Quality of Service over the Internet

1
Quality of Service over the Internet

• TUTORIAL
• International Conference on Information
  Technology
• CIT-2001
• Priyadarsi Nanda Dr Andrew Simmonds
• Department of Computer Systems
• Faculty of Information Technology
• University of Technology, Sydney, Australia
• pnanda, simmonds_at_it.uts.edu.au

2
Presentation Outline

• Introduction to QoS
• QoS via resource management
• QoS via generic switch architecture
• Internet structure Past, Present and Future
• Integrated Services Network (IntServ)
• Differentiated Services Network (DiffServ)
• Conclusion

3
Quality of Service (QoS)

• What is it?
• When you can measure what you are talking about
  and express it in numbers, you know something
  about it - Lord Kelvin

4
Quality of Service (QoS)

• Performance observed by the end user
• Combination of Delay, Jitter, Loss, Throughput
• Two most important for QoS Delay Loss
• QoS for different applications
• Real-time applications need assurance from the
  network for timely delivery
• Non-real-time applications need correctness of
  information delivery

5
Quality of Service (QoS)

• What causes poor QoS?
• Lack of resources in network or hosts
• Techniques to provide proper QoS
• Overprovisioning huge link bandwidth and high
  performance routers/switches
• Resource management (IntServ and DiffServ)
• Traffic control
• Generic switch architecture gt ATM

6
QoS via overprovisioning

• Fiber links can now support 1.6 Tbps
• Bottleneck is electronic switches/routers
• Optical switching currently switches paths, not
  packets (say av length 8000 bits)

Optical (1.6 Tbps)
Electrical (160 M packets/s)
Optical (1.6 Tbps)
7
QoS via resource management

• Traffic differentiation and prioritization
• Resource negotiation and Service Level Agreement
  (SLA)
• Network availability
• Guaranteed service level predictable QoS
• But guaranteed in a statistical sense

8
QoS via resource management

• Congestion control (reactive)
• Admission control (proactive)
• Traffic policing
• Leaky bucket
• Traffic shaping
• Improves predictability of flow, but introduces
  delay and reduces statistical multiplexing gains

9
Leaky bucket

• Virtual leaky bucket for traffic policing

Data flow unchanged
Capacity B bits
Current level
CLP bit set
av rate r bps
10
Virtual leaky bucket same as token bucket

• token allows x bits

Data flow unchanged
s tokens/s
Capacity C tokens
CLP bit set
(Capacity B Cx bits)
av rate r sx bps
11
Leaky bucket

• Real leaky bucket for traffic shaping

Capacity B bits
Current level
Packet discarded
Data flow smoothed but delayed
Data out r bps
12
Leaky bucket

• Virtual Max burst B rt bits in t s
• Token bucket Max burst (C st)x bits in t s
• Real Max delay td B/r s

Capacity B bits
r bps
13
Generic switch architecture

• e.g. Asynchronous Transfer Mode
• Cell switched technology gt low jitter
• Simple header H/W implementation
• gt Cut through switching gt low latency
• Connection oriented service at the network level,
  every flow has a defined QoS
• ATM switches have QoS and congestion management
  functions built in

14
Generic switch architecture

• ATM provides some of the backbone of the Internet
• ATM supports QoS
• Why not take advantage of these QoS islands in
  the core?
• ATM was a competitor to TCP/IP
• ATM to the desktop at 25 Mbps
• In the backbone at 4n x 155.52 Mbps

15
Generic switch architecture

• ATM Traffic classes
• CBR constant/committed bit rate
• Regular timing (e.g. 64 kbps speech)
• VBR variable bit rate
• Timing (e.g. compressed A/V)
• ABR available bit rate
• Best effort (data)
• UBR ubiquitous bit rate
• CLP bit set (low priority traffic)

16
Congestion control

• Reactive congestion control
• Differentiate traffic
• ATM gt CLP bit, FR gtDE bit, IP gt ToS field
• Act immediately to relieve congestion by
  discarding low value packets
• gt Utility of low priority traffic cushion
• Take steps to limit congestion by getting sources
  to reduce transmission rate

