An Evolutionary based Dynamic A E l ti b d D i Energy - - PowerPoint PPT Presentation

A E l ti b d D i An Evolutionary based Dynamic Energy Management Framework for Energy Management Framework for IP-over-DWDM Core Networks

Xin Chen, Chris Phillips School of Electronic Engineering & C t S i Computer Science

O tli Outline

Introduction

k d

Background New Energy Management Design New Energy Management Design Simulation Further work

2

Introduction Introduction

Benefits of Energy Saving Economical

Lower OPEX for ISPs Lower OPEX for ISPs

Environmental

Lower CO2 emissions

3

d i Introduction

Our energy management scheme combines

gy g infrastructure sleeping and virtual router migration together with automatic optical migration together with automatic optical layer connection forwarding to enable resources to be used in an energy efficient resources to be used in an energy-efficient manner.

4

k d Background

Network Architecture Network Architecture Energy Saving Approaches Infrastructure sleeping and Virtual

Router Migration Techniques Router Migration Techniques

5

Network Architecture

IP over DWDM

Network Architecture

IP over DWDM

We apply the wavelength continuity constraint. There is no wavelength conversion for through-traffic in the network.

6

S i A h Energy Saving Approaches

Static Mechanisms Static Mechanisms

Network planning, i.e. ILP

Dynamic Mechanisms

f l i d i Infrastructure sleeping, rate adaptation, network virtualization…..

7

Infrastructure Sleeping p g

Switch off unneeded equipment during off-peak periods periods Previous work [over 20% saving]

• L. Chiaraviglio, M. Mellia, and F. Neri, "Energy-Aware Backbone Networks: A Case Study" in

IEEE International Conference on Communications Workshops, 2009, pp. 1–5, June 2009. 8

f Sl i Infrastructure Sleeping

• Some issues and limitations
• Some issues and limitations

The problem of loss of connectivity due

p y to reconvergence

When to sleep / wake ? When to sleep / wake ?

• How to sleep / wake ?

9

Virtual Router Migration g

Move virtual routers among difference physical platform i h d di h i without degrading the service

Wang, Yi,; Keller, E.; Biskeborn, B.; Jacobus van der Merwe, Rexford, J.; ,"Virtual routers on the move: live router migration as a network-management primitive," SIGCOMM Comput. Commun. Rev. 38, 4 (August 2008), 231-242. 10

i l i i Virtual Router Migration

• Some issues and limitations
• When to trigger virtual router migration?
• When to trigger virtual router migration?
• Where to move virtual routers to?

11

Dynamic Energy Management Dynamic Energy Management Framework

Overall Energy Management Procedure Optical Connection Management VRM_MOEA

Virtual Router Migration – Multi-Objective Evolutionary Algorithm

12

Overall Energy Management Procedure Overall Energy Management Procedure

1 C ll t d l th t k t t

• 1. Collect and analyze the network status
• 2. Trigger VRM MOEA. (Quiet and Busy Thresholds)

gg _ (Q y )

• 3. Establish the new optical connections
• 4. Virtual router migration

5 Then Switch off (on) the corresponding physical

• 5. Then Switch off (on) the corresponding physical

platforms and removed the unneeded optical connections connections

• 6. Go back to step 1 to recheck the network status

13

Dynamic Optical Connection Management

Additi l ti l ti d d f Additional optical connections are needed for forwarding the traffic to the remote virtual t ( ) i th k t router(s) processing the packets Changes in the underlying physical network are hidden from the topology as seen by Layer-3 and p gy y y so reconvergence events are avoided

14

Example: Optical Connection Management

15

Destination Physical Platform Selection y Algorithm- VRM_MOEA

Individual chromosome representation:

Gene VR index Chromosome length set to sum of VRs

2

Gene, VR index Allele gives PP location

2 1 1

16

Destination Physical Platform Selection y Algorithm - VRM_MOEA

Initial Population: Pre screening procedure for selecting the variable solutions Evolutionary algorithm well- suited to real-time operation as search can be halted at any point and we only need a “good” and we only need a good solution

17

Destination Physical Platform Selection y Algorithm- VRM_MOEA

T bj ti f ti Two objective functions: 1.Power Consumption:

α 1

• total

base i lc base roadm i

P P N P P P

α

α θ β α β

=

= ⋅ + ⋅ + ⋅ ⋅ + ⋅

∑

（ ）

On PPs On PP Linecards Off PPs On ROADMs

to ta l

P P

α β

• ---- The power consumption of

the network

• ---- The number of active PPs
• ---- The number of ROADM in the

On PPs On PP Linecards Off PPs On ROADMs

ba se

P

lc

P

θ

• ---- The power consumption of

base system

• ---- The power consumption of

network

• ---- A percentage of the base system

power consumption a PP consumes h i i l i

ro ad m

P

i

N

a line card

• ---- The power consumption of

a ROADM when it is sleeping.

• ---- The number of active line cards in

the i-th PP

18

Destination Physical Platform Selection Al ith VRM MOEA Algorithm- VRM_MOEA

• 2. Virtual Router Migration Cost (Second objective function)

The first VRM cost component comes from hop count f i i l “h ” VR l ti t it d ti ti PP from an original “home” VR location to its destination PP. For i-th candidate solution:

β 1

_ ( ) ( , )

j j i j

Cost a i d g g

β =

= ∑

i

g

• ---- A function for obtaining the distance between two PPs: x1 and x2.
• ---- j-th gene in the default network configuration

( 1, 2 ) d x x

j i

g β

• ---- j-th gene in a candidate solution.
• ---- The number of VR

19

Destination Physical Platform Selection Al ith VRM MOEA Algorithm- VRM_MOEA

The second component comes from the virtual router The second component comes from the virtual router migration process

β 1

_ ( ) ( , )

j j current i j

Cost b i d g g

β =

=∑

i curren t

g

• ---- j -th gene in the current network

Therefore The overall cost of one possible solution

curren t

g

j i

g

j g configuration

• ---- j-th gene in a candidate solution.

