Computing, Information and Control ICIC International c⃝2011 ISSN 1349-4198
Volume 7, Number 2, February 2011 pp. 763–776
A NOVEL CACHE REPLACEMENT ALGORITHM FOR COOPERATIVE CACHING IN WIRELESS
MULTIMEDIA SENSOR NETWORKS
Vincent S. Tseng1, Ming-Hua Hsieh1 and Kawuu W. Lin2,∗
1Dep. of Computer Science and Information Engineering National Cheng Kung University
No. 1, University Road, Tainan 701, Taiwan tsengsm@mail.ncku.edu.tw; mhhsieh@idb.csie.ncku.edu.tw
2Dep. of Computer Science and Information Engineering National Kaohsiung University of Applied Sciences No. 415, Chien Kung Road, Kaohsiung 807, Taiwan
∗Corresponding author: linwc@cc.kuas.edu.tw Received October 2009; revised February 2010
Abstract. In recent years, integrated applications with multimedia devices and wireless
sensor networks promoted the evolution of wireless sensor networks, namely wireless mul-timedia sensor networks (WMSNs). The applications in WMSNs have to focus on both energy saving and application-level quality of service (QoS). Due to the characteristics in WMSNs, such as resource constraints and variable channel capacity, efficiently achieving the application-level QoS in WMSNs is a challenging task. To overcome this challenge, in this paper, we proposed a new kind of pattern named temporal region requesting pattern (TRRP) and a novel algorithm named TRRP-Mine for mining TRRPs efficiently. We also designed a temporal region requesting cost function of cache replacement, abbreviated as TRRC, for the cooperative caching multimedia content in WMSNs. Empirical eval-uations under various simulation conditions showed that the proposed method delivers excellent performance in terms of hit rate and the number of replacements.
Keywords: Multimedia sensor networks, Cache replacement, Temporal region request-ing pattern, Data minrequest-ing
1. Introduction. As wireless technologies progressed rapidly [18] and embedded micro-sensing MEMS technology facilitated wireless sensor networks (WSNs), the applications of wireless sensor network had attracted extensive attention in the past decade. With the capabilities of widespread surveillance, sensor networks are applied to a lot of applications, such as the environmental data collection [7,13], localization system [9] and pervasive health applications [8]. In recent years, the application that integrates the cheap CMOS cameras with microphones over WSNs is becoming a trend, and this kind of WSNs is named Wireless Multimedia Sensor Networks (WMSNs) [1].
The existing studies of WSNs mainly focus on the energy saving problem [10,14]. Nev-ertheless, the challenges of developing the applications in WMSNs are not only the energy saving but also the quality of service (QoS) [6] issue in application-level. A general net-work layer metric for QoS is the netnet-work latency. The QoS in multimedia content delivery over Internet can be achieved through Diffserv [6] or Intserv [6]. However, such solutions for providing QoS over Internet face the severe bottlenecks due to the limited power and memory space of sensor nodes. An intuitive way to provide QoS in WMSNs is applying the caching technique to sensor nodes. Assume that each sensor node is equipped with a local storage and capable of caching a small number of requests. When a sensor node
764 V. S. TSENG, M.-H. HSIEH AND K. W. LIN
issues a request, the sensor node will check whether it caches the requested item or not. If the requested item is held in the local cache, there is no need to obtain the item from other nodes and the latency is thereby minimized. Based on the caching strategy, a node importance-based cooperative caching method named NICoCa [5] was proposed. By in-corporating the node importance of the WMSNs and the residual energy of each sensor node in consideration, NICoCa can prolong the network lifetime and provide QoS in WM-SNs. Although the node importance of WMSNs plays a major role in cooperative caching, the attributes of multimedia item are essential factors in cache replacement that should be carefully treated. However, only taking the attributes of multimedia item, such as data size and the timestamp of the latest access into consideration for cache replacement is insufficient. For instance, although a multimedia item is temporally requested by some sensor nodes, it may suffer from a frequent replacement. This is because the multimedia item with larger data size has a higher priority to be selected as the candidate victim of cache replacement. In this way, the network latency would be increased because the sensor node must issue broadcast for the requested multimedia item. For efficiently decreasing the network latency to improve QoS in WMSNs, the selection of adaptive candidate of cache replacement is extremely important. Therefore, we will also focus on the problem of precisely determining a multimedia item that is temporally requested. In addition, the main goal of this work is to propose a pattern-based cost function for the design of cache replacement.
