Cooperative Caching in Wireless Multimedia Sensor Networksdelab.csd.auth.gr/papers/MONET08dkm.pdf · Cooperative Caching in Wireless Multimedia Sensor Networks ... is compared to

Mobile Netw ApplDOI 10.1007/s11036-008-0063-3

Cooperative Caching in Wireless MultimediaSensor Networks

Nikos Dimokas · Dimitrios Katsaros ·Yannis Manolopoulos

© Springer Science + Business Media, LLC 2008

Abstract The recent advances in miniaturization andthe creation of low-power circuits, combined withsmall-sized batteries have made the development ofwireless sensor networks a working reality. Lately, theproduction of cheap complementary metal-oxide semi-conductor cameras and microphones, which are able tocapture rich multimedia content, gave birth to what iscalled Wireless Multimedia Sensor Networks (WMSNs).WMSNs will boost the capabilities of current wirelesssensor networks, and will fuel several novel applica-tions, like multimedia surveillance sensor networks.WMSNs introduce several new research challenges,mainly related to mechanisms to deliver application-level Quality-of-Service (e.g., latency minimization).To address this goal in an environment with extremeresource constraints, with variable channel capacity andwith requirements for multimedia in-network process-ing, the caching of multimedia data, exploiting thecooperation among sensor nodes is vital. This article

Research supported by a Γ.Γ.E.T. grant in the context of theproject “Data Management in Mobile Ad Hoc Networks”funded by ΠY�AΓ OPA� II national research program.

N. Dimokas · Y. ManolopoulosDepartment of Informatics, Aristotle University,Thessaloniki, Greece

N. Dimokase-mail: [email protected]

Y. Manolopoulose-mail: [email protected]

D. Katsaros (B)Department of Computer & Communication Engineering,University of Thessaly, Volos, Greecee-mail: [email protected]

presents a cooperative caching solution particularlysuitable for WMSNs. The proposed caching solutionexploits sensor nodes which reside in “positions” ofthe network that allow them to forward packets orcommunicate decisions within short latency. These so-called “mediator” nodes are selected dynamically, so asto avoid the creation of hot-spots in the communicationand the depletion of their energy. The mediators arenot more powerful than the rest of the nodes, but theyhave some special role in implementing the cooperationamong the sensors. The proposed cooperative cachingprotocol includes components for locating cached dataas well as for implementing data purging out of thesensor caches. The proposed solution is evaluated ex-tensively in an advanced simulation environment, and itis compared to the state-of-the-art cooperative cachingalgorithm for mobile ad hoc networks. The results con-firm that the proposed caching mechanism prevails overits competitor.

Keywords cooperative caching · replacement policy ·multimedia · wireless sensor networks

1 Introduction

Wireless Sensor Networks [2, 18] (WSNs) haveemerged during the last years due to the advances inlow-power hardware design and the development ofappropriate software, that enabled the creation of tinydevices which are able to compute, control and com-municate with each other. A WSN consists of wirelesslyinterconnected devices that can interact with their en-vironment by controlling and sensing “physical” para-meters. WSNs attracted a huge interest from both the

Mobile Netw Appl

research community and the industry, that continuesto grow. This growing interest can be attributed to themany new exciting applications that were born as aresult of the deployment of large-scale WSNs. Suchapplications range from disaster relief, to environmentcontrol and biodiversity mapping, to machine surveil-lance, to intelligent building, to precision agriculture, topervasive health applications, and to telematics.

The support of such a huge range of applicationswill be (rather) impossible for any single realizationof a WSN. Nonetheless, certain common features ap-pear, in regard to the characteristics and the requiredmechanisms of such systems and the realization of thesecharacteristics is the major challenge faced by thesenetworks. The most significant characteristics shared bythe aforementioned applications concern [18]:

• Lifetime: Usually, sensor nodes rely on a batterywith limited lifetime, and their replacement is notpossible due to physical constraints (they lie inoceans or in hostile environments) or it is not in-teresting for the owner of the sensor network.

• Scalability: the architecture and protocols of sensornetworks must be able to scale up (or to exploit)any number of sensors.

• Wide range of densities: the deployment of sensornodes might not be regular and may vary signifi-cantly, depending on the application, on the timeand space dimension and so on.

• Data-centric networking: the target of a conven-tional communication network is to move bits fromone machine to another, but the actual purposeof a sensor network is to provide information andanswers, not numbers [17].

Recently, the production of cheap complementarymetal-oxide semiconductor (CMOS) cameras and mi-crophones, which can acquire rich media content fromthe environment, created a new wave into the evo-lution of WSNs. For instance, the Cyclops imagingmodule [29] is a light-weight imaging module whichcan be adapted to MICA21 or MICAz sensor nodes.Thus, a new class of WSNs came to the scene, theWireless Multimedia Sensor Networks (WMSNs) [1].These sensor networks, apart from boosting the existingapplication of WSNs, will create new applications: a)multimedia surveillance sensor networks which will becomposed by miniature video cameras [22] will be ableto communicate, to process and store data relevant tocrimes and terrorist attacks; b) traffic avoidance and

1http://www.xbow.com.

control systems will monitor car traffic and offer rout-ing advices to prevent congestion; c) industrial processcontrol will be realized by WMSNs that will offer time-critical information related to imaging, temperature,pressure, etc.

The novel applications of WMSNs challenged thescientific community because, as it is emphasizedin [1], these applications require us to rethink thecomputation-communication paradigm of traditionalWSNs. This paradigm has mainly focused on reduc-ing the energy consumption, targeting to prolong thelongevity of the sensor network. Though, the applica-tions implemented by WMSNs have a second goal, asimportant as the energy consumption, to be pursued;this goal is the delivery of application-level quality ofservice (QoS) and the mapping of this requirementto network layer metrics, like latency. This goal has(almost) been ignored in mainstream research effortson traditional WSNs.

The goal of Internet QoS in multimedia contentdelivery has been pursued in architectures like Diffservand Intserv, but these protocols and techniques do notface the severe constraints and hostile environmentof WSNs. In particular, WMSNs are mainly character-ized by:

• Resource constraints: sensor nodes are battery-,memory- and processing-starving devices.

• Variable channel capacity: the multi-hop natureof WMSNs, which operate in a store-and-forwardfashion because of the absence of base stations, im-plies that the capacity of each wireless link dependson the interference level among nodes, which isaggravated by the broadcasting operations.

• Multimedia in-network processing: is many appli-cations of WMSNs, a single sensor node is notable to answer a posed question, but several sensormust collaborate to answer it. For instance, a sensornode with a camera monitoring a moving group ofpeople, can not count their exact number and deter-mine their direction, but it needs the collaborationof nearby sensors in order to cover the whole extentof the group of people. Therefore, sensor nodesare required to store rich media, e.g., image, video,needed for their running applications, and also toretrieve such media from remote sensor nodes withshort latency.

Under these restrictions/requirements, the goal ofachieving application-level QoS in WMSNs becomesa very challenging task. There could be severalways to attack parts of this problem, e.g., channel-adaptive streaming [14], joint source-channel coding

http://www.xbow.com

Mobile Netw Appl

[11]. Though, none of them can provide solutions toall of the three aforementioned issues. In this paper,we investigate the solution of cooperative caching mul-timedia content in sensor nodes to address all threecharacteristics. In cooperative caching, multiple sensornodes share and coordinate cache data to cut com-munication cost and exploit the aggregate cache spaceof cooperating sensors. The plain assumption we make,is that each sensor node has a moderate local storagecapacity associated with it, i.e., a flash memory. Al-though, there exist flash memories with several giga-bytes storage capacity, e.g., the NAND flash [23] andthe trend is that they become cheaper, larger and morecommon, we do not assume extreme storage capabili-ties, so as to capture a broader field of applications andarchitectures.