Therefore, The overall cost of one possible solution is :

( ) ( ) (1- ) ( ) Cost i Cost a i Cost b i ϕ ϕ = ⋅ + ⋅ ( ) _ ( ) ( ) _ ( ) ϕ ϕ ϕ

• ---- Weight of two cost terms

20

Destination Physical Platform Selection y Algorithm- VRM_MOEA

Fi f i Fitness function: Strength Pareto Evolutionary Algorithm II (SPEA2) Selection Mechanism: To rn ment sele tion Tournament selection Crossover Operation: BLX-α crossover Mutation Operation: Mutation Operation: Mutation rate = 0.1

21

Strength Pareto Evolutionary Algorithm II (SP A2) II (SPEA2)

The relationship between two decision vectors : Dominance , indifference The relationship between two decision vectors : Dominance , indifference

22

Strength Pareto Evolutionary Algorithm II (SP A2) II (SPEA2)

SPEA2 P d SPEA2 Procedure: 1.Assign a strength score to each solution. The score is equal to the number of solutions it dominates. 2.Get the raw fitness value of a solution by summing up the strength score of solutions which dominate it. 3.Get the density value by K-th nearest neighbor method (K=1). 4.Add the raw fitness value and density value to obtain the fitness value. A non- dominate solution has fitness value 0.

23

Strength Pareto Evolutionary Algorithm II (SP A2) II (SPEA2)

24

BLX-α Crossover BLX α Crossover

It offers an opportunity that after crossover, the offspring’s genes come pp y p g g from a slightly larger range randomly selected between the two parents’ genes. Procedure: Procedure: For two Parets: G1, G2, the i-th gene of offspring is define:

[ ] h Uniform g I g I α α = − ⋅ + ⋅

i

h

min max

[ , ]

i

h Uniform g I g I α α = − ⋅ + ⋅

1 2 min

( , )

i i

g Min g g =

1 2

1

1 2 max

( , )

i i

g Max g g =

max min

I g g = −

1 i

g

2 i

g

• ---- i-th gene of G1
• ---- i-th gene of G2
• ---- An user define parameter

α

25

Simulation

Simulator Introduction Simulation Results

26

Simulator Introduction

• 1. Main simulation framework: Hybrid simulator
• 2. Network topology: a simple network topology generator

p gy p p gy g

• 3. Traffic model: fluid flow model and daily traffic model

4 D i i h i l l f l i bl

• 4. Destination physical platform selection problem:

VRM_MOEA

• 5. When to trigger VRM : Reactive mechanism

27

Simulation Results

Network 6N8L 11N14L Scheme Name Energy Energy Energy Energy Consumption / day Saving Consumption / day Saving No VRM 4057200.00 0.00% 7698240.00 0.00% Quick VRM 3223327.20 20.55% 6008048.00 21.96% VRM_MOEA(0 ,1) 3134507.20 22.74% 5457832.20 29.10% VRM_MOEA(0.2,0.8) 3125365.00 22.97% 5451220.20 29.97% VRM_MOEA(0.5,0.5) 3140139.80 22.60% 5471056.80 28.93% VRM_MOEA(0.8,0.2) 3160408.40 22.10% 5490892.40 28.67% VRM_MOEA(1 ,0) 3144768.20 22.49% 5517340.00 28.33%

A l 5 i l ti ith d d Average values over 5 simulations with random seeds Quick VRM chooses the best solution in a randomly generated population of candidate solutions without any evolutionary process

28

y y

Simulation Results Simulation Results

The optical infrastructure The energy saving is similar among the different VRM_MOEA schemes is “always on” In the off-peak hours, e.g. 12 to 24, the energy saving performance of VRM_MOEA is better than Quick VRM

29

Simulation Results

Per fibre, there are 40 channels each

• perating at
• pe a

g a 40Gb/s The PP switch fabric can accommodate 1Tb/s traffic load The occupied number of lightpaths fluctuate with the traffic load in the baseline scheme scheme. When VRs are moved to remote PPs , more optical channels are used for transmitting the packets to be processed by VRs than that of the baseline case. The discontinuities on the lines correspond to virtual router migrations.

30

The discontinuities on the lines correspond to virtual router migrations.

Simulation Results Simulation Results

Higher quiet and busy thresholds cause a higher energy saving. g q y g gy g

31

Simulation Results Simulation Results

60

1 γ ⎡ ⎤

50 60

)

It is more difficult to gain 1- ( ) (1 sin( )) 2

ij ij

T t f t γ δ γ ⎡ ⎤ = Α ⋅ + ⋅ + ⎢ ⎥ ⎣ ⎦

40 30

y Saving (%)

g the energy saving in a busier network.

20 10

Energy

10

32

Future Work Future Work

• 1. Reactive mechanism

Short term proactive mechanism 2 Add the VRM migration time into the simulation

• 2. Add the VRM migration time into the simulation

33

Thank you! Thank you!

34