EE382N Distributed Systems PublishSubscribe Systems Ayis Ziotopoulos Srivatsan Sridharan

1
EE382N Distributed SystemsPublish/Subscribe
SystemsAyis ZiotopoulosSrivatsan Sridharan
2
Contents

• Introduction
• Motivation
• System overview
• Research problems
• Contributions
• Conclusion

3
Introduction

• Internetwide applications
• Modern distributed systems
• Messaging services

4
Messaging Systems an example
5
Motivating Example

• Assume a car dealership for a large company
• Sales representative sells a car
• A request to manufacturing is sent
• Payment done electronically
• An update to sales department is sent
• Various supporting software applications involved

6
Motivation

• Lot's of small applications for different
  organizations combined due to e.g. vertical
  integration, merging, outsourcing etc.
• Internet wide applications e.g. stock trading,
  sports event broadcasting
• Loosely integrated distributed systems
• Need for a 'glue' technology

7
Messaging Oriented Middleware

• A new paradigm 'Message Oriented Middleware'
  (MOM)
• Real-time events 'messages' available through
  Internet e.g. stock quotations, buy/sell offers,
  weather reports, etc.

8
Pub/Sub Architecture

• Publishers
• Subscribers
• Brokers
• Information flow graph

9
Pub/Sub Architecture
10
Alternatives to pub/sub systems

• RMI
• Database Systems
• Group Communications Systems
• Work-flow Systems

11
Issues with broker networks

• Message Matching/Routing
• match at the publisher/broadcast and match
  locally
• main theme of this presentation!!!
• Ordering
• atomic broadcast / multicast
• Optimistic Delivery
• Reliability
• Topology Changes
• Security

12
Pub/Sub Systems

• Subject based
• Content based

13
Subject based pub/sub systems

• A fixed set of subjects is predefined e.g. for a
  stock trading system define 'IBM', 'Microsoft'
  etc.
• Each unit of information is classified based on
  subjects ltsubjectgt ltbodygt e.g. IBM some
  body

14
Subject based pub/sub systems advantages

• Advantages
• matching is trivial i.e. table lookup
• multicast is trivial i.e. one multicast group per
  subject
• simple, adopted by OMG for CORBA event channels
• Key problems
• 'coarse' resolution
• scalability issues for large number of subscribers

15
Content based Pub/Sub systems

• Define a number of
• information spaces
• and an associated event schema per space
• e.g. an information space for stock quotes could
  contain the following event schema
  issuestring, pricedollar, volumeinteger
• A subscription is a predicate against the event
  schema e.g. (issueIBM pricelt120 volumegt1000)

16
Content Based Pub/Sub Systems

• Advantages
• orthogonal criteria for searching.
• No administrative overhead for defining and
  maintaining a large number of groups.
• Key problems
• Efficient matching of event against large number
  of subscribers on a single message broker.
• Efficient multicast of messages within a network
  of message brokers.

17
Gryphon

• A content-based distributed messaging system
• Initiated in the spring of 1997 as a two-person
  exploratory project on publish/subscribe
  technology.
• In fall 1997, it was identified as one of 5
  projects to receive exploratory research funding
  via the prestigious ASSR program at IBM's T. J.
  Watson research center.
• The Gryphon system has been deployed over the
  Internet for real-time sports score distribution
  at the Grand Slam Tennis events, Ryder Cup, and
  for monitoring and statistics reporting at the
  Sydney Olympics.

18
Matching Events

• 'Inverse' database problem
• Matching-first
• generate a matching list, O(N).
• send message to all entries in this list.
• impractical for large networks, with multiple
  copies of the same message on a link.
• Flooding
• message is multicast to all destinations, O(N).
• perform matching at destination
• unsuitable for many subscriptions with a small
  percentage of subscribers in each one

19
Key Idea

• Use a more efficient data structure
• Partial Search Tree
• each subscription is a path from root to leaf of
  the tree.
• follow all paths from root to leaves satisfying
  the event.
• O(N) space, O(N(1-l)), 0 lt l lt 1, running time
• Still non-distributed !!!

20
Matching Tree
21
Link Matching

• Operate at broker level
• Do not use destination lists (!!!)
• On receiving an event, figure out which neighbour
  should receive it as well
• Neighbours can be other brokers or subscribers
• Use the PST to match the event

22
Trit Vector

• How to keep track of which neighbour should
  receive an event?
• Use a trit vector !!!
• For each outgoing link store a value indicating
  whether to send matched events down that link
• Possible values yes, no, maybe

23
Link Algorithm

• Each broker has a copy of all the subscriptions
  organized in a PST
• Each PST node in each broker is annotated with a
  trit vector
• During start-up an initialization trit is
  computed at every broker.
• At match time, the initialization mask for a
  broker is recursively refined until the broker
  has yes or no alone in its trits.

24
Updating the Trit Vector

• Alternative combine
• Apply to non edges
• Take the least specific result of two annotations
• Parallel combine
• Merge the results of the alternative combine
  operations
• Take the most liberal result of two annotations

25
Updating the Trit Vector
26
Optimizations
27
Optimizations

• Collapse a chain of edges to a single one
• For a path of n non- attributes compute and
  store the successor set
• Aggressive static analysis
• Use sub-trees for most common attributes

28
Research Problems

• Inequality tests are not addressed
• All attribute vocabularies are assumed to have
  same size
• Order of attributes is not taken into account
• Uniform distribution assumed for attributes
• Single storage for subscriptions

29
Contributions

• Distributions are rarely uniform
• Locality in references exists
• Serialize search starting from the most probable
  attribute
• Simulated results show gains

30
Atomic Broadcast/Multicast

• Order of messages is important too
• Nodes can fail/come up at any time, still no
  message should be lost
• Use a deterministic protocol to ensure this
• Protocol should be independent of network
  delays/reordering of messages
• Round-robin is inefficient
• An instance of the atomic broadcast algorithm

31
Loggers and Mergers
32
Bias Algorithm

• Integrity, Validity, Agreement, Total Order
• Economical
• Optimal w.r.t to total merge delay for two
  producers, asymptotically optimal for more
• Assumes
• FIFO between logger and merger
• no losses
• approximately synchronized clocks
• expected message rate from every producer can be
  estimated
• memoryless publisher processes

33
Bias Algorithm

• Combine timestamps and knowledge of expected
  message rate for each process
• Main idea use bias for each stream to 'delay'
  reception of messages

34
Bias Algorithm
35
References

• 'A case for Message Oriented Middleware', G.
  Banavar et. al. , Proceedings of DISC 99.
• 'Matching Events in a Content Subscription
  System', M. Aguilera et. al. , Principles of
  Distributed Computing 1999.
• 'An Efficient Multicast Protocol for
  Content-Based Publish-Subscribe Systems', G.
  Banavar et. al. , International Conference on
  Distributed Computing Systems 1999.
• 'Efficient Atomic Broadcast Using Deterministic
  Merge', M. Aguilera and R. Strom, Proceedings of
  PODC 2000
• http//www.research.ibm.com/distributedmessaging/g
  ryphon.html

36
Questions???