Received August 24, 2018, accepted September 24, 2018, date of publication October 8, 2018, date of current version November 8, 2018.

Digital Object Identifier 10.1109/ACCESS.2018.2874448

Online Caching and Cooperative Forwardingin Information Centric NetworkingKYI THAR 1, NGUYEN H. TRAN 1,2, (Member, IEEE), SAEED ULLAH1, THANT ZIN OO 1,AND CHOONG SEON HONG 1, (Senior Member, IEEE)1Department of Computer Science and Engineering, Kyung Hee University, Seoul 17104, South Korea2School of Information Technologies, The University of Sydney, Sydney, NSW 2006, Australia

Corresponding author: Choong Seon Hong (cshong@khu.ac.kr)

This work was supported in part by the Ministry of Science and ICT (MSIT), South Korea, through the Grand Information TechnologyResearch Center Support Program supervised by the Institute for Information & communications Technology Promotion (IITP) underGrant IITP-2018-2015-0-00742 and in part by the National Research Foundation of Korea (NRF) Grant funded by the KoreanGovernment (MSIT) under Grant NRF-2017R1A2A2A05000995.

ABSTRACT Information centric networking is one of the most promising future Internet architectures totackle the increasing network traffic by enabling in-network caching to cache popular contents. Althoughin-network caching reduces the network traffic by providing requested content locally to the users, severalchallenging issues are still unsolved. For example, identical contents are replicated in all routers alongeach request’s forwarding path, which incurs faster cache replacement and degrades cache utilization, andtemporally cached content’s locations are not easy to track or search in the network. Besides, it is tough tocorrectly predict future popularity of contents and decide which contents to store. Hence, in this paper,an online caching and cooperative forwarding scheme is proposed to enhance cache utilization and toreduce network delay, as well as reduce the workload on each router. The caching problem is formulatedas a Ski-Rental problem, which is a classical method for online decision making, in combination withconsistent-hashing to obtain an online coordinated caching solution. The proposed request forwardingscheme is based on consistent-hashing, where every router knows the potential location of the cached copyof the requested content and thereby avoiding the unnecessary forwarding. Finally, the proposed requestforwarding and caching schemes were validated by a chunk-level simulator. The simulation results showthat the proposed scheme outperforms the existing algorithms in terms of content hit rate, server load, theprocessing load on routers, and access delay.

INDEX TERMS Information centric networking, NDN online caching, Ski-Rental problem, cooperativeforwarding, consistent hashing.

I. INTRODUCTIONAccording to the Cisco Visual Networking Index, videostreaming from wireless devices has been generating mostof the Internet traffic and is forecast to continue increas-ing exponentially [1]. At present, the current Internet archi-tecture considers rigid end-to-end network flows betweennetwork nodes which cannot adapt quickly to the dynamicnetwork environments [2]. Increasing the network infrastruc-ture capacity to handle the exponential increase in usersdemand is economically expensive and impractical. Thus,new architectures are proposed to utilize caching in networknodes, such as routers in the core network and base-stations inedge network [3]. Information Centric Networking (ICN) isone such architecture which shifts the network paradigm from

location-based to information-based network [3]. ICN pro-poses to cache the transient data packets (contents) in everynetwork node which is used for full-filling users’ requests inthe near future. If the requested data packet is not found ina network node, the request is forwarded to its neighboringnodes. Named Data Networking (NDN) is one of the mostprominent ICN architecture [4].

A. WHAT IS NAMED DATA NETWORKING (NDN)?NDN is a network architecture in which, each router tem-porarily stores data packets (contents) to satisfy user requestsin the near future. In NDN, a network node (router, basestation) constitutes of i) content store, ii) Pending InterestTable (PIT), and iii) Forwarding Information Base (FIB).

VOLUME 6, 20182169-3536 2018 IEEE. Translations and content mining are permitted for academic research only.

Personal use is also permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

59679

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

Two types of packets manage the data flow: Data packetsand Interest packets. Data packets are stored in the contentstore in the form of a sequence of segments (called chunks).Interest packets are issued by the user, who then gets the Datapacket for the requested content. Thus, each router initiallychecks its content store when an Interest packet arrives.

If the requested content is found, the router retrieves therequested content, assembles the Data packet for the replyand discards the Interest packet. If the requested content isnot present in the content store, the router forwards Inter-est packet towards the potential content location or contentproviders (content servers) using the information from FIBwhich stores the name prefix of contents to forward the Inter-est packets in lieu of IP address. Routers keep the informationof the unresolved forwarded Interest packets in a PIT whichaggregates the identical Interest packets from different users,forwarding only one Interest packet and returning the relevantData packet to all the listed requesters when it arrives. If arouter receives a Data packet for which it has no entry inthe PIT, it considers the packet to be redundant or outdatedand discards it.

Every NDN node decides three important strategies:• Cache decision: the node decides whether to cache thecontent or not.

• Cache replacement: the node deletes old chunks for newincoming chunks when its cache space is full.

• Forwarding: the node chooses its logical and physicalinterfaces to forward the Interest packets.

B. CHALLENGES/ISSUESCache utlization: Caching every chunk passing throughNDN node is the simplest cache decision, however, thisleads to the redundant caching problem, i.e., identical contentreplicated in all routers along each request’s forwarding path,incurs faster cache replacement and degrades cache utiliza-tion. As a result, this type of caching decision reduces thecache availability for the whole network. Hence, duplicatecontents need to be removed. The elimination of duplicatecontents can improve cache utilization, but it requires someextra processing, i.e., exchanging control messages betweenneighboring routers and computation on each router. Hence,the major challenge is controlling the trade-off between cacheutilization and overhead incurred by computation and mes-sage exchange.

Content retrieval delay: Content retrieval delay variesdepending on its cached location. To achieve minimum delay,the most frequently requested contents (popular contents)should be stored at the router which is the nearest to users.Further, the router should prevent replacing old popular con-tents with new unpopular contents.

Content store processing load: Typically, all of the arriv-ing requests are searched in the content store to knowwhetherthe requested contents are located in the cache storage or notand that searching process increases the processing load oneach router. Thus, the content store processing load shouldbe distributed in order to reduce the processing load on eachrouter.

C. CONTRIBUTIONSIn this paper, the network is separated into two groupsto facilitate management; i) the self-governing backhaulnetwork termed as Regional Cache Cluster (RCC), andii) the access network which includes access nodes such asBase-Stations (BSs), Small-cell Base-Stations (SBSs) andaccess routers. Access nodes are located at the nearest pointsto the users; hence, access nodes will cache the popular con-tent to achieve improved performance by employing onlinecaching. RCC utilizes online coordinated caching and consis-tent hashing [5] based forwarding process to improve cacheefficiency, as well as to reduce the content store processingload on each router. We apply consistent hashing for RCCusing the virtual nodes concept from the Dynamo (key-valuestore) partition technique [6]. The consistent hash ring mapenables the routers to filter out the duplicate content chunks,and forward each Interest packet directly toward the custo-dian routers 1 2. Our contributions are summarized as follows:• We propose a cache decision and forwarding algorithmusing a consistent hash ring to improve cache hits,reduce the router’s processing load, and eliminate dupli-cate contents.

• We formulate the cache decision problem as aSki-Rental problem (discuss details in Sec. V-A), whichis solved using a randomized online caching and replace-ment algorithm to store valuable content instead ofkeeping all the content that passes through the node.

• Weevaluate our proposed scheme by usingCCNSim [7].The simulation results show that the proposed schemeimproves the cache hit probability up to 42% than overallexisting algorithms.

The rest of the paper is organized as follows: Sec. IIcovers the related work. The system model is describedin Sec. III, while the proposed scheme is presented in Sec. IV.The performance evaluation of the proposed scheme is givenin Sec. IV. Finally, the paper is concluded in Sec. VII.

II. RELATED WORKThere are many existing researches on caching and forward-ing in NDN.We classify some of these works by their cachingpolicies (i.e., coordinated or uncoordinated) and forwardingpolicies (i.e., non-cooperative or cooperative).Uncoordinated caching strategy3 stores the contents on the

return path4 without any predefined location [4], [8]–[14]NDN usually uses Leave Copy Everywhere (LCE) policy [4].When following LCE, routers cache each packet that passesthrough them. LCE is simple to implement and execute, butthe duplicate contents are stored in all the routers along thereturn path. Hence, LCE reduces overall cache utilization in

1Custodian routers are predefined routers that store relevant contentfiltered using the consistent hash ring.

2The hash ring map helps to reduce the processing load on the router byskipping the content store check

3Here thereafter, we use caching policy and caching strategyinterchangeably.

4The path which the Interest packet is forwarded.

