Title: IntServ, DiffServ, and TCP - What Does It All Mean?
1IntServ, DiffServ, and TCP - What Does It All
Mean?
- Glynn Rogers
- Research Leader - Advanced Networks Technology
- CSIRO Telecommunications and Industrial Physics
- glynn.rogers_at_tip.csiro.au
2Quality of Service (QoS)
- A major driving force in Internet evolution
- Not simply defined - means many things to many
people - Has sense of predictable network behaviour
- Central idea is provision of network resources
that an application requires to perform adequately
3QoS is Generating a Confusing Array of Acronyms
Diffserv
4But Its All Beginning to Fit Together
- Primary aim is to convey my emerging picture of
how - Secondary aim is to argue that something new and
important is happening here - a whole new area of networking is developing
- merging of traditional routing and addressing
IP world with telecommunications engineering - the technical consequence of convergence
- complex - wont happen overnight
5Firstly, a Caveat or Two
- What follows is based squarely on the
documentation of the relevant industry
standards organisations - the ATM Forum
- the Internet Engineering Task Force (IETF)
- These are my interpretations - any confusions are
mine. - Its the big picture that counts - dont worry
about the detail. - not a tutorial - convey general impression by
example -no attempt at completeness
6Why Do We Need Such a Revolutionary Change?
- Current best effort technology is essentially a
quarter of a century old - Two factors driving the development of a new
generation of multimedia applications - commercialisation of the Internet
- Increasing availability and decreasing cost of
bandwidth - No evidence of free bandwidth scenario emerging
- rejected in RFC1633 (1994) - still true
- demand always rises to meet supply
7QoS is Not New
- Telephone network has QoS
- economics and technology based on a single
application - highly developed engineering
- but one size fits all
- BISDN an attempt by telephony world to
generalise network to encompass diverse
applications - ATM technology - first full exploration of QoS on
demand concepts
8A Quick Look at ATM
- ATM is connection oriented
- end to end virtual connection established with
negotiated QoS characteristics - Service category - CBR, VBR etc
- traffic characteristics - peak rate, sustained
rate, burst size etc - QoS parameters - loss rate, delay, delay jitter
- SVC establishment requires both
- QoS routing (PNNI) and
- resource allocation in traversed switches
(signaling)
9Quality of Service and Resource Management
- Fundamental resource is output link rate
- Access managed via scheduling discipline
- Bursty input traffic held in buffers
- adds delay and jitter
- overflow causes packet loss
- These factors determine QoS at network level
- Optimise via buffer management and scheduler
parameter setting
10QoS in the Internet
- Internet Engineering Task Force (IETF) is
evolving QoS support mechanisms for the Internet
- two approaches - The Integrated Services Internet
- QoS for individual microflows
- perhaps too complex for large networks - wont
scale easily - Differentiated Services - more scaleable
- lose sight of individual microflows - Behaviour
Aggregates
11Why not Just Stick with ATM?
- Original ATM concept was QoS overkill
- end-to-end defined channel
- assumed long lived flows with specific
requirements - connection setup overheads relatively small
- OK for telephony, high quality VoD etc
- But Internet traffic is dominated by TCP
- significant proportion of short lived flows (eg
Web downloads of text and image pages - even streaming video applications are using TCP
12IETF IntServ Introduces Another Traffic Class
- Newer real time applications (Web based in
particular) are elastic or adaptable to modest
fluctuations in network performance - An example is streaming video over TCP
- TCP provides rate adaptation to network load
- application can respond to blocking at socket
calls - change frame rate (but careful with audio)
- hierarchical coding provides graceful degradation
- MPEG 4 supplies a formal framework
13IntServ Controlled Load Service
- Based on observation that for this class of
traffic the existing Internet works fine if it is
not heavily loaded - Use resource allocation to provide performance
equivalent to a lightly loaded network - Can base definition on qualitative specifications
as distinct from quantitative specifications of
ATM
14IntServ Also Provides for Established Traffic
Classes
- A growing number of demanding applications
- VoIP has stringent requirements on packet loss
and delay - Guaranteed Service designed for such applications
- Traditional best effort service class is still
required for non real time applications - IntServ provides a framework for defining new
service types
15The Integrated Services Concept
- Internal network resources are committed to
individual end-to-end microflows to provide the
QoS the service requires - connection setup - Applications must specify the traffic
characteristics of the microflow - token bucket model - rate and burst size specs.
