2017/2/14

1

Impact of Fluctuating Goals on Adaptability of Evolvable VNF Placement Method

Mari Otokura1, Kenji Leibnitz2, Yuki Koizumi1, Daichi Kominami1, Tetsuya Shimokawa2, Masayuki Murata1

1 Osaka University, Japan

2 NICT, Japan

1

Research Background

• Communication services are becoming more diverse and dynamic

• Internet of Things (IoT) is currently being realized

• Communication service providers will offer flexible and dynamic network functions

• Network functions process packets in the “middle” of networks

• Examples: firewalls, intrusion detection systems (IDS), caches

• Network functions have been implemented in hardware

• Not flexible and not dynamic

2

Edge Network

Core Network /

Internet

Cloud

IoT devices / users

Network Functions

• NFV implements network functions in software

• Virtual Network Functions (VNFs):

• Network functions virtualized by NFV

• Run on virtual machines (VMs)

• VNF placement problem

• Decision of VNF placement on physical machines (PMs) in physical networks

Network Function Virtualization (NFV)

VM

Large delay Small delay

3

VNF1 VNF2→

Request of a user

VNF1

VM

VNF2

VM

VNF1

VM

VNF2

VNF1 VNF2Request

of a user

Controller

(Step 3) Controllerplaces the components

on physical network

(Step 1)Request arrives

PM7

VM3

VNF2

Comp1

VNF2

Comp2

: Core

PM1 PM2 PM3 PM4

PM6PM5

PM10 PM9 PM8 PM7VM 3

VNF2

Comp 1

VNF2

Comp 2

VM 1 VM 2

VNF1

Comp 1

VNF1

Comp 2

VNF1 VNF2

VNF1

Comp 1

VNF1

Comp 2

(Step 2) Controller breaks down VNFsinto components

VNF2

Comp 1

VNF2

Comp 2

System Model

(Step 1) Request from user for a VNF chain arrives

(Step 2) Controller splits VNFs of chain into components

• Components: small VNF software modules [4]

(Step 3) Controller places components on PMs and assigns cores to them

(Step 4) Controller decides routes of traffic for chains through required components

4

(Step 4) Controller decides routes

Dynamic VNF Placement Problem

• Reconfiguration of VM/VNF placements on physical network when requests for VNF chains arrive/depart

• Main requirement for placement computation: short calculation time

• Solving optimization problem at every request change:

• Difficult to realize since even the static VNF placement problem is NP-hard

5

VM

VNF1 VNF2

VM

VNF3

Request arrive andreconfigure

placement

time

VNF2 VNF3

VM

VNF1 VNF2VNF1 VNF2

Request of a user

Objective

• We previously proposed an evolutionary method for dynamic VNF placement problems named Evolvable VNF Placement (EvoVNFP) [12]

• Utilizing knowledge from biological evolution under varying environments

• When organisms evolve in varying environments:

• Organisms become robust to environmental changes [13]

• Evolution of organisms speeds up [14]

• Objective

• Evaluating EvoVNFP in greater detail to clarify the influence of the parameter settings on the performance

6

[13] N. Kashtan and U. Alon, “Spontaneous Evolution of Modularity and Network Motifs,” PNAS, vol. 102, no. 39, pp. 13773–13778, Sep. 2005.

[14] N. Kashtan, E. Noor, and U. Alon, “Varying Environments Can Speed Up Evolution,” PNAS, vol. 104, no. 34, pp. 13711–13716, Aug. 2007.

[12] M. Otokura, K. Leibnitz, Y. Koizumi, D. Kominami, T. Shimokawa, and M. Murata, “Application of Evolutionary Mechanism to Dynamic Virtual

Network Function Placement,” in Proceedings of ICNP Workshop on Control Operation and Application in SDN protocols (CoolSDN), Nov. 2016.

2017/2/14

2

Evolvable VNF Placement (EvoVNFP)

• Dynamic VNF placement method addressing dynamic arrivals/departures of VNF chain requests

• Calculation of placements by a special type of Evolutionary Algorithm (described in detail later)

• When simulations generate placements which meet predetermined objectives, the controller implements these generated placements as real ones

7

time

EvoVNFP

simulation

Real placement

R3 arrives R1 departs

R1R2R3

R1+2 R1+2+3 R2+3

...

...

...

...

Special EA (described in detail later)

• An optimization algorithm as extension of EA, which imitates biological evolution in varying environments

• MVG changes its objective regularly every fixed number of generations

• Effects of regular changes:

• Individuals become robust to objective changes

• Evolution itself speeds up

• Because getting stuck in local solutions is prevented

no

Modularly Varying Goals (MVG) [13]

8

start

end?

mutation/crossover

fitness calc

init

select individuals for

next generation

end

yes

no

periodic objective change

periodicchange?

yes

[13] N. Kashtan and U. Alon, “Spontaneous Evolution of Modularity and

Network Motifs,” PNAS, vol. 102, no. 39, pp. 13773–13 778, Sep. 2005.

