Download Slides - Wireless System Research Group

Application of Evolutionary Algorithms
for Energy Efficient Grooming of
Scheduled Sub-Wavelength Traffic
Demands in Optical Networks
Ala Shaabana
University of Windsor
School of Computer Science
13-05-2013
2
3
Outline
•
•
•
•
•
•
•
•
Optical Communication Overview
Motivation
Solution Outline
Evolutionary Algorithms
Related Research
Energy Minimization for Scheduled Traffic
Results
Future work
4
Optical Communication
• Communication at a distance to carry information
using light.
• In 1880, earliest electrical device created to perform
optical communication is created.
– Photophone
• Optical communication today relies on optical fibers
to carry the information from point A to point B.
5
Optical Communication
• Communication at a distance to carry information using
light using optical fibers.
• Optical fibers: thin glass cylinders or filaments which
carry signals in the form of light.
• A Lightpath (LP): end-to-end optical communication
channel – may traverse multiple fibers.
May 13, 2013
Optical Communication
• An optical communication system uses:
– Transmitter: encodes a message into an optical signal.
– Channel: carries the signal to its destination
– Receiver: reproduces the message from the received
optical signal.
A. Shaabana — M.Sc.
Thesis Defense
6
7
Motivation
• Immense growth in high-bandwidth applications 
increase in energy consumption.
– 1% improvement in total energy consumption  $5M USD
savings per year in electricity cost [1].
• Existing approaches: put network interfaces and
components to sleep [2], switch off line cards [3][4], or
even entire links or nodes[5][6].
• Our approach: consider the applications that require
periodic use of bandwidth at predetermined times.
– Unlike static or dynamic traffic demands, this type of demands
(called scheduled traffic demands) is periodic and predictable
– Resource allocation can be optimized in both space and time.
8
Example
9
Motivation
• Integer Linear Programs (ILPs) are typically the go-to
solution for this kind of optimization problem.
• However, although ILPs achieve an optimal solution,
however they become computationally intractable
once the problem set becomes larger (larger network
sizes).
• It ends up taking too long and sometimes consuming
too many computational resources to find its
solution.
10
Solution Outline
• We present a Genetic Algorithm (GA) and a Memetic
Algorithm (MA) to route a set of periodic, subwavelength traffic demands over the network.
• The primary goal for these two approaches is to:
1) Route the traffic demands in such a way that the
maximum number of LPs can be switched off, hence
reducing the overall power consumption.
2) Reduce the total number of LPs needed to realize the
logical topology, such that the capacity constraints of the
LPs are not exceeded.
11
Solution Outline
• The GA has reduced the energy consumption more
so than the simple shortest path holding-time-aware
heuristic, which was presented in our previous work
in [7].
• In contrast, the MA was designed to build upon the
GA and improve on it further by adding local search
capabilities,.
– Solutions similar to the GA with considerably less time.
12
Solution Outline
• Note however, that evolutionary algorithms
themselves have been applied to many
computational problems.
• Computer simulations of evolution started as early as
1954 with the work of Nils Aal Barricelli.
• In 1989, Moscato et al. proposed Memetic
Algorithms based on Richard Dawkin’s notion of a
Meme [8].
13
Solution Outline
• At the time being, there are no applications of
evolutionary algorithms to optimize scheduled
demands in optical networks, and the current
approach seems to be to use a holding-time unaware
(HTU) approach.
14
A little Biology…
• “Evolution is the survival of the fittest” is a great
description of many evolutionary computation
systems.
• When one uses evolutionary computation to solve a
problem, it operates on a population (or a collection)
of data structures (or creatures/genes).
• This fundamental difference in the notion of fitness
is a key difference between biological evolution and
most evolutionary computation.
15
A little Biology…
• 2 opposing forces that drive evolution:
– Variation: process that produces new alleles/genes
– Selection: process whereby some alleles survive and
others do not
16
A little Biology…
• Evolutionary computing can accomplish variation by
making random changes in these data structures and
by blending parts of different data structures via
mutation and crossover (referred to as variation
operators).
