Anda di halaman 1dari 18

Traffic management and QoS

provisioning in
telecommunication networks
TNO Telecom

Hans van den Berg


J.L.vandenBerg@telecom.tno.nl

Overview

Centre of Excellence QoS within TNO Telecom


organisation, positioning
previous activities
Some current activities
Basic models for service integration
TCP performance modelling
Load adaptive scheduling in UMTS+
Overprovisioning in IP networks
Concluding remarks

Leidschendam 2

1
Centre of Excellence QoS within TNO Telecom

KITs BITs

EC
QoS
NAK
BNO

Main goals of CoEs


Build up expertise on specific subject
Exchange expertise/experience
Achieve synergie

Leidschendam 3

Centre of Excellence QoS


Main issue
How to handle the traffic generated by services/applications to
be supported by the networks/platforms?
realization of QoS requirements
efficient use of resources
manageable operations

Operator point of view

Leidschendam 4

2
Center of Excellence QoS
Main products

Quantitative models, methods and tools


stochastic models, simulation

Consultancy
guidelines for network dimensioning
guidelines for traffic management
guidelines (restrictions) for network- and system design
service implementation
SLA definition

Leidschendam 5

Centre of Excellence QoS


Current application areas (formerly: telephony, X25, FR, LANs, GSM, ATM, )
Fixed (IP) networks
Mobile/wireless networks (GPRS, UMTS, WLAN, ad-hoc)
ICT (distributed systems, middleware, web services, )

Motivation
Synergie
re-use of expertise and experience
Services over heterogeneous networks/platforms
end-to-end QoS

Close cooperation with other groups within TNO Telecom


E.g. network architectures, physical layer aspects,

Leidschendam 6

3
Centre of Excellence QoS
Cooperation
Universities/institutes
UT, CWI, TU/e, TUD, VU, CNET, VTT
National projects
Equanet, Beyond 3G, BR@H, RGE
European research programmes
(RACE, ACTS), IST, COST, Eurescom

Leidschendam 7

CoE QoS: previous activities


ATM Networks (1990s)
Network dimensioning
multi rate models (extension of Erlang model)
Traffic modelling
packet-, burst and flow level; self similarity, LRD
Statistical multiplexing
queueing models (fluid flow), effective bandwidths
Connection admission control (CAC)
static, measurement based
QoS differentiation
priority queues, resource sharing (WFQ, )

IP Networks (1996 - )
Intserv
Diffserv Leidschendam 8

4
Overview

Expertise Centre QoS within TNO Telecom


organisation, positioning
previous activities
Some current activities
Basic models for service integration
TCP performance modelling
Load adaptive scheduling in UMTS+
Overprovisioning in IP networks
Concluding remarks

Leidschendam 9

Some current activities


1. Basic models for service integration

Roughly two traffic types:

Stream traffic
real-time, delay sensitive applications, e.g. telephony, video
fixed transmission rates

Elastic traffic
non-real-time data applications (file transfer, WWW, )
rate can adapt to available link capacity
rate adaptation by e.g. TCP flow control
main QoS measures: transfer delay, throughput

Leidschendam 10

5
Basic models for service integration
elastic flows

Availablefor
available forelastic
elastictraffic
traffic

stream traffic
C

stream calls

Time

Stream traffic has priority over elastic traffic

Impact of stream traffic on elastic traffic performance?


Link dimensioning
for given traffic load, determine required link capacity C
blocking probability stream calls < P
mean throughput elastic flows > T

Leidschendam 11

Basic models for service integration


Stream calls
Poisson call arrivals s, mean holding time hs
fixed capacity requirement rs
blocking if # calls > C/rs

elastic flows
available for elastic traffic
Available for
C
elastic traffic
stream traffic
stream calls

Elastic flows (e.g. files)


Poisson arrivals e, mean file size fe
max. capacity consumption re (e.g. modem rate)
capacity not used by stream calls is equally shared by elastic
flows

Leidschendam 12

6
Basic models for service integration
Stream traffic
Stream traffic performance is not influenced by elastic traffic

Blocking prob. Ps is obtained from well known Erlang formula

Elastic traffic (without interfering stream traffic)


Elastic traffic behaviour can be modelled by M/G/1 GPS
queueing system analyzed in [Cohen79]

Explicit formula for mean file transfer delay D(x)


D(x) is independent of file size distribution!
D(x) is linear in file size x

Leidschendam 13

Basic models for service integration

Elastic traffic (with interfering stream traffic)