In order to provide QoS in WMSNs, in this paper, we further defined a new kind of pattern, named temporal region requesting pattern (TRRP), and proposed an algorithm named TRRP-Mine for mining TRRPs from the temporal requesting information of sen-sor nodes efficiently. In addition, we also designed a temporal region requesting cost function (TRRC) of the candidate cache victim to assist the cache replacement in reduc-ing replacements and increasreduc-ing hit rate. Through empirical evaluations, our proposed approach was shown to outperform other existing approaches in terms of hit rate and the number of replacements. The rest of this paper was structured as follows. Preliminaries were stated in Section 2. In Section 3, we described the problem definition, proposed pattern, mining algorithm, the proposed cost function of the candidate cache victim, and an elaborative example. Some simulation results were made in Section 4. A conclusion was given in Section 5.
2. Preliminaries. In the past decades, a lot of studies had probed into the caching techniques on many applications such as databases [12], web applications and wireless networks. Existing researches in caching on Web could be divided into two main archi-tectures: 1) cooperative and 2) non-cooperative [15]. Besides, there are a great number of caching approaches for wireless cellular network [11]. Some cooperative caching pro-tocols have been proposed for mobile ad hoc networks (MANETs) [2,4,17]. By utilizing geographical proximity or well-known clustering algorithms, a region concept has been em-ployed for the cooperative caching approaches [2,4,17]. In [19], the authors took both data and node locality into considerations and proposed three approaches, which were Cache-Data, CachePath and HybridCache respectively. However, the aforementioned caching techniques do not suit WMSNs due to the severe resource constraints like memory space and energy consumption. Most of existing studies on caching technique of wireless sensor networks focus on the placement of caches [16]. In [5], nevertheless, Dimokas et al. first proposed an approach of cooperative caching in WMSNs. A node importance-based co-operative caching (NICoCa) in [5] created a precedent of coco-operative caching on WMSNs. The plain assumption of NICoCa is that each sensor node has a moderate local stor-age capacity associated with it, i.e., a flash memory. NICoCa exploits the importance of
Yes, Local Hit, update the timestamp
Yes, Proximity Hit
Yes, Remote Hit
Item A V1
N (V1)
P (V1,D)
Global Hit
V1
Figure 1. Cache discovery flow chart
Yes, purging procedure Available Cache Space
is smaller than required
Candidate Victim Selection Cost Function
datum with the largest cost
Purge the datum Cost Function (Item)
Cost(i)= access K kth now T DataSize TTL u LRU mechanism Occurrence Concept
Figure 2. Cache replacement flow chart
sensors relative to the network topology in terms of their positions in the network and/or residual energy. A concept of mediator node [5], which plays a significant part of position in the network, is proposed to aid cooperative caching. Integrating the sensor’s position and residual energy into the caching strategy facilitates the prolongation of network life-time and short latency of mullife-timedia data retrieval. In our network model, we adopt the concept of NICoCa for the cache discovery component protocol [5] as shown in Figure 1. Four cases of the requested item may occur: 1) Local Hit (LH), 2) Proximity Hit (PH), 3) Remote Hit (RH) and 4) Global Hit (GH). When a sensor node issues a request of a
766 V. S. TSENG, M.-H. HSIEH AND K. W. LIN
multimedia item, it first searches its local cache. A Local Hit represents that the multi-media item is found on the local cache and a Proximity Hit denotes that the requested multimedia item is cached by a node in the 2-hop neighborhood of the original sensor node. If the multimedia item is not cached on the original sensor node, the original sensor node broadcasts the request to its 2-hop neighbors and mediator nodes for Proximity Hit. If a proximity miss occurs, the request’s route toward the Data Center proceeds. During the routing, if the requested item is cached by a node having at least one mediator node residing along the routing path, it is called Remote Hit. A Global Hit means that the requested item is acquired from the Data Center. As far as we know, the more the number of LH, PH and RH is, the better the QoS is. Considering the restrictions on hardware, most caching techniques require a cache replacement mechanism. An adaptive cache re-placement algorithm benefits the improvement of LH, PH and RH. Figure 2 illustrates the flow chart of cache replacement in NICoCa. A purging procedure of the cached items proceeds when the available space is smaller than the required space. For the selection of the candidate cache victim, Dimokas et al. proposed a cost function integrating many factors, such as data size, concept of Least Recently Used (LRU), Time-to-Live and the number of the multimedia item requested by the sensor node.
The concept of LRU and the number of the multimedia item requested by the sensor node may be insufficient to reveal the requesting information such as items are temporally requested by some sensor nodes. Considering the requesting information, a temporal region requesting pattern (TRRP) was proposed in our work. A TRRP represents that the multimedia item is temporally requested by a sensor region and reveals the relation between the multimedia items’s requesting information and the sensor node. In next section, the problem definition, the definition of TRRP, mining methodology, the proposed cost function and an elaborative example were described.