[14] N. Kashtan, E. Noor, and U. Alon, “Varying Environments Can Speed Up Evolution,”

PNAS, vol. 104, no. 34, pp. 13711–13 716, Aug. 2007.

MVG

Individuals robust to objective changes

MVG does not get stuck in local solutions

Conventional EA

Normal EAmay get stuck in local solutions

Detailed Behavior of EvoVNFP

• Intentionally change objectives every fixed number of generations (= period)

• Intentionally use EA without re-initialization of population when objectives change

9

S4

time

Example of

EvoVNFP

simulation

S1 S2 S1 S3 S1 S2 S1 S1 S3 S1 S2 S1 S3 S1 S2 S1 S1 S3 S1 S3 S1 S4

…

...

Requestsin system 1

3 113

13

1

1S1

S2

S3

S4

R3 departs R4 arrivesperiod

Requests

in system

R1

R2

R3

12

1

12

1S1

S2

S3

Requests

in system

R1

R2

S1

S2

S3

S4

13 1

13

13

1

1

R1

R2

R4

Real placement R1+2+3 R1+2 R1+2+4

Objective

state transition

in EvoVNFP

R1+2+3R1+2+3 R1+2R1+2

generate! generate!update! update! generate!

• Example structure of an individual (see figure below):

• Individual represents placement in network

• Connection between VM layer and component layer: allocation of component on VM

• Connection between PM layer and VM layer: allocation of VM on PM

• Mutation: randomly change one element of an individual

• Change connections or the number of cores saved in nodes

Individuals and Mutations

10

＝VM1

8

VM2

6

PM1

16

PM2

16

PM3

16

PM4

16

Comp1

2 Comp2

6

Comp3

3 Comp4

3

PM1

16

PM2

16

PM3

16

PM4

16

VM1

8

VM2

6

Comp1

2

Comp2

6

Comp3

3

Comp4

3

PM layer

VM layer

Componentlayer

VM3

4

VM4

4

number of cores

Fitness Function

• Evaluate how well placements adapt to objectives

• If individuals can be converted to placements:

• Small average delay of chains and small number of used cores high fitness

• Otherwise:

• Fitness is a negative value

• Small number of elements in individuals violating the constraints high fitness

11

𝐹 =

𝑑

𝑑𝑚𝑎𝑥+

𝑊( 𝑖,𝑘 𝑚𝑖,𝑘)

𝑐𝑚𝑎𝑥

−1

− Z

𝑑𝑚𝑎𝑥: reference value of delay𝑐𝑚𝑎𝑥: maximum number of cores𝑍: number of elements which violate the constraints

if individuals can be converted to placements

otherwise

Simulation Settings

• Physical network: 5 routers, 10 PMs, each PM has 16 cores

• Requests: tuples consisting of ingress router, egress router, VNF chain, and transmission rate

• Example: (𝑟1, 𝑟3, {𝑉𝑁𝐹1 → 𝑉𝑁𝐹2}, 200 Mbps)

• Reference methods for comparison:

• Conventional EA (Conv): normal EA that is rerun whenever there is an arival/departure of requests

• Random Immigrant GA (RandImm) [15]

• RandImm initializes randomly selected individuals after mutation step

• Parameters:

• Population size: 1000, elite size: 100, mutation probability: 0.8

• Replacement rate (RandImm): 0.3

12[15] J. Grefenstette, “Genetic Algorithms for Changing Environments,” in Proceedings of PPSN 1992. Elsevier, Sep.

1992, pp. 137–144.

2017/2/14

3

Types of Evaluations and Metrics

• Evaluations under varying parameters

• Evaluation with varying system load (EvoVNFP, Conv, RandImm)

• Evaluation of different period lengths (EvoVNFP)

• Evaluation metric

• Failure probability

• Probability of finding no feasible solution until the next objective change

13

Request

departs

Request

arrives

Request

departs

Request

arrives

EvoVNFPsimulation

Placement

FailSuccess Success

Evaluation with Varying System Load

• Failure probability of EvoVNFP is lowest

• Failure probability increases when the average load increases

14

EvoVNFP has the lowest failure

probability

Evaluation of Different Period Lengths

• 𝑇𝑝 = 5 is the best setting for the considered simulation setting

15

𝑇𝑝 = 5 is the

best setting

𝑇𝑝 = 5

R: average number ofrequests in the system

Summary and Future Work

• Summary

• Evaluating dynamic VNF placement method named EvoVNFP in greater detail by computer simulations

• EvoVNFP generates placements which meet user requests by MVG

• When requests arrive/depart, EvoVNFP runs EA without reinitializing population

• EvoVNFP switches between real objectives and relaxed sub-objectives every fixed number of generations

• Evaluation of EvoVNFP by computer simulations

• EvoVNFP can follow the dynamics of the request arrival/departure

• Specific parameter settings of EvoVNFP make its adaptability even better

• Future work

• Application of further evaluation metrics

16