Location Directory Services - PowerPoint PPT Presentation

1 / 33
About This Presentation
Title:

Location Directory Services

Description:

Papers we shall discuss today. Grid's Location Service (GLS) ... Monarch CMU's wireless extensions for ns. 802.11 Radio. Bandwidth:1Mbps. Radio range: 250m. ... – PowerPoint PPT presentation

Number of Views:11
Avg rating:3.0/5.0
Slides: 34
Provided by: vs3f
Category:

less

Transcript and Presenter's Notes

Title: Location Directory Services


1
Location Directory Services
  • Vivek Sharma
  • 9/26/2001

CS851 Large Scale Deeply Embedded
Networks
2
Overview
  • Problem Statement
  • Related work
  • Design Issues
  • Papers we shall discuss today
  • Grids Location Service (GLS)
  • Randomized Database Groups (RDG)
  • Comparison and Issues
  • Conclusion

3
Problem Statement
  • A directory service for a sensor network where
    nodes can lookup the geographical location of
    other nodes. The service implementation should be
  • Distributed among the nodes
  • Resilient to node failures
  • Scalable to a large number of nodes
  • Should have low memory and communication/power
    overheads

4
Related Work
  • Location Management in Mobile Systems
  • tracking mobility of users to route calls
    efficiently
  • the network has fixed nodes with much more
    resources
  • most of the architectures are hierarchical and
    thus not fault tolerant
  • Ad Hoc Networks
  • conditions closest to a typical sensor network
    (no fixed infrastructure)
  • additional power, communication and scalability
    issues apply
  • Smart Spaces
  • locating people and equipment in an office like
    environment
  • relative to a fixed set of wireless receivers

5
Related Work
  • Peer-to-Peer Applications
  • a distributed service to locate nodes with
    particular data items
  • no resource limitations or mobility in the system
  • Resource Location Problems
  • spatial gossip algorithms

6
Design Issues
Proactive vs.
Reactive (maintaining location
(on demand determination) informa
tion continuously)
Deterministic vs.
Non-Deterministic (e.g., hashing or ID mapping)
(randomized approaches in
choosing location servers)
Hierarchical vs.
Flat distributed set of arrangement of
location servers location
servers
7
Deterministic vs. Non-deterministic approaches
  • Non-deterministic approaches as opposed to
    deterministic approaches are usually inherently
    resilient and are capable of handling large
    degrees of node failure and mobility
  • The main problem while using a random approach is
    to control the randomization to provide desired
    behavior and to reduce the overheads of a random
    approach
  • In deterministic approaches, one has to
    especially work towards providing
    fault-tolerance. Generally, its extra work to
    ensure that a system is resilient to failures

8
Papers to be covered
  • Grids Location Service (GLS)
  • A scalable location service for geographic ad hoc
    routing Jannotti et al (MIT)
  • a location service based on selecting location
    servers based on node ID hash values
  • Randomized Database Groups (RDG)
  • Ad-hoc mobility management with Randomized
    Database Groups - Haas and Liang (Cornell)
  • a non-deterministic approach towards maintaining
    location information

9
Grids Location Service (GLS)

10
GLS Overview
  • The location service is used to enable
    geographical
  • ad-hoc routing
  • The network is divided into ordered grids or
    squares and each node is aware of the divisions
  • Each node determines its geographic position
    using a mechanism such as GPS
  • Every node maintains a table of its current
    neighbors identities and locations (each node
    broadcasts periodic HELLO packets)

11
GLS Overview
  • Location Servers Every node selects a group of
    nodes (location servers for that node)
    distributed throughout the network, where it
    maintains its current location.
  • Routing the location of the destination is
    determined by performing a location query and
    routing is then done using Geographic Forwarding.
  • Geographic Forwarding When a node needs to send
    a packet towards location P, the node forwards
    the packet to the node amongst its neighbors
    which is closest to P.

12
Example
Bs location servers
13
Selecting and Querying Location Servers
  • Selection A node recruits other nodes with IDs
    close to its own ID as its location servers.
    Location servers are selected in each sibling of
    a square that contains the node.
  • Querying A sends a request to the least node
    greater than B for which it has information. That
    node forwards the query in the same way.
    Eventually the query will reach a location server
    of B which will forward the query to B itself. B
    can now respond directly.