Since the battery lifetime can be extended if we man-age to reduce the “amount” of communication, cachingthe useful data for each sensor either in its local storeor in the near neighborhood can prolong the networklifetime. Additionally, caching can be very effectivein reducing the need for network-wide transmissions,thus reducing the interference and overcoming thevariable channel conditions. Finally, it can speed-upthe multimedia in-network processing, because, as itis emphasized in [1], the processing and delivery ofmultimedia content are not independent and their in-teraction has a major impact on the levels of QoS thatcan be delivered.

This paper investigates for the first time the tech-nique of caching in the context of WMSNs. The needfor effective and intelligent caching policies in sensornetworks has been pointed out several times [10, 25] inthe very recent literature, but no appropriate sophisti-cated policies have been proposed, although there arequite a lot of caching protocols in other fields (see rele-vant work in Section 2). This article proposes a noveland high-performance cooperative caching protocol,the NICoCa protocol named after the words NodeImportance-based Cooperative Caching, and comparesit with the state-of-the-art cooperative caching policyfor mobile ad hoc networks (MANETs), which is the“closer” competitor. Using the J-Sim simulation envi-ronment [34], we perform an experimental evaluationof the two methods, which attests that the proposedNICoCa cooperative caching policy prevails over itscompetitor.

The rest of this article is organized as follows: inSection 2 we review the relevant work and record thecontributions of the article. In Section 3 we present thedetails of the NICoCa protocol, and in Section 4 wepresent the results of the performance evaluation of themethods; finally, Section 5 concludes the paper.

2 Relevant work

The technique of caching has been widely investigatedin the context of operating systems and databases and isstill an attractive research area [24]. Similarly, cachingon the Web has been thoroughly investigated for coop-erative [12] and for non-cooperative [19] architectures.Wessels and Claffy [37] introduced the Internet cacheprotocol; Cache digests [30] and Summary Cache [12]enable proxies to exchange information about cachedcontent. In [6] a cooperative hierarchical Web cachingarchitecture was studied. However, the above archi-tectures and protocols usually assume a fixed networktopology and require powerful computation and com-munication capabilities.

In the context of wireless broadcast cellular net-works, a number of caching approaches have beenproposed [20]. These policies assume more powerfulnodes than the sensor nodes, and one-hop communi-cation with resource-rich base stations, which serve theneeded data.

A number of data replication schemes [15, 16] andcaching schemes [31, 35, 38] have been proposed inorder to facilitate data access in MANETs. Datareplication studies the issue of allocating replicas ofdata items to meet access demands. These techniquesnormally require a priori knowledge of the networktopology.

Caching schemes however do not facilitate data ac-cess based on the knowledge of distributed data items.In SimpleCache [38] the requested data item has alwaysbeen cached by the requester node. The node uses thecached copy in order to serve subsequent requests whenthey arrive. The requester node has to get the datafrom the data center in case of cache miss. Howeverincreasing the hop distance between the requester nodeand caching node will increase the response time for therequest.

The most relevant research works to our protocolare the cooperative caching protocols which have beendeveloped for MANETs. The main motive for thedevelopment of these protocols is the mobility of thenodes, and thus they all strive to model it or exploit it.Recently, a cooperative caching scheme, called CoCa,was proposed in [7, 8]. The CoCa framework facilitatemobile nodes to share their cached contents with eachother in order to reduce the number of server requestsand the number of access misses. The authors extendedCoCa with a group-based cooperative caching scheme,called GroCoCa, in [9]. According to GroCoCa, thedecision of whether a data item should be cacheddepends on two factors of the access affinity on thedata items and the mobility of each node. The mobile

Mobile Netw Appl

support station performs an incremental clustering al-gorithm to cluster the mobile nodes into tightly coupledgroups based on their mobility patterns. In GroCoCaalso the similarity of access patterns is captured byfrequency-based similarity measurement. GroCoCaimproves system performance at the cost of extra powerconsumption. Papadopouli and Schulzrinne [26] sug-gested the 7DS architecture. The authors deployed acouple of protocols in order to facilitate sharing anddissemination of information among users. It operateson two modes. The first one is a prefetch mode, basedon the information and user’s future needs and the sec-ond one is an on-demand mode, which searches for dataitems in a one hop multicast basis. Depending on thecollaborative behavior, a peer-to-peer and server-to-client mode are used. Unlike our approach, this strategyfocuses also on single-hop wireless environment and ondata dissemination, and thus the cache management in-cluding cache admission control and replacement is notwell explored. Sailhan and Issarny [32] proposed a col-laborative cache management strategy among mobileterminals interacting via an ad hoc network. The issuethat the authors addressed was on setting an ad hoc net-work of mobile terminals that cooperate to exchangeWeb pages. The proposed solution aims at improvingthe Web latency on mobile terminals while optimizingassociated energy consumption. It is implemented ontop of Zone Routing Protocol. The authors proposed afixed broadcast range based on the underlying routingprotocol.

The Zone Cooperative (ZC) [5], the Cluster Coop-erative (CC) [4] and the ECOR [33] protocols attemptto form clusters of nodes based either in geographicalproximity or utilizing widely known node clusteringalgorithms for MANETs [3]. In ZC, mobile nodes be-longing to the neighborhood (zone) of a given nodeform a co-operative cache system for this node sincethe cost for communication with them is low both interms of energy consumption and message exchanges.Each node has a cache to store the frequently accesseddata items. The data items in the cache satisfy not onlythe node’s own requests, but also the data requestspassing through it from other nodes. For a data missin local cache, the node first searches the data in itszone before forwarding the request to the next nodethat lies on a path towards the data center. As apart of cache management, a value-based replacementpolicy based on popularity, distance, size and time-to-live was developed to improve the data accessibil-ity and reduce the local cache miss ratio. Simulationsexperiments revealed improvements in cache hit ratioand average query latency in comparison with othercaching strategies. In CC, the authors present a scheme

for caching in MANETs. The goal of CC is to reducethe cache discovery overhead and provide better co-operative caching performance. The authors partitionsthe whole MANET into equal size clusters based onthe geographical network proximity. In each cluster,CC dynamically chooses a“super” node as cache statenode, to maintain the cluster cache state informationof different nodes within its cluster domain. The cachestate node is defined as the first node that enters thecluster. The cluster cache state for a node is the listof cached data items along with their time-to-live field.The cache replacement policy is similar to that in ZC.However, the ZC protocol is problematic in terms ofcommunication overhead and energy consumption, be-cause every node that lies on the path towards thedata center has to broadcast the packet to nodes in itszone in order to discover if there is a cached copy thatsatisfies the requested data item. In contrast with theZC protocol, the CC protocol reveal a smaller overheadin terms of messages exchange between nodes, becausethe intermediate node (the nodes between a requestingnode and the node which holds the requested data)broadcast the packet only to cluster state node. InECOR, each mobile node forms a cooperation zone(CZ) with mobile nodes in proximity by exchangingmessages to share their cached data items in orderto minimize bandwidth and energy cost for each dataretrieval. When a data request arrives, the node firstsearches the data in its CZ before forwarding the re-quest to the data center. According to ECOR eachnode broadcast every modification of the cached dataitems to nodes that belong to cooperation zone. Eachnode maintains a cache hint table for the cache infor-mation of all nodes in its proximity. However, in ECORappears a great number of exchanged messages andenergy consumption in case of large node density andbig cooperation radius.

The only protocols that deviated from such ap-proaches and tried to exploit both data and node lo-cality in an homogeneous manner are described in [38]and are the following: CachePath, CacheData, andHybridCache. In CacheData, intermediate nodes maycache data to serve future requests instead of fetchingdata from the “Data Center”. An intermediate nodecaches passing by data item locally when it finds thatdata item is popular and does not cache the data itemif all requests for it are from the same node. Thisrule is designed in order to reduce the cache spacerequirement because the mobile nodes have limitedcache spaces. In CachePath, a mobile node may cachethe information of a path to a nearby data requesterwhile forwarding the data and use the path informationto redirect future requests to the nearby caching site.