• There are good and bad mutations operating on a
population of data structures.
17
Evolutionary Algorithms
• Nearly three decades of R&D have demonstrated
that the mimicked search process of natural
evolution can yield very robust and direct computer
algorithms, even though these imitations are crude
simplifications of biological reality.
• The result of these efforts is Evolutionary Algorithms
(EAs).
18
Evolutionary Algorithms
• The population evolves towards improving regions of the
search space by means of randomized processes of mutation,
selection and recombination.
• The population is arbitrarily initialized.
• The environment delivers fitness information of the
individuals, and the selection process favors those individuals
of higher fitness to reproduce more often than worse
individuals.
• The recombination mechanism allows for the mixing of
parental information while passing it to their descendants,
while mutation introduces innovation into the population.
19
Related Research
• A number of recent works focus on reducing power
consumption in today’s core/transport networks.
• For example, Orgerie et al. (2012) developed an energyefficient framework for Bulk Data transfers in dedicated
networks with advance reservation.
• Meanwhile, Musumeci et al. (2012) evaluated the power
consumption of the various devices used at different network
architectures.
• Coiro et al. (2011) consider the dynamic traffic scenario, and
propose an energy aware routing scheme to improve the
energy efficiency by minimizing the number of active optical
amplifiers in the network [5].
20
Genetic Algorithms (GA)
• Most common of Evolutionary Algorithms.
• Generally, a population of candidate solutions to an
optimization problem is evolved towards better solutions.
• Iteratively, the evolution typically starts from a population of
randomly generated individuals, with the population in each
generation referred to as a generation.
• The more fit individuals are stochastically selected from the
current population
• The new generation of candidate solutions is then used in the
next iteration of the algorithm.
21
Genetic Algorithms (GA)
22
Memetic Algorithms (MA)
• While GAs have been inspired in trying to emulate
biological evolution, Memetic Algorithms (MAs) try
to mimic cultural evolution.
• Essentially, MAs are a marriage between populationbased global search and the local search heuristic
made by each of the individuals.
23
Memetic Algorithms (MA)
Energy Minimization in Scheduled
Traffic
• Using the GA and the MA, we aim to minimize the
energy consumption for traffic routing in a
Wavelength Division Multiplexed (WDM) network.
• Given a logical topology 𝐺 𝑁, 𝐿 .
– Each logical edge 𝑙 𝜖 𝐿 is comprised of one or more
lightpaths.
– The number of lightpaths needed to implement a given
logical edge is determined by:
• 1. The number of lightpaths needed to implement on that edge.
• 2. The capacity of a lightpath.
24
Energy Minimization in Scheduled
Traffic
• We are also given a set 𝑄 of the scheduled subwavelength demands, with known start and end
times.
• Each demand 𝑞 𝜖 𝑄 is represented by the tuple
(𝑠𝑞 , 𝑑𝑞 , 𝑛𝑞 , 𝛼𝑞 , 𝜔𝑞 , 𝜏𝑞 )
25
Energy Minimization in Scheduled
Traffic
• The goal is to route the traffic demands in such a way
that the maximum number of lightpaths can be
switched off at any given time, reducing the overall
power consumption.
• We also try to implement each logical edge using as
few lightpaths as possible.
26
27
Chromosome Representation
28
Fitness Function
• It is necessary to calculate the fitness value of each
new individual after we generate it.
• We use the following fitness function:
29
Fitness Function
• The number of active lightpaths 𝑛𝑙,𝑖 can vary,
depending on the traffic in interval 𝑖.
• Hence, the number of transceivers used for
implementing 𝑙 is determined by the maximum
number of active lightpaths needed for 𝑙 in any given
interval.
