A New VNE Method for More Responsive Networking in Many-to-Many Groups

PDF: http://bit.do/150708 Screencast: https://www.youtube.com/c/MaratZhanikeev

.

Evolution Towards VNE

1. OSPF Weight Setting Problem◦ normally, a single parameter : weight = cost

2. Multi-parameter e2e path optimization◦ multi-parameter problems are common

3. VNE : closer to OSPF -- multiple overlapping OSPFWSoptimizations

M.Zhanikeev -- [email protected] -- A New VNE Method for More Responsive Networking in Many-to-Many Groups -- http://bit.do/150708 2/19...

2/19

.

Basic VNE Formulation

• VNE : Virtual Network Embedding problem• P for physical and V for virtual

• have: physical graph G(NP,EP)• need to : map multiple user requests G∗(NV,EV)

◦ naturally, NV ∈ NP EV ∈ EP


3/19

.

VNE as a Technology

Physical layer

Virtual Layer 1

VN Request

2 Make VN

3 VN Reply

4 Virtual

Virtual

Physical


4/19

.

Node Profiling Methods

• mostly the same basic approach to node profiling (ranking) 01 02 04 04 05

• exception : method in 06 uses Page Rank -- dependency matrix◦ however, single node profiles are blurred by proliferated multihop dependency

.This paper......... proposes a better version of the most popular method

01 Y.Zhu+1 "Algorithms for Assigning Substrate Network Resources to Virtual Network Components" INFOCOM (2006)

02 J.Lu+1 "Efficient Mapping of Virtual Networks onto a Shared Substrate" Wash. Univ. (2006)

04 G.Even+3 "Competitive and deterministic embeddings of virtual networks" 13th ICDCN (2012)

04 G.Even+3 "Competitive and deterministic embeddings of virtual networks" 13th ICDCN (2012)

05 M.Yu+3 "Rethinking Virtual Network Embedding..." ACM SIGCOMM (2008)

06 X.Cheng+4 "Virtual Network Embedding Through Topology-Aware Node Ranking" ACM SIGCOMM (2011)


5/19

.

VNE Profiling Basics

• node computing CCPU and network CNET capacities

• respective UCPU and UNET utilizations by virtual graphs

• full notation: UCPU,i and UNET,j,i for i-th node and j-th link


6/19

.

VNE Profiling Basics (2)• node CPU and network (all links) profile:

RCPU =

(CCPU −

n∑i=1

UCPU,i

), (1)

RNET =

k∑j=1

(CNET,j −

n∑i=1

UNET,j,i

), (2)

• traditional profiling method:

Rold = RCPURNET. (3)

• proposed profiling method:

Rnew = logaRCPUlogaRNET, (4)

• ,... where a is scaling parameter


7/19

.

VNE Profiling: Example

48 8

28 9

Case 1 Case 2

• Rold ={24, 24

}• Rnew =

{3.17, 2.77

}

48 8

28 9

Case 1 Case 2

• change 9 to 8 in Case 2, then ...

• Rold ={24, 24

}(same)

• Rnew ={3.17, 3.4

}


8/19

.

Experiments


9/19

.

Experiments

• reality : many outcomes (decisions) are the same in old and new profiles

• finding : when different, the new profile tends to perform better

• evaluation method : detect decision diffs and visualize them


10/19

.

Decision Diffs: Same Link Count

0 20 40 60 80 100Node stress [1, 100-1]

0

100

200

300

400

500

Link

stre

ss (c

ount

* st

ress

) [1,

100

-1]

5

52

2

5

5

4

4

3

322

5

5

4

4

4

4

4

43

3

33

5

5

3

3

4

4

22

3

3

5

5

5

5

2

2

Proposed methodConventional versus Proposed line


11/19

.

Decision Diffs: Any/Diff Link Count

0 20 40 60 80 100Node stress [1, 100-1]

0

100

200

300

400

500

Link

stre

ss (c

ount

* st

ress

) [1,

100

-1]

5

4

3

2

4

2

4

4

5

53

5

32

5

3

5

44

4

5

2

5

5

5

2

4

4

2

4

3

45

5

2

4

5

52

2

Proposed methodConventional versus Proposed line


12/19

.

Modeling and Evaluation


13/19

.

Modeling

• low load : L ={0.01, 0.02, 0.03, 0.05, 0.07

}• high/heavy load : H =

{0.1, 0.15, 0.2, 0.25

}• medium load: M = L ∪ H• define a load tuple as XY where X is for node and Y for link utilizations

• example: LH type can be{0.02, 0.2

}


14/19

.

Simulation Setup

• method◦ compare between old A and new B picked randomly from{

LL, LH,HL,HH,MM}

◦ identical sequence of requests for both A and B◦ run until the first rejected (fully packed) request

• physical topology with 100 nodes, each node connects to 10 other nodesrandomly

• outcome metric : request count until rejection (capacity of the network)


15/19

.

Simulation Results

2300 2400 2500Existing (counts)

2300

2400

2500

Prop

osed

(cou

nts)

Type: LL

510 530 550 570Existing (counts)

510

530

550

570

Prop

osed

(cou

nts)

Type: LH


500

520

540

560

Prop

osed

(cou

nts)

Type: HL


440

460

480

500

Prop

osed

(cou

nts)

Type: HH


750

800

850

900

Prop

osed

(cou

nts)

Type: MM

LL LH HL HH MM

0

20

40

60

Avg

. diff

. (pr

op. -

exis

ting)


16/19

.

Yet Better Profiling


17/19

.

Yet Better Profiling

Log Decision Absolute Decision

Node Stress

Sum( Link Stresses)

Min Max

Max

Min

c N

c L Working space


18/19

.

That’s all, thank you ...


19/19