Mobile Netw Appl

By caching the data path for each data item, band-width and the query delay can be reduced since therequested data can be obtained through fewer numberof hops. The authors proposed also some optimizationstechniques. An intermediate node can save only thedestination node information because the path from thecurrent router to the destination node can obtained bythe underlying routing protocol. Also, the intermediatenode need to record the data path when it is closerto the caching node than the data center. One majordrawback of CachePath is that the cached path maynot be reliable and using it may increase the overhead.A cached path may not be reliable because either thedata item has become obsolete or the caching node cannot be reached. The hybrid protocol HybridCache com-bines CacheData and CachePath while avoiding theirweaknesses. In HybridCache, when a mobile node for-wards a data item, it caches the data or the path basedon some criteria. These criteria include the data itemsize, the time-to-leave of the data item and the numberof hops that a cached path can save and denoted asHsave. Hsave value is the difference between the distanceto the data center and the distance to the caching node.Hsave must be greater than a system tuning threshold.One drawback with these methods is that caching in-formation of a node cannot be shared if the node doesnot lie on the path between the data requester and thedata source. Moreover, the threshold values used inthese heuristics must be set carefully in order to achievegood performance. The only works on caching in WSNsconcern the placement of caches [28, 36] and thus theyare not strictly relevant to our considered problem.

2.1 Motivation and contributions

The protocols proposed so far for cooperative cachingin MANETs present various limitations. Those proto-cols, which first perform a clustering of the network andthen exploit this clustering (the cluster-heads, (CH)), inorder to coordinate the caching decisions, inherit theshortcomings of any bad CH selection. For instance,in [4, 5], the nodes which form the cluster are assumedto reside within the same communication range, i.e.,they are with on-hop distance from the other nodesof the cluster. Additionally, the nodes do not cachethe data originating from an one-hop neighbor. Thus,CHs which do no reside in a significant part of dataroutes can not serve efficiently their cluster members,because they do not have fast access (short latency)to requested data. The cooperation zone which isformed in [33] by selecting an optimal radius, impliesa large communication overhead, because every nodewithin that radius must send/receive any changes to the

caches of the other nodes within the radius. Finally, theHybridCache policy is tightly coupled to the underlyingrouting protocol, and thus if a node does not reside inthe route selected by the routing protocol can not cachethe data/path, or conversely, can not serve the requesteven if it holds the requested data.

Motivated by the weaknesses of the current coop-erative protocols and the unique requirements of theWMSNs, which are mainly static and not mobile, wepropose a cooperative caching policy which is based onthe idea of exploiting the sensor network topology, soas to discover which nodes are more important than theothers, in terms of their position in the network and/orin terms of residual energy. Incorporating both factorsinto the design of the caching policy we ensure bothnetwork longevity and short latency in multimedia dataretrieval. In summary, the article’s contribution are thefollowing:

• Definition of a metric for estimating the importanceof a sensor node in the network topology, which willimply short latency in retrieval.

• Description of a cooperative caching protocolwhich takes into account the residual energy of thesensor nodes.

• Development of algorithms for discovering therequested multimedia data, and maintaining thecaches (cache replacement policy).

• Performance evaluation of the protocol andcomparison with the state-of-the-art cooperativecaching protocol for MANETs, using an estab-lished simulation package (J-Sim).

3 The NICoCa cooperative caching protocolfor WMSNs

One of the main parts of the proposed protocol is theestimation of the importance of sensors relative to thenetwork topology. The intuition is that if we discoverthose nodes, which reside in a significant part of the(short) paths connecting other nodes, then these are the“important” nodes; then they may be selected as coor-dinators for the caching decisions, i.e., as “mediators”to provide information about accessing the requesteddata or even as caching points.

3.1 Measuring sensor node importance

A WMSN is abstracted as a graph G(V, E), where V isthe set of its nodes, and E is the set of radio connec-tions between the nodes. An edge e = (u, v), u, v ∈ Eexists if and only if u is in the transmission range of

Mobile Netw Appl

v and vice versa. All links in the graph are bidirec-tional, i.e., if u is in the transmission range of v, v isalso in the transmission range of u. The network isassumed to be in a connected state. The set of neigh-bors of a node v is represented by N1(v), i.e., N1(v) ={u : (v, u) ∈ E}. The set of two-hop nodes of node v, i.e.,the nodes which are the neighbors of node v’s neighborsexcept for the nodes that are the neighbors of nodev, is represented by N2(v), i.e., N2(v)={w : (u, w) ∈ E,

where w �= v and w /∈ N1and(v, u) ∈ E}. The combinedset of one-hop and two-hop neighbors of v is denoted asN12(v).

Definition 1 (Local network view of node v) The localnetwork view, denoted as LNv , of a graph G(V, E)

w.r.t. a node v ∈ V is the induced subgraph of G associ-ated with the set of vertices in N12(v).

A path from u ∈ V to w ∈ V has the common mean-ing of an alternating sequence of vertices and edges,beginning with u and ending with w. The length ofa path is the number of intervening edges. We de-note by dG(u, w) the distance between u and w, i.e.,the minimum length of any path connecting u and w

in G, where by definition dG(v, v) = 0, ∀v ∈ V anddG(u, w) = dG(w, u), ∀u, w ∈ V. Note that the distanceis not related to network link costs (e.g., latency), butit is a purely abstract metric measuring the numberof hops.

Let σuw = σwu denote the number of shortest pathsfrom u ∈ V to w ∈ V (by definition, σuu = 0 ). Letσuw(v) denote the number of shortest paths from u tow that some vertex v ∈ V lies on. Then, we define thenode importance index NI(v) of a vertex v as:

Definition 2 The NI(v) of a vertex v is equal to:

NI(v) =∑

u�=v �=w∈V

σuw(v)

σuw

. (1)

Large values for the NI index of a node v indicatethat this node v can reach others on relatively shortpaths, or that the node v lies on considerable fractionsof shortest paths connecting others. Illustration of thismetric is presented in Fig. 1 (with well formed andvague node clusters).

Apparently, when estimating the NI index for eachsensor node using the whole network topology weobtain a very informative picture of which nodes re-side in a large number of shortest paths betweenother nodes. Though, it would not be practical to usethis global information in the desing of a coopera-tive caching protocol, because it would require the

1

2

3

5

47

6

10

13

11

8

14

20

9

12 15

18

19

17

(68)

(0)

(0)

(0)(0)

(0)(0)

(0)

(0)

(0)

(0)

(0)

(0)

(1)

(156)(233) 16(96)

(26)

(131)

(97)

X (0)Y (0)

Z (0)

A (6.67)

B (13)

P (41)

Q (8)R (9.33)

U (54)

W (3.33)

V (1.33)

T (1.33)

Figure 1 Calculation of NI for two sample graphs. The numbersin parentheses denote the NI index of the respective node con-sidering the whole WMSN topology

exchange of a huge number of messages between thenodes to learn this information. Moreover, it wouldrequire a lot of communication rounds; for the estab-lishment of broadcast tree with a leader/root node,for learning the neighborhoods for the estimation andpropagation of the NI indexes and so on. These roundswould also increase the latency and compromise theprotocol’s robustness when changes in the topologywould take place. Therefore, if the idea of NI indexcould be successfully applied in neighborhoods to iden-tify nodes which reside in a large number of shortestpaths, then we could desing a localized algorithm basedon it. Indeed, using only the information contained inthe local network view of a node, we can still obtain thisinformation, quite accurately. For instance, computingthe NI indexes for all the nodes that are contained inLN8, we obtain the information of Table 1.

