A Framework for Distributed Event Composition

1
A Framework for Distributed Event Composition

• Middleware03, Rio de Janeiro, Brazil, June 2003

2
Motivation I

• Event-based systems provide loosely-coupled
  many-to-many communication
• This makes it easier to build large scale systems
• But subscribers may be overwhelmed by the vast
  number of primitive, low-level events
• They need a higher-level view
• Composite events(CEs) represent patterns of
  events
• Analogous to database triggers

3
Motivation II

• To date, event subscribers must do this
  themselves
• This makes them complex and error prone to
  implement
• They need to subscribe to all relevant events
• Common patterns can be reused
• CE Detectors can be close to publishers
• Reduces bandwidth consumption, minimizes latency
• Improves reliability
• Distributed detection increases scalability,
  efficiency
• Independence from publish/subscribe event system
• Supports any pub/sub system

? A generic middleware service for distributed
composite event detection
4
Overview

• Motivation
• Background to publish/subscribe systems
• Active Office Scenario
• Introduction to composite events
• Composite Event Detection Framework
• Architecture
• Composite Event Language
• CE Detectors
• Distributed Detection
• Evaluation
• Implementation over JMS
• Results

5
Publish/Subscribe Systems
?
publish
notify
?
Pub/Sub System
Event Publisher
Event Subscriber
subscribe

• Event publishers publish events
• Event subscribers subscribe to events using
  subscriptions
• Pub/sub system notifies subscribers of matching
  events
• Expressiveness of pub/sub system can be
  topic-based, content-based,

6
Active Office Scenario

• The Active Office is a ubiquitous computerized
  environment.
• Sensors monitor activity in the building.
• Users want context-aware notifications.

Location Sensor (Active Bat)
A
Temperature Sensor
T
Whiteboard Sensor
W
Door Sensor
D
Scenario Jean wants a PDA record of the
whiteboard from meetings she attends, sent to her
wireless PDA (but not if she returns to her
workstation afterwards).
7
Composite Events

• Composite events are patterns of simpler events
• e.g. Jean is seen in the kitchen, followed by Ken
• Constituent events could be primitive or
  composite
• Time model
• In distributed systems, a linear time model is
  insufficient
• Composite events are partially ordered by
  interval timestamps
• Intervals represent uncertainty or event duration

8
Architecture

• Pub/Sub System
• Minimal assumptions
• Composite events disguised as primitive

Distributed Application
sub(CE)
CE Detection
pub
sub
notify
pub
sub
notify

• CE Detection Framework
• Subscribes to primitive events (and simpler
  CEs)
• Automata detect event patterns, publish composite
  events
• Automata created dynamically, and managed by CE
  framework

Publish/Subscribe System
Network Transport

• Distributed Application
• Submits composite event subscriptions to CE
  framework
• Can also use underlying pub/sub system directly

9
Composite Event Language I

• CE detectors need bounded resource usage
• Automata should be statically analyzable
• Requirements for a CE language
• At least as expressive as other languages
• Easily factorizable, canonical representation
• We use automata which extend FSMs / REs
• Regular expressions are a well understood
  construct
• Extensions parallelisation, timing,
  parameterisation
• Cannot consider ALL events
• Higher level, domain-specific languages

10
Composite Event Language II

• Atoms
• Describable events are what can beexpressed by
  the pub/sub system
• E.g. Consider all events in input alphabet A,B,C,
  but accept only As

E
Examples
PersonEvent(roomFE01)
A ? A,B,C
11
Composite Event Detectors

• CE language expressions compile into detection
  automata
• The automaton for C1BP is

C
1
B

P
B
P
S0

B

P


B P
12
Distributed Detection

• Efficiency is improved by distributing detectors
  throughout the network
• An overlay network of event brokers supports CE
  detectors
• Distribution policies dictate detector placement
• Migrating subexpressions near to sources
• Excessive factorisation increases overheads
• Detector replication

13
Implementation over JMS

• Each CE type is associated with a JMS topic
• CE expressions can use JMS content-based
  filtering
• Broker nodes host CE detectors
• An administration topic is used to control
  detector instantiations

Scenario Jean wants a PDA record of the
whiteboard from meetings she attends, sent to her
wireless PDA (but not if she returns to her
workstation afterwards).
14
Results
2500
Wireless
2000
Wired
1500
Traffic (KB)
1000
500
0
1
6
5
4
3
2
Subscribers

• As the number of subscribers increases, so do the
  advantages of the CE framework
• Bandwidth is used efficiently, with CE detectors
  close to event sources especially wireless
  bandwidth
• CE notification delay remained low (lt220ms)

15
Conclusions

• Generic CE detection is a novel, useful
  middleware service
• Can be added to existing pub/sub systems
• Extended finite state automata provide a
  tractable computational model
• Distributed detection increases robustness for
  large scale ubiquitous systems
• Future work
• Higher level languages for expressing CEs
• CE detection for federating pub/sub frameworks

16
Thank You

• Any Questions?