13th IEEE Annual Consumer Commu-nications Networking Conference (CCNC), January 2016, pp. 73–78.

[26] A. Lindgren, F. B. Abdesslem, B. Ahlgren, O. Scheln, and A. M. Malik,“Design choices for the IoT in Information-Centric Networks,” in proc. of13th IEEE Annual Consumer Communications Networking Conference(CCNC), January 2016, pp. 882–888.

[27] Y. Huang and H. Garcia-Molina, “Publish/subscribe in a mobile envi-ronment,” Wireless Networks, vol. 10, no. 6, pp. 643–652, November2004.

50

[28] ——, “Publish/subscribe tree construction in wireless ad-hoc networks,”in proc of. International Conference on Mobile Data Management.Springer, December 2003, pp. 122–140.

[29] X. Cao and C.-C. Shen, “Subscription-aware publish/subscribe tree con-struction in mobile ad hoc networks,” in proc. of International Confer-ence on Parallel and Distributed Systems, vol. 2, December 2007, pp.1–9.

[30] G. Cugola, A. L. Murphy, and G. P. Picco, “Content-based publish-subscribe in a mobile environment.” in The Handbook of Mobile Middle-ware, October 2006, pp. 257–285.

[31] A. Detti, D. Tassetto, N. B. Melazzi, and F. Fedi, “Exploiting contentcentric networking to develop topic-based, publish-subscribe MANETsystems,” Ad Hoc Networks, vol. 24, pp. 115–133, January 2015.

[32] M. Amadeo, C. Campolo, and A. Molinaro, “Internet of things vianamed data networking: The support of push traffic,” in IEEE Interna-tional Conference and Workshop on the Network of the Future (NOF),December 2014, pp. 1–5.

[33] J. Francois, T. Cholez, and T. Engel, “CCN traffic optimization forIoT,” in proc. of IEEE Fourth International Conference on the Networkof the Future (NOF), October 2013, pp. 1–5.

[34] C. Tsilopoulos and G. Xylomenos, “Supporting diverse traffic types ininformation centric networks,” in Proceedings of the ACM SIGCOMMworkshop on Information-centric networking, August 2011, pp. 13–18.

[35] J. Chen, M. Arumaithurai, L. Jiao, X. Fu, and K. Ramakrishnan,“Copss: An efficient content oriented publish/subscribe system,” inproc. of IEEE Seventh ACM/IEEE Symposium on Architectures for Net-working and Communications Systems (ANCS), November 2011, pp.99–110.

[36] N. Fotiou, P. Nikander, D. Trossen, and G. C. Polyzos, “Developinginformation networking further: From PSIRP to PURSUIT,” in proc.of International Conference on Broadband Communications, Networksand Systems. Springer, 2010, pp. 1–13.

51

[37] A. Carzaniga, M. Papalini, and A. L. Wolf, “Content-based pub-lish/subscribe networking and information-centric networking,” in Pro-ceedings of the ACM SIGCOMM workshop on Information-centric net-working, August 2011, pp. 56–61.

[38] W. Drira and F. Filali, “A Pub/Sub extension to NDN for efficient datacollection and dissemination in V2X networks,” in proc. of IEEE 15thInternational Symposium on a World of Wireless, Mobile and Multime-dia Networks (WoWMoM), October 2014, pp. 1–7.

[39] A. V. Ventrella, L. A. Grieco, and G. Piro, “Information-Centric Net-working in Environmental Monitoring: an overview on publish-subscribeimplementations,” in Proc. of IEEE International Conference on Ad-vanced Video and Signal based Surveillance (AVSS) workshop, Lecce,Italy, August 2017.

[40] A. V. Ventrella, G. Piro, and L. A. Grieco, “Information-centric publish-subscribe mechanisms for Intelligent Transportation Systems,” in Proc.of IEEE International Conference on Service Operations and Logistics,and Informatics (SOLI 2017), Bari, Italy, Sep. 2017.

[41] FANTASTIC5G, “Deliverable D2.1 Air interface framework and speci-fication of system level simulations,” Call: H2020-ICT-2014-2, Projectreference: 671660, Flexible Air iNTerfAce for Scalable service deliv-ery wiThin wIreless Communication networks of the 5th Generation(FANTASTIC-5G), Tech. Rep. D2.1, May 2016.

[42] A. Medina, A. Lakhina, I. Matta, and J. Byers, “BRITE: An approachto universal topology generation,” in proc. of IEEE Ninth InternationalSymposium on Modeling, Analysis and Simulation of Computer andTelecommunication Systems, August 2001, pp. 346–353.

[43] T. Koponen, M. Chawla, B.-G. Chun, A. Ermolinskiy, K. H. Kim,S. Shenker, and I. Stoica, “A Data-oriented (and Beyond) Network Ar-chitecture,” SIGCOMM Comput. Commun. Rev., vol. 37, no. 4, pp.181–192, August 2007.