We observe that although the absolute values ofthe NI indexes are reduced as a consequence of thesmaller network size, the relative ranking of the 1-hopneighbors of the node 8 w.r.t. their NI index is the

Table 1 NI index of the nodes belonging to LN8 (from the localperspective of node 8)

n(ode) NI(n) n(ode) NI(n) n(ode) NI(n)

7 0 11 0 15 08 14 12 0 16 239 0 13 0 18 010 0 14 65

Mobile Netw Appl

same with that obtained when we considered the wholenetwork. Apparently, the larger the neighborhood weconsider the more accurate the picture of nodes’ im-portance we get. Although, we can not come up witha closed form that would describe when and at whatdegree the relative rankings change when we considersmall neighborhoods, instead of the whole network,because it depends on the graph characteristics, ourinvestigation showed that in more than 98% of thecases, the relative rankings remain unchanged.

At a first glance, the computation of the NI seemsexpensive, i.e., O(y ∗ x2) operations in total for a 2-hopneighborhood, which consists of x nodes and y links.For our undirected network, we can use breadth-firsttraversal to count the number and find the length ofshortest paths from any particular node (source) to anyother node (target) in time linear w.r.t. the number ofedges, i.e.,O(x ∗ y). Repeating this procedure for anypossible source, we conclude that computing the lengthand number of shortest paths between any possiblepair can be done in time O(y ∗ x2). Fortunately, wecan do better than this and exploit some redundancyin the computations. This procedure gives an O(x ∗ y)

method for computing the NI indexes for the nodes.For more information concerning the calculation andthe use of this metric in broadcasting protocols theinterested user can consult the work reported in [21].

3.2 Housekeeping informationin the NICoCa protocol

W.l.o.g. and adopting the model presented in [38], weassume that the ultimate source of multimedia data isa Data Center. This is not restrictive at all and simplyguarantees that every request, if it is not served byother sensor nodes and if does not expire, will finallybe served by the Data Center.

Firstly, it is assumed that each node is aware ofits 2-hop neighborhood. This information is obtainedthrough periodic exchange of “beacon” messages. Weassume that we are able to determine an assignment oftime slots to the sensor nodes such that no interferenceoccurs, i.e., no two nodes transmit in the same time slot.With this assignment, the nodes can transmit beaconmessages to discover their neighbors and also to acquirethe channel for transmitting their data requests. Thisassignment is determined using the D2-coloring algo-rithm [13]. Although more traditional methods can beused for channel arbitration, this approach contributesin reducing the latency by avoidind interference. Then,every node calculates the NI index of its 1-hopneighbors. The node uses this information in order tocharacterize some of its neighbors as mediator nodes;

the minimum set of neighbors with the larger NI which“cover” its 2-hop neighborhood are the mediator nodesfor that node; The node is responsible for notifyingits neighbors about which of them are its mediators.Thus, a node can be either a mediator or an ordinarynode. Notice here that there is no need for the node toknow the “importance” of its neighbors w.r.t. the wholenetwork; instead the node needs to know their impor-tance in its neighborhood. Of course, if the notion ofneighborhood is extended to cover the whole network(i.e., each node is aware of its k-neighborhood, where kequals the network diameter), then it is obvious thatthe node would have perfect knowledge of the exact NIindexes of all network’s nodes.

The sending of requests for data is carried out byan ordinary sensor (or ad hoc) routing protocol, e.g.,ad-hoc on-demand distance vector (AODV) [27]. Anode always caches a datum which has requested for.A node is aware of its remaining energy and of the freespace in its cache. Each sensor node stores the followingmetadata related to a cached multimedia item:

• the dataID, and the actual multimedia data item,• the data size (si),• a Time-To-Live (TTL) interval,• for each cached item, the timestamps of the K most

recent accesses to that item. Usually, K = 2 or 3.• each cached item is characterized either as O (i.e.,

own) or H (i.e., hosted). If an H-item is requestedby the caching node, then its state switches to O.

When a node acquires the multimedia datum he hasrequested for, then it caches it and broadcasts a smallindex packet containing the dataID and the associatedTTL, its remaining energy and its free cache space. Themediator nodes which are also 1-hop neighbors of thisnode store this broadcasted information. Notice herethat this set of mediator nodes includes the mediatorsthat the broadcasting node has selected, and also anyother mediators which have been selected by nearbynodes. In summary, every mediator node stores theremaining energy and the free cache space for each oneof its 1-hop neighbors, and for each dataID that hasheard through the broadcasting operation, the TTL ofthis datum and the nodes that have cached this datum.

3.3 The cache discovery component protocol

When a sensor node issues a request for a multimediaitem, it searches its local cache. If the item is foundthere (a local cache hit) then the K most recent ac-cess timestamps are updated. Otherwise (a local cachemiss), the request is broadcasted and received by the

Mobile Netw Appl

mediators. If none of them responds (a “proximity”cache miss), then the request is directed to the DataCenter.

When a non one-hop mediator node receives a re-quest, it searches its local cache. If it deduces thatthe request can be satisfied by a neighboring node (aremote cache hit), then stops the request’s route towardthe Data Center, and forwards the request to this neigh-boring node. If more than one nodes can satisfy therequest, then the node with the largest residual energyis selected. If the request can not be satisfied by thismediator node, then it does not forward it recursivelyto its own mediators. This is due to the fact that thesemediators will most probably be selected by the routingprotocol as well (AODV) and thus a great deal ofsavings in messages is achieved. Therefore, during theprocedure of forwarding a request toward the DataCenter, no searching to other nodes is performed apartfrom the nodes which reside on the path toward theData Center.

For example, suppose that sensor node SNi in Fig. 2issues a request for the data item x that is placedin data center DC1 and which has been cached bysensor nodes SNg and SNh. The black shaded nodesare the mediators nodes of the sensor network. Basedon the information presented in Fig. 1, we can easilyunderstand that node SNi has selected as its mediatorsthe nodes SNa and SNb . With their turn, these sensorswould select as their mediators the nodes SNc and SNd

(for node SNa) and SNd (for node SNb ).In the beginning sensor node SNi searches its own

cache. If it deduces that data item is not availablein local cache, it sends a proximity search request inneighboring mediator nodes SNa and SNb . Upon re-ceiving the search request, each mediator searches inthe proximity cache table. If data item found, eachmediator replies with an index packet that contains thedataID and the remaining energy of the sensor nodethat has the uppermost battery power and has cachedthe data item. SNi, upon the receipt of index packets,

selects the sensor node that has the smallest energy con-sumption and sends the request packet. Caching noderesponds with a reply packet containing the requesteddata item.

If no neighboring sensor node is caching the dataitem, SNi sends a request packet to the data centerDC1, as shown in Fig. 2. When SNx (x ∈ {d, e}) receivesa request packet, it searches in local cache and in prox-imity cache table. If the data item is not found, SNx for-wards the request through the path to the DC1. Whensensor node SNe found that the requested data itemhas been cached by some neighboring nodes, it choosesthe node that has the smallest energy dissipation andredirects the request packet to the caching node. Thecaching node sends a reply packet containing the dataitem x along the routing path until it reaches the orig-inal requester. Once the requester node receives thedata item, it notifies its one-hop mediators about thenew caching item by sending an index packet containingitem’s dataID. In case of not enough cache capacity, ittriggers the cache replacement protocol to determinethe data items that should be evicted from the cache.

For every issued request one of the following fourcases may take place:

1) Local hit: the requested datum is cached by thenode which issued the request. If this datum is valid(the TTL has not expired) then the NICoCa is notexecuted.

2) “Proximity” hit: the requested datum is cached by anode in the 2-hop neighborhood of the node whichissued the request. In this case, the mediator(s)return to the requesting node the “location” of thenode which stores the datum.