59680 VOLUME 6, 2018

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

the network. Leave Copy Down (LCD) policy [8] reduces theduplicate contents on the return path but still does not elimi-nate it completely. Probabilistic caching schemes [9]–[12] areproposed to reduce content access delay and improve cachespace utilization, in which contents are cached on the routerbased on its cache capacity and location of the router. Costaware caching is proposed in [13] and [14], in which contentsare cached on the router based on the cost of each content.Coordinated caching stores the contents off the path to

improve the overall network cache utilization [15]–[27]. Sev-eral content’s popularity prediction based caching schemessuch as deep learning based scheme [15], [16] are proposedto improve the cache utilization. Compared to the onlinecaching scheme [17], [18], deep learning based scheme needshigh computational resources and a huge amount of datafor training. To reduce the energy consumption for retriev-ing contents, [19] proposed energy aware caching scheme.Wang et al. [21] and Sourlas et al. [22] used to control mes-sage exchange to eliminate the duplicate contents or chunkswhich increase the network traffic. To avoid increasing net-work traffic while simultaneously eliminating duplicate con-tent, hash-based caching was proposed [22]–[25]. Althoughclassical hashing [22]–[25] is lightweight and efficient, it suf-fers from inconsistency and load balancing problems. In ourprevious work [26], [27], we used consistent hashing [5] asa foundation of forwarding and caching strategies to solvethese problems. So, we can significantly improve the cacheutilization in [26]. However, the delay in accessing the con-tents remains high. Thus, in [27], we used a controller tomake cache decisions, where the controller collects all con-tent information. This reduces access delay but increases thecontrol message overhead.Non-cooperative forwarding [28]–[30] schemes, such as

Interest flooding scheme, forward the Interest packets with-out exchanging the control messages. Flooding the Inter-est packets increases network traffic and degrades networkperformance [28]. To avoid network degradation, in [29], theInterest packets flooding is limited to the hop count threshold.In [30], routers rank the interfaces and choose the best pathto retrieve requested content. If the first ranked router doesnot provide the content, the router chooses the second-rankedinterface to forward an unsatisfied request.Cooperative forwarding reduces network traffic and

processing load at the routers due to Interest packetsflooding [17], [20]–[27], [31], [32]. Liu et al. [31] andZheng et al. [32] improved the throughput of the vehicularnetwork by considering cooperative Interest packet forward-ing mechanism among cluster’s members and the base sta-tion, where clusters are formed based on mobility behaviorsof nodes. In [17], [20], and [21], requests are forwardedcooperatively while the contents are cached in a coordinatedmanner, in which the router announces cached content infor-mation to its neighbors. The information exchange increasesthe control message overhead. To reduce this overhead, thehashing based forwarding scheme is used in [22]–[27] as incoordinated caching.

Cache replacement strategy informs the router whichold chunks to delete for new incoming chunks when itscache space is full. The cache replacement strategies canbe classified into i) recency-based, ii) frequency-based,iii) recency/frequency-based, iv) function-based [25], andv) randomized strategies [33]. The most widely used cachereplacement policy is the Least Recently Used (LRU) policywhere the LRU content is replaced with the new incomingcontent, in case the cache is full.Proposed scheme Hence, in this paper, we solve

the problem of simultaneously caching non-duplicatecontents and reducing control message overhead. We pro-posed the distributed caching scheme, which uses consistenthashing [5], [6] as a basis for non-duplicate caching wherethe actual cache decision is done by the Ski-Rental [35] basedrandomized online caching algorithm. Compared to the deeplearning based approach, our proposal does not need highcomputing resources and large dataset to train, where thenodesmake a cache decision based on the Ski-Rental problemwithout knowing the future popularity of each content or withlimited information. Also, the probabilistic caching schemecannot eliminate the duplicate content effectively. In orderto reduce the Interest packet flooding as well as to fit withthe consistent hashing based caching scheme, we proposedInterest packet forwarding scheme which used the sameconsistent hashing scheme as in cache decision.

FIGURE 1. System model.

III. SYSTEM MODELThe system model for our proposed caching and forward-ing scheme is shown in Fig. 1. The important abbre-viations and notations used in this paper are presentedin Table 1. R is denoted as the set of content routersand B is denoted as the set of access nodes. Thesetwo sets constitutes the set of nodes in the network(i.e., V = B ∪ R) such that each node v ∈ V is connectedto some other node v′ ∈ V via intermediate links. The setof these intermediate links or edges are denoted by E , andundirected graph G = (V, E) represents the whole network.The content routers can be classified into i) negotiator router,

VOLUME 6, 2018 59681

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

TABLE 1. Table of abbreviations and notations.

ii) bridge router, and iii) gateway router. The negotiator routercollects group information to generate a hash ring map, thebridge router is the connector between the router groups, andthe gateway router is the connector between the autonomoussystems. If Sv denotes the cache capacity of v, the total cachecapacity of the autonomous system is S =

∑v∈V Sv. The set

of content chunks F is stored at the content servers whereF , {fij | fij is the j-th chunk of i-th content}.The content routers are formed into several groups

A = {A1,A2, . . . ,An} inside an AS by the system adminis-trator, depending on their geographical location, i.e., r ∈ An,if router r is in groupAn. The number of routers in groupAnis An = |An|. Each router r in the group An possesses a setof keys Kn

r , which are defined as virtual routers. Each groupof routers constructs the consistent hash ring map Kn

r ⊆ Kn,where Kn

r is the keys of the router r of group n and Kn isthe total keys or the consistent hash ring map of the group n.The consistent hash ring map is used to filter the caching ofoverlapping content and Interest packet forwarding within agroup.

Initially, the network administrator manually forms groupsor clusters on the basis of the geographical locations of therouters. Then, each group elects a Negotiator Router by calcu-lating the closeness centrality using (1). Closeness centralityis one method to measure the importance of a router. Thehigher the closeness centrality, the closer the router is locatedto every router in the group, and the higher the importance ofthat router. The normalized closeness centrality κ(r) of each

router is calculated by

κ(r) =1

(N − 1)∑N

j=1d(r, j), (1)

where d(r, j) denotes the total number of shortest paths fromrouter r to router j; N denotes the number of routers or groupsize within one group; and the shortest paths are calculatedby Dijkstra’s algorithm. The router with the highest closenesscentrality, thus, becomes the Negotiator Router of the group.The Negotiator Router announces itself as the NegotiatorRouter to its group members. Then, all of the group-membersreply their information, such as cache capacity and the num-ber of interfaces, to the Negotiator Router to construct theconsistent hash ring map, Kn. Finally, the Negotiator Routerdistributes the constructed hash ring map to all the members.

IV. COOPERATIVE FORWARDING AND RANDOMIZEDCACHING IN CONTENT CENTRIC NETWORKINGIn this section, we first present the proposed modified NDNengine’s components and modified Interest packet. Second,we discuss the construction of a consistent hash ring map,which is the basic component of the forwarding and cachingprocess. Next, we present the forwarding process of accessnodes, and coordinated forwarding process of the regionalcontent centric cache cluster. Third, we discuss the cachingprocess, which applies the Ski-Rental problem to the cachedecision. At the end of this section, we discuss how tomanagerouter failures, removals, and additions.

FIGURE 2. Proposed modified NDN engine.

A. AUGMENTED NDNWe added new mechanisms to modify the original NDNto support our proposed cooperative forwarding and onlinecaching. The components inside the router are shown inFig. 2, which is a modified version of the original NDNengine [4]. The cluster routers need all the components tooperate, but the access nodes need only the components insidethe dotted box. The FIB and content store remain the same asin the original proposal. The major differences between theoriginal and proposed systems are as follows:• Hash ring is constructed to identify the custodian routerof the incoming Interest packet within a group. Using the

59682 VOLUME 6, 2018

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

hash ring and routing table, the non-custodian routerforwards the Interest packets directly to the custodianwithout checking its content store. The construction ofthe hash ring and its benefits are discussed in detailin Sec. IV-B.

• Cache decision bit, x ijr ∈ {0, 1}, is added to PIT in eachrouter. If x ijr = 1, then the chunk j of content i is storedin the content store at router r . Otherwise (i.e., x ijr = 0),the chunk, fij, is not stored. The cache decision is madeusing a randomized online caching algorithm beforeforwarding the unsatisfied Interest packet.

In our previous work [26], we used the hash ring mapfor both Interest and Data packet filtering, as shown inFigs. 3(a) and 3(b). Using the hash ring map in both casesreduced the performance of the router. The cache decision bitlets the router or access nodes to quickly identify whether thearriving content chunk should be stored in its content store,as shown in Figs. 3(c) and 3(d). In this case, the router doesnot need to reuse the consistent hash ring mapping in thecache decision process when a requested Data packet arrives.

