Alfio Lomb Alfio Lombard ardo o Anto tonio Manzalini Vincenzo - - PowerPoint PPT Presentation

SLIDE 1

Alfio Alfio Lomb Lombard ardo

• Vincenzo

Vincenzo Ricco iccobene ene Gio Giova vanni nni Sc Sche hemb mbra

DIEEI – University of Catania

Anto tonio Manzalini

Telecom Italia Strategy Future Centre

SLIDE 2

Pa

Paper er moti tivati tion an and d ref referen erence ce s scen cenario ario

Netw

twork analyti tical fr framewo mework

• Model of an NFV node
• Model of a non-NFV node
• Model of the whole network
• Derivation of performance parameters

Case

Case stu tudy

Conclusions

Conclusions and futu ture work

SLIDE 3

Service Providers and Netw

twork Operato tors ne need:

• Flexibility in network deployment and

management

• A flexible and optimal provisioning of network

functions and services could reduce equipment costs and allow to postpone network investments

• New network functionalities, services and policies

to increase dynamicity of the market

• Reducing OPEX and CAPEX

SLIDE 4

SDN

DN: Softw tware De Defined Netw tworks

• Decoupling the software control plane from the

hardware data plane (packets forwarding), and moving its logic to centralized controllers

NFV: Netw

twork Functi tion Virtu tualizati tion

• Virtualization of some network functions that can

run on standard HW, and that can be moved and instantiated in various locations of the network

SLIDE 5

Current Current approach approach

≡

SLIDE 6

NFV NFV approach approach

General purpose server

Virtual machines

≡

Data center

SLIDE 7

NFV NFV approach approach

General purpose server

Virtual machines

≡

Data center

SLIDE 8

An

An “NFV “NFV no node” ” is is characte terized by by: :

• A standard hardware architecture (x86 commodity

hardware)

• A virtualization capable software architecture
• A set of Virtual Machines (VMs) that run Network

Functions (e.g. Routers, Firewalls, Load Balancer, ...)

VM 1

Software Layer Hardware Layer

VM 2 VM 3 Virtual machines

SLIDE 9

Analy

Analysis sis of th the impact t of th the Netw twork Functi tion allocati tion

An

An analyti tical fr framewo mework fo for 
 perf perform

• rman

ance ce evaluati tion of

• f 


th the netw twork

SLIDE 10

• Network topology
• Network Function

allocation

• Traffic characterization

Routing Protocol Analytical Model

• Performance

parameters

E2e path for each flow

SLIDE 11

Let

t us consider th the netw twork represente ted by a directe ted graph G(V, E), whe where:

• V is a set of vertices
• E is a set of links among them

Let

t F be th the set t of functi tions deployed over th the netw twork

SLIDE 12

User tr

traffic is represente ted by a set t S of

• f f

flow lows, , each characte terized by th the following ite tems:

• ∈ V is the vertex that represents the source of

the flow s

• ∈ V is the vertex that represents the destination
• f the flow s
• fs is the mean bit rate characterizing the flow s
• funcs is the set of functions required by the flow s

s

σ

s

δ

SLIDE 13

) ( 1 , OUT i

Q

) ( 2 , OUT i

Q

) ( , OUT h i

Q

) ( ,

) (

OUT L i

OUT i

Q

) ( 1 , OUT i

Λ

) ( 2 , OUT i

Λ

) ( , OUT h i

Λ

) ( ,

) (

OUT L i

OUT i

Λ

non-NFV node NFV node

) ( 1 , OUT i

Q

) ( 1 , F i

Q

) ( 2 , F i

Q

) ( , F j i

Q

) ( ,

) (

F L i

F i

Q

) ( 2 , OUT i

Q

) ( , OUT h i

Q

) ( ,

) (

OUT L i

OUT i

Q

CPU

) ( , F j i

Λ

) ( ,

) (

F L i

F i

Λ

) ( 2 , F i

Λ

) ( 1 , F i

Λ

) ( 1 , OUT i

Λ

) ( 2 , OUT i

Λ

) ( , OUT h i

Λ

) ( ,

) (

OUT L i

OUT i

Λ

SLIDE 14

NFV node

An NFV node can be modeled as a set of queues, that belong to two categories:

• Functi

tions Queue Queue

They manage the access to the functions Their service rate depends

• n the CPU processing

s p e e d t o p r o c e s s t h e relative function

• Outp

tput t qu queu eues es

They manage the packet transmission on the output links Their service rate depends

• n the output bitrate

) ( , F j i

Q

) ( , OUT h i

Q

) ( 1 , OUT i

Q

) ( 1 , F i

Q

) ( 2 , F i

Q

) ( , F j i

Q

) ( ,

) (

F L i

F i

Q

) ( 2 , OUT i

Q

) ( , OUT h i

Q

) ( ,

) (

OUT L i

OUT i

Q

CPU

) ( , F j i

Λ

) ( ,

) (

F L i

F i

Λ

) ( 2 , F i

Λ

) ( 1 , F i

Λ

) ( 1 , OUT i

Λ

) ( 2 , OUT i

Λ

) ( , OUT h i

Λ

) ( ,

) (

OUT L i

OUT i

Λ

SLIDE 15

NFV node

k k F j i

j i

λ

∑

Φ ∈ ∀

= Λ

,

) ( ,

) ( , ) ( , CPU i j i F j i

C p ⋅ = µ

k k OUT h i

h i

λ

∑

Ψ ∈ ∀

= Λ

,

) ( ,

) ( , ) ( , NIC h i OUT h i

C = µ

Arrival Rate Service Rate

Functi tion Queues Queues Outp tput t Queues Queues

Arrival Rate Service Rate

) ( 1 , OUT i

Q

) ( 1 , F i

Q

) ( 2 , F i

Q

) ( , F j i

Q

) ( ,

) (

F L i

F i

Q

) ( 2 , OUT i

Q

) ( , OUT h i

Q

) ( ,

) (

OUT L i

OUT i

Q

CPU

) ( , F j i

Λ

) ( ,

) (

F L i

F i

Λ

) ( 2 , F i

Λ

) ( 1 , F i

Λ

) ( 1 , OUT i

Λ

) ( 2 , OUT i

Λ

) ( , OUT h i

Λ

) ( ,

) (

OUT L i

OUT i

Λ

SLIDE 16

NFV node

k k F j i

j i

λ

∑

Φ ∈ ∀

= Λ

,

) ( ,

) ( , ) ( , CPU i j i F j i

C p ⋅ = µ

k k OUT h i

h i

λ

∑

Ψ ∈ ∀

= Λ

,

) ( ,

) ( , ) ( , NIC h i OUT h i

C = µ

Arrival Rate Service Rate

Outp tput t Queues Queues

Arrival Rate Service Rate

Functi tion Queues Queues

: set of flows routed through the node i and requiring the function j

j i,

Φ

) ( 1 , OUT i

Q

) ( 1 , F i

Q

) ( 2 , F i

Q

) ( , F j i

Q

) ( ,

) (

F L i

F i

Q

) ( 2 , OUT i

Q

) ( , OUT h i

Q

) ( ,

) (

OUT L i

OUT i

Q

CPU

) ( , F j i

Λ

) ( ,

) (

F L i

F i

Λ

) ( 2 , F i

Λ

) ( 1 , F i

Λ

) ( 1 , OUT i

Λ

) ( 2 , OUT i

Λ

) ( , OUT h i

Λ

) ( ,

) (

OUT L i

OUT i

Λ

SLIDE 17

NFV node

k k F j i

j i

λ

∑

Φ ∈ ∀

= Λ

,

) ( ,

) ( , ) ( , CPU i j i F j i

C p ⋅ = µ

k k OUT h i

h i

λ

∑

Ψ ∈ ∀

= Λ

,

) ( ,

) ( , ) ( , NIC h i OUT h i

C = µ

Arrival Rate Service Rate

Functi tion Queues Queues Outp tput t Queues Queues

Arrival Rate Service Rate

: the CPU quota of i-th node assigned to VM (function) j

j i

p ,

) ( 1 , OUT i

Q

) ( 1 , F i

Q

) ( 2 , F i

Q

) ( , F j i

Q

) ( ,

) (

F L i

F i

Q

) ( 2 , OUT i

Q

) ( , OUT h i

Q

) ( ,

) (

