Active Coordination in Ad Hoc Networks

1
Active Coordination in Ad Hoc Networks

• Christine Julien
• Gruia-Catalin Roman
• Mobile Computing Laboratory
• Washington University in Saint Louis

2
Ad Hoc Networks

• Form opportunistically
• Change rapidly in response to mobility
• Rely on no fixed infrastructure
• Provide transient interactions
• Foster decoupled computing

3
Computational Model

• Ad hoc network
• Host (mobile or stationary)
• Defines location in physical space
• Agents/Application Components
• Unit of mobility residing on a host
• Data
• Owned and generated by each agent

Host 2
Host 1
Host 3
4
Context-Awareness

• Software and hardware components constantly move
  and change
• Applications must opportunistically adapt
• Context scope extends beyond the local host
• Perception of context varies by application
• Application specific context definition
• Asymmetric context interaction
• Multiple contexts evolving over time

5
Example Application
6
EgoSpaces Middleware

• Presents coordination model for easing
  application development
• Facilitates asymmetric coordination
• Allows applications to flexibly interact with the
  environment
• Based on tuple spaces
• Provides high-level abstractions of ad hoc
  network environment

7
The View Construct

• Maximal context contains all available data
• A view is a projection of the maximal context
• Egocentric abstraction of operating context
• Tailored to an applications individual needs
• Allow agent to control scope of views
• Ease program development
• Minimize performance penalties

Host 2
Host 1
Reference agents view yellow data within one hop
Host 3
8
Programming in EgoSpaces

• Application provides declarative view
  specifications
• EgoSpaces builds and
• maintains views
• Facilitates applications
• interactions with views
• Presents application
• with local-like data
• structure

Application
EgoSpaces
Network Support
9
Interacting with Context

• Present view contents as tuple space
• Use Linda-like operations
• Content-based pattern matching
• retrieve tuples (rd)
• remove tuples (in)
• Can affect overlapping views
• Atomic blocking, atomic probing, and scattered
  probing operations

4
2
1
rdgp(pattern)
match
3
5
10
Model Limitations

• Lacks natural adaptation mechanisms
• Provides only polling operations
• Allows interaction only with the system state
• Requires explicit interaction with individual
  data items
• Difficult to specify behaviors that affect
  general classes of data
• Does not cater to variations in programmer
  expertise

11
Reacting to Context

• Facilitate behavioral adaptation
• Agents respond to presence of certain tuple
• Application responds by removing or changing the
  tuple

Reactive view
F
C
react
change
?react to p remove and out modifiers(?)
12
Transactional Programming

• Atomic sequence of operations over a set of views
• Intermediate results are not visible
• Requires locking all view participants
• Cannot include blocking operations
• Reference agent provides view restriction
• Makes a deadlock-free implementation possible

Ttransaction over v1, v2, begin op1, op2, end
13
Augmented Reactions

• Extended reaction
• Local trigger
• Reaction and transaction are atomic
• Followed reaction
• Remote trigger
• Trigger, removal, and notification are atomic
• Transaction is a separate step

Reactive view
inpg
T
react
T
T
A
out
?react to p remove and out modifiers(?)
followed by T(?)
14
Active Coordination

• Applications benefit from high-level coordination
  with context
• Active views capture natural context interactions
• Apply lessons learned in other coordination
  middleware while accounting for ad hoc networks
• Reference agents attach behaviors to views
• Encountering specified conditions triggers an
  automatic response
• Generality and extensibility

15
Active CoordinationData Migration

• Replica management is impractical
• Transparently collect certain data
• Without explicitly reading each data item
• Migration automatically collects matching data
• Responds to changes in the view

Migration view
migrate
work order
work order
Mmigrate p modifiers(?)
16
Active CoordinationData Duplication

• Data availability is often more important than
  consistency
• Duplicate certain data items
• Duplicates available upon disconnection
• Originals are unaffected

Duplication view
4
3
1
duplicate
2
3
2
5
Dduplicate p modifiers(?)
17
Active CoordinationEvent Capture

• Response to events instead of state
• e.g., agent/host arrival, data access
• Special event generation mechanism
• Multiple registrations for same event
• Transaction specifies events callback

Event view
!!
inspector arrived event
Eevent p followed by Te(?)
18
Consistency Concerns

• Transactional semantics
• Strong application guarantees
• Can be expensive
• Can be guaranteed using special protocols
• Eager scheduling modality
• Best-effort semantics
• Allow more flexible implementation
• Increase performance
• Lazy scheduling modality

19
Simplifying Programming

• Programming abstractions
• Reduce programmer error
• Tailored to common needs of real applications
• Reduced programming effort
• Desirable changes in semantics

• structural data on this floor
• p ?pattern of data to collect?
• seenTuples new Vector()
• while(true)
• sleep(time)
• data ?.rdgp(p)
• if data ? null
• for i1 to data.length
• datai.newID()
• out(datai)
• seenTuples.add(datai)

• structural data on this floor
• p ?pattern of data to collect?
• d new Duplicate(p)
• ?.enable(d, eager)

20
Conclusions

• Coordination middleware must address constraints
  of ad hoc networks
• Allow adaptation to changing context
• Tailor context provision to applications needs
• Programming abstractions must satisfy application
  developers requirements
• Reactive and active constructs provide high-level
  coordination in EgoSpaces
• Required middleware support should be minimized
• EgoSpaces view abstraction defines context
• Behaviors are reducible to reactive construct

21
Thank You

• For more information
• http//www.cse.wustl.edu/mobilab