[44] D. Raychaudhuri, K. Nagaraja, and A. Venkataramani, “MobilityFirst:a robust and trustworthy mobility-centric architecture for the future

52

internet,” ACM SIGMOBILE Mobile Computing and CommunicationsReview, vol. 16, no. 3, pp. 2–13, July 2012.

[45] N. B. Melazzi, S. Salsano, A. Detti, G. Tropea, L. Chiariglione,A. Difino, A.-C. G. Anadiotis, A. S. Mousas, I. S. Venieris, and C. Z. Pa-trikakis, “Publish/subscribe over information centric networks: A Stan-dardized approach in CONVERGENCE,” in IEEE Future Network &Mobile Summit (FutureNetw), September 2012, pp. 1–8.

[46] L. Zhang, A. Afanasyev, J. Burke, V. Jacobson, P. Crowley, C. Pa-padopoulos, L. Wang, B. Zhang et al., “Named data networking,” ACMSIGCOMM Computer Communication Review, vol. 44, no. 3, pp. 66–73,2014.

[47] T. Yagyu, K. Nakamura, T. Asami, K. Sugiyama, A. Tagami,T. Hasegawa, and M. Arumaithurai, “Demo: Content-based push/pullmessage dissemination for disaster message board,” in proc. of ACM 2ndInternational Conference on Information-Centric Networking, 2015, pp.205–206.

[48] D. Tipper, “Resilient network design: challenges and future directions,”Telecommunication Systems, vol. 56, no. 1, pp. 5–16, May 2014.

[49] K. Katsaros, W. Chai, N. Wang, G. Pavlou, H. Bontius, and M. Paolone,“Information-centric networking for machine-to-machine data delivery:a case study in smart grid applications,” IEEE Network, vol. 28, no. 3,pp. 58–64, 2014.

[50] W. Xia, Y. Wen, C. H. Foh, D. Niyato, and H. Xie, “A survey onsoftware-defined networking,” IEEE Communications Surveys Tutorials,vol. 17, no. 1, pp. 27–51, First quarter 2015.

[51] M. B. Lehmann, M. P. Barcellos, and A. Mauthe, “Providing producermobility support in NDN through proactive data replication,” in proc. ofIEEE/IFIP Network Operations and Management Symposium (NOMS),April 2016, pp. 383–391.

[52] B. Feng, H. Zhou, and Q. Xu, “Mobility support in Named Data Net-working: a survey,” EURASIP Journal on Wireless Communicationsand Networking, vol. 2016, no. 1, p. 220, 2016.

53

[53] Z. Zhu, A. Afanasyev, and L. Zhang, “A new perspective on mobilitysupport,” Named-Data Networking Project, Tech. Rep, 2013.

[54] Y. Zhang, “Handoff performance in wireless mobile networks with unre-liable fading channel,” IEEE Transactions on Mobile Computing, vol. 9,no. 2, pp. 188–200, 2010.

[55] K. S. Munasinghe and A. Jamalipour, “Analysis of signaling cost for aroaming user in a heterogeneous mobile data network,” in proc. of IEEEGlobal Telecommunications Conference (GLOBECOM), 2008, pp. 1–5.

[56] K. L. Yeung and S. Nanda, “Channel management in micro-cell/macrocell cellular radio systems,” IEEE Transactions on VehicularTechnology, vol. 45, no. 4, pp. 601–612, 1996.

[57] Q. De Coninck, M. Baerts, B. Hesmans, and O. Bonaventure, “Observ-ing real smartphone applications over multipath TCP,” IEEE Commu-nications Magazine, vol. 54, no. 3, pp. 88–93, 2016.

[58] R. Ratasuk, J. Tan, and A. Ghosh, “Coverage and capacity analysis formachine type communications in LTE,” in IEEE Vehicular TechnologyConference (VTC Spring), 2012, pp. 1–5.

[59] Y. Wang, B. Xu, D. Tai, J. Lu, T. Zhang, H. Dai, B. Zhang, and B. Liu,“Fast name lookup for named data networking,” in proc. of IEEE 22ndInternational Symposium of Quality of Service (IWQoS), 2014, pp. 198–207.

[60] A. Afanasyev, C. Yi, L. Wang, B. Zhang, and L. Zhang, “SNAMP:Secure namespace mapping to scale NDN forwarding,” in proc. ofIEEE Conference on Computer Communications Workshops (INFO-COM WKSHPS), April 2015, pp. 281–286.

[61] V. Jacobson, J. Burke, L. Zhang, B. Zhang, K. Claffy, D. Krioukov,C. Papadopoulos, T. Abdelzaher, L. Wang, E. Yeh et al., “Named datanetworking (NDN) project 2013-2014 report,” 2014.

[62] A.-L. Barabasi, Network sciensce. Cambridge University Press, 2016.

54

publish-subscribe in mobile information centric networks ...€¦ · keywords: internet of things,...

Documents