FIGURE 3. Benefits of cache decision bit: Sub-figures (a) and (b) HashRing based filtering is used in both forwarding and caching process.Sub-figure (c) Hash Ring based filtering is only used to filter Interestpacket, and Sub-figure (d) cache decision bit is used to store Data packet.

To support our proposed mechanisms, we added four newfields to the Interest packet;

1) Interest Type (d ITij ) classifies the Interest packets forcoordinated forwarding

2) Custodian ID (dCIDij ) identifies which router will cachethe content chunk.

3) Group ID (dGIDij ) identifies the source of Interestpacket.

4) Key Update (dKUij ) is used to manage the router failuresor adding/removing routers in a group.

We will further discuss Interest packet types in Sec. IV-C andkey update in Sec. IV-E.

B. CACHE-CAPACITY-AWARE CONSISTENTHASH RING MAPHash functions eliminate duplicate content chunks withoutusing flooding strategy in Interest message packet forward-ing or extra control messages. Hence, it is the preferredmethod for coordinated caching and cooperative forward-ing schemes. Hashing can be categorized into classical or

consistent hashing. Previously, the classical hashinghave been widely used in the caching and forwardingprocess [23]–[25]. Although the classical-hash-function-based caching and forwarding process is lightweight, it suf-fers from inconsistency and load balancing problems becauseits hash value (usually the number of routers) is used formapping the contents to the routers.

FIGURE 4. Consistent hash ring and forwarding process.

Consistent hashing has several variants, and the originalconsistent hashing [5] uses a key range (a ring as shownin Fig. 4) and the nodes (servers) are allocated on the ringby the hash function. Incoming requests are mapped ontothe ring using the hash values. The router located first inthe clockwise direction on the ring handles the requests orcontent. The drawback here is that, if a router leaves thegroup, all the values assigned to it are shifted to the next routerin the clockwise direction, which can cause some seriousload balancing issues. Hence, we introduce the concept ofvirtual routers (or keys) to consistent hashing as given in [6]to construct the hash ring which forms the foundation for theInterest packet forwarding and caching process.

Each virtual router (or key) k is a non-negative randomvariable which follows a uniformly random distribution withk ∈ [kmin, kmax] ⊆ R+, where kmin and kmax are theminimumand maximum limits of the key range, respectively. The keysare allocated proportionally to the physical content routersbased on their cache sizes as shown in Fig. 4, i.e., a contentrouter with a larger cache size possesses more keys than acontent router with smaller cache size. Each content chunkcan then be filtered and cached in physical content routersusing the hash ring map. In the case of a router failure,the keys on the failed router are distributed proportionallyto the remaining routers. An example of key distribution isshown in Fig. 4 where routers A (Cache size: 500 chunks,5 keys), B (Cache size: 400 chunks, 4 keys), C (Cache size:300 chunks, 3 keys) and D (Cache size: 200 chunks, 2 keys)are allocated keys proportionally depending on their cachesizes. Thus, unlike the original consistent hashing, in our pro-posal, router failure will not cause all its keys to be dumpedonto a single router.

C. INTEREST PACKET FORWARDINGIn Interest packet forwarding, the node sends the request toneighboring nodes if the requested content is not found inits content store. The process of Interest packet forwardingin access nodes and content routers are different. Hence,we differentiate Interest forwarding into: i) forwarding in

VOLUME 6, 2018 59683

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

FIGURE 5. Forwarding process of (a) Base station/small-cell base station,(b) Content router.

access nodes as shown in Fig. 5(a), and ii) forwarding incontent router as shown Fig. 5(b).

1) INTEREST FORWARDING IN ACCESS NODESThe detailed process for interest forwarding in access nodesare described in Alg. 1 and Fig. 5(a). First, when the Inter-est packet dij arrives, the access nodes search the requestedcontent in its content store (line 3). If the requested contentchunk fij is found, the access nodes reply the content to theuser immediately (line 4). Otherwise, access nodes run Alg. 5to update the cache decision bit x ijr (line 6). This will furtherbe discussed in Sec. V).

Algorithm 1 Interest Forwarding in Access Nodes1: Input: Request dij arrives2: Output: Reply fij or forward dij3: if fij ∈ Fv then4: Reply fij5: else6: Run Alg. 5 to set x ijv7: if dij /∈ DPIT

v then8: DPIT

v ← DPITv ∪ {dij}.

9: DFIBv ← DFIB

v ∪ {dij}.10: Forward the request dij.11: else12: DPIT

v ← DPITv ∪ {dij}.

13: Stop forwarding the request dij.

Second, the access nodes check the PIT for the Interestpacket (line 7). If the Interest packet is not listed in the PITyet, it is added to the PIT (line 8), and access nodes forwardthe Interest packet to other routers by using FIB (line 9).Otherwise, it is added to the PIT (line 11), and access nodesstop forwarding the dij (line 12).

2) INTEREST FORWARDING IN ROUTERSThe detailed process for interest forwarding in content routeris described in Alg. 2 and Fig. 5(b). When the Interest packetdij arrives (Interest packet types will further be discussedin Sec. IV-D), the router checks the type of the Interest

Algorithm 2 Interest Forwarding in Router1: Input: Request dij arrives at router v2: Output: Reply fij or forward dij3: For router v ∈ R,4: if d ITij = 0 then5: Compute key by hashing k = h(dij) .6: Find router r that possesses key k ∈ Kn

r .7: if v = r then8: Repeat lines 3 to 8 from Alg. 19: d ITij ← 2, dGIDij ← n, dCIDij ← r .10: Forward dij to gateway dGWIDij11: else12: d ITij ← 1, dGIDij ← n, dCIDij ← r .13: Forward dij to custodian router r .

14: else if d ITij = 1 then15: if v = dCIDij then16: Repeat lines 3 to 8 from Alg. 117: d ITij ← 2. and forward dij.18: else19: Forward dij to custodian router r

20: else if d ITij = 2 then21: if dGIDij 6= n then22: Repeat lines 5 to 13, expect 1023: else24: Forward dij to gateway dGWIDij

packet d ITij . Depending on Interest type, router r performs oneof the following three actions.

In case d ITij = 0 (line 4), the router checks the arrivingInterest packet’s name with a hash function to get its corre-sponding key (line 5). If the router possesses the key, it isthe custodian of that Interest packet (line 6). For example,as shown in Fig. 1, r5 is the custodian of Interest packet‘‘youtube.com/A/01’’ whose corresponding key k = 0.On the other hand, r5 is not the custodian of Interest packet‘‘youtube.com/A/02’’ with its corresponding key k = 3.If the current router is the custodian (line 7), it searchesthe requested content in its content store. If the requestedcontent chunk fij is found, the router replies the contentto the user immediately. Otherwise, router runs Alg. 5 toupdate the cache decision bit x ijv . Then, the router assignsd ITij ← 2, dGIDij as current group ID, and dCIDij as router r(line 9). Next, the router forwards the Interest packet dij tothe dGWID

ij . If the current router is not custodian, it updatesd ITij ← 1, dGIDij as current group ID, and dCIDij as custodianrouter r (line 12). Then, the router forwards the Interestpacket dij to the custodian router by using the routing table(line 13).

In case d ITij = 1 (line 14), the router checks the dCIDij . If thecurrent router is the custodian (line 15), it repeats the process(lines 3 to 12) as in Alg. 1. If the current router is not thecustodian, it checks the dCIDij and forwards the Interest packetto the custodian router.

59684 VOLUME 6, 2018

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

In case d ITij = 2, the router searches the dGIDij . If thedGIDij from the Interest packet and current dGIDij is the same,the router just checks the dGWID

ij and forwards the Interestpacket to that dGWID

ij . If the dGIDij is different, the routerrepeats the same steps from lines 5 to 13, except line 10(the gateway selection process). The gateway selection is notneeded because the Interest carries the gateway informationin its dGWID

ij field. If the dGIDij is same, the router checks thedGWIDij from Interest packet and forwards it to the gateway.

D. INTEREST TYPES USED IN FORWARDING PROCESSWe introduce the different types of Interest packets to achievethe following properties: i) to enable direct Interest packetforwarding to the custodian router, ii) to reduce the usage ofthe hash function and the consistent hash ring, iii) to enablethe inter-cluster forwarding, and vi) to prevent the blockingof the Interest packet by an intermediate router.

The following example explains how the intermediaterouter blocks the Interest packet. In NDN, an intermediaterouter can block an unsatisfied Interest packet from the custo-dian router with any type of hash-based forwarding scheme orcoordinated forwarding. In Fig. 4, router A requests a chunkto router B (via router C), which is the custodian of Interestpacket I , and the requested chunk is not in router B’s contentstore. Thus, router B requests the requested chunk to thecontent server via router C . Therefore, router C receives thesame Interest packet: one from routerA, and another one fromrouter B. In the normal case, router C will hold the Interestpacket from custodian router B because it already has the PITentry for the first Interest packet forwarded to router B fromrouter A, which means the users never get the Data packet andneed to send the same Interest packet repeatedly when theirrequests time out. Enabling router C to forward the requestto the gateway (D) requires a new mechanism. Li et al. [23]solved this problem by using piggyback Interest packets. Inour proposal, we use the Interest Type d ITij field in the Interestmessage.

