GHT Geographic Hash Table for Data Centric Storage

1
GHT Geographic Hash Table for Data Centric
Storage

• Sylvia et. al
• Presented By - Jay

2
Concepts

• Data Centric Stroage
• Data are named and communication abstractions
  refer to these names rather than to node network
  addresses.
• Hash Table
• Distribute evenly and provide fast lookup.
• Geographic Hash Table
• Distribute over geography

3
Assumptions

• Large scale
• Geographic boundaries are known
• Nodes know their locations
• Access Points
• Minimize Energy Consumption
• Total Usage
• Hot Spot Usage

4
Sensor Net Data

• Observations
• Low level readings
• Events
• Derived from individual readings
• Tasks
• Users specified instructions
• Actions
• Taken after event occurence
• Queries
• Retrieve information from network

5
Data Dissemination

• Communication Costs
• Message Transmission for floods
• O(n)
• Point to Point Routing
• O(vn)

6
Canonical Methods

• External Storage
• Relevant data sent to central storage
• Costs O(vn) for each event
• No cost for user queries
• Local Storage
• Data stored locally, queries flood.
• Costs O(n) for queries
• Response costs O(vn)

7
Canonical Methods

• Data Centric Storage
• Store event by name, query by name
• O(vn) to store
• O(vn) to query
• O(vn) to return

8
Approximate Communication Costs

• n nodes
• T event types
• Dtotal total no of events detected
• Q no of event types for which queries are
  issued
• Dq no of events detected for the types of
  events queried for.

9
Cost

• External
• Total Dtotal ( vn)
• Hotspot Dtotal
• Local
• Total Q n Dq (vn
• Hotspot Q Dq

10
Data Centric Stroage

• Total
• List
• Q ( vn) Dtotal (vn) Dq (vn)
• Summary
• Q ( vn) Dtotal (vn) Q (vn)
• Hotspot
• List
• Q Dq (list)
• Summary
• 2 Q

11
Important Points

• As n get large, local storage incurs the highest
  packet count
• ES incurs a lower total message count but not by
  much
• If Dq gtgt Q and events are summarized, DCS has
  lowest load on ACCESS PATH
• If events are listed and Dtotal gtgt Dq than DCS
  and LS have lower access loads than external
  storage.

12
Where to Use DCS

• Sensor networks are large
• There are many detected events but not all are
  queried.

13
Design Criteria

• Node Failures
• Topology Changes
• Scalability
• Energy Constraints

14
Goals

• Persistence A (k, v) pair persists
• Consistency query to be routed to the correct
  (k,v) node
• Scalability Concentration of storage should be
  avoided

15
GHT

• Put()
• Get()
• Key k to a location k. Node nearest to k stores
  the (k,v)
• PRP Perimeter Refresh Protocol
• SR Structured Replication

16
GPSR

• Greedy Perimeter State Routing
• Greedy
• P
• erimeter

17
Greedy
18
Greedy Problem
19
Perimeter
20
Home Node and Home Perimeter

• Home Node
• Node closest to the geographic location.
• Home Perimeter
• The perimeter enclosing the location.
• What happens in case of mobility and failures.

21
Permiter Refresh Protocol

• Stores a copy of a key-value pair at each node on
  the home perimeter.
• Replica nodes.
• Th seconds, home node generates a refresh packet.
• Take over time , Tt runs at replicas.
• Td, death timer.
• In GHT, Td 3 Th and Tt 2 Th

22
PRP
23
PRP
24
PRP
25
Structured Replication

• Augment event name with hierarchy depth
• Given root r and given hierarchy depth d
• 4d 1
• Storage costs lower from
• O(v n) to O(v n/ 2d)
• Query costs go up
• O(v n) to O(2d v n)

26
Simulation

• Ns2
• 802.11 MAC
• 40 m lt 250m

27
Sim Results
28
SR
29
SR
30
Comparative Study

• n no of nodes
• T no of Event types
• Q no. of event types
• Di no of detected events of type i

31
(No Transcript)
32
(No Transcript)
33
(No Transcript)
34
(No Transcript)
35
Conclusion and Future Work

• Sounds good
• Effect of varying node density as DHT distributes
  by location
• Boundary knowledge to avoid hashing outside the
  boundary