Elastic traffic behaviour can be modelled by M/G/1 GPS


system with randomly varying service capacity
Nice properties (linearity, insensitivity) are lost
Very difficult problem! No explicit results available

For exponential file size and stream call holding time, D(x)
can be obtained as solution of set of differential equations
[Nunez, vd Berg, Mandjes 99]
impact of stream traffic fluctuations (time scale)
impact (max.) capacity requirements rs and re

Leidschendam 14

7
Basic models for service integration
Asymptotic result (i.e. for x ):

x
D( x) + const ,
C * e f e

where C* denotes the mean capacity not used by stream calls

asymptot
D(x)

without interfering
stream traffic

File size x

Leidschendam 15

Basic models for service integration

Recent results
Accurate approximation for mean elastic flow throughput
[Litjens, Vd Berg, Boucherie 03]

The greater the (elastic) call size variability the better the
performance! [Litjens & Boucherie 02]

Further Research
Explicit (approximate) formulas?
general flow size and (stream) call holding time distributions
[Bonald & Roberts 00], [VdMei, VdBerg 02]

impact stream traffic characteristics on elastic traffic performance?

Leidschendam 16

8
Leidschendam 17

2. TCP performance modelling


TCP has considerable impact on QoS
e.g. web page download times
flow control, slow start

Many papers, a lot of models

Roughly two categories:


flow level models
packet level models

TCP flows/packets

output Outgoing link (rate C)


buffer

Leidschendam 18

9
TCP performance modelling

Flow level models


Capture flow level stat. multiplexing effect
Poisson flow arrivals, limited peak rate

Neglect impact packet level parameters (ideal, fair sharing)


buffer size, RTT, W max , ...

Example: Processor Sharing models (M/G/N PS)


[Nunez et al.99], [Bonald&Roberts00], [VdBerg et al.00], [Riedl et al.00],..
nice, explicit results for e.g. (mean) file download time
typical application: link dimensioning

PS models yield (too) optimistic results and do not provide


packet level insights

Leidschendam 19

TCP performance modelling

Packet level models


Capture impact of (some) packet level parameters
RTT, W max , buffer size, ...
Do not take into account flow level dynamics
assume a fixed number of (persistent) flows
flow level stat. multiplexing effect is not modelled
Examples
[Padhye et al.98], [Kelly00], [Lassila et al.00], ...
typical application: tuning of buffer man. mech. like e.g. RED, ...

Packet level models yield (too) pessimistic results


(Too?) much focus on details, while important flow level
effects are not taken into account
synchronization? capture behaviour?

Leidschendam 20

10
TCP performance modelling

There is need for combined packet/flow level models


for TCP performance:

Stat. multiplexing effect on flow level


stochastic flow arrivals, limited peak rate

and

Impact of packet level parameters


buffer size, RTT, ...

Leidschendam 21

Packet/flow level TCP performance model


Approach [Lassila, V d B e r g & Mandjes 02]

1. Obtain throughputs T n when n flows are active from


(existing) packet level models, for n=1,2,...
captures packet level parameters

2. Analyze M/G/1 PS queue with state dependent service


rates nT n when n flows are present, n=1,2,....
captures flow level parameters
steady state probabilities
mean file transfer time

Fixed point iteration between step 1 and 2


T n s depend on packet loss probability p(T n ) at bottleneck link
use M/G/1/K model to determine p(T n )

Leidschendam 22

11
TCP performance modelling

Fi l eFile transfer
d o wnl o a d t itime
me [ s]
C = 10 Mb p s; r = 1 Mbps; me a n fil e si ze = 150 KB; K = 10
4. 0
3. 5
sisimulation
mul ati on
3. 0
M/M/G/N
G/ N PSPS
2. 5 refined
GPS PS)
(r efi ned
2. 0
1. 5
1. 0
0. 5
0. 0
0 0. 2 0. 4 0. 6 0. 8 1
l oa d

Leidschendam 23

Leidschendam 24

12
3. Load adaptive scheduling in UMTS+ (HSDPA)

Integration of stream and elastic traffic in UMTS+ (downlink)

HS-DSCHs for data traffic, DCHs for stream traffic (speech, video)

Exploit resources left over by prioritised stream traffic by dynamically


up/downgrading of the HS-DSCHs

T O T A L C A P A C IT Y !

D ATA TRANSFER!!!

