optimization of routing and resource allocation in ... · outline outline 1 curriculum vitae 2...

Optimization of Routing and ResourceAllocation in Communication Networks

HDR Defence Presented by

Fen ZhouLISITE, Institut Superieur d’Electronique de

Paris (ISEP)LIA, University of Avignon (UAPV)

Jury Members:Prof. Chadi Assi Concordia University ReviewerProf. David Coudert INRIA, Sophia Antipolis ReviewerProf. Abdelhamid Mellouk LiSSi, University of Paris 12 ReviewerProf. Marcelo Dias de Amorim LIP6, CNRS/Sorbonne University ExaminerProf. Abderrahim Benslimane LIA, University of Avignon PresidentProf. Nathalie Mitton INRIA, Lille ExaminerProf. Abderrezak Rachedi LIGM, University of Marne-la-Vallee Examiner

Avignon, Sept. 26th, 2018

Outline

Outline

1 Curriculum Vitae

2 Introduction

3 Delivery of Multiple Video Channels in Telco-CDNs

4 Maximizing Lifetime for Data Collection in Wireless Sensor Networks

5 Distance Spectrum Assignment in Elastic Optical Networks

6 Disaster-Resilient Service Provisioning in Optical Datacenters Networks

7 Conclusions and Perspectives

F. Zhou (ISEP/UAPV) HDR Defence Sept. 26, 2018 2 / 99

Curriculum Vitae

1 Curriculum Vitae

2 Introduction







Curriculum Vitae

Curriculum Vitae

2004 - 2007 Master on TelecommunicationsNorthwestern Polytechnic University, China

2007 - 2010 PhD on Optical NetworkingIRISA, INSA of Rennes, France

2010 - 2012 Post-doc on Content Delivery NetworksIMT Atlantique (ex Telecom Bretagne)

09/2012 - 09/2018 Associate Professor on NetworkingLIA lab, CERIUniversity of Avignon

09/2018 - present Associate Professor on NetworkingLISITE labInstitut Superieur d’electronique de Paris (ISEP)


Curriculum Vitae

ResumeResponsibilities + Awards

Awards: PEDR 2016-2020; IEEE senior member, etc

Responsibles: Master RISM; axis FR Agorantic

Conference organisations: ICNC, WCSP, Wimob, Netgcoop, MoWnet

Project coordinations: 5 Eiffel+CSC scholarships, regional projet, Sino-French project...

ResearchRouting, resource allocation, survivability and security in networking

Routing optimization in ITS

Advising: PhD (10), interns (6)

Publications: journals (25), conferences (44)

Teaching

1600h, 3rd year engineer, L2/L3/M1/M2

Networking, multimedia, CISCO, graph theory


Introduction

1 Curriculum Vitae

2 Introduction







Introduction

Communication Networks

https://reseaux.orange.fr/cartes-de-couverture/fibre-optique


Introduction

Communication Networks

Tx

Servers

Rx

Services

… …

Optical fiber

https://reseaux.orange.fr/cartes-de-couverture/fibre-optique


Introduction

Studied Problem in Communication Networks

Routing and Resource Allocation

Routing: path, tree, etc.


Introduction

Studied Problem in Communication Networks

Routing and Resource Allocation (RRA)

Routing

Resource allocation: spectrum, battery, bandwidth, capacity etc.

12.5GHz for one subcarrier

1 2 3 4 … …

Frequency Slot (FS)

320… …


Introduction

Optimization of Routing and Resource Allocation

Challenges

Coupled together

NP-Hard problems

Optimization techniques required


Introduction

Optimization techniques

Integer Linear Programming

Optimal solution

High computation complexity

Heuristics and Approximation Algorithms

Feasible solution, approximation ratio

Low computation complexity

Large-scale optimization: Column Generation

Near-optimal solution

Relatively low computation complexity


Introduction

RRA in Communication Networks

Routing and Resource allocation

Optimization techniques:

ILP, heuristic, approximation algorithm, column generation

Elastic Optical Networks (EONs)

Content Delivery Networks (CDNs)

Wireless Sensor Networks (WSNs)

Datacenter Networks (DCNs)


Delivery of Multiple Video Channels in Telco-CDNs

1 Curriculum Vitae

2 Introduction








MotivationOTT content providers: twitch, ustream, second screen video...

User-generated live video traffic: exploding

Telco-CDN: tight resource and deployment constraint

Problem: How to manage live video traffic? (CNG projet, thesis of Jiayi Liu)

links to IXP Telco-CDN linksentrypointsedge servers

s1 s2

s3

1

2

3

4

56

78 9

10

11 12 13

Figure 1: Topology of a French Telco-CDN



MotivationOTT content providers: twitch, ustream, second screen video...

User-generated live video traffic: exploding

Telco-CDN: tight resource and deployment constraint

Problem: How to manage live video traffic? (CNG projet, thesis of Jiayi Liu)

links to IXP Telco-CDN linksentrypointsedge servers

s1 s2

s3

1

2

3

4

56

78 9

10

11 12 13

Figure 1: Topology of a French Telco-CDNF. Zhou (ISEP/UAPV) HDR Defence Sept. 26, 2018 14 / 99


Multi-channel video delivery problemDescription

Teleco-CDNs G(V,E), entrypoints si ∈ S ⊂ V

Video channels i ∈ I, channel importance πi

Deliver channel i ∈ I from si to subscribing edge-servers Vi ⊂ V with rateless coding

Multi-tree overlay Fi: one distinct unit of rateless codes per tree

ConstraintsEdge-server capacity constraint (cv).∑

i∈I

∑T∈Fi

deg+T (v) ≤ cv,∀v ∈ V (1)

Rateless codes constraint K.

∀i ∈ I, ∀v ∈ Vi, |Fi(v)| ≥ K (2)

QoS (delay) constraint H.

∀i ∈ I, ∀T ∈ Fi,∀v ∈ T,∑

e∈SPT (v)

de ≤ H (3)



Multi-channel video delivery problemDescription

Teleco-CDNs G(V,E), entrypoints si ∈ S ⊂ V

Video channels i ∈ I, channel importance πi

Deliver channel i ∈ I from si to subscribing edge-servers Vi ⊂ V with rateless coding

Multi-tree overlay Fi: one distinct unit of rateless codes per tree

ConstraintsEdge-server capacity constraint (cv).∑

i∈I

∑T∈Fi

deg+T (v) ≤ cv,∀v ∈ V (1)

Rateless codes constraint K.

