nfv hanover zhaozhongliang v2 - wordpress.comoai)! ! full)softwarembased)networkfunctions...

45
Zhongliang Zhao, Andre Gomes, Eryk Schiller, Torsten Braun with related MCN partners Hanover, 27.10.2015 NFV Workshop 2015@Hanover EU FP7 Mobile Cloud Networking Project

Upload: buinhan

Post on 14-Jun-2018

215 views

Category:

Documents


0 download

TRANSCRIPT

Zhongliang  Zhao,  Andre  Gomes,  Eryk Schiller,  Torsten Braunwith  related  MCN  partners

Hanover,  27.10.2015

NFV  Workshop  2015@Hanover

EU  FP7  Mobile  Cloud  Networking  Project

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  2

n Motivationsn Introduction  to  the  MCN  project  (architecture,  services)n Radio  Access  Networks  as  a  Service  (RANaaS)n Information  Centric  Network  as  a  Service  (ICNaaS)n Mobility  Prediction  as  a  Service  (MOBaaS)n Conclusions

Outline

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  3

Motivations

n Large  growth  in  mobile  traffic  challenges  traditional  telecom  network  architecturesn Video,  data  traffics  require  high  bandwidthn Static  architectures  are  inflexible,  costly,  slow  to  innovate

n Today  MNOs  mostly  providen Traditional  connectivity  &  voice  servicesn Infrastructure  sharing,  pre-­sized,  pre-­provisioned,   pre-­customized

n Tomorrow  MNOs adopt  IaaS/PaaS  for  network  functionsn On-­demand,  Elastic,  Pay-­as-­you-­gon Reduces costs (move  from  CAPEX  to  OPEX)n Enables  seamless  integration  of  Mobile  Networks  +  Cloud  Computing  +  Storage

n Solution:  Bring  SDN+NFV  to  Telco  Industry

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  4

n Motivationsn Introduction  to  the  MCN  project  (architecture,  services)n Radio  Access  Networks  as  a  Service  (RANaaS)n Information  Centric  Network  as  a  Service  (ICNaaS)n Mobility  Prediction  as  a  Service  (MOBaaS)n Conclusions

Outline

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  5

Mobile  Cloud  Networking  (MCN)

Moving  Cloud  Computing  beyond  Data  Centers  towards  the  Air  Interface  and  Mobile  End  Users

n High  level  objectivesn Develop  a  novel  “mobile  network”  architecture,  using  proof-­of-­concept  prototypes,  to  lead  the  way  from  current  mobile  networks  to  a  fully  cloudified,  on-­demand,  and  elastic  provisioning  of  novel  mobile  service

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  6

MCN  Architecture

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  7

Service  Lifecycle

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  8

Key  Service  Components

n Service  Manager  (SM)n Provides  external  interface  to  the  usern Manages  the  SO  of  every  tenantn Receives  request  for  new  tenant  service  instances

n Service  Orchestrator  (SO)n Orchestration  of  a  service  instancen Manages  and  abstracts  underlying  resources  and  service  instancen “Runtime  &  Management”  step  of  the  service  lifecyclen Monitors  application  specific  metrics  and  scales

n Cloud  Controller  (CC)n Supports  the  deployment,  provisioning,  and  disposal  of  servicesn Manages  the  lifecycle  of  a  tenant  service  instancen Access  to  support/atomic  service,  configures  atomic  services

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  9

Key  Components  Overview  i

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  10

Key  Components  Overview  ii

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  11

MCN  and  NFV  Mapping  i

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  12

MCN  and  NFV  Mapping  ii

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  13

MCN  NFV  Scope  (11.2012)  &  Applicable  NFV  Use  Cases  (10.2013)

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  14

n Motivationsn Introduction  to  the  MCN  project  (architecture,  services)n Radio  Access  Networks  as  a  Service  (RANaaS)n Information  Centric  Network  as  a  Service  (ICNaaS)n Mobility  Prediction  as  a  Service  (MOBaaS)n Conclusions

Outline

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  15

Radio  Access  Network-­aaS (RANaaS)

n Motivations:  Today’s  RANs  face  problems  ofn Increasing  capacity  needs  vs.  falling  revenues  per  user.n Large  base  stations  deployments  of  high  costs  (CAPEX/OPEX)  and  power  consumption.

n Resources  dimensioned   for  peak  traffic  loads.n Large  fiber  connection  availability   in  urban  areas.