The router can easily identify the type of Interest packet bychecking d ITij field in the Interest packet. We use three typesof Interest packets. The default value of this field, d ITij = 0,represents the normal Interest, which is processed normallyas in the original NDN [4]. When a router receives this typeof packet, it searches the custodian of the Interest packetby using the hash function. On the other hand, d ITij = 1represents the custodian forward Interest, which the routeruses to forward the Interest packet directly to the custodianrouter without checking its content store. Finally, d ITij = 2represents the gateway forward Interest, which the router usesto forward the Interest packet directly to the gateway.

E. MANAGEMENT OF ROUTERS AND KEYSIn the case of router failure (or removal), the failed router willbe removed from the consistent hash ring map (line 1), andits key will be redistributed to remaining routers as describedin Alg. 3. The failure of a router is detected by using a NACKInterest packet as in [34]. When a router failure is detected,

Algorithm 3 Router Removal1: An = An \ {r}.2: Cacluate cache weight: ωi =

Si∑j∈An Sj

, ∀i ∈ An.3: for k ∈ Kn

r do4: Kn

i ← Kni ∪ {k} with probability ωi.

the detecting router informs to the Negotiator Router. Then,the Negotiator Router updates the consistent hash ring mapby redistributing the keys of the failed router to others insidethe group without affecting the fairness and then distributingthe new hash ring map to all routers as described in Alg. 3.The old key-values mapping to the failed router remains thesame in the new hash ring map (line 4).

For example, if router A fails in Fig. 4, the failure canbe detected by router C , which is the Negotiator Router.Then, router C updates the hash ring and redistribute keys ofrouterA, for example, key value 3 to routerC and key value 11to router D. Thus, with consistent hashing, incoming Interestpackets and cached content for the failed router can easily bemapped to other routers with minimal configuration changes.

Algorithm 4 Router Addition1: An = An ∪ {r}.2: for ∀i ∈ An do3: Cacluate cache weight: ωi =

Si∑j∈An Sj

.4: Cacluate key weight: νi = |Kn

i |/|Kn|.

5: while νi > ωi do6: Randomly select a key k ∈ Kn

i .7: Kn

i ← Kni \ {k}.

8: Knr ← Kn

r ∪ {k}.9: Cacluate νj = |Kn

j |/|Kn|, ∀j ∈ {i, r}.

In the case of router addition to the group, the NegotiatorRouter updates the keys distributed among the routers asdescribed in Alg. 4. First, it calculates the cache and keyweights of each router which is proportional to individualcache and key sizes, respectively (lines 3 and 4). Next, thetwoweights are compared, and the keys are redistributed untilthe weights become equal (lines 5 to 9). Then, the NegotiatorRouter distributes the updated new hash ring map to all therouters.

V. CACHE DECISION PROCESSOur objective is to improve cache hit probability and reduceaccess delay of content chunks. To achieve this, we proposea cache decision process which involves two steps: i) Cachedecision when the Interest packet arrives, and ii) Caching theData packet according to the cache decision. Cache decisionis made according to the randomized online algorithm basedon the Ski-Rental problem.

A. SKI-RENTAL PROBLEM PRELIMINARIESIn the Ski-Rental problem, a person decides whether to rent(paying the repeating cost) or buy (paying the one-time cost)skis. In the offline case (complete information is known), the

VOLUME 6, 2018 59685

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

person knows in advance how many days he will use the skis.In the online case (complete information is unknown), theperson does not know the number of skiing days in advance.Thus, the decision made by complete information can givethe optimal solution, whereas the online algorithm cannotbecause it lacks complete information. The performance ofan online solution can be evaluated by a competitive analysisas follows:Definition 1: Let υ(σ ) denote the cost of an online algo-

rithm υ on input sequence σ . Let υ∗(σ ) denote the cost of anoptimal, offline algorithm on input σ . An online algorithm υ

is θ -competitive if for any sequence,

υ(σ ) ≤ θυ∗(σ ). (2)The competitive ratio of the online algorithm is then

defined as

θ = maxσ

υ(σ )υ∗(σ )

. (3)

The Ski-Rental algorithm can be categorized into determin-istic and randomized approaches [35].

FIGURE 6. An example of deterministic and randomized Ski-Rentalproblem based cache decision process.

B. SKI-RENTAL-BASED CACHINGAND REPLACEMENT SCHEMEThe Ski-Rental problem can be mapped into the NDNcaching and replacement process, by considering a router asperson, ski buying as caching content chunk (the buying costis mapped to the cache memory usage cost) and ski rentingas not caching the chunk (the renting cost is analogous to thecost to download the content from the server). In the onlinecase, the router makes a decision whether to store the chunkor not, at the time of the Interest arrival without knowing thefuture popularity of each content chunk. Themain goal of thisalgorithm is to reduce the cache miss without knowing thefuture popularity of each content chunk. As shown in Fig. 6,pijv (the caching parameter) lies between 0 and 1 and is mono-tonically non-decreasing over time. In the deterministic solu-tion, the chunk caching decision (buying decision) is madewhen pijv value reaches to 1. Thus, the competitive ratio of thedeterministic solution is 2 according to (3), i.e., in the worstcase scenario, the router spends two times of the optimal costto cache the chunk. In the randomized solution, a threshold israndomly selected between 0 and 1 (γ ijv in the example shownin Fig. 6) for the pijv value. The caching (buying) decisionis made when pijv value is greater than or equal to γ ijv . Therandomized solution reduces the competitive ratio from 2 toe/(e − 1). In this paper, we use the randomized solution toimprove the performance (analysis is shown in section V-E)by keeping the competitive value less than 2. To obtain

the randomized solution, we use the threshold γ ijv , which isgenerated according to a uniform random distribution. If theretrieve cost pijv for the incoming chunk is less than γ ijv , therouter does not cache this incoming chunk. Otherwise, therouter caches the new incoming chunk.

Algorithm 5 Randomized Online Cache Decision

1: Input: k-th arrival of request dkij , where k = 1, . . . ,K2: Output: Cache the chunk fij or not3: Initialize: pijv , q

ijv ,C

ijv and x ijv ← 0

4: At every time t , updateC ijv (t) with (4) and q

ijv (t) with (10),

where t= 0, 1, 2, · · · , t − 15: if C ij

v is 0 then6: Update the C ij

v ← Cv(t) and qijv ← qv(t)

7: else8: Do not update the C ij

v and qijv9: Update pijkv with (5)

10: if pijkv ≥ γijkv then

11: x ijv ← 1.

C. CACHE DECISION PROCESS WHENINTEREST PACKET ARRIVESAccess nodes and routers use the same randomized onlinealgorithm (Alg. 5) for cache decision. We first define thecache space cost to store one chunk at the node v as follows:

C ijv = λv(t), (4)

where Cv(t) is initialized with the value of τ at time t = 0.Then, from time t = 1,Cv(t) is updated with λv(t)−λv(t−1),which is the number of inter-arrival requests within everytime frame. The probability to cache the chunk fij at thenode v is updated on every arrival of the requests. Cachingprobability at the k-th arrival of the request is updated asfollows:

pijv (k) = min

[1, pijv (k − 1)

(1+

1

C ijv

)+

1

qijvCijv

](5)

where qijv is updated according to (10), which is discussed insection V-E. Then, we define the threshold γ ijv (k) to makethe cache decision for chunk j of content i at the node v.In addition, γ ijv (k) has uniform distribution between 0 and 1.

Alg. 5 shows the randomized online algorithm for cachedecision and replacement, which works on every arrival ofthe Interest packet. Initially, pijv , x

ijv ,C

ijv and x ijv are set to 0

(line 3). At every time t , the value of Cv(t) is updated with (4)and qv(t) is updated with (10) (line 4). If an Interest packet fornew content arrives (line 5), the node sets C ij

v with the valueof Cv(t), and sets qijv with the value of qv(t) (line 6). Then,the probability to cache the chunk pijv (k) is updated with (5)(line 9). Next, node v runs the uniform random generator toset the value of γ ijv (k). If the probability to cache the chunkpijv (k) is greater than the threshold γ

ijv (k) (line 9), the node sets

cache decision bit x ijv ← 1 (line 10).