∀i ∈ I, ∀v ∈ Vi, |Fi(v)| ≥ K (2)

QoS (delay) constraint H.

∀i ∈ I, ∀T ∈ Fi,∀v ∈ T,∑

e∈SPT (v)

de ≤ H (3)F. Zhou (ISEP/UAPV) HDR Defence Sept. 26, 2018 15 / 99


Multi-channel video delivery problem

Objectives

Main: Max overall importance of the delivered video channels, i.e.,∑i∈I

Ri × πi

2nd: Min traffic load, i.e., number of links in the overlay

Telco-CDNlinks to IXP Telco-CDN links

entrypointsedge servers

s1 s2

s3

12

3

4

5 67

8 910

11 12 13

Example of a multi-tree overlayEntrypoint s1; subscribingedge-servers: 10 and 12, delay H = 2,and rateless codes parameter K = 2,upload capacity c8 = 2

s1

8

10 12

s1

9

12

s1

7

10



Proposed Solutions

Problem Decomposition

Overlay construction: capacity-and-delay-bounded min-cost forest

Bandwidth allocation: multi-dimension knapsack

ILP formulationsSOP-ILPs: ILP-Overlay + ILP-Bandwidth→ Suboptimal.ci

v: capacity reserved by edge server v ∈ V for forest Fi:∑i∈I

civ × Ri ≤ cv, ∀v (4)

JOP-ILP (overlay construction + bandwidth allocation)→ Optimal.Lik

uv ∈ 0, 1: Is 1 if arc (u, v) is used in T ik, 0 otherwise.∑i∈I

∑k∈1,...,Wi

∑u∈N+(v)

Likvu ≤ cv, ∀v (5)



Proposed Solutions

Problem Decomposition

Delivery Overlay: capacity-and-delay-bounded min-cost forest

Bandwidth allocation: multi-dimension knapsack

HeuristicsSOP1-heu: two-step optimizationHeu-Overlay for all first, then Heu1-Bandwidth (sorting πi)

SOP2-heu: two-step optimizationHeu-Overlay for all first, then Heu2-Bandwidth (knapsack and bin packing)

JOP-heu: Divide and ConquerRecursive overlay construction + Bandwidth allocation:



Heuristic vs. Optimal

Table 1: Heuristic Solutions vs MILP Solutions: 6 channelsvideo bit-rate JOP-ILP SOP-ILP JOP-heu SOP1-heu SOP2-heu

Profit ratio512 kbps 100% 100% 100% 100% 100%

1024 kbps 100% 90.7% 100% 100% 100%1536 kbps 100% 76.7% 100% 100% 100%2048 kbps 100% 67.4% 100% 93% 76.7%

Number of delivered channels512 kbps 6 6 6 6 6

1024 kbps 6 4.5 6 6 61536 kbps 6 3.5 6 6 62048 kbps 6 3.5 6 5 5

Ratio of used capacity512 kbps 12.7% 12.8% 14% 13.8% 13.8%

1024 kbps 21.7% 16.5% 23.7% 23.3% 23.3%1536 kbps 30.8% 18.1% 33.2% 32.7% 32.7%2048 kbps 40.4% 21.2% 42.6% 36% 34.3%

Average computing time (Seconds)2048 kbps 749.5 228 0.3 < 0.1 0.3



Results of Heuristics in French CDN - 1

40 60 80 1000.5

0.6

0.7

0.8

0.9

1

number of video channels

profi

trat

io

JOP-heu SOP1-heu SOP2-heu

(a) profit ratio vs. channels

40 60 80 100

40

60

80

100

number of video channelsde

liver

edvi

deo

chan

nels


(b) delivered channels vs. channels

Figure 2: Simulation Results in a French Telco-CDN



Results of Heuristic in French CDN - 2

40 60 80 1000

0.2

0.4

0.6

0.8

1


ratio

ofus

edca

paci

ty


(a) capacity ratio vs. channels

500 1,000 1,500 2,000 2,5000.5

0.6

0.7

0.8

0.9

1

video bitrate (kbps)pr

ofitr

atio


(b) profit ratio vs. bit-rate

Figure 3: Simulation Results-2 in a French Telco-CDN



Results of Heuristics in French CDN - 3

500 1,000 1,500 2,000 2,500

40

60

80

100

120

video bitrate (kbps)

deliv

ered

vide

och

anne

ls


(a) delivered channels vs. bit-rate

500 1,000 1,500 2,000 2,5000

0.2

0.4

0.6

0.8

1

video bitrate (kbps)ra

tioof

used

capa

city


(b) capacity ratio vs. bit-rate

Figure 4: Simulation Results in a French Telco-CDN



Computation Time of Heuristic

40 60 80 1000

2

4

6

8

10


com

putin

gtim

e(s

econ

ds)


(a) Computing time in French-CDN (16nodes)

Figure 5: Average Computing Time (seconds) on Real-Scale Systems: 30 ∼ 105channels, video bitrate 2048 kbps



Summary of Video Delivery in Telco-CDNs

Joint optimization of Video Delivery

Delivery of multiple video channels in Telco-CDNs

Joint optimization vs. two-step optimization

JOP serves 1/3 more channels, and also time-efficient


Maximizing Lifetime for Data Collection in Wireless Sensor Networks

1 Curriculum Vitae

2 Introduction








Wireless Sensor Networks (WSNs)

Applications

Environment monitoring: forest, pollution etc.