[GSMA,  The  Mobile  Economy  2013]

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  16

Today’s  4G  LTE  Network

n All  components  are  implemented  by  dedicated  hardware  platformn Expensiven Hard  to  configure  and  upgrade

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  17

Base  Station:  RRH  +  BBU

n In  Centralized-­RAN,  base  stations  are  split  into:n Remote  Radio  Head  (RRH),n Base  Band  Unit  (BBU).  

n RRHs  and  BBUs  are  linked  by  optical  cablesn Aims  fast  deployments.

   

 

 

 

   

   

 

BBUs

RRHs

 

 

 

       

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  18

Cloud-­RAN

n Currently,  BBUs  run  on  custom  dedicated  hardware  platforms.  

n RANaaS aims  to  extend  C-­RAN  by:n Taking  advantage  of  general  purpose  platforms.n Integrating  virtualisation  and  the  cloud  paradigms.n Using  shared  computation,  storage  and  networking  resources  of  cloud  infrastructures.

   

 

 

 

   

   

 

BBUs

RRHs

 

 

 

     

 

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  19

A  key  element:  Software  based  eNB

n A  commodity  hardware  computer  with  RAN  software  and  RF  equipment

n OpenAirInterface (OAI)n www.openairinterface.orgn Full  software-­based  network  functionsspanning  all  the  network  layers

n Open  source  from  project  partner

computer RF  card

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  20

OAI  LTE  Virtual  eNB and  EPC

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  21

C-­RAN  in  the  Data  Centre

: Server: Virtual Machine

: BBU

: Active RRH: Inactive RRH

Caption:

n RRH  are  connected  with  the  BBUs  in  the  Data  Centre.n With  few  users,  3  RRH-­BBU  pairs  cover  the  service  arean With  more  users,  more  RRHs  will  be  activated  and  extra  BBUs  will  be  instantiated  on  demand

BBU-Pool (in a Data Centre)

BBU Plane

End-user

Plane

RRH Plane

Switch

 

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  22

RANaaS Prototype

Based  on  a  Heat  orchestration  Template  (HoT),  Heat  orchestrator  uses  OpenStackcontrol  plane  to  build  Heat  stack,  a  composite  environment  (web  services,  HSS,  EPC  and  eNB).

LXC  provides  a  VE  in  OpenStack.  environment.  

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  23

BBU  Profiling  Evaluations

n Physical  Layer  is  the  most  deadline  criticaln BBU  processing  requirements  in  a  virtualized  environment  (processing  time  of  radio  frames  <  3  ms)

n Evaluated  Environments  for  BBU  Profilingn GPP/LXC/KVM/Public-­Private  Cloud

n Downlink(Transmission)  /  Uplink(Reception)

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  24

RANaaS Conclusions

n RANaaS aims  to  integrate  virtualisation  and  cloud  paradigms.  

n C-­RAN  software-­based  base-­stations  (BBUs)  have  critical  processing  requirements.

n The  feasibility  to  run  a  single  BBU  on  a  GPP  platform  was  evaluated  using  OAI  LTE  base  station  emulator:

n A  prototype  of  RANaaS was  presented,  based  on  Heat,  OpenStack  and  Linux  Containers.

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  25

n Motivationsn Introduction  to  the  MCN  project  (architecture,  services)n Radio  Access  Networks  as  a  Service  (RANaaS)n Information  Centric  Network  as  a  Service  (ICNaaS)n Mobility  Prediction  as  a  Service  (MOBaaS)n Conclusions

Outline

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  26

Information  Centric  Network-­aaS(ICNaas):  Follow  Me  Cloudn Motivations• Bandwidth  usage  is  high  and  mobile  users  are  dynamic• Follow-­Me-­Cloud   concept  has  a  deeper  integration  and  may  bring    additional   improvements• More  information  about  access  networks  (e.g.,  user  mobility,  bandwidth  usage,  etc.)