3. Problem Definition and Mining Approach. In our work, we adopted the network model as proposed in NICoCa [5]. In this model, it is assumed that the elemental source of multimedia data is a Data Center and sensor nodes are capable of caching the datum which have requested. In addition, it is supposed that each sensor node is aware of its 2-hop neighborhood. An ordinary sensor routing protocol accompanies the sending of requests. Each sensor node stores the related metadata of a cached multimedia item (datum) [5]. A request item is in the form as (dataID, timestamp) and represents that a sensor node launches a request for a multimedia item, dataID, at the timestamp. A request transaction of a sensor node consists of a series of request items. Therefore, a request transaction table from NICoCa recording the request logs of each sensor node [5] can be obtained as Table 1 shown. Without loss of generality, some multimedia items are requested temporally by some nodes which are 2-hop neighbors of each other. For the existing cache replacement in WMSNs, to determine whether a multimedia item is purged from the cache of the sensor node or not mainly depends on the size of the multimedia item and the last access timestamp. However, some multimedia items with large data size were temporally requested by the mediator nodes locating a specific region. Such multimedia items may suffer from a frequent replacement due to their large data size. To solve the above problem, one of our purposes was to discover temporal region requesting patterns. A temporal region requesting pattern represents that the multimedia item is temporally requested by the sensor nodes locating a 2-hop region. In our work, we named the temporal region requesting patterns as TRRPs. A mining problem of TRRPs was explained in next subsection.
In this way, the replacement mechanism had to calculate the cost of all the cached items to decide the candidate cache victim. Nevertheless, the cached item with the largest cost might be the one requested in the next K requests, where K was smaller than 70.
The experiments of adjusting the event probability were shown as Figures 8(a) and 8(b). A higher event probability represents that the sensor node more likely issues the requests which have been issued in the past. Therefore, with the increase in the event probability, the hit rate of the five methods had a gradual raise. In the meanwhile, the number of replacements of the five methods had correspondingly decreased. The cost functions of NICoCa, Belady’s Algorithm+NICoCa, and TRRC consider the occurrence of the requested items as well. As the event probability increased, therefore, the gap of the hit rate between the LRU and NICoCa enlarged.
5. Conclusions. This paper proposed a novel pattern named temporal region requesting pattern (TRRP) and a one-scanning mining approach, namely TRRP-Mine, to discover the TRRPs. In addition, by utilizing the discovered TRRPs, a pattern-based cost func-tion, named as Temporal Region Requesting Cost Function (TRRC), was designed for cache replacement of WMSNs. In TRRC, two novel factors, namely temporal reference rate and future discarding rate, were proposed to calculate the score of candidate cache vic-tim based on the requested probability built by temporal requesting information. To our best knowledge, this was the first work applying the pattern-based cost function to explore cache replacement of WMSNs. In the simulations, we compared TRRC with some well-known cache algorithms such as LRU, MRU, NICoCa and Belady’s Algorithm+NICoCa in terms of hit rate and the number of replacement.
In the meantime, the performance was evaluated by adjusting four key parameters, which are ItemNum, CacheSize, AvgItemSize and AvgEventProb respectively. Through empirical evaluations of different parameter settings, the hit rate of TRRC was higher than that of the other methods by 7% and the replacements of TRRC were fewer than those of the others by 25% at least. This result indicated that TRRC achieved efficiently the short latency in multimedia data retrieval. Although the proposed pattern-based cost function (TRRC) requires a period of collection of requesting information to train temporal reference rate and future discarding rate, for the future work, we will apply the stream mining technique for achieving the real-time application.
Acknowledgment. This research was partially supported by National Science Council,
Taiwan, under Grant No. 97-2218-E-151-003-MY2.
REFERENCES
[1] I. Akyildiz, T. Melodia and K. R. Chowdhury, A survey of wireless multimedia sensor networks,
Computer Networks, vol.51, pp.921-960, 2007.
[2] N. Chand, R. C. Joshi and M. Misra, Cooperative caching strategy in mobile ad hoc networks based on clusters, Wireless Personal Communications, vol.43, no.1, pp.41-63, 2007.
[3] N. Chand, R. C. Joshi and M. Misra, A zone co-operation approach for efficient caching in mobile ad hoc networks, International Journal of Communication Systems, vol.19, pp.1009-1028, 2006. [4] C.-Y. Chow, H. V. Leong and A. T. S. Chan, GroCoca: Group-based peer-to-peer cooperative
caching in mobile environment, IEEE Journal on Selected Areas in Communications, vol.25, pp.179-191, 2007.
[5] N. Dimokas, D. Katsaros and Y. Manolopoulos, Cooperative caching in wireless multimedia sensor networks, Mobile Networks and Applications, vol.13, pp.337-356, 2008.
[6] J. Harju and P. Kivimaki, Co-operation and comparison of diffserv and intserv: Performance mea-surements, Proc. of the 25th Conf. on Local Computer Networks, pp.177-186, 2000.