Health monitoring

Habitat monitoring



Data Collection and Transmission

Data Gathering Tree

Data collection in each time slot

Data transmission: many-to-one communications

Tree optimization: maximize lifetime

s

1

32

54

Figure 6: An example of a data-gathering tree



Motivations

LiteratureOnly heuristic algorithms are proposed

Only simple energy consumption model is considered

Exact solution is missing except exhaustive search

Either routing or data aggregation is proposed for improving network lifetime

This workThree data aggregation techniques

Exact solution using MILP formulations

Joint optimal routing and data aggregation

Gain on network lifetime



Maximum-Lifetime Data-Gathering Tree

Data Gathering Tree

Static sensor nodes: wireless + limited battery

Single sink: sensors report data to the sink in each time slot

Data aggregations

Data collection tree

Maximize network lifetime of tree (the number of time slots)

s

Figure 7: Data collection using a WSN



Maximum-Lifetime Data-Gathering Tree

Sensor Lifetime (number of time slots)

Emission energy unit et /packet

Reception energy unit er/packet

Computing equation

LTv =

Ev

ntv · et + nr

v · er (6)

Network Lifetime (number of time slots)

The time duration until the first sensor node is out of battery,i.e., L = min

v∈VLT

v

Critical for full monitoring in WSNs



Data Aggregation Techniques

Aggregation Modes

Full Aggregation Mode (FAM)

No Aggregation Mode (NAM)

Hybrid-Aggregation Mode (HAM) using Compressive Sensing

Figure 8: Comparison of different data compression methods



Energy Consumption Model

Sensor Lifetime

LTv =

Ev

ntv · et + nr

v · er (7)

Full Aggregation Mode (FAM)

Statistical queries: sum, max, ...

LTv =

Ev

et + dTv · er (8)

dTv is the number of incoming links of sensor v in

tree T . In Fig. 9, dT1 = 2 for node 1 in FAM

s

1

32

54

Figure 9: A data-gathering tree




Sensor Lifetime

LTv =

Ev

ntv · et + nr

v · er (9)

No Aggregation Mode (NAM)

No compression: Pictures, air parameters fordifferent regions, etc

LTv =

Ev

et + f Tv · (et + er)

(10)

f Tv denotes the number of sensors using v as a relay

to report data to s (including all descendants ofsensor v). In Fig. 10, f T

1 = 4 for node 1 in NAM

s

1

32

54





Sensor Lifetime

LTv =

Ev

ntv · et + nr

v · er (11)

Hybrid-Aggregation Mode (HAM)

Compressive Sensing (partial compression): oceandata

LTv =

Ev

minf Tv + 1, k · et + f T

v · er(12)

Let f Tv be the number of data units received from

the incoming links of sensor v. In Fig. 11, ifk = 4, then f T

1 = 4 for node 1.

s

1

32

54




Problem: Maximum-Lifetime Data-Gathering Tree

Given parameters

WSN topology G(V,E) and wireless transmission range

et and er

Data aggregation modes (FAM, NAM, HAM)

Energy consumption model

Objective

Find a data collection tree from sensors to the sink

Maximize network lifetime: max L



Problem: Maximum-Lifetime Data-Gathering Tree

Challenges

NP-Hard

Different aggregation modes

Non-linear relationship between lifetime and the number of transmitted and receivedmessages in Eq. (13)

LTv =

Ev

ntv · et + nr

v · er (13)

Maximizing network lifetime (gathering tree)

Three MILP formulations



Simulation Configurations

Parameters and ToolsInitial energy 1 J

Emitting energy et=6.4 µJ

Receiving energy et=12.8 µJ

100×100 m2 grid with an interval of 10 m

C++ and IBM ILOG CPLEX 12.2

EvaluationsNetwork lifetime vs. Number of sensors

Network lifetime vs. Transmission range

Network lifetime vs. Compression factor k



Topologies

Sensors on the grid (100×100 m2)

Configuration I: Sink in the center

Configuration II: Sink at a corner

10

10

s

s

(I) Sink in the center (II) Sink at one corner

10

10

Figure 12: WSN grid topology with different sink positions



Numerical Results

Results for Configuration I

20 25 30 35 40 450

2

4

6

8·104

number of sensor nodes (transmission range = 10)

netw

ork

lifet

ime

(min

utes

)

FAMNAM

HAM (k=2)HAM (k=3)

(a) Lifetime vs. # Sensors

10 15 20 25 30 35 40 450

2

4

6

8·104

transmission range (number of sensor nodes = 49)

netw

ork

lifet

ime

(min

utes

)

FAMNAM

HAM (k=2)HAM (k=3)

(b) Lifetime vs. Range

1 2 3 4 50

2

4

6

8·104

compressing factor k (transmission range =10)

netw

ork

lifet

ime

(min

utes

)

HAM (n=25)HAM (n=36)

(c) Lifetime vs. k

Figure 13: Simulation Results for Configuration I (sink at the center)



Data Accuracy vs. k

Table 2: Performance Evaluation of HAM (25 sensors, and sink node in center)

Metrics k = 1 k = 2 k = 3 k = 4 k = 5Lifetime 52083 26041 17361 13020 10416

Maximum Delay 10 10 10 8 6Data Accuracy 25% 36% 52% 72% 80%



Summary of Data Collection

Data gathering tree with maximum network lifetime

Joint routing and data aggregation

Three MILP formulations

Network Lifetime increases fivefold

Tradeoff between data accuracy and network lifetime


Distance Spectrum Assignment in Elastic Optical Networks

1 Curriculum Vitae

2 Introduction








Motivations

Elastic Optical Networks (EONs)

Huge bandwidth

Low latency

Spectrum flexibility

Figure 14: EONs



Motivations

Spectrum resource in optical fibers

Narrow-band (12.5GHz or less) frequencies→ Frequency Slots (FS’)

Flexibility

Limited number of FSs

12.5GHz for one subcarrier

1 2 3 4 … …

Frequency Slot (FS)

320… …

Figure 15: EONs and FSF. Zhou (ISEP/UAPV) HDR Defence Sept. 26, 2018 44 / 99


Distance Spectrum Allocation (DSA) in EONs

Guard Band (GB)

Two requests (lightpaths) with common links

Interference Mitigation for QoT

Physical layer security

Heterogeneous GB size→ Spectrum efficiency



Distance Spectrum Allocation (Thesis of Haitao Wu)

Spectrum Conflicts

Inside a lightpath: Spectrum continuity

Inside a lightpath: Spectrum contiguity

Between lightpaths: Guard Band (GB) separation



Conflict Graph G(V,E)v ∈ V: Vertices set with each node representing a request (lightpath)

e = (vi, vj) ∈ E: (vi, vj) exists if lightpaths vi and vj share common links

s ∈ N+: Index of an FS

v ∈ V: Vertices set with each node representing a request (lightpath)

vwi : Number of FSs required by request i

wvi : Set of contiguous FSs assigent for request i

vbi , v

ai : Begin/end index of FS assigned for request i

de: Least guard band size for separating lightpaths vi and vj

Figure 16: DSA conflict graph G(V,E, vwi , dvivj).