59686 VOLUME 6, 2018

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

D. CACHING PROCESS WHEN DATA PACKET ARRIVESThe randomized online caching algorithm updates x ijv at thearrival of the Interest packet at node v. The cache decisionprocess for the Data packet arrival is shown in Alg. 6. On thearrival of the Data packet or chunk, node v checks x ijv fromits PIT. If x ijv = 1 (line 4), node v caches the content chunk fijin its content store. In this case, if the cache of the node v isfree, node v directly cache the content chunk fij. Otherwise,node v removes the Least Recently Used (LRU) chunk from itcontent store (line 8) and keeps the new chunk fij at the mostrecently used position (line 9). If x ijv = 0, node v relays thechunk fij (line 6).

Algorithm 6 Cache Decision1: Input: The arrival of chunk dij2: Output: Relay and store chunk fij or not3: Relay fij to requested nodes.4: if x ijv = 1 then5: if Fv

i < sv then6: Fv

i ← Fvi ∪ {fij}

7: else8: Fv

i ← Fvi \ {LRU (fij)}

9: Fvi ← Fv

i ∪ {fij}

10: else11: Relay fij to requested nodes.

E. COMPETITIVE RATIOTo analyze the performance improvement of our proposal,we use the competitive ratio given in Def. V-A. We firstformulate the cache decision and replacement problem as aninteger linear program, where we minimize the cost to cachethe chunk fij or minimize the cache miss. The primal problemis given by:

minimize:p, z

C ijv p

ijv (k)+

K∑k=1

zijv (k)

subject to: pijv (k)+ zijv (k) ≥ 1,

pijv (k) ∈ {0, 1}, zijv (k) ∈ {0, 1}, ∀k. (6)

In (6), the first term of the objective function is the cachingcost and the second term is the requesting cost. There are twocontrol variables: pijv (k) which determines whether to cachethe chunk, and zijv (k) which determines whether to request thechunk. The variable x ijv (k) ← 1 when node v stores contentchunk fij and x

ijv (k) ← 0 otherwise. The variable zijv (k) ← 1

when the node v requests chunk dij and 1 6 k 6 K .Second, we relax the problem (7) to obtain an equivalent

linear program as follows:

minimize:p, z

C ijv p

ijv (k)+

K∑k=1

zijv (k)

subject to: pijv (k)+ zijv (k) ≥ 1, k = 1, . . . ,K ,

pijv (k) > 0, ∀k,

0 ≤ pijv (k) ≤ 1, 0 ≤ zijv (k) ≤ 1, (7)

Third, the dual problem of (7) can be written as follows:

maximize:y

K∑k=1

yijv (k)

subject to:K∑k=1

yijv (k) 6 C ijv ,

0 6 yijv (k) 6 1, k = 1, . . . ,K , (8)

where, the dual variable yijv (k) corresponds to the k-th requestof content chunk fij at node v. The first constraint of dualproblem (8) guarantees that the request cost of each chunk atnode v cannot exceed the caching cost. The primal variablespijv (k) and z

ijv (k) for the k-th request arrival are in the interval

between [0, 1]. pijv (k) is the geometric sequence and it isupdated by using (5). Thus, after k-th request,

pijv (k) =1

qijv·

(1+ 1

C ijv

)C ijv− 1

, (9)

where we set qijv as follows:

qijv =

(1+

1

C ijv

)C ijv− 1, (10)

to guarantee that pijv (k) will become 1 after k-th requests.The variable zijv (k) which is related to requesting cost for k-threquest of content chunk fij is denoted as follows:

zijv (k) = 1− pijv (k). (11)

Algorithm 7 Competitive Ratio Analysis

1: Initialize: Set pijv , zijv and y

ijv ← 0

2: The arrival of k th request d ijv (k), where k = 1, 2, · · · ,K3: while pijv (k) < 1 do4: update the pijv (k) with (5) and zijv (k) with(11)5: if pijv (k) ≥ γ

ijv (k) then

6: set pijv (k)← 17: else8: yijv (k)← yijv (k − 1)+ 1

Alg. 7 shows the steps to run the linear program to analyzethe competitive ratio of randomized online algorithm. Next,the cost of randomized algorithm can be analyzed using aprimal-dual analysis with the approximate complementaryslackness conditions given in [35]. On the arrival of the k-threquest, the probability of caching the chunk is pijv (k), and thecost of requesting the chunk fij isC

ijv∑K

k=1 pijv (k) = C ij

v pijv (k).

Then, the probability of forwarding the request for chunk fijon the k-th request is 1−

∑Kk=1 p

ijv (k). Therefore,

zijv (k) = 1−K−1∑k=1

pijv (k) ≥ 1−K∑k=1

pijv (k), (12)

and the probability of forwarding the request on the k-threquest is at most pijv (k), corresponding to the second term in

VOLUME 6, 2018 59687

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

the primal objective function. Thus, by the linearity of expec-tation, for any number of requests forwarding, the expectedcost of the randomized algorithm is at most the cost of thefractional solution. Thus, by [35, Th. 2.3], the algorithm ise/(e− 1)-competitive.

VI. PERFORMANCE EVALUATIONIn this section, we evaluate the performance of the proposedscheme using the chunk-level simulator, NDNSim [7], whichwas developed under the Omnet++ simulator. The simula-tion code is available in [39]. The parameters used in theexperiments are given in Table 2. We analyze the proposedscheme by deploying homogeneous group size, heteroge-neous cache size, and delay. We begin with a discussionof the configuration of the simulation environment. Next,we explain the performance metrics to compare the results ofour proposed scheme with other schemes. Finally, we discussthe simulation results.

A. CONFIGURATION1) TOPOLOGYFig. 7 shows the topologies used in the simulation. Theblack circles represent content routers, white circles representaccess nodes, and the yellow circle represents the contentserver. For the tree topology shown in Fig 7(a), all of thecontent routers are grouped as a single group and connectedwith 8 access nodes. For the grid topology displayed inFig 7(b), all four content routers are grouped into one groupand connected with 12 access nodes. For the grid topologyshown in Fig 7(c), all 11 content routers are formed into twoseparated groups. Grey color circles group is composed of5 content routers and black color circles group is composedof 6 content routers. There are ten access nodes which areevenly distributed among black and grey nodes.

FIGURE 7. Topology used in simulation.

2) CONTENT DISTRIBUTIONTo get a realistic estimation of the catalog size, we choosevideo contents similar to YouTube, which has an approxi-mate catalog size of about 108 [36], [37], the average contentsize of 10 megabytes space [38], and follows the geomet-ric distribution. Thus, the overall catalog size is about 1petabytes (PB).

3) CACHE SIZEConsiderations for the cache size of the routers are welldiscussed in [37]. In this paper, we use a heterogeneous cache

TABLE 2. Parameters used in the experiments.

size for the whole network to simulate consistent hashingkey distribution. Then, we consider how much cache sizeshould be allocated for the whole autonomous system. Weassign 5% of the catalog size for the whole network cache.

4) DELAYFor simplicity, we consider topologies with a uniformdelay (1ms). We also assume that the links are non-congestedand have infinite bandwidth. So, the network delay matchesthe propagation delay of each link.

5) GENERATING USER REQUESTSTo generate user requests, we first consider the object pop-ularity. YouTube traffic analysis [36]–[38] shows that thecontent requests (i.e., popularity) follow Zipf distribution.Hence, we use Zipf distribution for the content requests in oursimulation to reflect the real-world scenario. Zipf distributionis the discrete version of Pareto distribution which explainsthe famous 80-20 rule (also known as the Pareto principle)and has a variety of application in social, scientific, economicand many other types of observable phenomena. Thus, in thispaper, the object popularity follows the Zipf distribution,where the probability of content i is denoted as

P(i) =1/iα∑|F |j=1 1/j

α, (13)

where i is the rank of the i-th object.To run the simulation, we assume that the object popu-

larity follows a Zipf distribution range (α = 0.9 to 1.5).Requests arrive following a Poisson process, P {N (t) = n} =(λt)nn! e−λt . Also, user’s devices are implemented with

CCN protocol.

6) STATISTICS COLLECTIONThe simulation starts with empty caches. When the cachespaces become full, every router samples its hit rate at every0.1 secs simulated time and computes the variance of thecollected samples at every 60 window. Only when the cachehit variance falls under a given threshold (0.05) does a routerdeclare itself stabilized. We gather statistics after the routersreached the steady state.

59688 VOLUME 6, 2018

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

TABLE 3. Strategies compared in the experiments.

B. PERFORMANCE METRICSWe choose the probability of hit, probability of inner hitand workload on content store measured at the router asnetwork-centric Key Performance Indicator (KPI) metrics.We choose the access delay to retrieve the content measuredat the user device as user-centric KPI.

