Upload
alvin-norton
View
216
Download
0
Embed Size (px)
Citation preview
8 June 2005 TNC2005
JANET QoS Development JANET QoS Development Project: Project:
IP Premium and LBE Trials across IP Premium and LBE Trials across JANETJANET
Victor OliferUKERNANetwork Development Project Manager
8 June 2005 TNC2005
AgendaAgenda
• JANET QoS Project history• Test plan• Test results• Phase 1 recommendations• Phase 2 objectives and structure• QoS models to be explored
8 June 2005 TNC2005
QoS on JANETQoS on JANET
• Increasing volume of traffic leads to an increased probability of congestion– This could impact real-time services such as
VoIP, Videoconferencing and Content Delivery over JANET• The JANET QoS Development Project was
established to address these challenges
• JANET is a multi-domain, hierarchical network– Backbone (SuperJANET)– Regional Networks (RBCs, RNOs) – Site Networks (LEAs, Universities, Colleges, Schools etc)
8 June 2005 TNC2005
JANET QoS JANET QoS Development ProjectDevelopment Project
• Work in the QoS area commenced in 2001, with the formation of the QoS Think Tank– QoS Think Tank Report produced (stress on
DiffServ)• JANET QoS Development Project
commenced in 2002– Define the prototype framework and QoS
services– Call for project partners– Configure backbone and partner networks with
QoS (2003)– Conduct testing on the production network
with real applications (the first half of 2004)
8 June 2005 TNC2005
UKERNA dev.net
Manchester
Lancaster
Imperial
Soton
Swansea
LBE(Elastic bulky GRID)
Partners and QoS classesPartners and QoS classes
IP Premium(VC, VoIP)
All ingress interfaces of BARs have 5% limit for IP Premium
(DSCP 46)
SSDN
8 June 2005 TNC2005
Test HighlightsTest Highlights
• Aim of testing– To compare the behaviour of applications
when they are served as BE and non-BE (Premium or LBE)
• Comparison criteria– Subjective (observed quality of VC and VoIP
sessions during periods of congestion) – Objective (RTT, jitter and loss measured by
the monitoring infrastructure deployed)
8 June 2005 TNC2005
UKERNAUKERNA
ManchesterManchester
SwanseaSwansea
LancasterLancaster
Imperial College
Imperial College
SouthamptonSouthampton
Application / Site Edge Router Applicatio
n / Site Edge Router
Application / Site Edge Router
Application / Site Edge Router
Application / Site Edge Router
Collection Agent
SAA-responder
SQLDatabase
KeyProbes and responses between SAA probes and responders
Results from SAA probes going to SQL database
Application / Site Edge Router
Reading
SAA-probe
SAA-responder
SAA probe/responder
SAA-responder
SAA probe/responder
SAA probe/responder
Monitoring Monitoring infrastructureinfrastructure
8 June 2005 TNC2005
Router under test
1 2ABC
3DEF
4 5JKL
6MNOGHI
7 8TUV
9WXYZPQRS
* 0OPER
#
7940CISCO IP PHONE
imessages directories
settingsservices
1 2ABC
3DEF
4 5JKL
6MNOGHI
7 8TUV
9WXYZPQRS
* 0OPER
#
7940CISCO IP PHONE
imessages directories
settingsservices
Application-generator
Application-receiver
SAA-responder
SAA-probe (measures reflected traffic)
IPERF-client IPERF-server
Load2Load1
Load-destination
Hub/switch/router
Hub/switch/router
Database
QoS MIB objects
(every 1 min)
An example of a test An example of a test scheme scheme
Interface to be congested
23455
8 June 2005 TNC2005
Test ResultsTest Results
• In most cases both VC and VoIP traffic benefited from IP Premium service
• LBE unicast traffic received allocated percent of bandwidth during BE bursts
• Observed peculiarities:– POS OC3 interfaces of Cisco 6500: Premium traffic
had increased delays and loss during BE congestion – LBE + BE multicast traffic (Access GRID) behaviour
was unpredictable – LBE behaved like something better than BE, audio and video sessions failed
8 June 2005 TNC2005
Preliminary trial of RUDE without VC session
VC is served as BE, VoIP as EF, VC got frozen
VC is served as EF, VoIP as EF, no degradation of both.