Conflict Graph G(V,E)v ∈ V: Vertices set with each node representing a request (lightpath)

e = (vi, vj) ∈ E: (vi, vj) exists if lightpaths vi and vj share common links

s ∈ N+: Index of an FS

v ∈ V: Vertices set with each node representing a request (lightpath)

vwi : Number of FSs required by request i

wvi : Set of contiguous FSs assigent for request i

vbi , v

ai : Begin/end index of FS assigned for request i

de: Least guard band size for separating lightpaths vi and vj

Figure 16: DSA conflict graph G(V,E, vwi , dvivj).F. Zhou (ISEP/UAPV) HDR Defence Sept. 26, 2018 47 / 99


DSA ModelObjective: Minimize the maximum FS index assigned

Minimize maxs∈(∪

vi∈Vwvi

)(s) (DSA), (14)

ConstraintsBandwidth Requirement Constraint

|wvi | = vwi , ∀vi ∈ V, (15)

Spectrum Continuity Constraint: satisfied automatically

Spectrum Contiguity Constraint: wvi =vbi , v

bi + 1, ..., va

i − 1, vai

Spectrum Distance Constraint

distance(wvi ,wvj ) ≥ dvivj , ∀vivj ∈ E, (16)

where,distance(wvi ,wvj ) = min

s ∈ wvi t ∈ wvj

(|s− t| − 1).



DSA Exampple: Input

Table 3: Four requests in a 4-cycle EON

Bandwidth Capacity RouteRequest R1 3 FS’ B-A-DRequest R2 2 FS’ C-B-ARequest R3 3 FS’ A-D-C-BRequest R4 1 FS C-B-A-D

Figure 17: Explicit routes for the above table.



DSA Conflict Graph and Output



Comparison of related coloring problems

Classical Coloring Fractional Traditional DSA(e.g., WA ) Coloring SA (this work)

Vertex Color One color Set of colors Set of colors Set of colorsColor Contiguity N/A No need Required RequiredColor Distance of Disjoint Disjoint Identical positive Various positiveAdjacent Vertices Disjoint Disjoint integer integers



Hardness and Inapproximability

MHP: Minimum Hamitolion Path

Theorem 1

MHP ≤P DSA⇒ DSA is NP-hard.

Theorem 2

Unless NP ⊂ ZPP , no polynomial-time DSA heuristic algorithm APX can

guaranteeAPX(I)OPT(I)

within O(n1−ε) for all instances I, where n is the number

of vertices of I and ε > 0.



Upper and Lower Bounds

|opt(G)| : optimal solution of DSA problem

χ(G): chromatic number of G

∆(G): maximum nodal degree, χ(G) ≤ ∆(G) + 1

ψ ∈ Ψ(G): maximal clique / clique set

MHP(ψ): minimum Hamilton path length in ψ

ψw: sum of node weights in ψ

Theorem 3

Given any DSA conflict graph G, inequality (17) is held for the optimalsolution of the DSA problem.

maxψ∈Ψ(G)

|MHP(ψ)|+ ψw ≤ |opt(G)| ≤χ(G)−1∑

i=1

de′i+

χ(G)∑i=1

v′wi . (17)



Upper and Lower Bounds

|opt(G)|: optimal solution of DSA problem

∆(G): maximum nodal degree in G, χ(G) ≤ ∆(G) + 1

Corollary 4

If G(V,E, vwi , dvivj) is a DSA graph, then |opt(G)| ≤

∆(G)∑i=1

dei +

∆(G)+1∑i=1

vwi .



Two-phased Algorithm



Algorithm Analysis

First Phase Greedy Algorithm (FPGA):

Theorem 5

If G(V,E, vwi , dvivj) is a complete DSA graph with triangle inequality

then the approximation ratio of FPGA is not bigger than12(dlog2 |V|e+ 1).

Theorem 6

If a DSA graph G(V,E, vwi , dvivj) is a bipartite graph whose vertex label

is well, then FPGA can get the optimal solution of G.



The Numerical Results

Two-phase algorithm is compared with FPGA, ILP and PRA.

(a) 14 (b) 15 (c) 16

(d) 17 (e) 18 (f) 19

Figure 18: Six random graphs with 14-19 vertices.



Table 4: Numerical results for six random graphs

# vertex 14 15 16 17 18 19PRA 92.0 103. 101. 133. 125. 180.

FPGA 72.4 76.6 83.4 94.2 89.8 126.Two-phase 69.0 74.4 81.6 91.2 88.2 124.ILP-DSA 67.4 72.4 79.6 88.8 83.8 116.

14 15 16 17 18 190 %

10 %

20 %

30 %

40 %

50 %

60 %

70 %

Rel

ativ

ega

p

Two-phaseFPGAPRA

Figure 19: Relative gaps by Two-phase, FPGA and PRA.



Table 5: Numerical results for random complete graphs

# vertices 14 15 16 17 18 19PRA 168.8 193.8 237.6 238.6 283.0 290.3

FPGA 144.6 164.2 197.2 199.4 229.6 238.3Two-phase 143.2 163.2 196.2 196.6 226.8 233.6ILP-DSA 142.2 160.4 194.0 194.2 223.2 224.0

14 15 16 17 18 190 %

10 %

20 %

30 %

Rel

ativ

ega

p

Two-phaseFPGAPRA

Figure 20: Relative gaps by Two-phase, FPGA and PRA.



Figure 21: Six random graphs with 14 vertices cum edge number from 15 to 90

15 30 45 60 75 90

40

60

80

100

120

140

35

4858

82

102

136

35

4859

85

106

140

The number of edges

The

max

imum

FS’i

ndex

0 %

1 %

2 %

3 %

4 %

App

roxi

mat

ion

ratio

Two-phaseILP-DSAratio

Figure 22: Numerical results for Edge number scenario.



Table 6: Simulation results for NSFNET and US Backbone

NSFNET US Backbone

# request ILP-DSA Two-phased PRA # request ILP-DSA Two-phased PRA

10 29 29 29 10 33 33 33

20 72 72 76 30 186 189 197

30 153 153 177 50 351 363 462

40 200 201 252 100 — 1339 1898

50 420 423 500 150 — 2843 4666

60 — 469 602 200 — 3784 7743

70 — 598 805 250 — 6347 13020

80 — 890 1155 300 — 8140 17303



Summary

DSA in EONsNP-Hard and inapproximability

Lower and upper bounds

Near optimal two-Phase algorithm


Disaster-Resilient Service Provisioning in Optical Datacenters Networks

1 Curriculum Vitae

2 Introduction








Optical DCNs Survivability (Thesis of Min JU)Link Failure

Construction, damagedconnectors ... ...