1) PROBABILITY OF HIT (NETWORK CENTRIC)Typically, the performance of the cache is measured in termsof the probability that requested chunks are found at a givencontent store. The probability of a cache hit measures thehit and miss probabilities to determine how much trafficthe network can eliminate. The probability of a cache hit iscalculated as follows:

Phit =

∑v∈V N

hitv∑

v∈V Nhitv + Nmiss

v, (14)

where Phit is the probability of hit,∑

v∈V Nhitv is the total

number of hit at routers, access nodes, and∑

v∈V Nhitv +N

missv

is the total number of cache hit and cache miss.

2) PROBABILITY OF INNER HIT (NETWORK CENTRIC)We also consider inner hits, which measure how much work-load the network can handle. The inner hit is calculated asfollows:

Pinner-hit =

∑u∈U Fu −

∑w∈Vs Fw∑

u∈U Fu, (15)

where∑

u∈U Fu is the total chunk downloaded by all users,and

∑w∈Vs Fw is the total number of chunks that are down-

loaded from the content server.

3) CONTENT STORE WORKLOAD (NETWORK CENTRIC)The content store load measures the average workload effectby the Interest packets to the content store of router, BS,and SBS. The average load on content store is calculated asfollows:

W =

∑w∈Vs N

reqw∑

v∈V Nreqv

, (16)

whereN reqw is the total number of requests the nodew searches

in its content store, and N reqv is the total number of requests

that arrived at node v.

4) BANDWIDTH CONSUMPTION (NETWORK CENTRIC)Also, we measure the impact on bandwidth consumptionbetween content routers by measuring the average number

of data packets passing through each content router in eachevery one second.

5) AVERAGE DELAY (USER CENTRIC)In addition, we measure the latency to download the contentsfrom the user side, which is calculated as follows:

T =1U·

∑u∈U

∑j∈F t ijuFu

, (17)

where U is the total number of users, t iju is the delay inaccessing the content fij by the user u, and Fu is the totalnumber of downloaded content chunk by the user u.

TABLE 4. Performance improvement compared to benchmark (SPR-LCE).

C. SIMULATION RESULTSFig. 7 shows the network topologies used for the perfor-mance evaluation. In order to know which scheme con-tributes more, we separately run Ski-Rental scheme, Con-sistent Hashing scheme and Proposed (combined) scheme,then the results are shown in Table 4. Overall, as shown inresults from Table 4, the combined Ski-Rental and ConsistentHashing scheme showed better performance than individualSki-Rental scheme and Consistent Hashing scheme in termsof the cache hit, inner cache hit, Content Store workloadand average delay. This is because it reduces the storageof duplicate contents among neighbor routers. Moreover,the combined scheme only caches contents that seem tobe popular. However, the Consistent Hashing scheme andcombined scheme use more bandwidth than other schemes.This is because Consistent Hashing requires control over-head to remove duplicate contents in one-hop neighbors.

VOLUME 6, 2018 59689

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

FIGURE 8. Probability of cache hit comparison between proposed scheme and others.

This leads to more packet transmission in the entire net-work. Hence, Consistent Hashing increases cache hit atthe expense of increased bandwidth usage. The resultsfrom simulation have shown that Ski-Rental based cachingscheme is necessary to improve the performance of Consis-tent Hashing based caching and forwarding scheme. Then,we discuss network-centric and user centric-performanceresults compared to others scheme shown in Table 3 are asfollows.

1) PROBABILITY OF HIT (NETWORK CENTRIC)Fig. 8 shows the results of chunk level cache hit rate inthree network topologies. Higher probability of hit meansbetter performance. As we expected, in the Fig. 8, the hitrate of all schemes increase with α (Zipf parameter) val-ues from 0.9 to 1.5 because users shift requesting from alarge variety of content to requesting from a small amountof popular content. Overall, our proposed scheme showedbetter performance (up to 72%) than others. The resultsof the Tree topology, in Fig. 8(a), show that our pro-posed scheme outperforms SPR-LCD by 42%, SPR-LCEby 68% and SPR-PROB by 51%, respectively. The results ofGrid topology, in Fig. 8(b) show that our proposed schemeoutperforms SPR-LCD by 14%, SPR-LCE by 35% andSPR-PROB by 23%, respectively. The results of Abilenetopology, in Fig. 8(c), show that our proposed scheme outper-forms SPR-LCD by 43%, SPR-LCE by 72% and SPR-PROBby 53%, respectively. SPR-LCE gives the lowest cache hitbecause contents are cached at every node without con-sidering the popularity of contents. SPR-PROB gives thebetter performance than the SPR-LCE because it consid-ers probability to cache the content based on the size ofcache space and retrieving distance from the Content Storeof intermediate nodes or Content Server. SPR-LCD givesthe better performance than the SPR-PROB because popularcontents are replicated to the downstream nodes whenever

cache hit occurs. Our proposed scheme improves the perfor-mance of cache hit than others because of two facts. First,the node forwards the requests only to the potential locationusing coordinated forwarding. Second, the node stores onlythe popular contents using Ski-Rental based online cachingalgorithm.

2) PROBABILITY OF INNER CACHE HIT (NETWORK CENTRIC)The inner cache hit comparison is shown in Fig. 9, where wemeasure the percent of contents provided from the in-networkcache instead of provided from the content server. The resultsof the Tree topology, Fig. 9(a), show that the proposed schemeis 15% better than SPR-LCD, 23% better than SPR-LCEand 17% better than SPR-PROB. The results of Grid topol-ogy, in Fig. 9(b), show that the proposed scheme outper-forms 8% than SPR-LCD, 21% than SPR-LCE and 14% thanSPR-PROB. The results of Abilene topology, in Fig. 9(c),show that the proposed combined scheme is 19% better thanSPR-LCD, 24% better than SPR-LCE and 16% better thanSPR-PROB. Among different schemes, SPR-LCE gives thelowest inner hit probability because it stores duplicate contenton the request routes, and this affects the cache utilization ofeach router. SPR-PROB is better than the SPR-LCE becauseit stores the content depending on the location of the router,and reduces the storage of duplicate content. SPR-LCD isbetter than the SPR-LCE and SPR-PROB because it reducesthe storing of duplicate content and stores the popular itemsnearer to the users. Our proposed caching scheme outper-forms the others because it eliminates the duplicate content byusing consistent-hashing-based filtering technique and storesthe valuable content by using Ski-Rental based online cachedecision algorithm.

3) CONTENT STORE WORKLOAD (NETWORK CENTRIC)Fig. 10 shows the results of average workload on the nodesto find the relevant chunk in their content store. Lower

59690 VOLUME 6, 2018

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

FIGURE 9. Inner cache hit comparison between proposed scheme with others.

FIGURE 10. Content Store work load comparison between proposed scheme and others.

workload means better performance. The results of the Treetopology, in Fig. 10(a), our proposed scheme is 32% bet-ter than SPR-LCD, 39% better than SPR-LCE and 35%better than SPR-PROB. The results of the Grid topology,in Fig. 10(b), show that our proposed scheme is 9%better thanSPR-LCD, 17% better than SPR-LCE and 14% better thanSPR-PROB. The results of theAbilene topology, in Fig. 10(c),our proposed scheme is 33% better than SPR-LCD, 41%better than SPR-LCE and 37% better than SPR-PROB. FromFig. 10, the workload of the content store for finding therelevant content depends on the forwarding strategies thanthe caching strategies. SPR performance is lower than theproposed scheme because it forwards the Interest packetsto the content server via the shortest path. Thus, all of thenodes on the Interest packet forwarding path need to findthe requested content on their CS. Our proposed scheme hasbetter performance than the others because of the coordinatedInterest packets forwarding scheme, where the cluster of therouters shares the workload to search the requested contents.

4) BANDWIDTH USAGE (NETWORK CENTRIC)Fig. 11 shows the average bandwidth usage comparisonbetween the proposed scheme and others. A lower band-width usage means better performance. The results of Treetopology in Fig. 11(a), show that on average, our proposedscheme uses bandwidth 31% more than SPR-LCD, 2% lessthan SPR-LCE and 12% more than SPR-PROB. The resultsof Grid topology in Fig. 11(b), show that on average, ourproposed scheme uses bandwidth 38% more than SPR-LCD,5% less than SPR-LCE and 7% more than SPR-PROB. Theresults of Abilene topology, in Fig. 11(c), show that on aver-age, our proposed scheme consumes bandwidth 31% morethan SPR-LCD, 10% less than SPR-LCE and 3% more thanSPR-PROB. Overall, the bandwidth usage of our proposedscheme is lower than SPR-LCE because the proposed schemeimproves the cache utilization by eliminating the storingduplicate contents among neighbor routers. Therefore, thechance to satisfy the requested contents at the cache cluster ishigher than the others. On the other hand, due to Consistent

VOLUME 6, 2018 59691

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

FIGURE 11. Average bandwidth usage comparison between proposed scheme with others.

