Transcript Document

Policy-based QoS Framework for
Multi-service IP Networks
Hoon Lee
E-mail: [email protected]
Network and Service Assurance Lab.
Dept. of Information & Communications Engineering
Changwon National University
Changwon, Korea
Service Trends:
Triple Play=Voice+Data+Video
Voice from Phone
Data from PC
VoD/ TV
Videophone
There exists no killer applications! Pack them up!!
Triple Play Services:
- Italy: FastWeb
- Japan: NTT RENA, KDDI, SoftBank
- Korea
1. KT: All up prime (Megapass+VoIP+Videophone+Messaging+Broadcasting)
2. Dacom: Internet+VoIP+ Broadcasting
3. Hanaro Telecom: 2 Scenarios
PSTN: xDSL+POTS+SkyLife
Cable: Cable internet+VoIP+Broadcasting
Technologies for Internet QoS
Speed up & Over-provisioning (QoS-Free)
- Current BE service
- Applicable to any kind of future applications
* Almost zero delay if link speed is in the order of 10s of Mbps
Service Differentiation for Priority Traffic  SP, CBQ, Hybrid
- Wired network: IETF DiffServ + MPLS (Priority service +Tunneling)
Priority: EF > AF > BE
- Wireless ad hoc network: SWAN (Feedback control + CAC)
Priority: rt traffic > BE traffic
Policy-based QoS Guarantee
- Policies for
service differentiation / BW allocation / scheduling / routing
Policy-based QoS Framework of TEQUILA
Policy
Manageme
nt Tool
Policy
management
Policy
server
Policy consumer
From
Customer
SLS
Subscr.
Traffic
Estimation
Network
Provisioning
Static policy
(Long-term)
SLS
Req.
SLS
Inv.
SLS Management
Network
monitoring
Resource
manag.
Route
manag
Performance Manag. Traffic Engineering
Dynamic policy
(Short-term)
NTT RENA’s QoS Framework
 Separation of control and data transfer plane
 Flexible network control
 Centralized QoS management  e2e QoS
Service/Network Control Platform NIB
SCS
PS
BB
From NMS
PSTN
PC
Video server
PC
e2e Optical network
Web server
Phone
RENATM:
Resilient Network Architecture
SCS: Session control server
BB: Bandwidth broker
PS: Policy-server
NIB:Network information base
Phone
Policies for IP QoS
Principle for IP QoS: Be faithful to IP’s philosophy.
- Advantage of IP: Connectionless paradigm
 Simple & scalable
IP QoS Provisioning: via Policy-based networking
- Destination-based routing based on OSPF principle
- Treat QoS traffic with higher priority than the BE traffic
 SP does not sacrifice the lower class traffic
when the link speed exceeds 10s of Mbps!!
- Protection of QoS traffic: Class-based CAC
- Network–wide: Interoperation of Policy Server/NMS
 Dynamic CAC & bandwidth management
Policy-based Networking: Big Picture
Phones
Best effort IP
network
VoIP
G/W
PBX
…
Current
AN
Router
PCs/Servers
Core
Access Premium backbone
network
Router
Router
Voice buffer
MPLS Tunnel
Policy server farm
QoS
Server
(SLA)
Policy Base
VPN
Bandwidth
Broker
Voice
Access network
Data
NIB
Traffic
meter
PCs/Phones/TEs/Serv
ers
Core node: DiffServ-based CBQ + PHB-based Scheduling + MPLS-TE
Edge node: SLA negotiation, UPC, Packet classification /QoS mapping, CBQ, Packet-scheduling
Packet Level SLS
Service
type
Data
service
Voice
service
Video
service
Bundle
service
Attributes
Conventional
BE service
QoS compatible
to PSTN
New
Interactive
Application
services
QoS Requirements
(ITU-T)
Email, ftp,
low quality
video
None
Internet
telephony,
Interactive
multimedia
E2E delay < 150ms for
99.99% of packets,
PLR < 10-3
TV,
Videoconfere
ncing
E2E delay < 150ms for
99.99% of packets,
PLR < 10-4
IP VPN ,www,
on-line game,
streaming
multimedia
Minimum contracted BW,
E2E delay < 1~4sec
PLR < 10-6
Mapping between DiffServ & MPLS
QoS
Services
Premium
service
Assured
service
Better than BE
service
Best-Effort
service
DiffServ
PHB
EF
AF 1/2
AF 3/4
BE
MPLS
Label
Platinum
Gold
Silver/ Bronze
Steel
ITU-T QoS
Class
0/1
2
3~4
5
Typical
Applications
VoIP
VPN
Signaling,
VoD
WWW,
telnet,
streaming
service
e-mail
Bandwidth Allocation Alternatives
Bandwidth reservation model
- Absolute QoS guarantee
- Low efficiency
- e.g.: IntServ architecture
- Application to: Videophone service
Bandwidth share with priority scheme model
- Statistical QoS guarantee
- High utilization
- e.g.: DiffServ architecture
- Application to: Multi-service
Bandwidth Reservation Model :
Videophone Service Architecture
ISP
LAN
Video Phone
Internet traffic
Cd
IP Network
(DiffServ)
E-S/W
Router
…
Cv = ?
Cv=C
C: Number of videophone connection (channel)
: Bandwidth of a videophone connection
E-S/W
…
Input to The System
Parameters:
 Number of subscribers: M (Tens of thousand)
 Fraction of active connections at busy hour:  (10%~20%)
 Mean session duration: 1/ ( 1,000seconds)
 Mean session arrival rate:  (0.01 ~ 1 )
 Session broking probability:  (0.5~1%)
 Bandwidth requirement of a Videophone session: 
 = 2Mbps (For basic rate service)
(8bits/pixel250200pixels/frame5frames/sec=2Mbps)
Analytic System Model
Assumption on the session:
Session arrival: Poisson arrival
Session duration: Exponential distribution
System model:
Infinite number of traffic sources
Full availability link
M/M/c/c Queuing model with C concurrent channels
Erlang B-formula for GoS of videophone service
C
E (C ,  ) 
C!
C
i
i 0
i!