Node FailureNode equipment failure(transponder, switching) ... ...

Large Area DisasterDatacenter system damage,earthquake, flood ... ...

→ →

Most common!! Threat on Datacenter networks!!

Cables were cut by ship, 2017,Somalia. $10 million/day

Weather and climate disasters,2017, USA. $306 billion

A ”power surge” in one BritishAirways’ datacenter, 2017



Large Area Disaster Failure in Datacenter Networks

Disasters: Volcan, Tsunamis, Hurricane, Flood, etc

Disaster zones: Set of OXC nodes and fibers

DC content survivability

Disaster-disjoint primary path and backup path

5

4

61

3

2

Disaster zone

Primary path 1-2

Backup path 1-3-5

Request at node 1 with content



Path Protection against Disaster Failure

5

4

61

3

2

Disaster zonePrimary path 1-2

Backup path 1-3-5 Content

Primary path 1-4

Backup path 1-3-5

5

4

61

3

2

DEBPP SEBPP

r1 r2

1

1

21 21

1

1

11

Data center with content replia

The same content can be replicated at multiple DC nodes.

Anycast routing, i.e., from one to one of many

One to one of many: the service can be provisioned from any of the DC withcontent replica.

DEBPP (Dedicated end-to-content backup path protection)

SEBPP (Shared end-to-content backup path protection)



Disaster-aware Path Protection Schemes in the Literature

ILP Heuristics

SLR WDM

[C. Develder et al, Trans. Netw., 2014] [C. Develder et al, Trans. Netw., 2014][M. F. Habib et al, J. Lightw. Technol., 2012] [M. F. Habib et al, J. Lightw. Technol., 2012]

[S. S. Savas et al, Photon. Netw. Commun., 2014] [S. S. Savas et al, Photon. Netw. Commun., 2014][S. S. Savas et al, Photon. Netw. Commun., 2016] [S. S. Savas et al, Photon. Netw. Commun., 2016]

EON

[R. Xu et al, Optoel. Global Conf., 2016][C. Ma et al, Photon. Netw. Commun., 2015] [C. Ma et al, Photon. Netw. Commun., 2015]

[X. Li et al, Opt. Express, 2016] [X. Li et al, Opt. Express, 2016][B. Chen et al, Photon. Netw. Commun., 2017]

ILP formulation: not scalable

Heuristics: not tractable

Impacts of No. DC and replica: not yet studied

Comparison of protection techniques: missing



Objective and Constraints

Inputs:

Set of disaster zones (DZs)

Set of requests and their requiredcontent

Number of DCs and content replica (k)

EON topology and set of FSs

Output:

DC location

Placement of content replica

Disaster-disjoint primary and backuppaths

FS allocation

Objective: Minimize total spectrum usage for disaster protection

Single disaster zone protection

DEBPP vs. SEBPP

Impacts of No. Dc and replica

Scalable and tractable approach



Methodology 1: Joint ILP Formulation

Ojective: Link-FSs Max-FSs

Minimize θ1 · (∑a∈A

∑r∈R

pWra · φr +

∑a∈A

Ta) + θ2 ·∆

Constraints:(1) Datacenter and content assignment

(2) Disaster-disjoint path generation

(3) Spectrum allocation

Computational complexity

DEBPP

No. of dominant variablesO(|R|2, |R||A|, |R||Z|, |C||D|)No. of dominant constraintsO(|R|2|A|, |R||Z||A|)

SEBPP

No. of dominant variablesO(|R|2, |R||A|, |R||Z|, |C||D|)No. of dominant constraintsO(|R|2|A|, |R|2|Z|, |R||Z||A|)



Methodology 2: Column Generation (CG)

Step 2: Initial solution

Dijkstra’s

shortest path

algorithm

Initial columns

Step 3: Master problem

Solving linear relaxation of

master problem

Get low bound and

dual variables

Step 4: Pricing problem

Solving pricing problem

with ESWBP algorithm

Reduced cost < 0 ?Yes No

Step 5: Final solution

Solving master

problem

Integer solution

Add

new

columns

Integerizing all the

relaxed variables

Step 1: DC assignment

and content placement

Content placement with

content protection ILP model

in [Habib et al., 2012]

DC nodes and content replicas

Column generation

procedure

Advantages of CG

Scalable iterative optimization

Lower bound z∗

ρ-optimum, ρ ≤ (z− z∗)/z∗




2

68

1

4

75

3

DC

DC

DC

Contents

Step 1: DC assignment and content placement

K DC nodes: Average minimum distance

Place content replica in DCs closer to its popular region

Habib et al. Design of disaster-resilient optical DC networks. J. Lightw. Technol., 2012.F. Zhou (ISEP/UAPV) HDR Defence Sept. 26, 2018 70 / 99



2

68

1

4

75

3

DC

DC

DC

Contents

R1 at node 8: 3 spectrum slots

Primary path: 8-1-2 with spectrum

Backup path: 8-7-6-5 with spectrum

1 2 3

1 2 3


Primary path and backup path via Dijkstra’s shortest path

Spectrum usage allocation with coloring heuristic




2

68

1

4

75

3

DC

DC

DC

Contents R1 at node 8: 3 spectrum slots

Output:

Configuration 1:



1 2 3

1 2 3

Input:

Configuration 1:



1 2 3

1 2 3

… …

Step 3: Master problemA configuration ωr for request r:

1 Primary path and backup path2 Spectrum usage for primary and

backup paths

Input: ∀r ∈ R, ωr

Output: Final primary and backup pathsand associated spectrum




2

68

1

4

75

3

DC

DC

DC

Contents R1 at node 8: 3 spectrum slots

Output:

Configuration 1:


Backup path: 8-7-5 with spectrum

1 2 3

1 2 3


Backup path: 8-7-5 with spectrum

2 3 4

1 2 3

New configuration 2:

New configuration 3:

… …



1 2 3

1 2 3

Step 4: Pricing problem (k shortest paths)

Input: R, G(V,A), Disaster zones

Output: Configuration ωr, ∀r ∈ R

1 Primary path and backup path of ∀r ∈ R

2 Spectrum usage on primary path of ∀r ∈ R

3 Spectrum usage on backup path of ∀r ∈ R





Dijkstra’s

shortest path

algorithm

Initial columns


Solving linear relaxation of

master problem

Get low bound and

dual variables

Step 4: Pricing problem

Solving pricing problem

with ESWBP algorithm

Reduced cost < 0 ?Yes No


Solving master

problem

Integer solution

Add

new

columns

Integerizing all the

relaxed variables

Step 1: DC assignment

and content placement

Content placement with

content protection ILP model

in [Habib et al., 2012]

DC nodes and content replicas

Column generation

procedure


Convert all variables to integer

Solve the master problem again



Numerical results: Simulation settings

NSFNET network14 nodes, 44 links3.1 nodal degree, 14 DZs

6

87

54

3

1

14

13

1112

9

10

2

US Backbone network28 nodes, 90 links3.2 nodal degree, 15 DZs

26

28

27

1094 61

1423

2115 7

24

1215

25

8

13 18

1619

21

22

20

317

DZ

Hardware: 3.5 GHz CPU, 8 GBytes RAM

Software: CPLEX 12.06

Traffic

FSs: randomly [1, 10]No. requests: [10, 100]

Parameters:

300 Available FSs10 Available contentcontent replicas K



Spectrum usage of DEBPP and SEBPP vs. # requests

No. RequestsJoint ILP models CG approach

GapzILP

1h FStotal FSlink tILP zLP zCG1h FSlink FSmax tCG tCG

LPR tCGILP

10DEBPP 290 260 30 8s 312 312 270 42 1s 1s 0s 7.59%SEBPP 253 233 20 448s 275 275 243 32 5s 5s 0s 8.70%

20DEBPP 448 412 36 3600s 488 493 431 62 6s 6s 0s 10.04%SEBPP 379 345 34 3600s 424.733 432 388 44 7s 7s 0s 13.98%

30DEBPP 663 613 50 3600s 691.998 700 630 70 7s 7s 0s 5.58%SEBPP 614 570 44 3600s 608 630 572 58 23s 23s 0s 2.61%

40DEBPP 867 792 75 3600s 898.389 912 826 86 54s 8s 46s 5.19%SEBPP 832 765 67 3600s 761.612 792 719 73 49s 44s 5s -



SEBPP spectrum usage vs. # DCs and # content replica K

US BackboneNo. of DCs Location of DCs

4 Nodes 1, 12, 21, 286 Nodes 1, 7, 14, 19, 21, 288 Nodes 1, 7, 9, 12, 14, 19, 21, 28

10 Nodes 1, 3, 7, 9, 12, 14, 19, 21, 24, 28

20 40 60 80

1,000

2,000

3,000

No. of Requests

Tota

lFSs

Usa

ge

DC4DC6DC8

(a) FSs vs. # DCs

20 40 60 80

1,000

2,000

3,000

No. of Requests

Tota

lFSs

Usa

ge

K=2K=3K=4K=5

(b) FSs vs. # K (6 DCs)

20 40 60 80

1,000

2,000

3,000

No. of Requests

Tota

lFSs

Usa

ge

K=2K=4K=6K=8

(c) FSs vs. # K (8 DCs)



Summary of Disaster Protection

Comparison of Protection Techniques

Spare capacity efficiency: SEBPP > DEBPP > p-Cycle

Recovery speed: DEBPP > p-Cycle > SEBPP

DEBPP vs. SEBPP, less spectrum usage, higher computation complexities.

Impacts on Spectrum usage

Reasonable # DCs,→ spectrum savings.

Reasonable # content replica k→ spectrum savings.

Optimization

ρ-optimum CG approach

Scalable for large disaster failure in DCNs

Path generation

Spectrum allocation


Conclusions and Perspectives

1 Curriculum Vitae

2 Introduction








Conclusions

1: Video Delivery in Telco-CDNs

Overlay construction (Thesis of J. Liu)

Bandwidth allocation

2: Data Collection in WSNsGathering tree

Data aggregation

Energy allocation (lifetime)

3: DSA in EONsSpectrum allocation (Thesis of H. Wu)

Distance constraint

4: Protection in Optical DCNs

Replica placement (Thesis of M. Ju)

Path protection against diasters

Frequency slot allocationwwOptimization of routing and resource allocation

Tools: ILP, CG, heuristic, approximation algorithm

Optimization techniques do help!



Conclusions





Data aggregation



Distance constraint




Frequency slot allocation

wwOptimization of routing and resource allocation





Conclusions





Data aggregation



Distance constraint




Frequency slot allocationwwOptimization of routing and resource allocation





Perspectives

Similar optimization techniques will also be helpful to solve:⇓

Physical-layer security in EONs

Threats: attacks, interferences

Routing, spectrum assignment andprotection

QoE for 360 video streaming

QoE

360 video delivery

Intelligent transportation systems

Road security via VANETs

Car redistribution for carsharing

Hazardous transportation

Network function virtualizationNetwork embedding

Service function chain provision

+ new directionBi-level optimization for ITS and green networking.


Overview of Research Activities


Routing optimization and resource allocation

Optical networks, RF/VLC HetNets

Wireless ad-hoc networks (VANET, WSN et MANET)

Content delivery networks

Intelligent transportation systems



Optical Networks and Datacenter Networks

ProblemsRouting and spectrum assignment (RSA)

Survivability and security

Multicast, network functionvirtualization,



Wireless Ad-Hoc Networks

ProblemsVANETs: Routing for thedissemination of safety messages

WSNs: Data gathering andaggregation

ProblemsMANETs: Malicious nodedetection

RF/VLC HetNets: Resourceallocation



Content Delivery Networks

ProblemsVideo streaming using rateless coding

Overlay optimization




Intelligent Transportation Systems (ITS)

ProblemsCar relocation optimization forcar-sharing systems

Personalized visit tours

Hazardous transport design


Additional Slides for Data Collection Trees in WSNs

MILP for FAM

max L (FAM-MIP) (18)

s.t. constraints (19)-(27).