14
Querying Location Servers Example
15
Updating Location Information
  • A node updates its order-2 location servers every
    time it moves a threshold distance d, its order-3
    servers when it moves a threshold distance 2d,
    and so on. So, a node sends out updates
    proportional to its speed and updates are sent to
    distant servers less often than to local servers
  • Forwarding Pointers are used at the order 1 grid
    to let farther nodes route correctly when a node
    moves out of its square

16
Simulation Scenario
  • Monarch CMUs wireless extensions for ns.
  • 802.11 Radio
  • Bandwidth1Mbps
  • Radio range 250m.
  • 100 nodes/km2
  • Order-1 square side 250 m
  • Mobility random waypoint model
  • Network of 600 nodes the scale of a campus or
    city

17
Results
  • Scalability of GLS

18
Results
  • Performance of GLS in
  • the presence of mobility

19
Results
  • Performance of GLS
  • with node failures

20
Pros and Cons
  • Pros
  • Each node has to maintain a small amount of state
  • The querying technique is not paralyzed by
    failure of location servers
  • Cons
  • Prone to performance degradation due to node
    failures and high degrees of mobility
  • Fixed size squares nodes in high density areas
    have to maintain more state information so there
    is much more stress on these nodes in terms of
    power
  • The nodes should know the GRID structure
    beforehand

21
Randomized Database Groups (RDG)

22
RDG Overview
  • A set of location databases form a virtual
    backbone, which is dynamically formed and
    distributed among the nodes.
  • Location update a node writes its location to a
    randomly chosen group of k databases
  • Location lookup A randomly chosen group of k
    databases is queried.
  • The destination node location is provided to the
    source by the databases at the intersection of
    the queried database group and the group last
    written to by the destination node.

23
The virtual backbone
  • Formation During initial setup, network flooding
    could be used to find the set of nodes that best
    serve as the backbone (e.g. uniformly
    distributed)
  • Maintenance When a backbone node is detached
    from the network, a nearby non-backbone node is
    recruited to take its place

24
Randomized Database Groups
  • Given a virtual backbone with n location
    databases, any combination of k databases forms a
    RDG
  • When a node needs to update its location
    information, it uses any accessible RDG out of
    the nCk possible. Same for location query
  • k could be different for different nodes
    depending on the nodes traffic and mobility
    patterns
  • With appropriately chosen k, the probability of
    non-intersection between the set of databases
    queried and the set of databases updates can be
    made sufficiently small

25
Example
n 6 databases e.g. of RDG all combinations of
size k3 1,2,3,1,2,4,1,2,5,. A
node accesses the set of databases through the
database nearest to it.
Virtual Backbone and the Location Databases
26
Mobile Location Updates
  • Call-origination update the querying node writes
    its current location into the queried databases.
  • Location-change update When a node changes
    location, it updates its new location in a RDG.
  • Periodic Update Apart from the above, a node
    sends location information at every interval.

27
Mobility Management Costs
  • pe probability that a database is inaccessible
    at any time instant.
  • fo(t) PDF for the length of time between any
    two consecutive call originations
  • fm(t) PDF for the length of time between any
    two consecutive location change updates.
  • Tp Periodic update interval
  • cu expected cost of accessing a database
  • cl expected miss penalty.
  • Cupdate k cu
  • Closs cl X Expected number of lost calls per
    unit time

28
Optimal RDG size determination
  • We can see that even for high pe, optimal cost is
    achieved with low k due to the tradeoff in the
    cost metric

29
Pros and Cons
  • Pros
  • Allows tuning of performance based on expected
    parameter values for the system
  • Expected to handle large degrees of node failures
    well
  • Can be made adaptive to each nodes traffic and
    mobility patterns
  • Cons
  • Communication overheads could be significant with
    respect to other approaches due to maintaining
    redundant location info
  • Greater load on the location databases so life
    time could be low for those nodes (although these
    nodes need not be on all the time)
  • Analytical results, a lot of assumptions.
    Unfortunately no simulations to get an idea of
    performance in scenarios

30
Comparison
  • RDG
  • Non-deterministic selection of location databases
  • Scalability
  • k is likely to be high implying storing more
    state information
  • Location servers are especially marked out and
    hence greater load on them (power)
  • Inherently fault resilient due to the random
    approach
  • Expected to handle high degrees of node mobility
    and node switch-offs better (though maybe at a
    higher cost?)
  • GLS
  • Deterministic ID based technique to select
    location servers
  • Scalability
  • State maintenance overheads are low
  • Location information is spread out on all nodes
    (Asm density)
  • Reasonably resilient to node failures due to less
    state info and robust querying method
  • Performance degradation in the presence of a high
    degree of mobility and node switch-offs could be
    significant

31
Conclusions
  • A randomized approach is attractive because of
    its inherent capacity to handle high degrees of
    mobility and provide high degrees of resilience
  • But some of these advantages could be offset by
    the amount of overheads due to redundancy in the
    state information maintained
  • The GLS technique uses techniques similar to
    hashing to distribute location information evenly
    on the set of nodes and uses intelligent
    heuristics to provide a robust location querying
    service

32
Some Issues
  • The implementation of a location directory
    service could impose significant overheads on the
    system
  • Questions to ask -
  • Do we really need a location directory service?
  • ID-less routing, Directed Diffusion
  • Is the value added more than the costs?
  • It might not sound feasible or necessary to have
    a global location service for sensor networks.
    One could consider having
  • a higher level directory service to map Data or
    Tasks to locations, and
  • a lower level directory service to map node-IDs
    to locations within groups

33
Thanks!!
  • Vivek Sharma
Write a Comment
User Comments (0)
About PowerShow.com