1. 1 Cooperative Caching in Wireless Multimedia Sensor Nets Nikos Dimokas 1 Dimitrios Katsaros 1,2 (presentation) Yannis Manolopoulos 1 3 rd MobiMedia Conference, Nafpaktos, Greece, 27-29/August/2007 1 Informatics Dept., Aristotle University, Thessaloniki, Greece 2 Computer & Communication Engin. Dept., University of Thessaly, Volos, Greece

2. 2 Wireless Sensor Networks (WSNs) Wireless Sensor Networks features Homogeneous devices Stationary nodes Dispersed network Large network size Self-organized All nodes acts as routers No wired infrastructure Potential multihop routes

3. 3 Communication in WSNs Communication between two unconnected nodes is achieved through intermediate nodes Every node that falls inside the communication range r of a node u, is considered reachable

4. 4 WSNs - Applications Applications Habitat monitoring Disaster relief Target tracking Agriculture

5. 5 Wireless Multimedia Sensor Nets (WMSNs) Cheap CMOS cameras: Cyclops imaging module is a light-weight imaging module which can be adapted to MICA2 or MICAz sensor nodes

6. 6 WMSNs - Applications Boost the existing application of WSNs Create new applications multimedia surveillance sensor networks : will be composed by miniature video cameras and will be able to communicate, to process and store data relevant to crimes and terrorist attacks traffic avoidance and control systems : will monitor car traffic and offer routing advices to prevent congestion industrial process control : will be realized by WMSNs that will offer time-critical information related to imaging, temperature, pressure, etc

7. 7 What’s so special about WMSNs ? [ Ian Akyildiz: Dec’06 ] We have to rethink the computation-communication paradigm of traditional WSNs which focused only on reducing energy consumption WMSNs applications have a second goal, as important as the energy consumption delivery of application-level quality of service (QoS) mapping of this requirement to network layer metrics, like latency This goal has (almost) been ignored in mainstream research efforts on traditional WSNs

8. 8 What’s so special about WMSNs ? Resource constraints sensor nodes are battery-, memory- and processing- starving devices Variable channel capacity multi-hop nature of WMSNs implies that wireless link capacity depends on the interference level among nodes Multimedia in-network processing sensor nodes store rich media (image, video), and must retrieve such media from remote sensor nodes with short latency

9. 9 Our proposal … Cooperative Caching: NICOCA protocol multiple sensor nodes share and coordinate cache data to cut communication cost and exploit the aggregate cache space of cooperating sensors Each sensor node has a moderate local storage capacity associated with it, i.e., a flash memory

10. 10 Relevant work (1/2) Caching in operating systems, in databases, on the Web No extreme resource constraints like WMSNs Caching for wireless broadcast cellular networks More powerful nodes, and one-hop communication with resource-rich base stations Most relevant research works: cooperative caching protocols for MANETs GroCoca : organize nodes into groups based on their data request pattern and their mobility pattern ECOR, Zone Co-operative, Cluster Cooperative : form clusters of nodes based either in geographical proximity or utilizing widely known node clustering algorithms for MANETs

11. 11 Relevant work (2/2) Protocols that deviated from such approaches: CacheData : intermediate nodes cache the data to serve future requests instead of fetching data from their source CachePath : mobile nodes cache the data path and use it to redirect future requests to the nearby node which has the data instead of the faraway origin node Amalgamation of them: the champion HybridCache cooperative caching for MANETs One caching work on WSNs concerns the placement of caches

12. 12 Our contributions … Definition of a metric for estimating the importance of a sensor node, which will imply short latency in retrieval Description of a cooperative caching protocol which takes into account the residual energy Datum discovery and cache replacement component subprotocols Performance evaluation of the protocol and comparison with the state-of-the-art cooperative caching for MANETs, with J-Sim

13. 13 Terminology and assumptions WMSN is abstracted as a graph G(V,E) An edge e=(u,v) exists if and only if u is in the transmission range of v and vice versa. All links in the graph are bidirectional. The network is assumed to be connected N 1 (v) : the set of one hop neighbours of v N 2 (v) : the set of two hop neighbours of v N 12 (v) : combined set of N 1 (v) and N 2 (v) LN v : is the induced subgraph of G associated with vertices in N 12 (v) d G (v,u) : distance between v and u

14. 14 A measure of sensor importance Let σ uw =σ wu denote the number of shortest paths from u  V to w  V (by definition, σ uu = 0) Let σ uw (v) denote the number of shortest paths from u to w that some vertex v  V lies on We define the node importance index NI(v) of a vertex v as: Large values for the NI index of a node v indicate that this node can reach others on relatively short paths, or that v lies on considerable fractions of shortest paths connecting others

15. 15 The NI index in sample graphs In parenthesis, the NI index of the respective node; i.e., 7(156): node with ID 7 has NI equal to 156. Nodes with large NI:  Articulation nodes (in bridges), e.g., 3, 4, 7, 16, 18  With large fanout, e.g., 14, 8, U Therefore: geodesic nodes

16. 16 The NI index in a localized algorithm For any node v, the NI indexes of the nodes in N 12 (v) calculated only for the subgraph of the 2-hop (in general, k -hop) neighborhood reveal the relative importance of the nodes in covering N 12 For a node u (of the 2-hop neighbourhood of a node v ), the NI index of u will be denoted as NI v (u)

17. 17 NI computation At a first glance, NI computation seems expensive, i.e., O( m * n 2 ) operations in total for a 2-hop neighbourhood, which consists of n nodes and m links: calculating the shortest path between a particular pair of vertices (assume for the moment that there exists only one) can be done using bfs in O( m ) time, and there exist O( n 2 ) vertex pairs Fortunately, we can do better than this by making some smart observations. The improved algorithm ( CalculateNodeImportanceIndex ) is quite complicated and beyond the scope of this presentation THEOREM. The complexity of the algorithm CalculateNodeImportanceIndex is O( n * m ) for a graph with n vertices and m edges