17
Congestion notification

• Frame Relay has 2 bits in its header
• FECN Forward Explicit Congestion Notification
• BECN Backward Explicit Congestion Notification
• FR is an access protocol so supports all 3
  congestion notification methods, including
  TCP/IPs Forward Implicit

18
Congestion notification
Forward Implicit
BECN
FECN
19
Internet structure

• Past classic IP with no QoS
• Present IntServ, DiffServ
• IntServ end-to-end QoS but does not scale
• DiffServ scales well, but coarse grained
• Future IPv6, all optical core, IntServ over
  DiffServ, Internet2, and things coming out of
  this conference!

20
Internet structure

• Some definitions
• Packet layer 3 protocol data unit (PDU) of
  variable length
• Cell fixed length PDU
• Datagram/connectionless independently routed
  packet, as in IP, or layer 4 type of service e.g.
  UDP
• Connection oriented follows fixed route, as in
  ATM at network layer, or layer 4 type of service
  e.g. TCP

21
Classic IP

• Interconnection of networks
• Core network is Best Effort
• No guarantee of service, no QoS
• No differentiation amongst traffic in the core of
  the network
• Internet Protocol IPv4

22
Classic IP

• TCP is a connection oriented service built upon
  IP
• TCP sits in the hosts Source Address (SA) and
  Destination Address (DA)

23
Classic IP

• Routing and routed protocols
• Dynamic routing e.g. Routing Internet Protocol
  (RIP)
• Routed protocol provided by IP which is a
  connectionless protocol
• RIP responds dynamically to a fault in the
  networks

24
RIP

• Before fault
• Routing tables updated every 30 s

25
RIP

• Fault occurs
• At some time t (0 lt t lt 30s)

26
RIP

• Inconsistent state of the network
• t 30, 60, 90, 120, 150 s

Routing tables
27
RIP

• State of the network after fault recognized
• t gt 180s

28
RIP

• Network has responded dynamically to fault
• We want the network to respond in a similar way
  to congestion
• but faster!

29
TCP congestion control

• Congestion occurs when too few network resources
  are available
• QoS is redundant if there is never any congestion
• Relies on all users behaving socially
• Aims to share resources equally amongst hosts

30
TCP congestion control

• TCP can only use forward implicit congestion
  notification
• TCP predicts and guesses network state at certain
  time
• Reduces transmission rate if guesses congestion

31
TCP congestion control

• Imagine 10 TCP flows (each 10Mbps) through 10
  Mbps bottleneck link
• Each flow averages 1 Mbps

32
TCP congestion control

• Now add a 10Mbps UDP stream
• UDP (User Datagram Protocol) is a connection less
  service built upon IP
• There is no flow to manage!

33
Social stand points

• Classic IP
• no differentiation between traffic or users
• relies on the social conscience of users, backed
  by peer pressure and ultimate threat of
  disconnection
• aims to share out resources equally regardless of
  cost/use

34
Social stand points

• QoS
• differentiates between traffic
• differentiates between users
• users need to be policed
• aims to make a profit by allocating scarce
  resources to premium customers

35
Present QoS Integrated Services Network
(IntServ)

• Refers to a body of work produced by Internet
  Engineering Task Force (IETF 1994b draft)
• Designed to provide a set of extensions to the
  best-effort traffic delivery model
• Provide QoS guarantees on a per-flow basis
• - A flow is defined as a set of packets
  associated with a single application sharing
  common requirements

36
IntServ

• Application requirements - Real time
• Tolerant real time
• can tolerate some extra induced jitter, loss or
  lateness of data still producing a reasonable
  signal quality when played back, e.g. packetized
  audio and video streaming applications
• Intolerant real time
• unacceptable signal quality if more than allowed
  minimum jitter, loss or lateness, e.g. two-way
  telephony applications and circuit-emulation