Downlink

PRIORITISED LOAD

time

Leidschendam 25

Load adaptive scheduling in UMTS+ (HSDPA)

Example

Wraparound network of omnidirectional BTSs with budget p max

Speech calls handled on 12.2 kbits/s DCHs

performance measure: outage probability P O T G

Data flows handled on HS-DSCHs

performance measure: expected throughput E{R}

([Litjens 02], [Litjens, VdBerg 02])

Leidschendam 26

13
Load adaptive scheduling in UMTS+ (HSDPA)

PO
RW EE
AT R FFA
AIIR
RNNEESSSS
90 1
90 1
E{R} ADAPTIVE
E{R} PPOTG
OTG
ADAPTIVE
(kbits/s)
throughput(kbits/s)

60 0.1

outage
60 0.1

outageprobability
expectedthroughput

probability
expected

30 0.01
30 0.01

E{R} FIXED
E{R}
0 FIXED 0.001
0 0.001
0 2 4 6 8 10
0 2 4 6 8 10
m inimum HS-DSCH power assignment
m inim u m H S - D S C H p o w e r a s s i g n m e n t

Leidschendam 27

Load adaptive scheduling in UMTS+ (HSDPA)

Further research
Throughput variance
Take into account impact of TCP
interaction with TCP flow control
(un)fairness

Leidschendam 28

14
Leidschendam 29

4. Overprovisioning in IP networks

(Cooperation UT, TNO Telecom, Surfnet)

Background
IP Diffserv: Appropriate network (over)dimensioning
What is a suitable level of IP network overdimensioning?
QoS requirements
not too much overcapacity
Operator has to decide whether or not to extend network
capacity!

Obvious approach: use traffic measurements

Leidschendam 30

15
Overprovisioning in IP networks

A v e r a g e l o a d i n m e a sure m e n t inte rva l s

0.9

0.8

0.7

0 .1 s e c
0.6
1 sec
load

0.5 1 min

0.4 5 min
15 m in
0.3

0.2

0.1

0
0 10 20 30 40 50 60 70 80 90

tim e (in se c o n d s)

Leidschendam 31

Overprovisioning in IP networks
What is the relation between traffic characteristics on different
time scales?

Can we use e.g. standard 5 minute measurement intervals to


predict the traffic on a much smaller time scale?
large time scale: simple/cheap measurements
small time scale: relevant for QoS

Goal
Standard traffic measurements + suitable traffic model +
analysis

Traffic characteristics on smaller time scale required capacity

Leidschendam 32

16
Required capacity:
- based on 1 second interval traffic measurements

- based on 1 second interval traffic estimations (using 5 minutes measurements)

M e a su r e m e n t s o f m e a n l o a d b e st fit curve v s. Estim a te c u r v e


10

9
Short Term m e an quantile (in Mbps)

6
99.5% fit-curve
5
99.5% estimate

9 9 % f it-curve
4
99% estimate
3 9 5 % f it-curve

95% estimate
2
9 0 % f it-curve
1 90% estimate

0
0 1 2 3 4 5 6 7 8 9 10
L o n g t e r m m e an (in M b p s )

Leidschendam 33

Overprovisioning in IP networks
Our approach is based on
Flow level traffic model
Poisson flow arrivals, general flow size distribution

Transient analysis of aggregate rate during (small) time interval

Insight in impact of different system parameters


interval length
file/flow size distribution
access rates

Further research
Does relatively simple approach work in other scenarios?
current scenario: public internet access, relatively small access rates
what about s cenarios with large access rates (e.g. university campus)?
other, more advanced (packet level) traffic models required, e.g. self similar?

Leidschendam 34

17
5. Other topics
Q o S of Voice over IP [ Kooij et al. 99 02]

Determine packet loss and delay (variation) on network level


traffic modelling (upperbound for traffic burstiness)
queueing models
Mapping of network level QoS to perceived QoS
Network Quality Mean Opinion Score
ITU-T standard by KPN Research
Impact CODEC, jitter buffer etc.

WLANs

Ad-hoc networks

Leidschendam 35

Concluding remarks
E n d-to-end QoS modelling becomes more and more complex
Interaction between OSI layers
in particular in mobile/wireless access networks
physical layer, MAC layer, transport layer
applications (e.g. adaptivity) VoIP
Heterogeneous networks
Different network technologies
Multiple domains
Unbundling
SLAs!
Service platforms and information systems

New project: EQUANET


End-to-end QoS in next generation networks
ICT Doorbraak
KPN (TNO Telecom), Lucent, UT, CWI, Tu/e

Leidschendam 36

18

Anda mungkin juga menyukai