18. 18 Pseudocode for CalculateNodeImportanceIndex (1/2)

19. 19 Pseudocode for CalculateNodeImportanceIndex (2/2)

20. 20 Housekeeping information in NICoCa Ultimate source of multimedia data is a Data Center Each node is aware of its 2-hop neighborhood Uses NI to characterize some neighbors as mediators A node can be either a mediator or an ordinary node Each sensor node stores the dataID, and the actual multimedia datum the data size, TTL interval for each cached item, the timestamps of the K most recent accesses each cached item is characterized either as O (i.e., own) or H (i.e., hosted)

21. 21 The cache discovery protocol (1/2) A sensor node issues a request for a multimedia item Searches its local cache and if it is found ( local cache hit ) then the K most recent access timestamps are updated Otherwise ( local cache miss ), the request is broadcasted and received by the mediators These check the 2-hop neighbors of the requesting node whether they cache the datum ( proximity hit ) If none of them responds ( proximity cache miss ), then the request is directed to the Data Center

22. 22 The cache discovery protocol (2/2) When a mediator receives a request, searches its cache If it deduces that the request can be satisfied by a neighboring node ( remote cache hit ), forwards the request to the neighboring node with the largest residual energy If the request can not be satisfied by this mediator node, then it does not forward it recursively to its own mediators, since this will be done by the routing protocol, e.g., AODV If none of the nodes can help, then requested datum is served by the Data Center ( global hit )

23. 23 The cache replacement protocol Each sensor node first purges the data that it has cached on behalf of some other node Calculate the following function for each cached datum i The candidate cache victim is the item which incurs the largest cost Inform the mediators about the candidate victim If it is cached by a mediator, the metadata are updated If not, it is forwarded and cached to the node with the largest residual energy

24. 24 Evaluation setting (1/2) We compared NICOCA to: Hybrid, state-of-the-art cooperative caching protocol for MANETs Implementation of protocols using J-Sim simulation library

25. 25 Evaluation setting (2/2) Measured quantities number of hits (local, remote and global) residual energy level of the sensor nodes average latency for getting the requested data the number of packets dropped Present here only results for number of hits representative of: latency, collisions and energy consumption A small number of global hits less network congestion, fewer collisions and packet drops. Large number of remote hits  effectiveness of cooperation Large number of local hits ≠ effective cooperation the cost of global hits vanishes the benefits of local hits

26. 26 Cache vs. hits (MB files & uniform access) in a sparse WMSN (d = 4)

27. 27 Cache vs. hits (MB files & uniform access) in a dense WMSN (d = 7)

28. 28 Cache vs. hits (MB files & uniform access) in a very dense WMSN (d = 10)

29. 29 Observe: MB files & uniform access For all network topologies (sparse, dense and very dense), NICoCa achieves more remote hits and less global hits than HybridCache This performance gap widens in favor of NICoCa as we move from sparse to denser WMSNs For very dense sensor deployments, NICoCa achieves double the remote hits of HybridCache and only half of its global hits For sparse WMSNs HybridCache achieves slightly more local hits than does NICoCa, but this gap vanishes completely when moving to denser network This small gain of HybridCache for sparse topologies is not advantageous at all, since it incurs global hits as many as twice the number of its local hits

30. 30 Cache vs. hits (KB files & Zipfian access) in a sparse WMSN (d = 4)

31. 31 Cache vs. hits (KB files & Zipfian access) in a dense WMSN (d = 7)

32. 32 Cache vs. hits (KB files & Zipfian access) in a very dense WMSN (d = 10)

33. 33 Observe: KB files & Zipfian access For all network topologies (sparse, dense and very dense), NICoCa achieves more remote hits and less global hits than HybridCache For very dense WMSNs, the requests reaching Data Center for NICoCa are less than half those of HybridCache! NICoCa's global hits do not vary significantly with varying network topologies and varying local sensor storage Global hits of HybridCache are severely affected by the topology and the cache size For cache equal to 1% of the total data, HybridCache's global hits increase at a pace of 50%! The results for Zipfian access on megabyte-sized data more impressively in favor of NICoCa

34. 34 Summary Wireless Multimedia Sensor Networks (WMSNs) Unique features of WMSNs call for protocol designs that provide application-level QoS Cooperative caching protocol, NICoCa, suitable for WMSNs NICOCA evaluation with J-Sim and comparison to the state-of-the-art protocol NICOCA can: reduce the global hits at an average percentage of 50% increase the remote hits (due to the effective sensor cooperation) at an average percentage of 40%

35. 35 Important references 1.I. Akyildiz, T. Melodia, and K. R. Chowdhury. A survey of wireless multimedia sensor networks. Computer Networks, 51:921-960, 2007 2.Y. Diao, D. Ganesan, G. Mathur, and P. Shenoy. Rethinking data management for storage-centric sensor networks. Proceedings of the Conference on Innovative Data Systems Research (CIDR), pp. 22- 31, 2007 3.S. Nath and A. Kansal. FlashDB: Dynamic self-tuning database for NAND flash. Proceedings of the ACM International Conference on Information Processing in Sensor Networks (IPSN), pp. 410-419, 2007 4.L. Yin and G. Cao. Supporting cooperative caching in ad hoc networks. IEEE Transactions on Mobile Computing, 5(1):77-89, 2006

36. 36 Thank you for your attention! Any questions?