∑v∈V

∑a∈δ+(v)

Xa = n, (19)

∑a∈δ−(s)

Fa = n, (20)

∑a∈δ+(v)

Fa −∑

a∈δ−(v)

Fa = 1, ∀v (21)

Xa ≤ Fa, ∀a (22)

Xa ≥1

nFa, ∀a (23)∑

a∈δ−(v) Xa − i + 12

dv≤ Bi

v, ∀v, ∀i ∈ Jdv − 1K (24)

∑a∈δ−(v) Xa − i + 1

2

dv≥ Bi

v − 1, ∀v, ∀i ∈ Jdv − 1K (25)

Ev − (et + i · er)Lv

M1v

≥ Biv − 1, ∀v, ∀i ∈ Jdv − 1K (26)

L ≤ Lv, ∀v (27)



MILP for NAM

max L (NAM-MIP) (28)


∑v∈V

∑a∈δ+(v)

Xa = n, (29)

∑a∈δ−(s)

Fa = n, (30)

∑a∈δ+(v)

Fa −∑

a∈δ−(v)

Fa = 1, ∀v (31)

Xa ≤ Fa, ∀a (32)

Xa ≥1

nFa, ∀a (33)∑

a∈δ−(v) Fa − i + 12

n≤ Bi

v, ∀v, ∀i ∈ Jn− 1K (34)∑a∈δ−(v) Fa − i + 1

2

n≥ Bi

v − 1, ∀v, ∀i ∈ Jn− 1K (35)

Ev −(

et + i(et + er))

Lv

M2v

≥ Biv − 1, ∀v, ∀i ∈ Jn− 1K (36)

L ≤ Lv, ∀v (37)



MILP for HAM (part 1)

Flow Constraintsmax L (HAM-MIP) (38)


∑v∈V

∑a∈δ+(v)

Xa = n, (39)

∑a∈δ−(v)

Fa ≤ (k − 1) + (n− k)Hkv , ∀v (40)

∑a∈δ−(v)

Fa ≥ k · Hkv , ∀v (41)

∑a∈δ+(v)

Fa −∑

a∈δ−(v)

Fa ≥ 1− (n− k)Hkv , ∀v (42)

∑a∈δ+(v)

Fa ≥ (k − 1)Hkv + 1, ∀v (43)

Xa ≥1

kFa, ∀a (44)

Xa ≤ Fa, ∀a (45)



MILP for HAM (part 2)

Energy Constraints (Linearization)max L (HAM-MIP) (46)


Ev − (i · et + j · er)Lv

M3v

≥ Biv + Bj

v − 2, ∀v, ∀i ∈ JkK,

∀j ∈ Jn− 1K (47)∑a∈δ+(v) Fa − i + 1

2

k + 1≤ Bi

v, ∀v, ∀i ∈ JkK (48)

∑a∈δ+(v) Fa − i + 1

2

k + 1≥ Bi

v − 1, ∀v, ∀i ∈ JkK (49)

∑a∈δ−(v) Fa − j + 1

2

n≤ Bj

v, ∀v, ∀j ∈ Jn− 1K (50)∑a∈δ−(v) Fa − j + 1

2

n≥ Bj

v − 1, ∀v, ∀j ∈ Jn− 1K (51)

L ≤ Lv, ∀v (52)


Additional Slides for Disaster Failure Protection (DFP)

Why not p-Cycle Protection?

2

68

1

4

75

3

P-Cycle 1-2-3-6-7-8

Working path 7-5

Protection path 7-6-3-2

p-Cycle vs. DEBPP vs. SEBPP

Drawbacks: Waste spare capacity

Backup Datacenter on p-cycle

Inefficient spare capacity sharing



Disaster Failure: Joint ILP Formulation—DEBPP & SEBPP



∑r∈R

pWra · φr +

∑a∈A

Ta) + θ2 ·∆

Constraints1.DC assignment and contentplacement constraints

∑d∈D

ΛWrd = 1, ∀r (53)

∑d∈D

ΛBrd = 1, ∀r (54)

∑d∈D

Rcd ≤ K, ∀c (55)

ΛWrd + Λ

Brd ≤ Rcr

d , ∀r, ∀d (56)

2.Flow-conservation constraints

∑a∈Ψ

+v

pWra −

∑a∈Ψ−v

pWra =

1, v = sr

− ΛWrv , v ∈ D, ∀r, ∀v.

0, otherwise

(57)

∑a∈Ψ

+v

pBra −

∑a∈Ψ−v

pBra =

1, v = sr

− ΛBrv, v ∈ D, ∀r, ∀v.

0, otherwise

(58)



Disaster Failure: Joint ILP Formulation—DEBPP+SEBPP



∑r∈R

pWra · φr +

∑a∈A

Ta) + θ2 ·∆

Constraints

3.Disaster-zone-disjoint pathconstraints

αWrz ≤

∑a∈z

pWra, ∀r, ∀z (59)

αWrz ≥ pW

ra, ∀r, ∀z, ∀a ∈ z (60)

αBrz ≤

∑a∈z

pBra, ∀r, ∀z (61)

αBrz ≥ pB

ra, ∀r, ∀z, ∀a ∈ z (62)

αWrz + α

Brz ≤ 1, ∀r, ∀z (63)

4.Spectrum allocation constraintspW

ra + pWr′a − 1 ≤ γW

rr′ , ∀r, r′, r > r′, ∀a (64)

γWrr′ = γ

Wr′r, ∀r, r′, r > r′ (65)

pBra + pB

r′a − 1 ≤ γBrr′ , ∀r, r′, r > r′, ∀a (66)

γBrr′ = γ

Br′r, ∀r, r′, r > r′ (67)

pWra + pB

r′a − 1 ≤ γWBrr′ , ∀r, r′, r 6= r′, ∀a (68)

βWrr′ + β

Wr′r = 1, ∀r, r′, r > r′ (69)

βBrr′ + β

Br′r = 1, ∀r, r′, r > r′ (70)

gWr + φr ≤ ∆, ∀r (71)

gBr + φr ≤ ∆, ∀r (72)



Disaster Failure: Column Generation (CG)Step 1: DC assignment

Minimize∑d∈V′

∑r∈R

Hs,d · Rc,r

∑d∈V′

Rc,r = K, ∀c (73)

∑d∈z

Rc,r ≤ 1, ∀c, ∀z (74)

a aFarhan Habib, et al, J.Lightw. Technol., 2012.


