Modelling and Dimensioning of Optical Burst Switching Networks

1
Modelling and Dimensioning of Optical Burst
Switching Networks

• Jolyon White
• Supervisor Rod Tucker

2
Overview

• Brief intro to Optical Burst Switching (OBS)
• OBS network modelling
• Analysis, simulation ? blocking probability
• OBS network design
• Allocating capacity to minimise costs
• Optimal routing to maximise revenue

3
Optical Burst Switching

• A brief introduction

4
OBS Network Overview
OBS Cross-connect
Burst assembling edge routers
WDM Link
Client hosts (Access network)
5
What is OBS?

• Compromise between
• Optical packet switching
• Statistical multiplexing
• No optical buffering
• Wavelength routing (optical circuit switching)
• Decoupled control plane (control packets)
• Reservation messages sent ahead of burst data

6
OBS Reservation protocols
Source
Destination
OXC1
OXC2
OXC3
Control packet processing delay
Initial offset time,
Channel free for other reservations
Final offset time,
Burst transmission
Channel released immediately after burst clears
OXC
Channel holding time,
OXC Optical cross-connect
Channel reservation period
Time
7
Modelling
Topology
Link, Route, Network Blocking probability
Link capacities
Performance model
Traffic matrix
Routing
8
Dimensioning
Performance model
Topology
Routing
Routing and dimensioning model
Service level
Link capacities
Traffic matrix
9
Modelling OBS Networks

• Simulation and Analysis

10
OBS Link Models
11
OBS Link Model

• Most common link model
• Blocking probability given by Erlangs formula
• Good model if processing delay is NEGLIGIBLE
• Denoted by in this talk

12
Modelling OBS Networks

• Network modelling reduced load approx.
• Given
• Topology
• Routes
• External traffic
• Link capacities
• Find
• Link, route, and network blocking probabilities

Rosberg et al., Performance analyses of optical
burst switched networks, IEEE JSAC, vol. 21 no.
7, 2003
13
Network Description

• A set of links
• A set of routes
• External route loads ,
  
• Link capacities ,
  
• (Number of wavelengths supported on each link)

14
OBS Reduced Load Approximation
External traffic load (Erlangs)
15
OBS Reduced Load Approximation

• Reduced load for link 1

16
OBS Reduced Load Approximation

• In general, the link load is
• Circuit switching equivalent

17
OBS Reduced Load Approximation

• Given the link load, link blocking probability
  is
• However, depends on .
• ? fixed point equation
• Solution by successive substitution
• Note We could use any link model here

18
End-to-end blocking probability

• We now have , link blocking
• Assume link independence
• Route blocking probability is

19
Model Assumptions

• Arrivals to each link are Poisson
• Links block bursts independently
• The burst length is an exponential random
  variable
• Renewed at each link in the route
• In a real network, burst length does not change
  at each link

20
Simulation Results Ring network
21
Burst length propagation
Real networks (and our simulation)
This is also what a Markov chain simulator does!!
22
Simulation Results ring network
23
Correlation in OBS

• With no external traffic, bursts are accepted on
  each hop in sequence
• ? no blocking
• If there is some external or cross traffic
• Correlation still exists, but less

24
NSFNET North America

• 12 routes. Example route

25
NSFNET North America

• 12 routes. Example route

26
NSFNET North America
4
8
16
32 wavelengths
27
NSFNET Effect of Processing Delay
Mean burst holding time
Channels
Load per route
Shortest path routing (156 routes)
28
Designing OBS Networks

• Minimising cost and maximising revenue

29
Network optimisation

• Dimensioning
• How many wavelengths required on each link?
• How much does the network cost?
• Optimal routing
• How to balance traffic flows?
• How to maximise revenue earning rate?
• Minimise cost maximise revenue
• Satisfy grade-of-service (GoS) constraints

30
Cost and revenue

• We assume cost is a function of link capacity
• Network Cost
• Revenue is a function of the carried load on each
  route
• Revenue

revenue earning rate on route r
31
How much does a network cost?

• Advantages of this approach
• Simplicity
• Leads to link dimensioning decoupling (more
  later)
• Drawbacks
• Does not account for more complex economics,
  e.g.
• Chassis cost
• Switch fabric cost
• Flat rate service fees

32
Overview of the problems
Routing
Dimensioning
Non-load sharing
Load sharing
33
Dimensioning with GoS Constraints

• Problem definition

Subject to
A. Girard, Revenue Optimization of
Telecommunications Networks, IEEE Trans. Commun.
Vol. 4 no. 1, 1993
34
Optimality Link sub-problem

• Link problem decouples
• Independent sub-problems for each link
• Equivalent to uni-variate minimization problem

A. Girard B. Liau, Dimensioning of adaptively
routed networks, IEEE Trans. Networking, vol. 1
no. 4, 1993
35
Optimality Operating point

• Operating point equation

Definitions
Load carried on link j that comes from route r
36
Optimality Operating point
OBS
Circuit switched
Definitions
Load carried on route r (circuit switched)
Load carried on link j that comes from route r
37
Revenue maximisation

• Problem definition

Subject to
38
Optimality Routing

• If there is flow on route r, then at optimality
• If there is no flow, then
• ? Marginal net revenue earning rate is negative

39
Optimality Routing

• If there is flow on route r, then at optimality
• Equivalent for circuit switching

40
Comparisons Circuit Switching

• Regime of interest low blocking probability
• Dimensioning/routing similar to circuit switching
• Differences in optimal capacity lt 1
• No real difference because capacities are integer
• Good for OBS!
• Because we can leverage existing design tools
• But is this the end of the story?

41
Comparisons Circuit Switching

• How do we comparing blocking events?
• Circuit blocking user waits, reattempts
• No data lost
• OBS blocking user retransmits
• Data is lost
• Fair comparisons must include retrans. delay
• But not always obvious how to include
• Interactions with TCP

42
Summary

• Presented a model for OBS network performance
• Reduced load approximation
• Identified sources of discrepancies between
  predictions of the RLA and simulation
• Many discrepancies are overlooked by Markov chain
  simulation of OBS
• Generalised circuit switching dimensioning
  techniques to OBS
• Showed that resulting designs are similar