Title: OPTICAL PACKET SWITCHING OVER ARBITRARY PHYSICAL TOPOLOGIES USING THE MANHATTAN STREET NETWORK
1- OPTICAL PACKET SWITCHING OVER ARBITRARY PHYSICAL
TOPOLOGIES USING THE MANHATTAN STREET NETWORK - AN EVOLUTIONARY APPROACH
- Olufemi Komolafe - University of Strathclyde, UK.
- David Harle - University of Strathclyde, UK.
- David Cotter - Corning Research Centre, UK.
2OVERVIEW
- INTRODUCTION
- MSN WITH CLOCKWORK ROUTING
- DEPLOYMENT
- GENETIC ALGORITHMS
- RESULTS
- CONCLUSION
3OVERVIEW
- INTRODUCTION
- MSN WITH CLOCKWORK ROUTING
- DEPLOYMENT
- GENETIC ALGORITHMS
- RESULTS
- CONCLUSION
4OPTICAL PACKET SWITCHING
Optical logic devices immature (w.r.t.
electronics)
Unavailability of static random access optical
memory devices
Avoid contentions in optical domain (or resolve
minimising buffering)
Simple routing processing
Multiprocessor Interconnection Architectures
5OVERVIEW
- INTRODUCTION
- MSN WITH CLOCKWORK ROUTING
- DEPLOYMENT
- GENETIC ALGORITHMS
- RESULTS
- CONCLUSION
6Manhattan Street Network (MSN)
- Directed graph
- N X N nodes
- N is even
- Degree 2
- Toroidal network
- all nodes topologically equivalent
- Traditional routing scheme
- based on computing relative addresses
- contentions in network
7MSN - Clockwork Routing
- Operation
- All nodes synchronised to global clock
- Timeslots arranged in Modulo N sequence of frames
(for N x N MSN) - Automated switch state changes at nodes
- No additional routing processing at intermediate
nodes - only for me or not for me?
- No contention within network
- peripheral electronic buffering used
8OVERVIEW
- INTRODUCTION
- MSN WITH CLOCKWORK ROUTING
- DEPLOYMENT
- GENETIC ALGORITHMS
- RESULTS
- CONCLUSION
9DEPLOYMENT
How can the MSN be used to reduce routing
complexity avoid optical domain buffering over
topologies such as these?
10Deploy as Virtual Topology
- Use MSN as virtual topology in WDM networks
- Nodes of MSN mapped onto physical topology nodes
- Nodes interconnected by lightpaths established
according MSN connectivity - Clockwork Routing used to switch packets
optically at lightpath terminals - Lightpath length will vary according to mapping
of MSN nodes onto physical topology nodes
11Deployment Cost
- Cost Mean lightpath length
- impacts end-to-end number of hops packets
traverse - affects number of ? needed
- indicates number of optical cross-connects
traversed between adjacent MSN nodes - affects deployment of optical amplifiers
consumption of other network resources
128 Costs
- Maximum mean
- lightpath length
- inter-nodal distance
- Shortest Path routing discipline
- inter-nodal distance
- Random Choice routing discipline
- inter-nodal distance
- Minimum number of physical hops
- post-embedding route selection
of ? physical hops (i.e. OXCs) traversed by
lightpaths
optical cross connects traversed by packets
13OVERVIEW
- INTRODUCTION
- MSN WITH CLOCKWORK ROUTING
- DEPLOYMENT
- GENETIC ALGORITHMS
- RESULTS
- CONCLUSION
14GENETIC ALGORITHMS (GA)
- Adaptive computational models inspired by nature
- population of solutions evolve to yield better
solutions - Features of problems to which GA applied
- Readily accurately encodable as a binary string
- No ordering dependencies
- Frequency of value inconsequential
15Problem Encoding
- Each individual in population represents an
embedding of the MSN in physical topology
MSN node 4 mapped onto physical topology node 1
Fitness/Cost Resulting mean lightpath length
16Crossover Techniques
- Crossover
- Used to produce offspring for next generation
from 2 parents - Simplest form
- selecting random point in parents and then
swapping tails
17Simple Crossover
2 Parents..
A
B
18Simple Crossover
A
B
Randomly selected crossover point
19Simple Crossover
PARENTS
OFFSPRING
20BUT.