.
Constraint on the Service level: E(C,)  .
where  = (M     /)/3600
Results and Discussion
Typical Assumptions:
M = 30,000 residential subscribers
 = 0.1 (Residential= 10%, Business=20%)
1/ =1,000 seconds
 =0.36/0.72 sessions / Busy hour / Person (Residential / Business)
 =1%
 = 2Mbps (basic rate)
Result of computation:
 Input traffic in Erlang: 300 Erlang
 Computed number of channel: 323 Channels
 Required bandwidth: Cv=C   = 323  2Mbps = 646Mbps
 To provide the safety margin, we have to take into account the
traffic from alternate route of the neighboring nodes:
Cv Final= 2  Cv=1.3Gbps  Final result.
Comparison: Residential vs. Business
When the subscribers are business customers
-  = 0.2
-  = 0.72
(The offered load increases to 4 times that of the residential subscribers!)
Total required bandwidth for a number of subscribers:
Number of
subscriber
Required number of channel
(residential / business)
Total Required Bandwidth
(residential / business)
30,000
323 / 1292
1.3 / 5.2 Gbps
60,000
650 / 2600
2.6 / 10.4 Gbps
90,000
928 / 3712
3.7 / 14.8 Gbps
Bandwidth Share Model :
Strict Priority Scheduling Scheme
System model: DiffServ-aware MPLS
Service model: Strict priority (SP) to voice over data1 over data2
Router model: M/G/1 queue with non-preemptive service
Objectives: Evaluation of delay for class1, 2, and 3 packets
Voice
packet
SP
Data1
packet
C
Data 2
packet
Our concern:
1. Can we apply the SPSS in a DiffServ router for BcN?
2. How about the behavior of delay with respect to the system parameters?
System Model
 System parameters:
- Mean arrival rate for voice/data1/data2: 1, 2 , 3
- Mean service time for voice/data1/data2 : 1/1, 1/2 , 1/3
- Second moment of service time: E[k2],k=1,2,3
- Offered load for voice/data1/data2 : 1, 2, 3
- Link capacity: C
 Source models:
- Voice: Poisson arrival, fixed packet size
- Data1 & data2: Poisson arrivals, Pareto distributions
Delay Performance
Mean waiting times for M/G/1 queue with SP service:
2

E
[

k 1 k k ]  R
3
W1 
2(1  1 )
W2 
R
,
(1  1 )(1  1   2 )
W3 
R
,
(1  1   2 )(1  1   2   3 )
Mean waiting time for M/G/1 queue with FIFO service:
2=squared coefficient of variation for
 2 1  CS2
C
S
WFIFO  (

).
 1 
2
service time of a packet
1
Numerical Experiments
Source traffic profile:
- Voice source: G.711 Voice coder, 216bytes
- Data source: Ethernet frame, Pareto distribution,
m
F (l )  Pr{ L  l}  1  ( ) , (l  m,   0).
l
Minimum packet size, m: 500~1500bytes
Tail index: =3
Link capacity per output port: 1M, 10M, 100Mbps
Traffic Load Type
Load
Type
1
2
3

A
0.1
0.4
0.4
0.9
B
0.3
0.3
0.3
0.9
C
0.5
0.2
0.2
0.9
D
0.7
0.1
0.1
0.9
Light-voice
Heavy-data
Heavy-voice
Light-data
Waiting Time of Voice Packets for
Different Link Capacities
m2=500bytes, m3=1500bytes
Under SP scheduling scheme, delay of voice packet is
almost negligible for high-speed links!
Waiting Time of Voice Packets for
Different Service Schemes
m2=m3=1,000 bytes, C=1Mbps
The conventional wisdom of
“SP isolates voice traffic from non- voice traffic”
does not hold!
This is more evident for the WFQ-families.
Delay Performance of Data Traffic
Performance comparison between different classes:
W2
1

.
W1 1  1   2
W3
1  1

.
W2 (1  1   2   3 )
1=0.2
1=0.4
1=0.2
2=0.4
2=0.2
Summary
Policy is important for QoS provisioning in future Internet.
Network provisioning is dependent on the policy.
Reservation model over-estimates the network resources.
Shared bandwidth model will prevail.
Accurate dimensioning of network resources saves cost.
References
[Lee] Hoon Lee, “Strategies for the construction of Policy-based
managed IP QoS”, Final Report of NCA II-RER-04041,
November 30, 2004.
[Lee] Hoon Lee et al., “Dimensioning NGN for QoS guaranteed voice
services”, Jr. of IEEK, Vol. TC-40, No.12, December 2003.
[Lee] Hoon Lee, “Delay analysis of DiffServ/MPLS network”,
Industrial Mathematics Initiative 2004, August 26-28, Korea.
[Lee] Hoon Lee et al., “Delay performance of non-realtime services for the strict priority scheduling scheme”, Jr. of
the research institute of industrial technology, Vol.18, May 2004.
[Trimintzios] P. Trimintzios et al., An architectural framework for
providing QoS in IP differentiated services networks, TEQUILA
Project report.