Scaling of QoS and Business Models

1
Scaling of QoS and Business Models

• Andy Reid
• Chief Network Services Architect, Group CTO

TRIS-TISPANWorkshop on NGN Interconnection18th
January 2006
2
Status of QoS Management in the Industry

• Elastic traffic controlled by TCP
• Settled solution
• Proven robust and scalable
• No settled solution for non-elastic traffic, eg
  voice and video
• Diff-serv with QoS based traffic engineering
  advocated by some
• Bandwidth management advocated by some
• Variety of new innovative solutions
• Unresolved over many years now
• No settled view of network builder services

3
QoS Requirements of Different Services

• Elastic services
• Browsing, file transfer
• Can tolerate large changes in allocated bit rate
• Total traffic demand can be effectively
  controlled by degrading bit rate which
  discourages future browsing/transfers
• Non-elastic services
• Voice and video
• Minimum bit rate below which transfer becomes
  ineffective
• Total traffic can be effectively controlled by
  denying new flow requests
• Network builder services
• Service needs to look like physical infrastructure

Network Stability is Paramount
4
Some Starting Observations

• Congestion points are normally at the network
  edge only
• Network core does not congest under normal
  traffic loadings
• Core links tend to be either lightly/moderately
  loaded, or in heavy congestion
• Links are very rarely on the knee of the
  congestion curve
• Need to consider abnormal traffic loading
• Network Failures
• Normal conditions, urgent ie real time traffic
• Failure conditions, important ie mission
  critical traffic

5
Sources of Variance and Correlation

• Primary sources
• Adverse traffic events, eg festivals, natural
  disasters, occasional sporting events, etc
• Rolling out of new services
• Secondary sources
• Different pricing strategies between competitive
  services, eg geographically averaged alongside
  geographically de-averaged
• Changes in pricing structures and pricing
  differentials
• Scaling of variance
• ?ij2 ? E(Tij) - statistical independence
• ?ij ? E(Tij) - self similar
• ?ij ? E(?jTij) - adverse shifts in traffic

6
Scaling of Traffic Demand

• General scaling of traffic matrix with number of
  end points
• Metcalfes law
• Tij ? n? or T(n) ? n2? , ? close
  to zero
• Distribution of size of source end points
• Zipfs law
• Ti ? i? , ? close to -1 when i is in
  rank order
• Distribution of size of destination end points
  from source end points
• Zipfs law
• Tij ? n?(ij)? , -1 lt ? lt 0 depending on
  correlated targeting of rollout
• Conclusion
• T(n) ? n? 2 2? or T(n) ? n?ln(n) when ? -1

Total traffic will grow between ln(n) and n2
depending on correlated targeting of rollout
7
Probability Distributions

• Normal distribution
• Pro central limit theorem for traffic
  aggregation
• Con not positive only, assumes no
  correlation/self similarity
• Log normal distribution
• Pro basic model for geometric growth
• Con no real basis from traffic aggregation
• Power law
• Pro associated with self similar traffic
  aggregation (eg Pareto, Zipf)
• Con requires truncations to be meaningful
• Gamma distribution
• Pro similar to power law but mathematically
  tractable
• Con limited basis from traffic aggregation

8
Probability Distributions
Normal distribution
Gamma distribution
Normal distribution
p(x/µ)
Log normal distribution
Gamma distribution
p(x/µ)
x/µ
Log normal distribution
s ¼µ is illustrated
x/µ
9
Probability Distributions
Log normal distribution
Power law distribution
Gamma distribution
p(x/µ)
Gamma distribution
Power law distribution
Normal distribution
p(x/µ)
Log normal distribution
x/µ
Normal distribution
s 4µ is illustrated
x/µ
10
Mapping Traffic to Network Structure
Concentration of many small traffic matrix
elements on this link
Few large elements of traffic matrix with short
routing
Many small elements of traffic matrix with long
routing
Moderate number of moderately sized elements of
traffic matrix with medium routing
11
Use of Diffserv for Non-Elastic Traffic
-Conclusions

• Traffic matrix elements with high variance
  conspire to concentrate on thinner, expensive
  network links
• Network cost optimisation will but larger traffic
  matrix elements on short routing and cheaper
  links
• Large numbers of very small traffic matrix
  elements on thinner, expensive links
• Probability of non-elastic traffic severely
  overloading with itself scales very poorly with
  the topological size of a network
• Effect of abnormal traffic loads, not expected
  traffic loads

12
Observations and Consequences

• Shorten diff-serv network topology
• Continuing role for transmission layer
  technologies
• Equalise route lengths for all traffic routes
• Aided by death of distance
• Consider use of non-blocking network structures
• Limited scalability
• Avoid diff-serv interconnects where possible (for
  non-elastic traffic)
• Use admission control at inter-operator
  boundaries for non-elastic traffic
• May also be required at domain boundaries within
  an operators network

More information http//www.springerlink.com/ BT
Technology Journal Volume 23 No 2 April 2005
13
Mapping Downstream Retail Applications to
Upstream Infrastructure

• Key retail application markets
• Consumer
• fixed telephony, mobile telephony, broadband
  Internet, broadcast TV, premium live broadcast
  TV, films, music, retailing
• Business and non-profit sector
• integrated business solutions eg office
  automation, b2c systems, b2b system, finance
  systems, production automation systems, sector
  specific applications, etc
• Key upstream infrastructure
• Copper access, cable access, satellite,
  terrestrial broadcast, DVD/video/CD in the high
  street, DVD/video/CD through the mail, cinema
• Retail Packaging price, convenience, and/or
  brand
• Cable/DSL triple play
• Tesco broadband and Tesco mobile
• Virgin mobile
• Amazon video retail

14
Mapping Downstream Retail Applications to
Upstream Infrastructure (cont)

• Model of close vertical integration voluntarily
  breaking down
• Mobile virtual network operators (MVNOs)
• iPOD
• Skype telephony
• Video rental through amazon.com
• Cross industry supply
• Enabling retail packaging
• Practical and evolutionary convergence
• New broadband applications are normally
  entering other industries competitive markets
• VoD must be price competitive with current
  cinema, and video rental pricing
• Music download must be price competitive with
  current CD pricing

15
Dangers of Open Upstream Markets

• Different retail applications have different
  demand side and supply side competitive
  characteristics
• Often very different price structures
• Retail packaging allows recovery of common cost
  against several sources of revenue
• There is no uniform price for the upstream
  resource
• price discrimination at the retail level for use
  of the upstream resource
• Similar to life cycle pricing where early
  adopters pay a premium over laggards
• An open upstream market will normally reduce or
  eliminate downstream price discrimination
• This may make upstream investment uneconomic
• Compare with the economics of public health care

16
Conclusions on Business Models

• Trying to create a clean business model
  separation between services and universal network
  bandwidth has a tension
• Stimulates innovation and convergence
• Stagnates investment incentives in the network
• Need some flexibility in the business model
  assumptions
• Admission control at an interconnection is a
  technical requirement of real time services
• Gives opportunity for flexible business model