FIGURE 12. Average content access delay comparison between proposed scheme with others.

Hashing, our proposed scheme consumes more bandwidththan SPR-LCD and SPR-PROB.

5) AVERAGE ACCESS DELAY (USER CENTRIC)Fig. 12 compares the average access delay, which measuresthe average round trip time to receive the requested content.A lower access delay means better performance. In Treetopology, Fig. 12(a) shows that the proposed scheme is9% better than SPR-LCD, 25% better than SPR-LCE and16% better than SPR-PROB. In Grid topology, Fig. 12(b)shows that the proposed scheme is 9% better than SPR-LCD,25% better than SPR-LCE and 17% better than SPR-PROB.In Abilene topology, Fig. 12(c) shows that the results showthat our proposed scheme is 2% better than SPR-LCD, 21%better than SPR-LCE and 11% better than SPR-PROB. Over-all, SPR-LCE suffered from the highest latency to downloadthe requested content because LCE caching policy storesall of the passing contents on every node. Thus, the cacheutilization is degraded, and that effect leads to increase in the

average delay to retrieve contents. SPR-PROB scheme givesbetter results than SPR-LCE because it reduces the storing ofduplicate contents and caches the popular items nearer to theusers. SPR-LCD is better than the SPR-LCE and SPR-PROBbecause it stores the content depending on the location ofthe router, and reduces the storage of duplicate contents. Ourproposed scheme is slightly better than the SPR-LCDbecausethe proposed forwarding strategy supports router to find therequested contents which are stored off-path and proposedSki-Rental based caching scheme keeps the potential popularcontent on each node.

VII. CONCLUSIONIn this paper, the joint caching and Interest forwardingscheme for NDN is proposed to store digital content effi-ciently and effectively by utilizing the advantages of coop-erative and non-cooperative caching. Moreover, the proposedforwarding scheme is able to forward requests directly to thecustodian router without flooding. Furthermore, any nodes(core routers and access routers) that are not the custodian

59692 VOLUME 6, 2018

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

of the arriving request can skip the Content-Store-Checkingstep. Besides, the proposed scheme prevents the replacingof non-valuable content with valuable content with the helpof Ski-Rental-problem-inspired randomized online cachingpolicy. As a result, the proposed scheme improves the cachehit ratio, and inner cache hit. In addition, it reduces theworkload of the router’s content store and average delay todownload the requested content. The proposed mechanismwas intensively simulated, and the experimental results showthat with our proposed mechanism, the cache hits becomehigher while the workload on router’s content store becomeslower. Further, the inner hits are higher while the averagewaiting time to get contents is much lower than the othersimilar proposals in the literature. In this paper, our focus isto improve the cache utilization without knowing the futurepopularity of contents. Therefore, as future work, we willfocus on the caching based on content’s popularity prediction.

REFERENCES[1] CISCO Visual Networking Index (VNI). Accessed: Jan. 10, 2018. [Online].

Available: www.cisco.com[2] R. Jain, ‘‘Internet 3.0: Ten problems with current Internet architecture

and solutions for the next generation,’’ in Proc. IEEE Mil. Commun.Conf. (MILCOM), Washington, DC, USA, Oct. 2006, pp. 1–9.

[3] I. U. Din, S. Hassan, M. K. Khan, M. Guizani, O. Ghazali, and A. Habbal,‘‘Caching in information-centric networking: Strategies, challenges, andfuture research directions,’’ IEEE Commun. Surveys Tuts., vol. 20, no. 2,pp. 1443–1474, 2nd Quart., 2018.

[4] V. Jacobson, D. K. Smetters, J. D. Thornton, M. F. Plass, N. H. Briggs, andR. L. Braynard, ‘‘Networking named content,’’ in ACMProc. 5th Int. Conf.Emerg. Netw. Exp. Technol., Rome, Italy, Dec. 2009, pp. 1–12.

[5] D. Karger, E. Lehman, T. Leighton, R. Panigrahy,M. Levine, andD. Lewin,‘‘Consistent hashing and random trees: Distributed caching protocols forrelieving hot spots on the World Wide Web,’’ in Proc. 29th Annu. ACMSymp. Theory Comput., El Paso, TX, USA, May 1997, pp. 654–663.

[6] G. DeCandia et al., ‘‘Dynamo: Amazon’s highly available key-valuestore,’’ ACM SIGOPS Operating Syst. Rev., vol. 41, no. 6, pp. 205–220,Oct. 2007.

[7] G. Rossini, D. Rossi, and G. Rossini, ‘‘ccnSim: An highly scalableCCN simulator,’’ in Proc. IEEE ICC, Budapest, Hungary, Jun. 2013,pp. 2309–2314.

[8] N. Laoutaris, H. Che, and I. Stavrakakis, ‘‘The LCD interconnection ofLRU caches and its analysis,’’ Perform. Eval., vol. 63, no. 7, pp. 609–634,Jul. 2006.

[9] I. Psaras, W. K. Chai, and G. Pavlou, ‘‘Probabilistic in-network cachingfor information-centric networks,’’ in Proc. ACM 2nd ICN Workshop Inf.Centric Netw., Helsinki, Finland, Aug. 2012, pp. 55–60.

[10] I. Psaras, W. K. Chai, and G. Pavlou, ‘‘In-network cache managementand resource allocation for information-centric networks,’’ IEEE Trans.Parallel Distrib. Syst., vol. 25, no. 11, pp. 2920–2931, Nov. 2014.

[11] M. Zhang, J. Tang, Y. Rao, H. Luo, and H. Zhang, ‘‘Degree-based proba-bilistic caching in content-centric networking,’’ China Commun., vol. 14,no. 3, pp. 158–168, Mar. 2017.

[12] K. Li, C. Yang, Z. Chen, and M. Tao, ‘‘Optimization and analysis ofprobabilistic caching in N -tier heterogeneous networks,’’ IEEE Trans.Wireless Commun., vol. 17, no. 2, pp. 1283–1297, Feb. 2018.

[13] A. Ndikumana et al., ‘‘In-network caching for paid contents in content cen-tric networking,’’ in Proc. IEEE Global Commun. Conf. (GLOBECOM),Dec. 2017, Singapore, pp. 1–6.

[14] H. Wang, J. Hu, G. Min, W. Miao, and N. Georgalas, ‘‘Cost-aware opti-misation of cache allocation for information-centric networking,’’ in Proc.IEEE Global Commun. Conf., Singapore, Dec. 2017, pp. 1–6.

[15] E. Zeydan et al., ‘‘Big data caching for networking: Moving from cloud toedge,’’ IEEE Commun. Mag., vol. 54, no. 9, pp. 36–42, Sep. 2016.

[16] W.-X. Liu, J. Zhang, Z.-W. Liang, L.-X. Peng, and J. Cai, ‘‘Contentpopularity prediction and caching for ICN: A deep learning approach withSDN,’’ IEEE access, vol. 6, pp. 5075–5089, Jan. 2018.

[17] S. Ullah, K. Thar, and C. S. Hong, ‘‘Management of scalable video stream-ing in information centric networking,’’ Multimedia Tools Appl., vol. 76,no. 20, pp. 21519–21546, Oct. 2017.

[18] S. Ullah, K. Thar, M. G. R. Alam, J. H. Son, J. W. Lee, and C. S. Hong,‘‘Delivering scalable video streaming in ICN enabled long term evolutionnetworks,’’ in Proc. IEEE 18th Netw. Oper. Manage. Symp. (APNOMS),Kanazawa, Japan, Oct. 2016, pp. 1–4.

[19] Y. An and X. Luo, ‘‘An in-network caching scheme based onenergy efficiency for content-centric networks,’’ IEEE Access, vol. 6,pp. 20184–20194, Apr. 2018.

[20] J. M. Wang, J. Zhang, and B. Bensaou, ‘‘Intra-AS cooperative caching forcontent-centric networks,’’ in Proc. 3rd ACM SIGCOMM Workshop Inf.Centric Netw., Hong Kong, Aug. 2013, pp. 61–66.

[21] V. Sourlas, P. Flegkas, and L. Tassiulas, ‘‘A novel cache aware rout-ing scheme for information-centric networks,’’ Comput. Netw., vol. 59,pp. 44–61, Feb. 2014.

[22] V. Sourlas, I. Psaras, L. Saino, and G. Pavlou, ‘‘Efficient hash-routing anddomain clustering techniques for information-centric networks,’’ Comput.Netw., vol. 103, pp. 67–83, Jul. 2016.

[23] Z. Li and G. Simon, ‘‘Time-shifted TV in content centric networks: Thecase for cooperative in-network caching,’’ in Proc. IEEE Int. Conf. Com-mun. (ICC), Kyoto, Japan, Jun. 2011, pp. 1–6.