37
IntServ

• Application requirements
• Elastic applications
• Waits for packets to arrive before processing
• Interactive burst, e.g. telnet, chat
• Interactive bulk, e.g. ftp
• Asynchronous, e.g. email

38
IntServ

• Service classes
• guaranteed service
• packets never arrive late
• early packets are buffered
• controlled load service
• applications as if on lightly loaded networks
• better than best effort service

39
IntServ Mechanism

• Use of controlled load service
• User asks the network to provide a connection
• User informs the network about what is going to
  be injected into the network with a flowspec
• Admission control
• Network decides when to say NO to a service class

40
IntServ Mechanism

• Signaling
• exchanging information related to requests for
  connection, flow specs and admission control
  decisions between users and network components
• Packet scheduling
• management and scheduling of packets in queues
  in the switches and routers

41
IntServ Mechanism

• Flowspecs
• Tspec describes the flows traffic
  characteristics, more complicated
• Rspec describes the service requested from the
  network, simpler to describe

42
IntServ Mechanism

• Admission control
• exerted by the network components on new flows
  by monitoring their Tspecs and Rspecs
• ensure no degradation in service quality of old
  flows
• says Yes or No to new flows

43
IntServ Mechanism

• Admission control depends on various factors
• Type of request, e.g. urgent
• Type of service, CBR or bursty
• Queuing discipline, e.g. FIFO, FQ, WFQ,
  pre-emptive?
• Admission control policy
• simply discard the request
• allow the flow but with a note non confirming
  which means may drop the flow in case of
  interference with other flows in progress

44
IntServ Mechanism- Signaling

• Resource Reservation Protocol (RSVP)
• developed by IETF IETF 1997f draft
• RSVP is a signaling mechanism and QoS control
  information is the signal content
• does not specify routing but is designed to
  interoperate with existing IP routing protocols
• receiver oriented approach built on
  connectionless service
• robustness achieved by periodic refreshing of
  soft state in the routers

45
IntServ Mechanism and RSVP

• Simple reservation scheme (unicast)

46
IntServ Mechanism and RSVP

• Multicast Reservation
• PATH
• PATH
• 
  RESV
• 
  (merged)
• 
  
  
• 
  
  RESV
• 
  
• 
  
  RESV

Sender 1
Sender 2
Receiver A
Receiver A
47
IntServ Mechanism - Packet scheduling

• Packet classification
• associate each packet with the appropriate
  reservation
• Examining SA/DA, protocol number, SP/Dport and
  flow level field (in IPv6)
• Managing the packets in the queue in order to
  provide proper QoS as agreed upon, called
  scheduling

48
IntServ Queue management

• Queuing is the technique of storing packets using
  some discipline (FIFO, FQ, WFQ, CBQ) for
  subsequent transmission on a link
• Components of a queuing system
• Buffer Length
• 
  
• Arrival Process
  Departure
  Process
  
• 
  Queuing Discipline

49
IntServ Queue management

• FIFO queuing
• also called First Come First Serve (FCFS)
• highly efficient when no resource constraint and
  adequate levels of switching capability
• inefficient when load increases resulting
  congestion in the network
• pushes all responsibility for congestion control
  and resource allocation out to the edges of the
  network thereby dropping packets
• does not discriminate between different traffics

50
IntServ Queue management

• Fair queuing (FQ)
• maintain a separate queue for each flow handled
  by the router
• queues are served in a round-robin manner
• each of the flows contained in the queue are
  served fairly
• designed to be used in conjunction with an
  end-to-end congestion control mechanism
• any bandwidth that is not used by one flow is
  automatically allocated to other flows
• FQ provides guaranteed minimum share of bandwidth
  to each flow with a possibility to get more
  bandwidth

51
IntServ Queue management

• Fair queuing
• Flow 1
• Flow 2
  
  Outgoing link
• Flow 3
  Router selects
  the flows
• 
  in
  round-robin fashion