– That is, 𝑛𝑙 = max 𝑛𝑙,𝑖 𝑖 = 𝑖1 , 𝑖2 , … , 𝑖𝑚𝑎𝑥
30
GA/MA Selection and Crossover
• The selection of individuals from the initial
population as parents is carried out using the
Roulette-Wheel selection method.
• We have used k-point crossover for each crossover
operation in order to produce new offspring from the
selected parents.
31
GA/MA Selection and Crossover
May 13, 2013
Mutation
– Single candidate chromosome is first selected randomly for mutation
– Change a single allele of an individual, with a 𝑃𝑙𝑠 probability of this
event happening.
• For our simulation probability of mutation =0. 08
• Obtained by trial and error; does not converge the
population too quickly, losing potentially better
chromosomes, nor does it ruin the population’s diversity.
33
Local Search
• In order to simplify the way local search works, let us
first take an example.
• Population 𝑃 of chromosomes (candidate solutions)
• Divided into 𝑁 neighborhoods of chromosomes.
34
Local Search
• When local search is implemented, it searches
through the neighborhoods within population and
chooses the most locally optimal chromosome from
each neighborhood.
• That chromosome is then guaranteed to make it
through to the next generation.
• The local search can be integrated within the
evolutionary cycle mainly in two ways.
35
Local Search

The first way is the application of the local search to a
candidate solution, called lifetime learning.

We have implemented our method using the second way.
•
The application of the local search during the solution generation
phase, that is, the generation of a perfect child.
36
Local Search Procedure
37
Local Search Procedure
38
Results
• Simulations were run with different demand sets on a
number of well known networks such as the 14-node
NSFNET and 20-node ARPANET.
• For each network topology and size of demand set, the
results reported in this section represent the average
values of at least five runs.
• We also experimented with different values of the
constant 𝑎 for 𝑎 = {0, 3, 10 , 20}.
– Changing the value of 𝑎 did not produce significant change in
the results.
39
Results
• Although the GA is less computationally demanding
than an ILP, an Amazon “Elastic Cloud” server with
8GB of RAM memory and 4 Amazon EC2 Compute
Units (ECU) was required.
– Where each ECU is equivalent to a 1.0 – 2.0GHz 2007
Opteron or Xeon processor).
• In contrast, the MA experiments, although operating
on the same data sets, required only a 2GB RAM
memory server and utilizing 1 Amazon ECU.
May 13, 2013
Results
• knowledge of demand
holding times result in
significant energy
improvements over HTU
approaches (26%-40%).
• Additional improvements
(8% - 13%) compared to the
HTA shortest path heuristic
in [8].
A. Shaabana — M.Sc.
Thesis Defense
• MA we can reduce the
amount of computational
resources and time used
while achieving similar
results.
40
41
Results
• The simulation results show that knowledge of demand
holding times result in significant energy improvements
over HTU approaches.
• The proposed GA leads to further improvements
compared to the HTA shortest path heuristic we have
proposed in [8].
• Interestingly, they also demonstrate that using the MA
we can reduce the amount of computational resources
and time used while similar outputs to the solutions
presented by the GA.
42
Results
• Specifically, knowledge of demand holding times
significantly reduced energy consumption, with
improvements between 26% and 40%, even using a
simple shortest path routing approach.
• This reduction is achieved by simply switching off
lightpaths when they are not carrying any traffic.
43
GA vs. MA Results
• Tested for 𝑎 = {0, 3, 10, 20} and an optical fiber
bandwidth capacity of 𝐺 = 160 Gbit/s
• For each value of A in the figures, we took the
average of five test cases, and plotted the
improvement percentage over the value of A.
44
GA vs. MA Results
45
GA vs. MA Results
46
GA vs. MA Results
47
GA vs. MA Results
48
GA vs. MA Results
49
GA vs. MA Results
50
Chronological Analysis
51
Chronological Analysis
• We speculate that this is not due to the MA doing
less computations than the GA.
• Likely due to the MA’s capability of detecting when
there are no more better solutions to be obtained.