3) Remote hit: the requested datum is cached by anode and this node has at least one mediator re-siding along the path from the requesting node tothe Data Center.

4) Global hit: the requested datum is acquired fromthe Data Center.

Figure 2 A request packetfrom sensor node SNi isforwarded to the cachingnode SNg

b

SNc

SNd

SNe

SNg

SNh

SNj

SNa

DC 1DC 2

proximity searchSN

request

iSN

Mobile Netw Appl

Table 2 Simulation parameters

Parameter Default value Range

# items (N) 1000Smin (KB) 1Smax (KB) 10Smin (MB) 1Smax (MB) 5Zipfian θ 0.8# nodes (n) 500 100–1000Bandwidth (Mbps) 2Waiting interval (tw) 10 s for items with KB size

100 s for items with KB sizeClient cache size (KB) 800 200 to 1200Client cache size (MB) 125 25 to 250Zipfian skewness (θ) 0.8 0.0 to 1.0

3.4 The cache replacement component protocol

Even though the cache capacity of individual sensorsmay be in the order of gigabytes (e.g., NAND flash)the development of an effective and intelligent replace-ment policy is mandatory to cope with the overwhelm-ing size of multimedia data generated in WMSNs.The NICoCa protocol employs the following four-steppolicy:

STEP 1. In case of necessity, before purging fromcache any other data, each sensor node firstpurges the data that it has cached on behalfof some other node. Each cached item ischaracterized either as O (i.e., own) or H(i.e., hosted). In case of a local hit, then itsstate switches to O. If the available cachespace is still smaller than the required, exe-cute Step 2.

STEP 2. Calculate the following function for eachcached datum i: cost(i)= si

TT Li∗ now−tK−th access

K .The candidate cache victim is the item whichincurs the largest cost.

STEP 3. Inform the mediators about the candidatevictim. If it is cached by some mediator, thenthis information returns back to the nodeand purges the datum. If the datum is notcached by some mediator(s), then it is for-warded to the node with the largest residualenergy and the datum is purged from thecache of the original node. In any case, themediators update their cached metadataabout the new state.

STEP 4. The node which caches this purged datum,informs the mediators with the usual broad-casting procedure.

The pseudocode for the complete algorithm NICoCais presented in the Appendix.

4 Performance evaluation

We evaluated the performance of the NICoCa proto-col through simulation experiments. We conducted alarge number of experiments with various parameters,and compared the performance of NICoCa to a state-of-the-art cooperative caching policy for MANETs,namely HybridCache [38].

4.1 Simulation model

We have developed a simulation model based on theJ-Sim simulator [34]. In our simulations, the AODV[27] routing protocol is deployed to route the data

0

2000

4000

6000

8000

10000

12000

14000

16000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)

NICoCa - LHHybrid - LH

NICoCa - RHHybrid - RH

NICoCa - GHHybrid - GH

0

2000

4000

6000

8000

10000

12000

14000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)




Figure 3 Impact of sensor cache size on hits (MB-sized files, θ = 0.0 and θ = 0.8) in a sparse WMSN (d = 7) with 100 sensors

Mobile Netw Appl

0

2000

4000

6000

8000

10000

12000

14000

16000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)




0

2000

4000

6000

8000

10000

12000

14000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)




Figure 4 Impact of sensor cache size on hits (MB-sized files, θ = 0.0 and θ = 0.8) in a dense WMSN (d = 10) with 100 sensors

traffic in the WSN. We use IEEE 802.11 as the MACprotocol and the free space model as the radio propa-gation model. The wireless bandwidth is 2 Mbps.

We tested the protocols for a variety of sensornetwork topologies, to simulate sensor networks withvarying levels of node degree, from 4 to 10. We alsoconducted experiments by choosing the number ofnodes between 100 and 1000. In addition, in our exper-iments we evaluate the protocol efficiency under twodifferent set of data item sizes. Each data item has sizethat is uniformly distributed from 1KB to 10KB for thefirst set, and from 1MB to 5MB for the second.

The network topology consists of many square gridunits where one or more nodes are placed. The numberof square grid units depends on the number of nodesand the node degree. The topologies are generated asfollows: the location of each of the n sensor nodes isuniformly distributed between the point (x = 0, y = 0)

and the point (x = 500, y = 500). The average degree dis computed by sorting all n ∗ (n − 1)/2 edges in thenetwork by their length, in increasing order. The gridunit size corresponding to the value of d is equal to

√2

times the length of the edge at position n ∗ d/2 in thesorted sequence. Two sensor nodes are neighbors ifthey placed in the same grid or in adjacent grids. Thesimulation area is assumed of size 500 m × 500 m andis divided into equal sized square grid units. Beginningwith the lower grid unit, the units are named as 1, 2, . . . ,in a column-wise fashion.

The client query model is similar to what have beenused in previous studies [38]. Each sensor node gener-ates read-only queries. After a query is sent out, if thesensor node does not receive the data item, it waits foran interval (tw) before sending a new query. The accesspattern of sensor nodes is: a) location independent,that is, sensor nodes decide independently the data of

25

30

35

40

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

20

25

30

35

40

45

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

Figure 5 Impact of sensor cache size on latency (MB-sized files, θ = 0.0 and θ = 0.8) in a sparse WMSN (d = 7) with 100 sensors

Mobile Netw Appl

20

25

30

35

40

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

20

25

30

35

40

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

Figure 6 Impact of sensor cache size on latency (MB-sized files, θ = 0.0 and θ = 0.8) in a dense WMSN (d = 10) with 100 sensors

interest; each sensor node generates accesses to thedata following the uniform distribution, and b) Zipfianwith θ = 0.8, where groups of nodes residing in neigh-boring grids (25 grids with size 100 m × 100 m) have thesame access pattern. We tested the protocols both forzipfian access pattern and for uniform access pattern.In case of zipfian access pattern we conducted experi-ments with varying θ values between 0.0 and 1.0.

Similar to [38], two data centers are placed at op-posite corners of the simulation area. Data Center 1is placed at point (x = 0, y = 0) and Data Center 2 isplace at point (x = 500, y = 500). There are N/2 dataitems in each data center. Data items with even idsare stored at Data Center 1 and data items with oddids are stored at Data Center 2. The size of each dataitem is uniformly distributed between smin and smax.We assumed that data items are not updated. The datacenters serve the queries on a first-come-first-servedbasis. The system parameters are listed in Table 2.

4.2 Performance metrics

The measured quantities include the number of hits(local, remote and global), the average latency for get-ting the requested data and the message overhead. Itis evident that a small number of global hits impliesless network congestion, and thus fewer collisions andpacket drops. Moreover, large number of remote hitsproves the effectiveness of cooperation in reducing thenumber of global hits. A large number of local hitsdoes not imply an effective cooperative caching policy,unless it is accompanied by small number of global hits,since the cost of global hits vanishes the benefits oflocal hits.

4.3 Evaluation

We performed a large number of experiments varyingthe size of the sensornet (in terms of the number of itssensor nodes), varying the access profile of the sensornodes, and the cache size relative to the aggregate sizeof all data items. In particular, we performed exper-iments for 100, 500, and 1000 sensors, for cache sizeequal to 1%, to 5% to 10% of the aggregated sizeof all distinct multimedia data, for access pattern withθ equal to 0.0 (uniform access pattern) to 1.0 (highlyskewed access pattern), for average sensor node degreeequal to 4, 7 (very sparse and spare sensornet) and 10(dense sensornet), and for data item size equal to afew kilobytes (KB) and also equal to a few megabytes(MB). For each different setting we measured thenumber of hits (local, remote, global), the latency2 andthe message overhead. In the sequel of the article wewill present only a representative set of the results,since there are many of independent parameters andthree dependent performance metrics. We partition thegraphs in two large groups w.r.t. whether they deal withsmall KB-sized files or large MB-sized multimedia files.