52
IntServ Queue management

• Weighted Fair queuing
• simple variation to FQ called Weighted Fair
  Queuing (WFQ)
• each flow is associated with some weight
  specifying how many bits to transmit from the
  queue
• for simple FQ, weight is 1, indicating 1 bit
  to transmit from each queue each time around
• router must learn how much weight to assign
  to each queue either manually or through
  signaling
• can be implemented using RSVP

53
IntServ Queue management

• Class Based Queuing (CBQ)
• Output queuing in which traffic allocation to
  several possible queues
• groups traffic into classes and assigns a
  ratio or metric to each class
• offers the benefits of traffic equity and
  prioritization without bandwidth starvation
• increased level of resource allocation to the
  higher- precedence queues and a relative
  decrease to the lower-precedence queues
• primitive method of differentiating traffic
  into various classes of service

54
IntServ reference model (IETF 1994b)

• Background
  Process
• Traffic Input
  
  Output Queues
• Forward Driver
  Packet Sceduler
• Forwarding Table
• 
  
  Output Driver

Routing Agent
Reservation Setup Agent
Management Agent
Admission Control
Routing Table
Traffic Control Database
Classifier
55
IntServ Summary

• IntServ is an end-to-end model
• Better than best effort service model
• Scalability is a major issue
• Every flow passing through a router needs
  reservation in terms of a soft state which is a
  problem as number flow increases
• the states must be refreshed periodically as long
  as the flow is to be maintained

56
DiffServ - Differentiated Services network

• Effort by IETF to provide end-to end QoS
• Result of reaction against the scalability
  problem faced by IntServ/RSVP
• Aims to provide differentiation among variety of
  traffics by aggregating flows into a small number
  of groups
• Eliminates storing soft state information about
  individual flows in the router

57
DiffServ

• IPv4 Packet header format

58
DiffServ

• DiffServ Code point (DSCP)
• Use of ToS field of IPv4,Traffic class field of
  IPv6
• Defined in early 1980s, but was largely unused
  until introduction of prioritization into IP
  header
• 0
  5
  7
• CU Currently Unused
• Pool Codepoint space Allocation
• 1 XXXXX0 Standard
  activity
• 2 XXXX11
  Experiment/local action
• 3 XXXX01
  Experiment/local action

DSCP
CU
59
DiffServ

• A DiffServ domain (DS) is defined as a set of
  contiguous DS compliant networks having DS
  compliant nodes
• DS is implemented in individual routers by
  queuing and forwarding packets based on the DSCP
• DS is not based on priority, application or flow
  but on possible forwarding behavior of packets
  called Per Hop Behavior (PHB)

60
DiffServ

• PHB defines the set of rules for a particular
  class of traffic
• Does not require a particular queuing discipline
• 6-bit DSCP identifies a particular PHB to be
  applied to a packet

61
DiffServ

• PHB defines behavior of individual routers rather
  than end-to-end services
• PHB provides a particular service level
  (bandwidth, queuing and dropping decisions) in
  accordance with network policy
• 2 PHBs are under active discussion with the
  DiffServ working group EF and AF
• - Expedited and Assured Forwarding

62
DiffServ

• Expedited Forwarding (EF), Premium service
• meant for traffic sensitive to delay and loss
• router guarantees performance if arrival rate of
  packets less than the forwarding rate
• Implementation strategy
• strict priority over all other packets
• use WFQ with highest weight given to EF
• drop out-of-profile packets
• need for admission control and traffic shaping

63
DiffServ

• Assured Forward (AF)
• more complex than EF
• AF divides traffic into four different classes in
  three different classes of service (Gold, Silver
  and Bronze)
• Drop Pref. Class 1 Class 2
  Class 3 Class 4
• Low 010000
  011000 100000 101000
• Medium 010010 011010
  100010 101010
• High 010100
  011100 100100 101100
• During congestion, router may discard packets
  based on their drop preferences