– This causes it to stop, saving multiple useless computations
that would lead to the same solution.
52
Future Work
• While the proposed GA and MA perform better than
ILPs in terms of computational time and resources,
there is still potential for future improvement.
• One of the fundamental strengths of GAs and MAs is
the diversity of their parameters. It is very possible to
achieve better results after optimizing the
parameters for specific topologies.
53
Future Work
• MAs in particular have great potential in energy
optimization problems in optical networks.
– In terms of parameters, the way the local search
mechanism defines neighborhoods can be changed to
something more complex than “adjacent chromosomes”.
– While this is a common and viable implementation, it may
not be the most optimized option for our purposes.
54
Future Work
• We have also only used one of many “move”
mechanisms in Local Search, there are countless
other algorithms that can be explored and exploited
in order to achieve more optimized results.
55
Future Work
• With respect to our current implementation, more
experimentation and data analysis should be applied,
not only with different size and complexity of
instances of networks, but also with other network
optimization problems, as this has proven to be a
promising direction in optical network optimization
problems.
56
Thank you for listening.
57
Bibliography
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
G. Shen and R. S. Tucker, “Energy-minimized design for ip over wdm networks," Optical Communications and Networking, IEEE/OSA Journal
of, vol. 1, no. 1, pp. 176-186, 2009.
M. Gupta and S. Singh, “Greening of the internet," in Proceedings of the 2003 conference on Applications, technologies, architectures, and
protocols for computer communications, pp. 19-26, ACM, 2003.
J. Chabarek, J. Sommers, P. Barford, C. Estan, D. Tsiang, and S. Wright, “Power awareness in network design and routing," in INFOCOM 2008.
The 27th Conference on Computer Communications. IEEE, pp. 457-465, IEEE, 2008.
F. Idzikowski, S. Orlowski, C. Raack, H. Woesner, and A. Wolisz, “Saving energy in ip-over-wdm networks by switching off line cards in low
demand scenarios," in Optical Network Design and Modeling (ONDM), 2010 14th Conference on, pp. 1-6, IEEE, 2010.
A. Coiro, M. Listanti, A. Valenti, and F. Matera, “Reducing power consumption in wavelength routed networks by selective switch off
of optical links," Selected Topics in Quantum Electronics, IEEE Journal of, vol. 17, no. 2, pp. 428-436, 2011.
B. G. Bathula and J. M. Elmirghani, “Green networks: Energy effcient design for optical networks," in Wireless and Optical Communications
Networks, 2009. WOCN'09. IFIP International Conference on, pp. 1-5, IEEE, 2009.
A. Shaabana, F. Luo, Y. Chen, and A. Jaekel, “A genetic algorithmb-ased approach for energy effcient grooming of scheduled subwavelength
traffc demands in optical networks," Submitted to IEEE Globecom 2013.
P. Moscato, “On evolution, search, optimization, genetic algorithms and martial arts: Towards memetic algorithms," Caltech concurrent
computation program, C3P Report, vol. 826, p. 1989, 1989.
D. Ashlock, Evolutionary computation for modeling and optimization. Springer Science+ Business Media, 2006.
P. Datta, M. Sridharan, and A. K. Somani, “A simulated annealing approach for topology planning and evolution of mesh-restorable optical
networks," in 8th IFIP Working conference on optical networks design and modeling (ONDM), vol. 16, Citeseer, 2003.
N. Krasnogor, “Memetic algorithms," in Handbook of Natural Computing (G. Rozenberg, T. Bck, and J. Kok, eds.), pp. 905{935, Springer Berlin
Heidelberg, 2012.
M. Oca, C. Cotta, and F. Neri, “Local search," in Handbook of Memetic Algorithms (F. Neri, C. Cotta, and P. Moscato, eds.), vol. 379 of Studies
in Computational Intelligence, pp. 29{41, Springer Berlin Heidelberg, 2012.