1. Network Kernel Architectures and Implementation (01204423) Data-Centric and Content-Based Networking Chaiporn Jaikaeo chaiporn.j@ku.ac.th Department of Computer Engineering Kasetsart University Materials taken from lecture slides by Karl and Willig

2. 2 Overview Interaction patterns and programming model Interaction patterns and programming model Data-centric routing Data-centric routing Data aggregation Data aggregation Data storage Data storage

3. 3 Interaction Patterns Standard networking interaction Standard networking interaction  Client/server, peer-to-peer  Explicit or implicit partners, explicit cause for communication Desirable properties for WSN (and other applications) Desirable properties for WSN (and other applications)  Decoupling in space – neither sender nor receiver need to know their partner  Decoupling in time – “answer” not necessarily directly triggered by “question”, asynchronous communication

4. 4 Publish/Subscribe Interaction Achieved by publish/subscribe paradigm Achieved by publish/subscribe paradigm  Idea: Entities can publish data under certain names  Entities can subscribe to updates of such named data Software bus Publisher 1 Publisher 2 Subscriber 1Subscriber 2 Subscriber 3 Variations Variations  Topic-based:  Inflexible  Content-based  use general predicates over named data

5. 5 P/S Implementation Options Central server: mostly not applicable Central server: mostly not applicable Topic-based: group communication (multicast) Topic-based: group communication (multicast) Content-based: does not directly map to multicast Content-based: does not directly map to multicast  Needs content-based routing/forwarding for efficient networking t < 10 t > 20 t < 10 t 20 t > 20 t > 40 t > 20 t 20 t = 35!

6. 6 Overview Interaction patterns and programming model Interaction patterns and programming model Data-centric routing Data-centric routing Data aggregation Data aggregation Data storage Data storage

7. 7 One-Shot Interactions Too much overhead to explore topology Too much overhead to explore topology For small data sets For small data sets  Flooding/gossiping For large data sets For large data sets  Idea: exchange data summary with neighbor, ask whether it is interested in data  Only transmit data when explicitly requested  Nodes should know about interests of further away nodes  E.g., Sensor Protocol for Information via Negotiation (SPIN)

8. 8 SPIN example ADV (1) REQ (2) DATA (3) (4) ADV (5) REQ (6) DATA

9. 9 Repeated Interactions More interesting: Subscribe once, events happen multiple times More interesting: Subscribe once, events happen multiple times  Exploring the network topology might actually pay off  But: unknown which node can provide data, multiple nodes might ask for data How to map this onto a “routing” problem? How to map this onto a “routing” problem?

10. 10 Repeated Interactions Idea: Put enough information into the network so that publications and subscriptions can be mapped onto each other Idea: Put enough information into the network so that publications and subscriptions can be mapped onto each other  But try to avoid using unique identifiers One option: Directed diffusion One option: Directed diffusion  Try to rely only on local interactions for implementation

11. 11 Directed Diffusion Two-phase Pull Two-phase Pull Phase 1: nodes distribute interests Phase 1: nodes distribute interests Interests are flooded in the network Interests are flooded in the network  Apparently obvious solution: remember from where interests came, set up a convergecast tree  Problem? Sink 1 Sink 2 Sink 3 Source X Sink Source X

12. 12 Direction Diffusion Option 1: Node X forwarding received data to all “parents” in a “convergecast tree” Option 1: Node X forwarding received data to all “parents” in a “convergecast tree”  Not attractive, many needless packet repetitions over multiple routes Option 2: node X only forwards to one parent Option 2: node X only forwards to one parent  Not acceptable, data sinks might miss events

13. 13 Direction Diffusion Option 3: Only provisionally send data to all parents, but ask data sinks to help in selecting which paths are redundant, which are needed Option 3: Only provisionally send data to all parents, but ask data sinks to help in selecting which paths are redundant, which are needed  Information from where an interest came is called gradient  Forward all published data along all existing gradients

14. 14 Gradient Reinforcement Gradients express not only a link in a tree, but a quantified “strength” of relationship Gradients express not only a link in a tree, but a quantified “strength” of relationship  Initialized to low values  Strength represents also rate with which data is to be sent Intermediate nodes forward on all gradients Intermediate nodes forward on all gradients  Can use a data cache to suppress needless duplicates

15. 15 Directed Diffusion Second phase: Nodes that contribute new data (not found in cache) should be encouraged to send more data Second phase: Nodes that contribute new data (not found in cache) should be encouraged to send more data  Sending rate is increased, the gradient is reinforced  Gradient reinforcement can start from the sink  If requested rate is higher than available rate, gradient reinforcement propagates towards original data sources Adapts to changes in data sources, topology, sinks Adapts to changes in data sources, topology, sinks

16. 16 Directed Diffusion - Example

17. 17 Directed Diffusion – Extensions Two-phase pull suffers from interest flooding problems Two-phase pull suffers from interest flooding problems  Mitigated by combining with topology control, in particular, passive clustering Geographic scoping & directed diffusion Geographic scoping & directed diffusion