4.3.1 Experiments with MB-sized data items

The purpose of this set of experiments is to examinethe performance of the caching algorithms when theyhave to deal with large multimedia files, e.g., video files,queried by the sensornet.

2The latency is measured in seconds, which does not correspondsto the usual time metric, but to internal simulator clock.

Mobile Netw Appl

0

10000

20000

30000

40000

50000

60000

70000

80000

90000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)




0

10000

20000

30000

40000

50000

60000

70000

80000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)




Figure 7 Impact of sensor cache size on hits (MB-sized files, θ = 0.0 and θ = 0.8) in a sparse WMSN (d = 7) with 500 sensors

Figures 3 and 4 show the number of hits achievedby the two protocols for a small (sparse and dense,respectively) sensornet for both uniform and skewedaccess pattern. The first observation is that, as expected,the number of local hits increases for both protocols asthe access pattern becomes more skewed. The interest-ing point is that, although for uniform access patternsHyrbidCache is slightly better than NICoCa w.r.t. thelocal hits, the situation is reversed when the requestsare concentrated to a smaller number of files, whichcan be attributed to the more efficient replacementand admission policy of the NICoCa. With respectto the number of global hits, NICoCa achieves halfthat of hybrid and the performance gap widens as wemove to dense sensor deployments; actually NICoCamaintains almost constant the number of global hits.The reason behind this is the relative performance

of the algorithms w.r.t. the remote cache hits. Forsparser deployments NICoCa is two times better thanHyrbidCache, and this difference becomes more ev-ident for denser networks. Thus, it proves to be amore effective cooperation scheme, due to the factthat it strives to exploit the network topology. Theserelative performance results are straightforwardly re-flected to the access latency incurred by the algorithms(Figs. 5 and 6).

When we move to larger sensornets with 500 (Fig. 7–8) and with 1000 nodes (Fig. 12), the superiority ofthe NICoCa caching algorithm in terms of hits is stillevident, but the results are not so impressive, becausewe constrain more sensor nodes to be dispersed in thesame geographical region, thus creating replicas of thesame data. This superiority is reflected to the accesslatency as well (Figs. 9 and 10). We observe that as

0

10000

20000

30000

40000

50000

60000

70000

80000

90000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)




0

10000

20000

30000

40000

50000

60000

70000

80000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)




Figure 8 Impact of sensor cache size on hits (MB-sized files, θ = 0.0 and θ = 0.8) in a dense WMSN (d = 10) with 500 sensors

Mobile Netw Appl

50

55

60

65

70

75

80

85

90

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

50

55

60

65

70

75

80

85

90

95

100

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

Figure 9 Impact of sensor cache size on latency (MB-sized files, θ = 0.0 and θ = 0.8) in a sparse WMSN (d = 7) with 500 sensors

we move to larger sensornets, the latence graduallyincreases, because the denser deployment (more nodesin the same region) has a negative effect on the effi-ciency of communication, aggravating the collisions andpacket drops.

At this point it is interesting to note the total numberof messages that are communicated between the sensornodes, which is also the metric that models the total net-work energy dissipated (Figs. 11). For a dense sensornetwith uniform and skewed access pattern, NICoCa sendsat most half of the messages sent out by HyrbidCacheand the situation becomes more favorable for NICoCaas the access pattern becomes more skewed, which isexpected. These results are confirmed for larger sensor-nets with 1000 nodes (Figs. 12, 13, and 14).

In summary, for all network topologies NICoCaachieves more remote hits and less global hits thanHyrbidCache. This performance gap widens in favor

of NICoCa as we move from sparse to denser WMSN.It is striking that for very dense sensor deployments,NICoCa achieves double the remote hits of HyrbidCache and only half of its global hits. Examining thelocal hits, we observe that for sparse sensor networksHyrbidCache achieves slightly more local hits thanNICoCa, but this gap vanishes completely when movingto denser network topologies. Besides, this small gain ofHyrbidCache for sparse topologies is not advantageousat all, since it incurs global hits as many as twice thenumber of its local hits.

4.3.2 Experiments with KB-sized data items

A significant question arises whether these relativeresults still hold when the sensornet has to deal withsmaller multimedia files, with size equal to a few kilo-bytes. Although we expect that WMSNs will deal with

40

45

50

55

60

65

70

75

80

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

40

45

50

55

60

65

70

75

80

85

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

Figure 10 Impact of sensor cache size on latency (MB-sized files, θ = 0.0 and θ = 0.8) in a dense WMSN (d = 10) with 500 sensors

Mobile Netw Appl

400000

600000

800000

1e+006

1.2e+006

50 100 150 200 250

Tot

al n

umbe

r of

mes

sage

s

Cache Size (MB)

NICoCaHybrid

300000

400000

500000

600000

700000

800000

900000

1e+006

1.1e+006

50 100 150 200 250

Tot

al n

umbe

r of

mes

sage

s

Cache Size (MB)

NICoCaHybrid

Figure 11 Impact of sensor cache size on number of messages (MB-sized files, θ = 0.0 and θ = 0.8) in a dense WMSN (d = 10) with500 sensors

0

20000

40000

60000

80000

100000

120000

140000

160000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)




0

20000

40000

60000

80000

100000

120000

140000

160000

50 100 150 200 250

Cac

he H

its

Cache Size (MB)




Figure 12 Impact of sensor cache size on hits (MB-sized files, θ = 0.8) in a dense WMSN (d = 10) with 1000 sensors

80

100

120

140

160

180

200

220

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

80

100

120

140

160

180

50 100 150 200 250

Avg

Lat

ency

(se

c)

Cache Size (MB)

NICoCaHybrid

Figure 13 Impact of sensor cache size on latency (MB-sized files, θ = 0.8) in a dense WMSN (d = 10) with 1000 sensors

Mobile Netw Appl

1e+006

1.2e+006

1.4e+006

1.6e+006

1.8e+006

2e+006

50 100 150 200 250

Tot

al n

umbe

r of

mes

sage

s

Cache Size (MB)

NICoCaHybrid

800000

1e+006

1.2e+006

1.4e+006

1.6e+006

1.8e+006

50 100 150 200 250

Tot

al n

umbe

r of

mes

sage

s

Cache Size (MB)

NICoCaHybrid

Figure 14 Impact of sensor cache size on the number of messages (MB-sized files, θ = 0.8) in a dense WMSN (d = 10) with 1000sensors

2000

4000

6000

8000

10000

12000

14000

16000

18000

200 400 600 800 1000 1200

Cac

he H

its

Cache Size (KB)




2000

4000

6000

8000

10000

12000

14000

16000

18000

200 400 600 800 1000 1200

Cac

he H

its

Cache Size (KB)




Figure 15 Impact of sensor cache size on hits (KB-sized files, θ = 0.8) in a sparse and dense WMSN (d = 7 and d = 10) with 100 sensors

0.05

0.1

0.15

0.2

0.25

200 400 600 800 1000 1200

Avg

Lat

ency

(se

c)

Cache Size (KB)

NICoCaHybrid

0

0.05

0.1

0.15

0.2

0.25

0.3

200 400 600 800 1000 1200

Avg

Lat

ency

(se

c)

Cache Size (KB)

NICoCaHybrid

Figure 16 Impact of sensor cache size on latency (KB-sized files, θ = 0.8) in a sparse and dense WMSN (d = 7 and d = 10) with 100sensors

Mobile Netw Appl

0

10000

20000

30000

40000

50000

60000

70000

80000

200 400 600 800 1000 1200

Cac

he H

its

Cache Size (KB)




0

10000

20000

30000

40000

50000

60000

70000

80000

200 400 600 800 1000 1200

Cac

he H

its

Cache Size (KB)