• Greater  granularity  on  decision  making  (up  to  small  groups  of  users)

• Integration  with  ICN/CCN  enables  greater  proximity  of    caching  and  repository  migrations

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  27

Follow-­Me  Cloud  (FMC):  Architecture

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  28

n Key  Conceptsn When  users  move,  content  moves  with  them  to  the  new  locationn Such  principle  may  also  be  applied  to  service  instancesn Monitoring  can  trigger  content  relocationn Proactive  content  relocation  is  the  most  desired  scenarios– Mobility  prediciton  based– Not  always  accurate

n Challengesn How  can  resources  be  virtualized  to  allow  such  relocation?n What’s  the  role  of  caching?n What  to  relocate?n When  should  relocation  be  performed?  Under  which  conditions?

Follow-­Me  Cloud

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  29

n Gather  Information  at  the  FMC  Controllern Mobility  prediction  and  detectionn Content  popularity  over  time  intervals,  cache  hit  ratios,  etcn Content  availability  at  the  networksn Resources  availability  (compute  power,  storage,  etc)

n Decision  Makingn Decide  IF migration  should  occur,  WHERE to  and  WHAT content  should  be  transferred  at  WHEN.

n Goals:  improve  content  access  times,  reduce  bandwidth  usage  and  minimize  migration  costs.

n Perform  Content  Migrationn Use  the  ICN/CCN  protocols  to  perform  content  migrationn Selective  sync  of  repositories  based  on  decisions  from  the  FMC  controller

Follow-­Me  Cloud:  Objectives

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  30

n MADM:  Multiple-­Attribute  Decision-­Makingn Not  tied  to  specific  problemsn Supports  multiple  weighted  attributen Finite  number  of  solutions  for  problems  which  involve  selection/assessment

n Clost  to  ideal  solutions,  as  ideal  are  usually  not  obtained:

n Attributen Content  popularity,  content  sizen User  mobility  prediction  information/accuracyn Free  storage  space  at  possible  destinationsn Relative  size  of  mobility  group

Decision-­Making  Methods

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  31

ICN/FMC  Conclusions

• Follow-­Me  Cloud  brings  benefits  to  very  dynamic  mobile  networks

• MCDM  methods  can  be  applied  to  the  problem  if  deciding  where  to  relocate  content.

• It  is  possible  to  obtain  a  very  good  decision  without  much  information  and  complexity.

• Cost  of  content  migration  is  important  for  the  networks,  but  tradeoff  with  benefit  must  be  carefully  analyzed.

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  32

n Motivationsn Introduction  to  the  MCN  project  (architecture,  services)n Radio  Access  Networks  as  a  Service  (RANaaS)n Information  Centric  Network  as  a  Service  (ICNaaS)n Mobility  Prediction  as  a  Service  (MOBaaS)n Conclusions

Outline

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  33

Mobility  Prediction-­aaS (MOBaas)

n Motivations• Traffic  redirection   in  EPCaaS to  optimize  bandwidth  usage• Content  relocation   in  ICNaaS to  optimize  content  migration

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  34

Implementation  Architecture

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  35

Run-­time  Prediction  Procedure

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  36

Mobility  Prediction  Algorithm  I  

• Dynamic  Bayesian  Networks-­based  mobility  prediction

• Conditional   distribution  of  the  next  cell  of  a  user  comprises  location and  time dependent   distributions

• current  cell  (   )• movement  time  (          )• movement  date  (          )• n  is  the  time  step  (in  minutes)  of  a  day  

CiTnDl

P( c(n+1) = ci,| c(n) = ci,T (n) = Tn,D = Dl ) =

β ×P( c(n+1) = ci,| c(n) = ci )

location dependency

+ (1−β)×P( c(n+1) = ci,|T (n) = Tn,D = Dl )

time dependency

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  37

Mobility  Prediction  Algorithm  II

• Location  dependency   is  a  first-­order  Markov  Chain• Encodes  the  transition  probability  among  locations  (cells)• Calculate  the  probability  of  moving  from                                    to                                            by  counting  the  number  of  transitions  among  cells  from  the  trace  file