18. 18 Directed Diffusion – Extensions Push diffusion – few senders, many receivers Push diffusion – few senders, many receivers  Same interface/naming concept  Do not flood interests, but flood the data  Interested nodes will start reinforcing the gradients Pull diffusion – many senders, few receivers Pull diffusion – many senders, few receivers  Still flood interest messages, but directly set up a real tree

19. 19 Overview Interaction patterns and programming model Interaction patterns and programming model Data-centric routing Data-centric routing Data aggregation Data aggregation Data storage Data storage

20. 20 Data Aggregation Packets may combine their data into fewer packets Packets may combine their data into fewer packets  Data Aggregation Depending on network, aggregation can be useful or pointless Depending on network, aggregation can be useful or pointless

21. 21 Metrics for Data Aggregation Accuracy Accuracy  E.g., differences, ratios, statistics Completeness Completeness  A form of accuracy Latency Latency Message overhead Message overhead

22. 22 Expressing Aggregation Request One option: Use database abstraction One option: Use database abstraction  Request by SQL clauses E.g., E.g., SELECT AVG(temperature), floor WHERE temperature > 30 GROUP BY floor HAVING floor > 5

23. 23 Intermediate Results Aggregation requires partial information to represent intermediate results Aggregation requires partial information to represent intermediate results  Called Partial State Records  E.g., to compute average, sum and number of previously aggregated values is required  Expressed as  Expressed as  Update rule:  Final result is simply s/c

24. 24 Categories of Aggregation Classified by properties of aggregation functions Classified by properties of aggregation functions  Duplicate sensitive  E.g., median, sum, histograms  Insensitive: maximum or minimum  Summary or examplary  Composable: for aggregation function f, there exists a function g such that f(W 1  W 2 ) = g( f(W 1 ), f(W 2 ) )  Monotonic

25. 25 Categories of Aggregation Behavior of partial state records Behavior of partial state records  Distributive – end results directly as partial state record, e.g., MIN  Algebraic – p.s.r. has constant size; end result easily derived  Content-sensitive – size and structure depend on measured values (e.g., histogram)  Holistic – all data need to be included, e.g., median  Unique – only distinct data values are needed

26. 26 Placement of Agg. Points Convergecast trees provide natural aggregation points Convergecast trees provide natural aggregation points But: what are good aggregation points? But: what are good aggregation points?  Ideally: choose tree structure that minimizes size of aggregated data  Figuratively: long trunks, bushy at the leaves  In fact: a Steiner tree problem in disguise Good aggregation tree structure can be obtained by slightly modifying Takahashi- Matsuyama heuristic Good aggregation tree structure can be obtained by slightly modifying Takahashi- Matsuyama heuristic

27. 27 Placement of Agg. Points Alternative: look at parent selection rule in a simple flooding-based tree construction Alternative: look at parent selection rule in a simple flooding-based tree construction  E.g., first inviter as parent, random inviter, nearest inviter, …  Result: no simple rule guarantees an optimal aggregation structure

28. 28 Broadcasting Aggregated Value Goal: distribute an aggregate of all nodes’ measurements to all nodes Goal: distribute an aggregate of all nodes’ measurements to all nodes  Construcing |V| convergecast trees not appropriate Idea: Use gossiping combined with aggregation Idea: Use gossiping combined with aggregation  Compute new aggregate when receiving new value  If significant, gossip

29. 29 Overview Interaction patterns and programming model Interaction patterns and programming model Data-centric routing Data-centric routing Data aggregation Data aggregation Data storage Data storage

30. 30 Data-Centric Storage Goal: Store data for later retrieval Goal: Store data for later retrieval  Difficult in absence of gateway nodes/servers Question: Where to put a certain datum? Question: Where to put a certain datum?  Avoid a complex directory service

31. 31 Data-Centric Storage Idea: Let name of data describe which node is in charge Idea: Let name of data describe which node is in charge  Data name is hashed to a geographic position  Node closest to this position is in charge of holding data  Similar to peer-to-peer networking (DHT)  Geographic Hash Tables (GHT)  Use geographic routing to store/retrieve data at this “location” (in fact, the node)

32. 32 Geographic Hash Tables Nodes not available at the hashed location Nodes not available at the hashed location  Use nearest node Failing and new nodes Failing and new nodes  Replicate data to neighbors  New nodes learn existing value-key pairs from neighbors Limited storage per node Limited storage per node  Distribute data to other nodes in the vicinity Key location Timeout New key location

33. 33 Conclusion Data-centric networking considerably changes communication protocol design paradigms Data-centric networking considerably changes communication protocol design paradigms  Nicely supports an intuitive programming model – publish/subscribe Aggregation is a key enabler for efficient networking Aggregation is a key enabler for efficient networking