IP Premium UKERNA – Manchester test, 16th March 2004
8 June 2005 TNC2005
50%
Averagetransmission delay
Utilisation
100%
Very bursty traffic
Almost regular traffic
Fixed delay(signal propagation etc)
Explored window of Explored window of QoS benefitsQoS benefits
Voice & video tolerance
Overpovisioned
area
10%-20% 80%-90%
8 June 2005 TNC2005
JANET QoS Development Project Phase 2
• Recommendations of Phase 1 participants:– QoS benefits were noticeable, it works!– Establish Phase 2– Define production QoS Service Model for JANET– Conduct large scale piloting activities
• Some technical areas need to be further investigated
• Guidance documentation requirements• We should keep pace with GEANT2
8 June 2005 TNC2005
JANET QoS Development Project Phase 2 (2005- 2007)
QoS Architecture Group
Interworkingwith other
technologies:IPv6, Multicast,
Firewalls
LowBandwidth
Connections:new partnersamong FE
Policy &
Management
Monitoring and
Management
Applicationsrequirements
8 June 2005 TNC2005
Types of DiffServ models, Types of DiffServ models, to be explored in Phase 2to be explored in Phase 2
Factors:– Destination awareness
– Static vs. dynamic reservations
– Trust relationships between domains
8 June 2005 TNC2005
Static destination-unaware Static destination-unaware DifferServ model DifferServ model
Domain B(Regional
Network)
SL
A
A-B
SL
A
B-A
Domain A(Campus
Network)
Domain J(SuperJANET,
National Backbone
Network)
Domain F(Campus
Network)
Domain C(Campus
network)
ISP Domain G(GEANT,
Pan-European Backbone
Network)
Domain E(Regional
Network)
SLA
B-J SLA J-
B
User A1
User C1
User F1
R1
R2
- Only edge routers of domain do admission control and traffic policing:
to protect domain form excess of privileged traffic
- Neighbouring domains conclude SLA to process traffic classes in similar manner
and according common marking, so called DSCP values (e.g. 46 for Premium)
No reservation for flows!Each domain is responsible for
proper processing of QoS classes (aggregates)
8 June 2005 TNC2005
Pros and cons of static destination-Pros and cons of static destination-unaware DiffServ modelunaware DiffServ model
• Scalability – excellent! • Flexibility – excellent!• Maintainability - excellent!• Guarantees – poor
– Clash of flows in some output interface is allowed (flows’ routes are not under control)
– Reasonable hierarchical design of a network decreases the clash probability
8 June 2005 TNC2005
Applications of static DiffServ modelApplications of static DiffServ model
• A few well-known sources of IP Premium traffic (e.g. VC studios) which are allowed to communicate only with each other
• Restriction of a number of egress flows from origin domain at application level: – E.g. by VoIP gatekeepers
8 June 2005 TNC2005
Dynamic destination-aware DiffServ with Bandwidth Brokering
-Admission control-- Traffic policing
Domain B(Regional
Network)
BA-ADomain A
(Campus Network)
-Admission control-- Traffic policing
Domain J(SuperJANET,
National Backbone
Network)
Domain F(Campus Network)
Domain C(Campus network)
ISP Domain G(GEANT,
Pan-European Backbone
Network)
Domain E(Regional
Network)
SLA B-
J SLA J-
B
BA-E
BA-G
BA-B
BA-J
BA-C
R2
1. Check of total throughput
2. Configure admission control to pass a flow
-Admission control-- Traffic policing
•Relieve routers from counting and storing state information
•Bandwidth Brokers (BBs) do this job for routers
•Hybrid of IntServ and pure DiffServ
•Very new – GEANT2 is going to deploy it in next 4 years
•Needs re-configuring of policers for every flow – not scalable for high-level domains
8 June 2005 TNC2005
The same + trust relationships between domains
Domain B(Regional
Network)
BA-ADomain A
(Campus Network)
- Admission control- Traffic policing
Domain J(SuperJANET,
National Backbone
Network)
Domain F(Campus Network)
Domain C(Campus network)
ISP Domain G(GEANT,
Pan-European Backbone
Network)
Domain E(Regional
Network)
BA-E
BA-G
BA-B
BA-J
BA-C
trust
trust
•Reconfiguring of policers and admission control tools happen only in origin domain – good scalability
•Additional risk of misconfiguring devices within low-level domains
8 June 2005 TNC2005
Applications of dynamic Applications of dynamic destination-aware DiffServ destination-aware DiffServ modelmodel
• Effective for long-lived flows, e.g. VC
• Non-effective for short-lived flows, e.g. VoIP conversations
8 June 2005 TNC2005
A combination of static and dynamic A combination of static and dynamic DiffServ models is possible…DiffServ models is possible…
• However it needs different DSCP for marking static and dynamic flows
8 June 2005 TNC2005
Any questions?
Further Information:
JANET QoS Development www.ja.net/development/qos
Project Manager: Victor Olifer, [email protected]