Figure 17 Impact of sensor cache size on hits (KB-sized files, θ = 0.8) in a sparse and dense WMSN (d = 7 and d = 10) with 500 sensors

0.2

0.4

0.6

0.8

1

1.2

200 400 600 800 1000 1200

Avg

Lat

ency

(se

c)

Cache Size (KB)

NICoCaHybrid

0.2

0.4

0.6

0.8

1

1.2

200 400 600 800 1000 1200

Avg

Lat

ency

(se

c)

Cache Size (KB)

NICoCaHybrid


0

20000

40000

60000

80000

100000

120000

140000

200 400 600 800 1000 1200

Cac

he H

its

Cache Size (KB)




0

20000

40000

60000

80000

100000

120000

140000

200 400 600 800 1000 1200

Cac

he H

its

Cache Size (KB)




Figure 19 Impact of sensor cache size on hits (KB-sized files, θ = 0.8) in a sparse and dense WMSN (d = 7 and d = 10) with 1000sensors

Mobile Netw Appl

1

1.5

2

2.5

3

200 400 600 800 1000 1200

Avg

Lat

ency

(se

c)

Cache Size (KB)

NICoCaHybrid

0.6

0.8

1

1.2

1.4

1.6

1.8

2

2.2

200 400 600 800 1000 1200

Avg

Lat

ency

(se

c)

Cache Size (KB)

NICoCaHybrid


MB-sized images of video files, it might be the casethat the sensor nodes will exchange smaller images aswell. To investigate the performance of the cooperativecaching protocols for this case, we performed the sameset of experiments but for KB-sized files and here wedemonstrate a subset of the results obtained.

The general observations that we recorded for thecase of large MB-size files, still hold for this case;NICoCa achieves significantly smaller number of globalhits and larger number of remote hits than HyrbidCache does. It is not worthy to comment on eachindividual performance graph (Fig. 15, 16, 17, 18, 19,and 20), since in all cases NICoCa is the clear winner;it achieves again 25% more remote hits and 50% lessglobal hits than HyrbidCache, which is only marginallybetter than NICoCa in terms of local hits.

5 Summary and conclusions

The recent advances in miniaturization, the creationof low-power circuits, and the development of cheapCMOS cameras and microphones, which are able tocapture rich multimedia content, gave birth to what iscalled WMSNs. WMSNs are expected to fuel many newapplications and boost the already existing. The uniquefeatures of WMSNs call for protocol designs that willprovide application-level QoS, an issue that has largelybeen ignored in traditional WSNs. Taking a first steptoward this goal, this article develops a cooperativecaching protocols, the NICoCa protocol, suitable fordeployment in WMSNs. The protocol “detects” whichsensor nodes are most “central” in the network neigh-borhoods and gives to them the role of mediator in or-

der to coordinate the caching decisions. The proposedprotocol is evaluated with J-Sim and its performanceis compared to that of a state-of-the-art cooperativecaching protocol for MANETs. The obtained resultsattest the superiority of the proposed protocol which isable to reduce the global hits at an average percentageof 50% and increase the remote hits due to the effectivesensor cooperation at an average percentage of 40%.The performance of the protocol is particularly high forthe delivery of large multimedia data.

Appendix

The NICoCa cooperative caching protocol

// di: data item i, i ∈ [1 . . . 1000]// request(di): Request for data item i// Ni: Node i// FS: Free cache space// RE: Remaining energy// PCT: Proximity Cache Table// ipacket: An index packet that contains di’s id, FS and RE

(A) Cache Discovery Algorithmif( di is in local cache of requester node ) then

send ipacket to CHs;return;

if( requester node is CH and di’s id in PCT ) thenselect caching node with largest RE;send request(di) to caching node;

elserequester node sends request(di) to CHs;

when CHs answers or time elapsedif( caching nodes found ) then

select caching node with largest RE;send request(di) to caching node;

Mobile Netw Appl

elsesend request(di) to data center;

when Ni receives request(di)if( Ni has a valid copy ) then

send di to requester node;else if( Ni is CH and di’s id in PCT ) then

select caching node with largest RE;redirect request(di) to caching node;

elseforward request(di) to caching node;

(B) Replacement Policywhile( current node has not enough FS )

Select a valid di with largest value and store it temporary;Send to CHs di’s id;Remove the valid di;

when a CH gets di’s idif( CH gets di’s id and di’s id not in PCT ) then

select caching node with largest RE and FS;send answer to requester node;when current node get answers from CHsforeach( temporary stored di )

if( there is no other caching node ) thenSelect caching node with least RE and largest FS;Send di to new caching node;

Remove temporary stored di;

(C) Cache Admission Policywhen the packet with di obtained from current nodeif( current node is packet’s destination ) then

if( there is enough FS ) thencache di;send ipacket to CHs;

elsecall Replacement Policy ;

when CH gets an ipacketif( CH get ipacket ) then

store di’s id, RE and FS in PCT;

References

1. Akyildiz I, Melodia T, Chowdhury KR (2007) A survey ofwireless multimedia sensor networks. Comput Netw 51(4):921–960

2. Akyildiz I, Su W, Sankarasubramaniam Y, Cayirci E (2002)A survey of wireless sensor networks. IEEE Commun Mag40(8):102–116

3. Basagni S, Mastrogiovanni M, Panconesi A, Petrioli C (2006)Localized protocols for ad hoc clustering and backbone for-mation: a performance comparison. IEEE Trans ParallelDistrib Syst 17(4):292–306

4. Chand N, Joshi RCRC, Misra M (2006) A zone co-operationapproach for efficient caching in mobile ad hoc networks. IntJ Commun Syst 19:1009–1028

5. Chand N, Joshi RCRC, Misra M (2007) Cooperative cachingstrategy in mobile ad hoc networks based on clusters. WirelPers Commun 43(1):41–63

6. Che H, Tung Y, Wang Z (2002) Hierarchical web cachingsystems: modeling, desing and experimental results. IEEE JSel Areas Commun 20(7):1305–1314

7. Chow C-Y, Leong HV, Chan ATS (2004) Cache signaturesfor peer-to-peer cooperative caching in mobile environments.In: Proceedings of the IEEE international conference onadvanced information networking and applications (AINA),vol 1. IEEE, Piscataway, pp 96–101

8. Chow C-Y, Leong HV, Chan ATS (2004) Peer-to-peer co-operative caching in mobile environments. In: Proceedingsof the IEEE international conference on distributed com-puting systems workshops (ICDCSW). IEEE, Piscataway,pp 528–533

9. Chow C-Y, Leong HV, Chan ATS (2007) GroCoca: group-based peer-to-peer cooperative caching in mobile environ-ment. IEEE J Sel Areas Commun 25(1):179–191

10. Diao Y, Ganesan D, Mathur G, Shenoy P (2007) Rethinkingdata management for storage-centric sensor networks. In:Proceedings of the conference on innovative data systemsresearch (CIDR). Asilomar, 7–10 January 2007, pp 22–31

11. Eisenberg Y, Luna CE, Pappas TN, Berry R, KatsaggelosAK (2002) Joint source coding and transmission power man-agement for energy efficient wireless video communications.IEEE Trans Circuits Syst Video Technol 12(6):411–424

12. Fan L, Cao P, Almeida AZ, Broder JM (2000) Summarycache: a scalable wide-area web cache sharing protocol.IEEE/ACM Trans Netw 8(3):281–293

13. Gandhi R, Parthasarathy S (2004) Fast distributed wellconnected dominating sets for ad hoc networks. ComputerScience Department, University of Maryland at College Park,Tech. Rep. CS-TR-4559

14. Girod B, Kalman M, Liang YJ, Zhang R (2002) Advancesin channel-adaptive video streaming. Wirel Commun MobComput 2(6):573–584