64
DiffServ

• Implementation strategy
• Some minimum amount of bandwidth and buffering
  reserved for each class
• resource management for different classes needs
  to done beforehand looking at the demands of
  various classes of traffic
• Always marked as non-conformant in policing rule

65
DiffServ

• DiffServ Architecture
• DS Domain 1
• 
  
  
• 
  
  DS Domain 2
• Edge Routers mark packets
• (Ingress Router)
• 
  Core Routers only forward
  Egress Router

66
DiffServ

• A service Level Agreement (SLA) between different
  domains established
• Both inbound and outbound packets in a DS domain
  are marked according to SLA and traffic
  conditioning agreements
• DS may trigger accounting mechanism at network
  boundaries to track each service usage for
  quality level monitoring and billing purposes
• DS is rule based, hence good for policy based
  network management

67
DiffServ

• Traffic entering into a DS domain are conditioned
  prior to assignment of DSCP
• Traffic classification and conditioning at the
  Edge Router (Ingress)
• Packets
  Forward
• 
  
  Drop
  
  
  
• 
  
• 
  
  

Meter
Classifier
Marker
Shaper/Dropper
68
DiffServ

• Resource management challenges
• How to decide what users get which type of
  service?
• Where to implement bandwidth sharing policy?
• Who will ensure proper resource management?
• Solution
• many (ongoing research) but we will emphasize
  only one gt Bandwidth Broker (BB)

69
DiffServ

• Bandwidth Broker (BB), initially proposed by Van
  Jacobson
• BB is a logical entity residing in each
  administrative domain
• Internally keeps track of QoS requests from
  individual users and applications and allocates
  resources according to some policy (intra-domain)
• working as an agent between BBs of neighbor
  domains for setting up maintaining bilateral
  agreement (inter-domain)

70
DiffServ

• BB also manages resources for each class
• by keeping track of current allocation of marked
  traffic
• interpreting new requests in the light of
  policies and current allocation
• Configures edge routers to deliver a particular
  service to flows

71
DiffServ

• Bandwidth
  Brokers

Domain 1
Domain 2
BB
APP
APP
BB
72
DiffServ Summary

• DiffServ provides a scalable solution for QoS
  over the Internet
• Provides coarse-grain end-to-end QoS
• Can not provide per-flow (fine-grain) QoS
• Edge routers are configured to classify and mark
• Bandwidth Brokers play central role at the
  control level of DiffServ architecture

73
Conclusion

• Is classic IP good enough today?
• No lt Resource hungry applications
• Solutions
• IPv6
• All optical core gt no intelligence in core
  routers
• IETF QoS effort, Internet2 QBONE
• and what else . . . ?

74
Conclusion - QoS Song

• We dont need no reservation
• We dont need no admission control
• All applications must be adaptive
• The Net works just fine, so leave it alone
• Hey! Professor! Leave the Net alone!
• We dont need no traffic management
• Over-provision bandwidth for all
• The only true god is TCP/IP
• The Net isnt broken, so leave it alone
• Hey! Professor! Leave the Net alone!
• All we want is flat rate pricing for all
• Keshav An Engineering Approach to Computer
  Networking

75
References

• Paul Ferguson, Geoff Huston, Quality of Service
  Delivering QoS on the Internet and in Corporate
  Networks, John Wiley Sons, Inc.
• Larry L. Peterson Bruce S. Davie, Computer
  Networks A systems approach, Morgan Kaufmann,
  Second edition
• S. Blake, D. Black, M. Carlson, E. Davies, Z.
  Wang and W. Weiss, An architecture for
  differentiated services, RFC 2475, December 1998
• http//www.seas.upenn.edu/ross/book/emerge/diffse
  rv, Differentiated services
• W. Stallings, Differentiated Services, CSD
  February 2000
• http//www.allot.com/html/products_white.shtm,
  Policy based network architecture
• A.Terzis, L.Wang, L.Zhang, A scalable Resource
  management framework for Differentiated Services
  Internet, Internet draft

76
Thank You All !!!!