OUT L i

OUT i

Q

CPU

) ( , F j i

Λ

) ( ,

) (

F L i

F i

Λ

) ( 2 , F i

Λ

) ( 1 , F i

Λ

) ( 1 , OUT i

Λ

) ( 2 , OUT i

Λ

) ( , OUT h i

Λ

) ( ,

) (

OUT L i

OUT i

Λ

: the mean packet processing rate

• f the processor in the i-th NFV node

) (CPU i

C

SLIDE 18

NFV node

k k F j i

j i

λ

∑

Φ ∈ ∀

= Λ

,

) ( ,

) ( , ) ( , CPU i j i F j i

C p ⋅ = µ

k k OUT h i

h i

λ

∑

Ψ ∈ ∀

= Λ

,

) ( ,

) ( , ) ( , NIC h i OUT h i

C = µ

Arrival Rate Service Rate

Functi tion Queues Queues Outp tput t Queues Queues

Arrival Rate Service Rate

) ( 1 , OUT i

Q

) ( 1 , F i

Q

) ( 2 , F i

Q

) ( , F j i

Q

) ( ,

) (

F L i

F i

Q

) ( 2 , OUT i

Q

) ( , OUT h i

Q

) ( ,

) (

OUT L i

OUT i

Q

CPU

) ( , F j i

Λ

) ( ,

) (

F L i

F i

Λ

) ( 2 , F i

Λ

) ( 1 , F i

Λ

) ( 1 , OUT i

Λ

) ( 2 , OUT i

Λ

) ( , OUT h i

Λ

) ( ,

) (

OUT L i

OUT i

Λ

: the set of flows crossing the node i and leaving it through the NIC h

h i,

Ψ : the transmission rate of the h-th

• utput link of the i-th NFV node

) ( , NIC h i

C

SLIDE 19

) ( 1 , OUT i

Q

) ( 2 , OUT i

Q

) ( , OUT h i

Q

) ( ,

) (

OUT L i

OUT i

Q

) ( 1 , OUT i

Λ

) ( 2 , OUT i

Λ

) ( , OUT h i

Λ

) ( ,

) (

OUT L i

OUT i

Λ

A non-NFV node can be modeled as a set of output queues, one for each output link

non-NFV node

k k OUT h i

h i

λ

∑

Ψ ∈ ∀

= Λ

,

) ( ,

) ( , ) ( , NIC h i OUT h i

C = µ

Outp tput t Queues Queues

Arrival Rate Service Rate

SLIDE 20

where is equal to:

• The whole netw

twork can be modeled as a netw twork of queues

Model definiti

tion: an N-dimensional conti tinuous- ti time Markov chain whose sta tate te is defined as fo follo llows: ws: ( )

) ( , ), ( ) (

1 ) (

t S t S t S

N

… =

Σ

( )

) ( , ), ( ), ( , ), ( ) (

) ( , ) ( 1 , ) ( , ) ( 1 ,

) ( ) (

t S t S t S t S t S

OUT L i OUT i F L i F i i

OUT i F i

… … =

( )

) ( , ), ( ) (

) ( , ) ( 1 ,

) (

t S t S t S

OUT L i OUT i i

OUT i

… =

) (t S i

(NFV Node) (non-NFV Node)

SLIDE 21

Assumpti

tions:

• Exponentially-distributed interarrival times
• Exponentially-distributed service times in both NF and OUT

queues

• the routing algorithm is able to avoid closed loops

hypotheses of the Jackson theorem the equilibrium probability distribution of the network has a product-form solution:

[ ]

N N

t π π π π π ⋅ ⋅ = =

Σ

… …

1 1 ) (

, , ) (

( ) ( )

) ( , ) ( 1 , ) ( , ) ( 1 ,

) ( ) (

OUT L i OUT i F L i F i i

OUT i F i

π π π π π ⋅ ⋅ ⋅ ⋅ ⋅ = … …

( )

