Transcript TE Solutions

Seminar
Introduction to Traffic Engineering
October 2009
Ernst Nordström
[email protected]
1
Traffic levels
2
Traffic characterization
• Source traffic parameters
–
–
–
–
Peak packet rate
Mean packet rate
Maximum burst size
Minimum packet rate
• Call class characterization
–
–
–
–
–
Orgin-destination node pair
Inter-arrival time distribution
Holding time distribution
Source traffic parameters
Charging rule
3
Definition of traffic processes
• Point process:
Tn 
n 0

• Count process: N( t )t 0
N(t )  max{n : Tn  t }
• Inter-arrival time process:
A n 
n 1
A n  Tn  Tn 1
4
Inter-arrival time distributions
• An inter-arrival time distribution P(An≤t) is
lightly tailed if its variance is finite
• An inter-arrival time distribution P(An≤t) is
heavy tailed if its variance is infinite
5
Traffic models
•
•
•
•
Renewal traffic models
Markov-based traffic models
Self-similar traffic models
Autoregressive models
6
Markov-based traffic models
• A.A. Markov and A. Kolmogorov pioneered
the theory of Markov processes
• Markov property: the current state
summarizes all relevant information about
past states
• Non-zero autocorrelations in {An} allow for
modeling of traffic burstiness
7
Markov-modulated
Fluid process
• Views traffic as a stream of fluid, characterized
by a flow rate (e.g. bits per second)
• Appropriate when the individual traffic units are
numerous relative to the chosen time scale
• A continuous-time Markov chain modulates
traffic arrival (fluid) rate in states 1, 2, .., m of the
state space
8
FIFO fluid simulation
9
Quality of Service (QoS)
• Objective performance measure
• Performance metrics:
– Packet loss probability
– Mean packet delay
– Maximum packet delay (e.g. 95 % quantile)
– Packet delay variation
– Throughput (bits/second)
10
Quality of Experience (QoE)
• Subjective performance measure
• Desribes users satisfaction of all
imperfections affecting the service
• Performance metrics
– Video quality
– Channel change time
– Blocking probability for VoD requests
11
Grade of Service (GoS)
• Call blocking probability
• Call set up delay
12
Traffic engineering (TE)
• TE objective is to deliver desired Quality of
Service (QoS) with minimum consumption
of network resources
• Optimize effectivness in terms of proximity
to optimality
• Optimize simplicity in terms of time and
space complexity
13
TE functions
•
•
•
•
•
•
Traffic management
Capacity management
Traffic measurement
Traffic modeling
Network modeling
Performance analysis
14
Traffic and capacity management
• Traffic and capacity management relies on
a relationship between three models:
– traffic model
– network model
– performance model
15
Importance of TE
• Expansion of network capacity driven by
increase in traffic demand
• With an effective TE solution fewer call
requests need to be rejected leading to an
increased revenue
• An effective TE solution allows longer time
period between capacity upgrades
16
TE planning
• TE Planner software tool for automated
selection of TE algorithms
• TE complexity is restricted by system
reponse time requirements
• Find set of TE algorithms with maximal
effectivness that provides the desired TE
complexity
17
TE planning problems
1. Call admission control and QoS
evaluation
2. QoS routing
3. Data center design
4. Network design and GoS evaluation
18
Packet traffic models
• Short-range depedent (SRD) models
– Superposition of Markov ON/OFF fluid sources
– Discrete Autoregressive (DAR) source
• Long-range dependent (LRD) models
– Superposition of heavy-tailed ON/OFF fluid sources
– Fractional Brownian Motion (FBM) source
• Hurst parameter H, 0 ≤ H≤ 1, measures self
similarity of traffic arrival process
19
QoS evaluation
• By analysis
– Determine model for traffic and network
resources
– Compute analytical QoS solution
• By simulation
– Use same traffic and network model as in
analysis
– Simulate random pattern of traffic arrivals,
service completions, and network resource
occupancy
20
Call admission control
• Accepts/rejects call requests based on
expected end-to-end QoE/QoS
• Flooding of link states at regular time
intervals (1s- 30s)
• Generic CAC decision based on uncertain
(aged) link state information
• Actual CAC decision at each node along
the selected routing path
21
Unicast routing
• Native IP network
– Best effort – no QoS guarantees
– Shortest path routing via IS-IS or OSPF
• IP/MPLS network
– QoS guarantees by resource reservation
– Constraint-based routing with multiple QoS
constraints
– Hop-by-hop (OSI Layer 3) routing or explicit (OSI
Layer 2) routing
22
Multicast routing
• Native IP network
–
–
–
–
Best effort – no QoS guarantees
PIM SSM
PIM SM
PIM DM
• IP/MPLS network
– QoS guarantees by resource reservation
– P2MP LSP
– PIM MPLS
23
Data center design
• Design of central (VHE) and regional
(VHO) video server systems
– Number of video servers
– Video server allocation rule
– In-advance transfer of stored video from
central VHE to to regional VHOs
24
Core and metro
physical network design
• Global physical network (PN) configuration
– Assign user community (population) to network nodes
– Dimension PN link capacities
• Global PN re-conguration
– Assign network nodes to new or expanded user
communities
– Adjust PN link capacities
• Problem input parameters include user
population, viewing preference vector, VoD
content duration statistics, and viewer request
rate vector
25
Core and metro
virtual network design
• Virtual networks (VNs)
– Overlay network on top of PN
– Built by TE-LSPs or ATM VPs
– Topology of VN can be different than for PN
– Many VN links can share a PN link
– VN link can consist of multiple successive PN
links
• Global VN configuration
• Global VN re-configuration
26
Correlation in arrival process
• Buffer distribution is a function of the
autocorrelation function (ACF)
• Impact of correlation in arrival process becomes
nil beyond a time scaled known as correlation
horizon
• Correlation horizon is a function of the maximal
buffer size
• Only necessary to chose a model of video traffic
that captures the correlation structure up to the
given correlation horizon
27
Network operation modes
• Network operating under packet scale
congestion
– Enough resources are allocated to keep the risk of
packet-level overload at the output multiplexer within
tolerable limits
– Short term correlations are most important
• Network operating under burst scale
congestion
– Enough resources are allocated to keep the risk of burstlevel overload at the output multiplexer within tolerable
limits
– Both short- and long-term correlations are important 28
Handling of congestion
• Large buffers are helpful to significantly
reduce the loss rate only for SRD traffic
• So for video traffic which is LRD, large
buffers will not decrease the loss
significantly, but may cause exessive
delys, which is not tolerated in IPTV
networks
29
Burst- versus
packet-level simulation
• Burst level simulation can be implemented
by Markov fluid traffic models
• Fluid simulation on the network is subject
to a ripple effect
• For very small buffers, packet-level
simulation will be more accurate
• Packet-level simulation can be
implemented by MMPP traffic models
30
Conclusions for IPTV networks
1. IPTV networks are most likely to operate
under packet scale congestion
2. Markov traffic models are sufficient for IPTV
networks
3. Traffic smoothing or shaping is
recommended and will improve the
statistical multiplexing gain
4. Simple metods like Chernoff bound could
be used by CAC in IPTV networks
31