Computing Initial primary path and backup path

1 for each r2 Traditional Dijkstra’s shortest path to compute primary path3 Calculate threat disaster zones4 remove all the links in the threat disaster zones5 Traditional Dijkstra’s shortest path to compute backup path

Allocation spectrum usage

1 Construct the conflict graph among the primary path and backup paths2 Coloring heuristic



Disaster Failure: Column Generation (CG)


Spectrum channel H: a set of contiguous spectrum.For example, H for a request with 2 FSs:1, 1, 0, 0, 0, ..., 0, 0,0, 1, 1, 0, 0, ..., 0, 0,0, 0, 1, 1, 0, ..., 0, 0,· · · · · · ,0, 0, 0, 0, 0, ..., 1, 1.A configuration ωr for request r:

1 primary path and backup path2 spectrum channels for primary and backup paths

Variables:

1 qωr ∈ 0, 1: Equals 1 if ω-th configuration for request r is used including theworking path, backup path and assigned channels, and 0 otherwise.

2 eaf ∈ 0, 1: Equals 1 if link a with FS f is used by the working or backup pathsof all the requests, and 0 otherwise.

3 of ∈ 0, 1: Equals 1 if FS f is used, and 0 otherwise.




Step 3: Master problemOjective: Link-FSs Max-FSs

Minimize θ1 ·∑a∈A

∑f∈S

eaf + θ2 ·∑f∈S

of

Constraints:

1.DEBPP

∑ω∈Ωr

qωr = 1, ∀r (75)

∑r∈R

∑ω∈Ωr

(pWωrah + pB

ωrah′ ) · qωr ≤ eaf , ∀a, ∀f ∈ Bh, Bh′

(76)

of ≥ eaf , ∀a, ∀f (77)

of ≥ of+1, ∀f (78)

2.SEBPP

∑ω∈Ωr

qωr = 1, ∀r

(79)∑r∈R

∑ω∈Ωr

pWωrah · qωr ≤ eaf , ∀a, ∀f ∈ Bh (80)

∑r∈R

∑ω∈Ωr

pWωrah · (1− αW

ωrz) + pBωrah′ · α

Wωrz · qωr ≤ eaf ,

∀a, ∀f ∈ Bh, Bh′ , ∀z (81)

of ≥ eaf , ∀a, ∀f(82)

of ≥ of+1, ∀f(83)



Disaster Failure: Column Generation (CG)Step 4: Pricing problem

Dual variables:

1 µr : dual variable for DEBPP and SEBPP2 ζaf ≥ 0 : dual variable for DEBPP3 ρaf ≥ 0 : dual variable for SEBPP4 ηafz ≥ 0 : dual variable for SEBPP

Reduced cost for DEBPP:

cDEBPPωr

= 0− µr −∑a∈A

∑f∈Bh∪Bh′

(pWωrah + pB

ωrah′ ) · ζaf = −µr +∑a∈A

pWωrah · CostWωrah +

∑a∈A

pBωrah′ · CostBωrah′

CostWωrah =∑

f∈Bh

ζaf CostBωrah′ =∑

f∈Bh′

ζaf (84)

Reduced cost for SEBPP:

cSEBPPωr

= 0−(µr −

∑a∈A

∑f∈Bh

pWωrah · ρaf −

∑a∈A

∑f∈Bh∪Bh′

∑z∈Z

pWωrah · (1− αW

ωrz)− pBωrah′ · α

Wωrz

· ηafz

)

= −µr +∑a∈A

pWωrah · CostWωrah +

∑a∈A

pBωrah′ · CostBωrah′ (85)

CostWωrah =∑

f∈Bh

ρaf +∑

f∈Bh∪Bh′

∑z∈Z

(1− αWωrz) · ηafz CostBωrah′ =

∑f∈Bh∪Bh′

∑z∈Z

αWωrz · ηafz (86)




Step 4: Pricing problemConfiguration ωr in DEBPP

1 For each d ∈ D2 For each d′ ∈ D, d′ 6= d3 For each h ∈ Hr4 For each h′ ∈ Hr5 Calculate shortest path with

CostWωrah

6 Calculate threat disaster zones7 remove all the links in the threat

disaster zones8 Calculate shortest path with

CostBωrah′

Configuration ωr in SEBPP

1 For each d ∈ D2 For each d′ ∈ D, d′ 6= d3 For each h ∈ Hr4 For each h′ ∈ Hr5 Calculate K=2 shortest path with

hop count6 Calculate threat disaster zones7 remove all the links in the threat

disaster zones8 Calculate shortest path with

CostBωrah′


Convert all variables to integer

Solve the master problem again


Additional Slides for DSA Problem

Ordered Distance Spectrum Assignment (ODSA)

Giving a G(V,E, vwi , dvivj) with a vertex order Oi,

the beginning FS indices for each vertex are sorted in the increaseing order inthe final solution i.e.,

Oi = (vi1 , vi2 , ..., vin) : vbij ≥ vb

ik , ∀j > k (87)

where, vbij and vb

ik is the beginning FS’ index of the contiguous FS’ setassigned to vij and vik respectively.



Input : A DSA graph G(V,E, vwi , dvivj), and a vertex order Oi = (vi1 , vi2 , ..., vin)

Output: An assignment strategy for the beginning index Seq = vbij : 1 ≤ j ≤ n and the maximum

FS’ indexvb

i1 ← 1;va

i1 ← vwi1 ;

Seq← vbi1 ;

j← 2;while j ≤ n do

s1 ← max∀k<j,vik vij∈E

vaik + dvik vij

+ 1;

s2 ← vbij−1

;vb

ij ← maxs1, s2;va

ij ← vbij + vw

ij − 1;Seq← Seq ∪ vb

ij;j← j + 1;

endreturn Seq and max

1≤j≤n(va

ij)

Algorithm 1: Procedure O-L



The time complexity of Procedure O-L is O(|E|)

Theorem 7

The Procedure O-L results in an optimal spectrum assignment strategy for theODSA.

Corollary 8A DSA problem can be optimally solved by the Procedure O-L under certainvertex order.

Now, DSA has been transform to a Permutation-based Optimization Problem(POP).


optimization of routing and resource allocation in ... · outline outline 1 curriculum vitae 2...

Documents