Bob Braden

1
Internet Quality of Service -- Fantasy and
Reality

• Bob Braden
• USC Information Sciences Institute
• 30th Anniversary
• Sept 9, 2002

2
What is QoS?

• Qwoss -- No, its not an exotic vegetable...
• Fundamental network capability Allowing an end
  user (or collection of users, e.g., a campus or
  corporation) to control the attributes of
  communication service.
• It sounds like almost a no-brainer, but it is
  actually exceedingly difficult and a bit fuzzy.
• It cuts across many basic technical and economic
  aspects of the Internet.

3
Early History

• Internet designers called it Type of Service
  (TOS)
• 1981 Jon Postel defined Internet Protocol (IP)
  RFC 791
• He took a SWAG to define a TOS byte in the IP
  header
• 3 bits of priority (Mandated by DoD --
  precedence)
• 3 bits of TOS attributes
• Low delay?
• High throughput?
• High reliability?
• This sounded plausible, but ...

4
But...

• What do these 3 TOS attributes MEAN?
• How can routers implement them?
• How/when should they be set?
• Is this the right set of attributes?
• Need quantifiers?
• How can you prevent a tragedy of the commons?
• Administrative control (e.g., military/corporation
  ), or
• Economic control -- charging

5
Why do We Need QoS?

• To match network service to application
  requirements?
• Resolve conflicts when network is overloaded,
  e.g.,
• Interactive services want minimum delay
• File transfer email want maximum throughput
• Web users want both (!?)
• Or to make some users more equal than others?

6
Making Users Unequal

• Basic Internet design all users are equal!
• DoD said No way!
• Required service priority, linked to military
  hierarchy.
• ISPs want to sell premium service to corporations
  and government agencies that can afford it.
• Free-market approach -- QoS Knob.
• Download too slow? Move the slider up, increase
  your cost per minute.

7
Why is QoS Hard (1)?

• 1. How do you define the service?
• By its effect? Service attributes observable by
  users in end systems
• ISPs want to write contracts, called
  Service-level agreements (SLAs), and charge
  money.
• Or by mechanism? Specific queueing mechanisms in
  routers
• Users and providers care about effect, not
  mechanism -- need to pick a service model.

8
Service Model...

• User-observable attributes might be
• Reliable delivery of data? (How reliable?)
• Ordered delivery of data?
• Max bandwidth (measured over what interval?)
• Max end-to-end queueing delay
• Max jitter (delay variation)
• Which ones matter?

9
Effect vs. Mechanism

• Suppose your ISP says For an extra 10/month I
  will give your packets priority over packets from
  users who dont pay extra.
• Priority is a mechanism what would its service
  effect be?
• Its quite hard to connect effect ltgt mechanism
  many academic papers have been generated on this
  topic.

10
Service Model...

• Its quite hard to connect service ltgt mechanism
• Internet traffic does not follow any simple
  statistical laws it can be very BURSTY.
• Generally, to build a mechanism matching a useful
  service model requires traffic shaping/policing
  mechanisms to place a bound on the burstiness.
• Delay or drop non-conformant packets in each user
  stream
• Most common token bucket shaper/policer.

11
Why is QoS Hard (2)?

• 2. You can build mechanisms that operate strictly
  packet/packet, but to make any QoS guarantee
  requires per-flow state in routers.
• Violates the Internet religion
• Thou shalt forward IP datagrams using stateless
  routers.
• This religion provided simplicity, robustness,
  generality, and scalability -- not to be given up
  lightly.
• There are also resource and scalability issues.

12
Why is QoS Hard (3)?

• 3. QoS requires accounting/feedback (e.g.,
  charging) to avoid a tragedy of the commons.
• Many technical, business, legal, social
  problems...

13
A Case Study of QoS

• 1991 Internet research community believed that
  multimedia teleconferencing would become the 1000
  pound gorilla (killer app) on the Internet.
• (The Web had not happened yet!).
• ISI research had played a significant role...
• Packet speech experiments on ARPAnet Danny Cohen
  _at_ISI.
• VTC research _at_ ISI Steve Casner _at_ BBN.
• VTC technology -- the MBONE tools -- developed
  on DARPA Research Testbed Network (DARTnet) built
  operated _at_ ISI.

14
Realtime...

• Packet voice, and to a lesser extent packet
  video, require realtime service -- bounded E2E
  delay.
• The Internet research community set to work on
  the technical problems ...
• Developed Internet Integrated Service.
• Mostly DARPA-funded.
• 1994 Integrated Services in the Internet
  Architecture RFC 1633 Braden_at_ISI, Clark_at_MIT,
  Shenker_at_PARC

15
Internet Integrated Service (IIS)

• How define service?
• Two service models
• Guaranteed tight bound on E2E Q delay
• Controlled Load loosely defined good service
• How do users request IIS?
• RSVP signaling protocol to request and set up
  QoS state in routers.
• Initial design Zhang Shenker _at_PARC
• Prototyping and standardization Estrin, Braden,
  Berson, Lindell, Herzog _at_ ISI.

16
What Happened to IIS?

• The Web happened, and it became the Internet
  gorilla, not VTC.
• Some IETF opinion-leaders excommunicated IIS as
  heretical to the Internet religion.
• Microsoft bought into IIS
• Windows 2000 implements it.
• The RSVP signaling protocol component has been
  widely adopted/adapted to other Internet
  signaling applications.

17
What Happened?

• ISPs ignored IIS.
• No business case or other incentive.
• E2E IIS requires more collaboration than ISPs can
  muster
• Another problem, which I have no time to discuss
  Multicast!
• The payment problem is not much closer to
  solution.

18
What Happened?

• ISPs have pushed another QoS approach
  Differentiated Service.
• Redefine 6 bits of TOS byte to select 64 service
  classes.
• QoS within ISP clouds rather than E2E.
• Classify packets in boundary routers between
  ISPs, set these bits for use by the interior
  routers.
• They can charge large customers for preferred
  service.

19
Current Status of Internet QoS

• We know much more than we did 20 years ago about
  QoS service models and mechanisms.
• There is still no Internet-wide deployment of
  QoS.
• However, some forms of QoS are implemented by
  router vendors and are deployed in private
  intranets.
• The IP telephony gorilla is crashing in the
  underbrush ...