) ( , ) ( 1 ,

) (

OUT L i OUT i i

OUT i

π π π ⋅ ⋅ = …

if NFV if non-NFV

SLIDE 22

Let us indicate:

• Utilization coefficient of the

j-th NF queue in the node i

• Utilization coefficient of the

h-th OUT queue in the node i

) ( , ) ( , ) ( , F j i F j i F j i

µ ρ Λ =

) ( , ) ( , ) ( , OUT j i OUT j i OUT h i

µ ρ Λ =

{ } [

] [ ]

k F j i F j i F i t F k i

k t S

) ( , ) ( , ) ( ) ( ,

1 ) ( Prob lim ρ ρ π ⋅ − = = ≡

→∞

{ } [

] [ ]

k OUT j i OUT j i OUT i t OUT k i

k t S

) ( , ) ( , ) ( ) ( ,

1 ) ( Prob lim ρ ρ π ⋅ − = = ≡

→∞

SLIDE 23

Probability that the VM j

in the node i is not using the CPU quota assigned to it:

Mean number of packets

in the queueing systems

Mean sojourn time in

the queueing system

) ( , ) ( , ) ( , ) ( ,

1 1

F j i F j i F j i F j i

P µ ρ Λ − = − =

) ( , F j i

Q

) ( , ) ( , ) ( ,

1

F j i F j i F j i

ρ ρ ν − =

) ( , ) ( , ) ( , F j i F j i F j i

W Λ = ν

) ( , OUT j i

Q and

) ( , F j i

Q

) ( , OUT j i

Q and

) ( , ) ( , ) ( ,

1

OUT j i OUT j i OUT j i

ρ ρ ν − =

) ( , ) ( , ) ( , OUT j i OUT j i OUT j i

W Λ = ν

SLIDE 24

End-to-end delay for each flow

⎥ ⎥ ⎦ ⎤ ⎢ ⎢ ⎣ ⎡ ⋅ + ⋅ =

∑ ∑ ∑

= = =

) ( ) (

) ( , ) ( , 1 ) ( , ) ( , 1 1 ) 2 (

) ( ) (

k I W k I W W

OUT h i OUT h i L h F j i F j i L j N i e e k

OUT i F i

⎩ ⎨ ⎧ =

• therwise

node in the function the uses flow the if 1 ) (

) ( ,

i j k k I

F j i

where:

⎩ ⎨ ⎧ =

• therwise

NIC he through t node the leaves flow the if 1 ) (

) ( ,

h i k k I

OUT H i

SLIDE 25

CASE E STUDY DY

SLIDE 26

TARGET

ET

• finding the end-to-end path for each flow

REQ

EQUIREM EMEN ENTS

• the first and the last nodes for each flow are the

ingress and the egress nodes specified for that flow

• the path for each flow has to cross nodes

implementing the functions requested by that flow

ROUTING ALGORITHM

SLIDE 27

C:

: reference link capacity ty

• defined as the bandwidth of the link with the

highest capacity in the network

All th

the link capaciti ties are normalized with th respect t to to C

SOME E NOTATION

ROUTING ALGORITHM

SLIDE 28

: Boolean characte

terizati tion of th the netw twork functi tion distr tributi tion

t v

I

⎩ ⎨ ⎧ =

• therwise

Function Network the implements node the if 1 t v I

t v

SOME E NOTATION

ROUTING ALGORITHM

SLIDE 29

: Boolean characte

terizati tion of th the netw twork functi tion distr tributi tion

: Boolean characte

terizati tion of th the functi tion requirements ts for netw twork tr traffic

t v

I

t s

a

⎩ ⎨ ⎧ =

• therwise

Function Network the implements node the if 1 t v I

t v

SOME E NOTATION

⎩ ⎨ ⎧ =

• therwise

Function Network the requires flow the if 1 t s

t s

α

ROUTING ALGORITHM

SLIDE 30

ROUTING ALGORITHM

Routi

ting algorith thm outp tput t

⎩ ⎨ ⎧ → =

• therwise

link

• n the

allocated is flow the if 1 w v s y

s vw

Routi

ting algorith thm ta target t

∑∑∑

= ∈ ∈

⋅

S s V v V w s s vw f

y

1 Sum of loads of all the links in the network

SLIDE 31

Possible values of the variables

ROUTING ALGORITHM

Sub

Subject to to: :