- flows are policed to ensure conformance
- Network performs Connection Admission Control
- Method of resource allocation up to implementor
16Why Not Just Extend ATM?
- ATM is based on Layer 2 switching
- IntServ retains Layer 3 forwarding mechanism
- essentially a connectionless environment
- flows are more abstract than a VC - akin to
traffic trunk concept in MPLS - IntServs signalling protocol - RSVP - is
receiver driven and soft state based - much greater compatibility with multicast
17So What Went Wrong?
- RSVP is dead reports are exaggerated
- QoS is complex
- requires systems rather than individual protocol
approach - more time required for development and acceptance
- Nevertheless there is a problem
- IntServ inconsistent with Internet philosophy of
keeping complexity to the network edge - requires interior nodes to retain state for each
microflow - state explosion problem in interior of big
networks
18Enter Differentiated Services
- DiffServ distinguishes between end-to-end
services and the behavior of the individual
network components required to support them - DiffServ is based on a set of defined Per Hop
Behaviours (PHBs) specified via an IP header
byte, the DS byte - 3 types of PHB so far defined in RFCs
- Class Selectors - priority based - cf IP
priority - Expedited Forwarding (EF)
- Assured Forwarding (AF)
19Diffserv Emphasis is on Individual Interfaces
- State explosion problem is avoided by
aggregating traffic requiring the same QoS at
each interface - Each Behaviour Aggregate experiences the node
performance specified in the required Per Hop
Behaviour - The behaviour aggregate and PHB are determined by
the DiffServ Code Point (DSCP) carried in the DS
Byte
20Expedited Forwarding and Assured Forwarding PHBs
- about bandwidth allocation - via schedulers
such as weighted fair queuing as well as buffer
management - Expedited Forwarding - reserved resources
(aggregated) - signalling (RSVP?) - VoIP - Assured Forwarding - 4 classes
- intended for controllable sources such as TCP
- controlled packet drops - 3 levels of drop
precedence with a separate DSCP for each level
21Example of an Assured Forwarding Mechanism
AF class 1
AF class 2
Weighted Fair Queuing Scheduler
drop probability
buffer with 3 level RED mechanism
AF class 3
AF class 4
22Seeing the Woods for the Trees- Diffserv Domains
- Domain - collection of nodes under one
administration with common policies for routing,
QoS, etc - Domains interact via Service Level Agreements
- traffic policy written as Service Level
Specifications - traffic managed using Traffic Conditioning
Specifications - Domains interconnnect via boundary nodes which
contain Traffic Conditioning Elements - packet filters, meters, shapers, policers etc
- note these all act on aggregates specified by the
DSCP
23Management Issues - Provisioning Diffserv Domains
- Both EF and AF PHBs require explicit resource
allocation - bandwidth, buffer space etc - Mechanisms for allocating resources over domain a
research issue - static allocation - management systems
- dynamic allocation - bandwidth broker - active
networks - Routing implications - traffic engineering
- constrained routing
- MPLS
24Of Microflows and Macroflows - IntServ over
DiffServ
- Policing at domain boundaries on aggregates
- Without individual CAC all flows in an aggregate
can suffer from over commitment - IETF Integrated Services over Specific Lower
Layers (ISSLL) working group proposes using
DiffServ network as akin to, say, ATM link - Aggregation of RSVP requests into single RSVP
action - mapping of IntServ services onto Diffserv Per
Domain Behaviours - determined by node PHBs
25Example Scenario - TCP based Streaming Video
- Assume a properly resourced Diffserv domain
- Assume a Bandwidth Broker which can interact with
RSVP to provide IntServ admission control - Combine Controlled Load with Assured Forwarding
- both in spirit of elastic flows on lightly loaded
network - Require policing to control average TCP flow rate
- nonconforming packets marked down to a DSCP
giving higher drop probability in AF class - we have experimentally demonstrated that this
works!
26Traffic Generator 1
Traffic Sink
CISCO 7505 Router
Linux Router
Traffic Generator 2
Traffic Generator 3
IntServ Domain
IntServ Domain
DiffServ Domain
27TCP Rate Control Using Source Policing and
Assured Forwarding
28Video On Demand Server
Video On Demand Client
Accelar Switch Router
Linux Router
Traffic Generator
IntServ Domain
DiffServ Domain
IntServ Domain