15. Hara T (2003) Replica allocation methods in ad hoc net-works with data update. ACM/Kluwer Mob Netw Appl 8(4):343–354

16. Hara T, Madria SK (2006) Data replication for improv-ing data accessibility in ad hoc networks. IEEE Trans MobComput 5(11):1515–1532

17. Huang GT (2003) Casting the wireless sensor net. TechnolRev 106:51–56

18. Karl H, Willig A (2006) Protocols and architectures for wire-less sensor networks. Wiley, New York

19. Katsaros D, Manolopoulos Y (2004) Caching in webmemory hierarchies. In: Proceedings of the ACM sympo-sium on applied computing. Cyprus, 14–17 March 2004,pp 1109–1113

20. Katsaros D, Manolopoulos Y (2004) Web caching in broad-cast mobile wireless environments. IEEE Internet Comput8(3):37–45

21. Katsaros D, Manolopoulos Y (2006) The geodesic broadcastscheme for wireless ad hoc networks. In: Proceedings of theIEEE international symposium on world of wireless, mobilemultimedia (WoWMoM). IEEE, Piscataway, pp 571–575

22. Kulkarni P, Ganesan D, Shenoy P, Lu Q (2005) SensEye: a multi-tier camera sensor network. In: Proceedingsof the ACM international conference on multimedia (MM).Singapore, 6–11 November 2005, pp 229–238

Mobile Netw Appl

23. Mathur G, Desnoyers P, Ganesan D, Shenoy P (2006) Ultra-low power data storage for sensor networks. In: Proceedingsof the ACM international conference on information process-ing in sensor networks (IPSN). Nashville, 19–21 April 2006,pp 374–381

24. Megiddo N, Modha DS (2003) ARC: a self-tuning, low over-head replacement cache. In: Proceedings of the USENIXconference on file and storage technologies (FAST). SanFrancisco, 31 March–2 April 2003, pp 115–130

25. Nath S, Kansal A (2007) FlashDB: dynamic self-tuningdatabase for NAND flash. In: Proceedings of the ACMinternational conference on information processing in sensornetworks (IPSN). Cambridge, 25–27 April 2007, pp 410–419

26. Papadopouli M, Schulzrinne H (2001) Effects of power con-servation, wireless coverage and cooperation on data envi-ronments. In: Proceedings of ACM symposium on mobile adhoc networking and computing (MOBIHOC). Long Beach,4–5 October 2001, pp 117–127

27. Perkins CE, Royer E (1999) Ad hoc on-demand dis-tance vector routing. In: Proceedings of the IEEE work-shop on mobile computing systems and applications. IEEE,Piscataway, pp 90–100

28. Prabh KS, Abdelzaher TF (2005) Energy-conserving datacache placement in sensor networks. ACM Trans SensorNetw 1(2):178–203

29. Rahimi M, Baer R, Iroezi OI, Garcia JC, Warrior J,Estrin D, Srivastava M (2005) Cyclops: in situ image sensingand interpretation in wireless sensor networks. In: Proceed-ings of the ACM international conference on embedded net-worked sensor systems (SenSys). San Diego, 2–4 November2005, pp 192–204

30. Rousskov A, Wessels D (1998) Cache digests. Comput NetwISDN Syst 30(22–23):2155–2168

31. Sailhan F, Issarny V (2002) Energy-aware web caching formobile terminals. In: Proceedings of the IEEE interna-tional conference on distributed computing systems work-shops (ICDCSW). IEEE, Piscataway, pp 820–825

32. Sailhan F, Issarny V (2003) Cooperative caching in adhoc networks. In: Proceedings of the IEEE internationalconference on mobile data management (MDM). IEEE,Piscataway, pp 13–28

33. Shen H, Das SK, Kumar M, Wang Z (2004) Cooperativecaching with optimal radius in hybrid wireless networks. In:Proceedings of the international IFIP-TC6 networking con-ference (NETWORKING), ser. Lecture Notes on ComputerScience, vol 3042. Athens, 9–14 May 2004, pp 841–853

34. Sobeih A, Hou JC, Kung L-C, Li N, Zhang H, Chen W-P,Tyan H-Y, Lim H (2006) J-Sim: a simulation and emulationenvironment for wireless sensor networks. IEEE WirelessCommun Mag 13(4):104–119

35. Takaaki M, Aida H (2003) Cache data access system in ad hocnetworks. In: Proceedings of the IEEE spring semiannualvehicular technology conference (VTC), vol 2. IEEE,Piscataway, pp 1228–1232

36. Tang B, Das S, Gupta H (2005) Cache placement in sensornetworks under update cost constraint. In: Proceedings of the(ADHOC-NOW), ser. Lecture Notes on Computer Science,vol 3738. Springer, Heidelberg, pp 334–348

37. Wessels D, Claffy K (1998) ICP and the Squid Web cache.IEEE J Sel Areas Commun 16(3):345–357

38. Yin L, Cao G (2006) Supporting cooperative caching in adhoc networks. IEEE Trans Mob Comput 5(1):77–89

Nikos Dimokas was born in Giannitsa, Greece in 1978. He re-ceived B.Sc. and M.Sc. in Computer Science from University ofCrete, Greece, in 2001 and 2004, respectively. Between May 2004and December 2005 he worked as a research and developmentengineer in the Institute of Computer Science at Foundation ofResearch and Technology Hellas (FORTH). Currently, he is aPh.D. candidate at the Department of Informatics of AristotleUniversity of Thessaloniki, Greece. His research interests includewireless sensor networks and vehicular ad hoc networks.

Dimitrios Katsaros was born in Thetidio (Farsala), Greecein 1974. He received a BSc in Computer Science fromAristotle University of Thessaloniki, Greece (1997) and a Ph.D.from the same department on May 2004. He spent a year (July1997–June 1998) as a visiting researcher at the Department ofPure and Applied Mathematics at the University of L’Aquila,Italy. Currently, he is a lecturer at the Department of Computerand Communication Engineering of University of Thessaly(Volos, Greece). He is editor of the book “Wireless InformationHighways” (2005), co-guest editor of a special issue of IEEEInternet Computing on “Cloud Computing” (2009–2010), andtranslator for the greek language of the book “Google’sPageRank and Beyond: The Science of Search Engine Rank-ings”. His research interests are in the area of distributed systems,including the Web and Internet, social networks analysis, mobileand pervasive computing, flash storage devices, mobile/vehicularad hoc networks, wireless sensor networks, and delay/disruptiontolerant networks.

Mobile Netw Appl

Yannis Manolopoulos was born in Thessaloniki, Greece in 1957.He received a B.Eng. (1981) in Electrical Eng. and a Ph.D.degree (1986) in Computer Eng., both from the Aristotle Univ.of Thessaloniki. Currently, he is Professor at the Departmentof Informatics of the same university. He has been with theDepartment of Computer Science of the Univ. of Toronto, theDepartment of Computer Science of the Univ. of Maryland atCollege Park and the Univ. of Cyprus. He has published over 200papers in journals and conference proceedings. He is co-author ofthe books “Advanced Database Indexing”, “Advanced SignatureIndexing for Multimedia and Web Applications” by Kluwer andof the books “R-Trees: Theory and Applications”, “NearestNeighbor Search: a Database Perspective” by Springer. He hasco-organized several conferences (among others ADBIS2002,SSTD2003, SSDBM2004, ICEIS2006, ADBIS2006, EANN2007).His research interests include Databases, Data mining, WebInformation Systems, Sensor Networks and Informetrics.

Documents

Cooperative Caching in Wireless Multimedia Sensor Networksdelab.csd.auth.gr/papers/MONET08dkm.pdf · Cooperative Caching in Wireless Multimedia Sensor Networks ... is compared to