21Crossover Techniques
- Duplications and omissions will occur which
produce invalid offspring - Similar problems faced when applying GA to
Travelling Salesman Problem - GA encoding
- Individual Touring order of city
- Fitness/Cost Length of tour
- Crossover Techniques developed and adapted for use
22Crossover Techniques
- 4 different crossover techniques used
- Crossover Correct
- correct any omissions and duplications
- Partially Mapped Crossover
- swap elements between 2 randomly chosen points
- Order Crossover
- seeks to preserve parents order in offspring
- Cycle Crossover
- ensures each element in offspring comes from same
position in a parent
23Selection of Optimum GA Parameters
- GA parameters empirically selected
- Population size
- speed memory limitations important
- sizes of 250, 500, 1000 and 2000 considered
- Crossover probability
- parents may survive unaltered in next generation
- probability of 0.5 and 1 considered
- What is the best population size and crossover
probability for each crossover technique?
24Crossover Correct (XC)
Mean lightpath length
Generations
25Partially Mapped Crossover (PMX)
Mean lightpath length
Generations
26Order Crossover (OX)
Mean lightpath length
Generations
27Cycle Crossover (CX)
Mean lightpath length
Generations
28Selection of Optimum GA Parameters
- Crossover Correct, Cycle Crossover
- Population 2000
- P(Crossover) 1
- Partially Mapped Crossover
- Population 1000
- P(Crossover) 1
- Order Crossover
- Population 2000
- P(Crossover) 0.5
29OVERVIEW
- INTRODUCTION
- MSN WITH CLOCKWORK ROUTING
- DEPLOYMENT
- GENETIC ALGORITHMS
- RESULTS
- CONCLUSION
30RESULTS
- Optimum parameters applied for each crossover
technique and used to deploy MSN in arbitrary
physical topologies - Different MSN sizes used
- 4x4, 6x6, 8x8,10x10
3116 Node Physical Topology
Mean lightpath length
Generations
3236 Node Physical Topology
Mean lightpath length
Generations
3364 Node Physical Topology
Mean lightpath length
Generations
34100 Node Physical Topology
Mean lightpath length
Generations
35Explanation Exploration Vs Exploitation
- 2 fundamental operations of search algorithms
36Explanation Exploration Vs Exploitation
- Order Crossover gives worst result
- good at exploration but exploits wrong
information - seeks to preserve order NOT position
37Order Crossover (OX)
P(Crossover) 1 yields no improvement hence
little relevant exploitation done
Mean lightpath length
P(Crossover) 0.5 means that there is some
chance of important characteristics being passed
down generations
Generations
38Explanation Exploration Vs Exploitation
- Order Crossover gives worst result
- good at exploration but exploits wrong
information - seeks to preserve order NOT position
- Crossover Correct obtains good results despite
relative simplicity - Cycle Crossover arguably best at exploitation
- all positions in offspring filled with element
from same position in parent - Partially Mapped Crossover best combines
exploration and exploitation synergistically
39Comparison with Other Approaches
- Other techniques may be used to find good
embeddings of MSN in physical topology - Simulated Annealing (SA)
- Hill Climbing (HC)
- Random Search (RS)
- Results obtained compared to best GA result found
40Results of All Approaches
41Comparison with Other Approaches
- GA gives best result for most network sizes
- Evaluating cost function time-consuming
- indicative of time taken for solution to be found
- GA and HC consider similar number of solutions,
but SA and RS consider 10 and 20 times more
solutions - ? GA typically gives the best result and does so
significantly quicker than the other techniques
42OVERVIEW
- INTRODUCTION
- MSN WITH CLOCKWORK ROUTING
- DEPLOYMENT
- GENETIC ALGORITHMS
- RESULTS
- CONCLUSION
43CONCLUSION
- Manhattan Street Network with Clockwork Routing
- no optical domain contention
- simple routing scheme
- Genetic algorithms proposed and implemented as
method to deploy Manhattan Street Network in
arbitrary physical topologies - Genetic algorithms seen to out-perform other
heuristics and yet still be significantly quicker