[7] R. Holman, J. Stanley and T. Ozkan-Haller, Applying video sensor networks to nearshore environ-ment monitoring, IEEE Pervasive Computing, vol.2, no.4, pp.14-21, 2003.
CACHE REPLACEMENT ALGORITHM FOR COOPERATIVE CACHING IN WMSNS 775 Event Probability 0.0 0.1 0.2 0.3 0.4 0.5 0.6 H it R a te (% ) 0 10 20 30 40 50 60 70 80 90 100 LRU MRU NICoCa Belady's Algorithm+NICoCa TRRC Event Probability 0.0 0.1 0.2 0.3 0.4 0.5 0.6 N u m b e r o f R e p la c e m e n t 0.0 4.0e+5 8.0e+5 1.2e+6 1.6e+6 2.0e+6 2.4e+6 LRU MRU NICoCa Belady's Algorithm+NICoCa TRRC
Average Data Size of Multimedia Item (mb) 0.0 2.5 5.0 7.5 10.0 12.5 15.0 H it R a te ( % ) 0 10 20 30 40 50 60 70 80 90 LRU MRU NICoCa Belady's Algorithm+NICoCa TRRC
Average Data Size of Multimedia Item (mb) 0.0 2.5 5.0 7.5 10.0 12.5 15.0 N u m b e r o f R e p la c e m e n t 4.0e+5 8.0e+5 1.2e+6 1.6e+6 2.0e+6 2.4e+6 2.8e+6 LRU MRU NICoCa Belady's Algorithm+NICoCa TRRC 7(a) 7(b) 8(a) 8(b)
Figure 7. (a) Hit rate for LRU, MRU, NICoCa, Belady’s algorithm+NICoCa and
TRRC with the average data size of multimedia item varied; (b) Number of replacement for LRU, MRU, NICoCa, Belady’s algorithm+NICoCa and TRRC with the average data size of multimedia item varied
Figure 8. (a) Hit rate for LRU, MRU, NICoCa, Beladys algorithm+NICoCa and TRRC with the event probability varied; (b) Number of replacement for LRU, MRU, NICoCa, Beladys algorithm+NICoCa and TRRC with the event probability varied
[8] F. Hu and S. Kumar, Multimedia query with QoS considerations for wireless sensor networks in telemedicine, Proc. of Int. Conf. on Internet Multimedia Management Systems, 2003.
[9] M. Ito and M. Tanaka, Localization of a moving sensor by particle filters, International Journal of
Innovative Computing, Information and Control, vol.4, no.1, pp.165-173, 2008.
[10] Y. Jin, J. Jo and Y. Kim, Energy-efficient multi-hop communication scheme in clustered sensor networks, International Journal of Innovative Computing, Information and Control, vol.4, no.7, pp.1741-1749, 2008.
[11] D. Katsaros and Y. Manolopoulos, Web caching in broadcast mobile wireless environments, IEEE
Internet Computing, vol.8, no.3, pp.37-45, 2004.
[12] N. Megiddo and D. S. Modha, ARC: A self-tuning, low overhead replacement cache, Proc. of the
USENIX Conf. on File and Storage Technologies, pp.115-130, 2003.
[13] H. Nakano, A. Utani, A. Miyauchi and H. Yamamoto, An efficient data gathering scheme using a chaotic PCNN in wireless sensor networks with multiple sink nodes, ICIC Express Letters, vol.3, no.3(B), pp.805-812, 2009.
[14] Y. Obashi, H. Chen, H. Mineno and T. Mizuno, An energy-aware routing scheme with node relay willingness in wireless sensor networks, International Journal of Innovative Computing, Information
and Control, vol.3, no.3, pp.565-574, 2007.
[15] G. Pallis, A. Vakali and J. Pokorny, A clustering-based prefetching scheme on a Web cache environ-ment, Computers & Electrical Engineering, vol.34, no.4, pp.309-323, 2008.
[16] K. S. Prabh and T. F. Abdelzaher, Energy-conserving data cache placement in sensor networks,
ACM Transactions on Sensor Networks, vol.1, no.2, pp.178-203, 2005.
[17] H. Shen, S. K. Das, M. Kumar and Z. Wang, Cooperative caching with optimal radius in hybrid wireless networks, Proc. of the Int. IFIP-TC6 Networking Conference, LNCS, vol.3042, pp.841-853, 2004.
[18] K. Yamada, K. Kimura, H. Yuki and K. Yoshida, The home network system by mutual complement of wireless and wired communications, ICIC Express Letters, vol.2, no.1, pp.73-79, 2008.
[19] L. Yin and G. Cao, Supporting cooperative caching in ad hoc networks, IEEE Transactions on