• Weekday/weekend• Same  probability   of  moving  from  one  cell  to  two  neighbors

• Integrate time  stamp  (the  time  when  a  user  is  in  the  cell)  and  CellID  to  drive  the  M.C.  states.  State  includes  (Time,  CellID)

C(n) =Ci C(n+1) =Ci,

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  38

Mobility  Prediction  Algorithm  III  

• Modified  DBN  model  for  mobility  prediction  algorithm

• Updated  conditional  distribution  of  the  next  cell

• Second  part  is  day-­dependency,  which  expresses  the  transition  probability   to  next  cell  given  a  specific  day• Derived  by  averaging  the  number  of  transitions  to  a  specific  cell• Weekday/weekend

P( c(n+1) = ci,| c(n) = ci,T (n) = Tn,D = Dl ) =

β ×P(c(n+1) = ci,| c(n) = ci, T (n) = Tn )+ (1−β)×P(c(n+1) = ci, |D = Dl )

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  39

Algorithm  Evaluation

• Nokia  Data  Traces• 100  users  over  2-­6  months,  Switzerland• 30%  for  testing  and  evaluate  the  accuracy• 70%  for  learning  to  derive  the  M.C.  states  and  calculate  the  M.C.  state  transition  probability  matric

• e.g.,  a  trace  of  2.5  months  includes  10  Monday,  we  use  the  first  7  for  learning,  the  last  3  for  testing

• Parameters  and  Metrics• Randomly  select  50  states  out  of  the  M.C.  states  derived  for  each  particular  day  of  a  week  (MON/TUE/…/SUN)

• Prediction  in  next  20  minutes• Mean  absolute  error  (MAE)  in  which                            define  the  predicted  and  actual  probability

MAE = 1n

| Pi ,L − Pi ,T |, n = 50i=1

n

∑Pi ,L ,Pi ,T

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  40

Evaluation  Results:Overall  Accuracy  of  100  Users

• Prediction  accuracy  =  (1  -­ MAE)  x  100

34%User  ID:  6026

73%User  ID:  5960

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  41

Evaluation  Results:Quality  of  Movement  Traces

User  ID:6026

User  ID:5960

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  42

Evaluation  Results:Accuracy  of  Specific  User

User  ID:  6026User  ID:  5960

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  43

MOBaaS Conclusions

• Trace  quality  has  significant  impact  on  prediction  accuracy

• Prediction  accuracy  result  is  better  than  NOKIA  paper  (DBN-­based  model  best  accuracy  is  around  52%)• Improved  DBN  model,  which  integrates  time-­step  with  Cell-­ID  to  drive  a  new  M.C.  states

• Weekday  is  easier  to  be  predicted  than  weekend• Because  it’s  weekday?

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  44

Final  Remarks

• MCN  =  Mobile  Networks  +  Cloud  Computing• It  brings  the  concepts  of  on-­demand,  elastic,  and  pay-­as-­you-­go  into  Telco  industry.

• RANaaS aims  to  integrate  virtualization  and  cloud  paradigms  to  provide  a  Cloud-­RAN,  which  can  reduces  cost  and  provides  service  flexibly.

• ICNaaS/FMC  facilitates  the  migration  of  content  caches  located  at  the  edge  of  mobile  network.

• MOBaaS provides  user  movement  prediction  on  demand,  which  can  be  used  to  optimize  bandwidth  usage  and  content  relocation.

© 2012-­2015    MCN.  All  rights  reserved.  /  Page  45

Thanks  For  Your  Attention!http://cds.unibe.ch

Project  duration:  1-­11-­2012  /  31-­10-­2015  (extended  to  31-­4-­2016)  Project  budget:  15.7  M-­€ /  10.5  M-­€ funding  Number  of  person/months:  1408  person  monthsDetails:  https://www.mobile-­cloud-­networking.eu