[24] L. Saino, I. Psaras, and G. Pavlou, ‘‘Hash-routing schemes for informationcentric networking,’’ in Proc. 3rd ACM SIGCOMMWorkshop Inf. CentricNetw., Hong Kong, Aug. 2013, pp. 27–32.

[25] S. Wang, J. Bi, and J. Wu, ‘‘Collaborative caching based on hash-routingfor information-centric networking,’’ ACM SIGCOMMComput. Commun.Rev., vol. 43, no. 4, pp. 535–536, Aug. 2013.

[26] K. Thar, S. Ullah, and C. S. Hong, ‘‘Consistent hashing based cooperativecaching and forwarding in content centric network,’’ in Proc. IEEE 16thAsia–Pacific Netw. Oper. Manage. Symp., Hsinchu, Taiwan, Sep. 2014,pp. 1–4.

[27] K. Thar, S. Ullah, R. Haw, T. LeAnh, T. Z. Oo, and C. S. Hong, ‘‘Hybridcaching and requests forwarding in information centric networking,’’ inProc. IEEE 17th Asia–Pacific Netw. Oper. Manage. Symp. (APNOMS),Busan, South Korea, Aug. 2015, pp. 203–208.

[28] G. Rossini and D. Rossi, ‘‘Coupling caching and forwarding: Benefits,analysis, and implementation,’’ in Proc. 1st ACM Int. Conf. Inf. CentricNetw., Paris, France, Sep. 2014, pp. 127–136.

[29] L. Wang, S. Bayhan, J. Ott, J. Kangasharju, A. Sathiaseelan, andJ. Crowcroft, ‘‘Pro-diluvian: Understanding scoped-flooding for contentdiscovery in information-centric networking,’’ in Proc. ACM 2nd Int. Conf.Inf. Centric Netw., San Francisco, CA, USA, Sep. 2005, pp. 9–18.

[30] D. Posch, B. Rainer, and H. Hellwagner, ‘‘SAF: Stochastic adaptive for-warding in named data networking,’’ IEEE/ACM Trans. Netw., vol. 25,no. 2, pp. 1089–1102, Apr. 2017.

[31] X. Liu, K. Zheng, J. Zhao, X.-Y. Liu, X. Wang, and X. Di, ‘‘Information-centric networks with correlated mobility,’’ IEEE Trans. Veh. Technol.,vol. 66, no. 5, pp. 4256–4270, May 2017.

[32] K. Zheng, Y. Cui, X. Liu, X. Wang, X. Jiang, and J. Tian, ‘‘Asymp-totic analysis of inhomogeneous information-centric wireless networkswith infrastructure support,’’ IEEE Trans. Veh. Technol., vol. 67, no. 6,pp. 5245–5259, Jun. 2018.

[33] S. Podlipnig and L. Böszörmenyi, ‘‘A survey of Web cache replacementstrategies,’’ ACM Comput. Surv., vol. 35, no. 4, pp. 374–398, Dec. 2013.

[34] C. Yi, A. Afanasyev, L. Wang, B. Zhang, and L. Zhang, ‘‘Adaptive for-warding in named data networking,’’ ACM SIGCOMM Comput. Commun.Rev., vol. 42, no. 3, pp. 62–67, Jun. 2012.

[35] N. Buchbinder and J. S. Naor, ‘‘The design of competitive online algo-rithms via a primal–dual approach,’’ Found. Trends Theor. Comput. Sci.,vol. 3, nos. 2–3, pp. 93–263, May 2009.

[36] M. Cha, H. Kwak, P. Rodriguez, Y.-Y. Ahn, and S. Moon, ‘‘I tube, youtube, everybody tubes: Analyzing the world’s largest user generated con-tent video system,’’ in Proc. 7th ACM SIGCOMM Conf. Internet Meas.,San Diego, CA, USA, Oct. 2007, pp. 1–14.

[37] G. Rossini and D. Rossi, ‘‘A dive into the caching performance of contentcentric networking,’’ in Proc. IEEE 17th Int. Workshop Comput. AidedModeling Design Commun. Links Netw. (CAMAD), Barcelona, Spain,Sep. 2012, pp. 105–109.

[38] P. Gill, M. Arlitt, Z. Li, and A. Mahanti, ‘‘Youtube traffic characterization:A view from the edge,’’ inProc. 7th ACMSIGCOMMConf. InternetMeas.,San Diego, CA, USA, Oct. 2007, pp. 15–28.

[39] Simulations Source Code. Accessed: Aug. 20, 2018. [Online]. Available:github.com/kyithar/Ski-Rental

VOLUME 6, 2018 59693

K. Thar et al.: Online Caching and Cooperative Forwarding in ICN

KYI THAR received the Bachelor of ComputerTechnology degree from the University of Com-puter Studies, Yangon, Myanmar, in 2007. He iscurrently pursuing the Ph.D. degree with theDepartment of Computer Science and Engineer-ing, Kyung Hee University, South Korea, forwhich he was awarded a scholarship for his grad-uate study in 2012. His research interests includename-based routing, in-network caching, multi-media communication, scalable video streaming,

wireless network virtualization, deep learning, and Future Internet.

NGUYEN H. TRAN (S’10–M’11) received theB.S. degree in electrical and computer engineeringfrom the Hochiminh City University of Technol-ogy in 2005 and the Ph.D. degree in electrical andcomputer engineering from Kyung Hee Universityin 2011. He was an Assistant Professor with theDepartment of Computer Science and Engineer-ing, Kyung Hee University, from 2012 to 2017.Since 2018, he has been with the School of Infor-mation Technologies, The University of Sydney,

where he is currently a Senior Lecturer. His research interests include apply-ing analytic techniques of optimization, game theory, and machine learningto cutting-edge applications, such as cloud and mobile-edge computing,datacenters, resource allocation for 5G networks, and Internet of Things.He received the Best KHU Thesis Award in engineering in 2011 and severalbest paper awards, including IEEE ICC 2016, APNOMS 2016, and IEEEICCS 2016. He receives the Korea NRF Funding for Basic Science andResearch from 2016 to 2023. He has been an Editor of the IEEETRANSACTIONS

ON GREEN COMMUNICATIONS AND NETWORKING since 2016.

SAEED ULLAH received the B.S. degree in infor-mation technology from IBMS/CS, Peshawar,in 2006, and the M.S. degree in information tech-nology from the National University of Sciencesand Technology, Islamabad, Pakistan, in 2010.He is currently pursuing the Ph.D. degree withthe Department of Computer Science and Engi-neering, Kyung Hee University, South Korea.His research interests include multimedia com-munication, scalable video streaming routing and

in-network caching, and Future Internet.

THANT ZIN OO received the B.Eng. degree inelectrical systems and electronics at MyanmarMaritimeUniversity, Thanlyin,Myanmar, in 2008,the B.S. degree in computing and information sys-tem from London Metropolitan University, U.K.,in 2008, and the Ph.D. degree in computer scienceand engineering at Kyung Hee University, SouthKorea, in 2017. He received scholarships fromKyung Hee University for his graduate studies andfromBritish Council for his undergraduate studies.

His research interests include wireless communications, network virtualiza-tion, data centers, sustainable energy, optimization techniques, and artificialintelligence.

CHOONG SEON HONG (S’95–M’97–SM’11)received the B.S. and M.S. degrees in electronicengineering from Kyung Hee University, Seoul,South Korea, in 1983 and 1985, respectively, andthe Ph.D. degree from Keio University in 1997.In 1988, he joined KT as a Member of Techni-cal Staff, where he was involved in broadbandnetworks. In 1993, he joined Keio University,Japan. He was with the Telecommunications Net-work Laboratory, KT, as a SeniorMember of Tech-

nical Staff and as the Director of the Networking Research Team until 1999.Since 1999, he has been a Professor with the Department of ComputerScience and Engineering, Kyung Hee University. His research interestsinclude Future Internet, ad hoc networks, network management, and networksecurity. He is a member of the ACM, the IEICE, the IPSJ, the KIISE, theKICS, the KIPS, and the OSIA. He has served as the general chair, the TPCchair/member, or an organizing committee member for international con-ferences, such as NOMS, IM, APNOMS, E2EMON, CCNC, ADSN, ICPP,DIM, WISA, BcN, TINA, SAINT, and ICOIN. He is currently an AssociateEditor of the IEEE TRANSACTIONS ON NETWORK AND SERVICE MANAGEMENT, theInternational Journal of Network Management, and the IEEE JOURNAL OF

COMMUNICATIONS AND NETWORKS, and an Associate Technical Editor of theIEEE Communications Magazine.

59694 VOLUME 6, 2018