, V w v M f y

vw S s s s vw

∈ ∀ ≤ ⋅

∑

∈

S s w v V v y y

s V w s wv V w s vw

∈ ∀ ≠ ∈ ∀ =∑

∑

∈ ∈

} , , { and

s

δ σ

It ensures that no link carries more traffic flow than its capacity

1 ≤ ≤

s vw

y

Flow-conservation constraint: it ensures that no flow is lost or created except for at the ingress and the destination nodes

SLIDE 32

They ensure that the flow s enters the network through only one node, and leaves the network from only one node

ROUTING ALGORITHM

Sub

Subject to to: :

S s y

V w s w

s

∈ ∀ =

∑

∈

1

σ

S s y

V v s v s

∈ ∀ =

∑

∈

1

δ

F t S s I a y

V v V w t w t s s vw

∈ ∀ ∈ ∀ ≥ ⋅ ⋅

∑∑

∈ ∈

, 1

It ensures that each traffic flow crosses the nodes which implement the required functions

SLIDE 33

Capacity = C*10-1 3

4

5 6 1 2 Reference Capacity C Capacity = C*10-2 Core
 Network Access and 
 aggregation 
 network Data Center A Data Center B A1 A2

SLIDE 34

3

4

5 6 1 2 Data Center A Data Center B A D F1 F1 F2 F2 F3 F3 F4 F4 F5 F5 F6 F6 F7 F7 F8 F8 F1 F1 F2 F2 F3 F3 F7 F7 F8 F8 F4 F4 F5 F5 F6 F6

A A B B A A B B A A B B A A B B C C D D

C B Fi Fi

Ingress node for the flow i Ingress node for the flow i

Ei Eight flo flows ws

kpps 1200

) ( 1

=

CPU

C

[ ]kpps

4500 , 2000

) ( 1

∈

CPU

C kpps 10

5 ) ( 5

=

CPU

C kpps 10

5 ) ( 6

=

CPU

C

kpps 5 . 99 =

s

f

F1, F2, F3, F4, F5 F1, F2, F3, F4, F5

kpps 67 , 132 =

s

f

F6 F6

kpps 33 , 666 =

s

f

F7, F8 F7, F8

A1 A2

SLIDE 35

3

4

5 6 1 2 Data Center A Data Center B A D F1 F1 F2 F2 F3 F3 F4 F4 F5 F5 F6 F6 F7 F7 F8 F8 F1 F1 F2 F2 F3 F3 F7 F7 F8 F8 F4 F4 F5 F5 F6 F6

A A B B A A B B A A B B A A B B C C D D

C B All the functions are allocated on the aggregation nodes. This case stresses the aggregation nodes processing and does not stress the network. A1 A2

SLIDE 36

2 2.5 3 3.5 4 4.5 x 10

6

0.5 1 1.5 2 2.5 3 3.5 4 x 10

• 8

Mean packet processing rate of Node 2 [pkt/s] Mean end-to-end per flow delay [s] Flow 1 Flow 2 Flow 3 Flow 4 Flow 5 Flow 6 Flow 7 Flow 8

Only F7 and F8 flows are affected by the Node 2 processing rate because they require functions C and D (that reside on the node 2)

F1 F2 F3 F4 F5 F6 F7 F8

SLIDE 37

3

4

5 6 1 2 Data Center A Data Center B A B D F1 F1 F2 F2 F3 F3 F4 F4 F5 F5 F6 F6 F7 F7 F8 F8 F1 F1 F2 F2 F3 F3 F7 F7 F8 F8 F4 F4 F5 F5 F6 F6

A A B B A A B B A A B B A A B B C C D D

C The functions are partially allocated

• n the aggregation

nodes and partially

• n the Data

Centers. This case stresses both the aggregation nodes processing capacity and the network. A1 A2

SLIDE 38

Only F7 is (lightly) influenced by the Node 2 processing rate because it requires the function C F4, F6 and F8 suffer a higher delay because they have to reach destination in the A2 cloud and, at the same time, need to be processed by the aggregation nodes.

2 2.5 3 3.5 4 4.5 x 10

6

1 2 3 4 5 6 7 x 10

• 8

Mean packet processing rate of Node 2 [pkt/s] Mean end-to-end per flow delay [s] Flow 1 Flow 2 Flow 3 Flow 4 Flow 5 Flow 6 Flow 7 Flow 8 F1 F2 F3 F4 F5 F6 F7 F8

SLIDE 39

3

4

5 6 1 2 Data Center A Data Center B A B D F1 F1 F2 F2 F3 F3 F4 F4 F5 F5 F6 F6 F7 F7 F8 F8 F1 F1 F2 F2 F3 F3 F7 F7 F8 F8 F4 F4 F5 F5 F6 F6

A A B B A A B B A A B B A A B B C C D D

C In this case we have stressed:

• Network portion

between aggregation nodes and core network

• Processing

capacity of Node 2 This case stresses both the aggregation nodes processing capacity and the network. A1 A2

SLIDE 40

Now [F2, F3, F5, F6] and [F7, F8] flows are influenced by the Node 2 processing rate because:

• [F2, F3, F5, F6]

require function B

• [F7, F8] require

functions C and D [F7, F8] suffer the same delay

2 2.5 3 3.5 4 4.5 x 10

6

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 x 10

• 8

Mean packet processing rate of Node 2 [pkt/s] Mean end-to-end per flow delay [s] Flow 1 Flow 2 Flow 3 Flow 4 Flow 5 Flow 6 Flow 7 Flow 8 F1 F2 F3 F5 F4 F6 F7 F8

SLIDE 41

3

4

5 6 1 2 Data Center A Data Center B A D F1 F1 F2 F2 F3 F3 F4 F4 F5 F5 F6 F6 F7 F7 F8 F8 F1 F1 F2 F2 F3 F3 F7 F7 F8 F8 F4 F4 F5 F5 F6 F6

A A B B A A B B A A B B A A B B C C D D

C B In this case we reduced the processing load of node 2, more stressing the network A1 A2

SLIDE 42

2 2.5 3 3.5 4 4.5 x 10

6

2 4 6 8 10 12 14 16 x 10

• 9

Mean packet processing rate of Node 2 [pkt/s] Minimum end to end per flow delay [s] Flow 1 Flow 2 Flow 3 Flow 4 Flow 5 Flow 6 Flow 7 Flow 8

Now all the flows are less influenced by Node 2 processing capacity variation because Node 2 is less overloaded.

SLIDE 43

1 2 3 4 1 2 3 4 5 6 7 x 10

• 8

Function allocation case Mean end-to-end per flow delay [s] Flow 1 Flow 2 Flow 3 Flow 4 Flow 5 Flow 6 Flow 7 Flow 8

kpps 4290

) ( 2

=

CPU

C

Let us use the model to find the best function allocation Cases 3 and 4 are the best cases. The case 2 it the most unfair and present the worst case in terms of mean end-to-end delay

SLIDE 44

A

A te telecommunicati tions netw twork with th N NFV FV capabiliti ties ha has been been con considered sidered

An analyti

tical framework of th the netw twork has been been def defin ined ed

The model applicability

ty has been demonstr trate ted in a case stu tudy

SLIDE 45

Accurate

te mo model of

• f a s

a sin ingle N le NFV FV no node

• Markov model of all function queues capturing their

correlated behaviors

De

Definiti tion and evaluati tion of routi ting algorith thms specific for NFV netw tworks

• A centralized constrained routing algorithm could
• ptimize the traffic allocation with respect to the

function allocation

Functi

tion allocati tion policies policies

SLIDE 46

Functi

tion Migrati tion te techniques

Analyti

tical mo model of th the tr transient period period du durin ring functi tion migrati tion

De

Definiti tion of

• f g

green reen te techniques fo for N NFV FV netw tworks

• Global approach (e.g. path aggregation and specific

function allocation)

• Local approach (e.g. frequency scaling in node

processors)

SLIDE 47

QUES ESTIONS?