Quality of Service - applications

1
Quality of Service - applications

• Henning Schulzrinne
• with Wenyu Jiang
• Dept. of Computer Science
• Columbia University
• NSF QoS workshop, April 2002

2
Application requirements

• What kind of applications?
• What matters?
• What doesn't?
• How do we know?

3
Applications the diminishing set (or the cynics
view of QoS)

• video-on-demand ? 80 GB disk P2P-over-TCP
• voice-over-IP ? where?
• multimedia conferencing ? still waiting, after 50
  years...
• network games ? latency kills

4
Why is QoS unpopular?

• need to admit failure bandwidth too cheap to
  meter
• undemocratic some traffic is more equal than
  other
• reminds you of your mom no, you cant have that
  10 Mb/s now
• socialist administer scarcity - we like SUVs (or
  to drive 100 mph)!
• risky scheme security (DOS)
• only displacement applications (such as
  telephony) need QoS
• requires cooperation edge-ISP, transit ISPs, end
  systems
• snake oil add QoS, lose half your bandwidth

5
QoS evaluation

• Problem no tool for reliably evaluating quality
  end-to-end
• network (PSTN Internet) hardware operating
  system
• depends on packet loss, jitter (playout buffer),
  FEC, loss correlation
• need subjective evaluation, but too expensive
• applicability of objective measures (PSQM, etc.)
• use speech recognition as measure
• e-model

6
Voice quality vs. packet loss
7
MOS vs. loss for FEC and LBR
8
Myth TCP loves lossy networks
Lakshman/Madow/Suter ToN 2000
9
Rrel as Universal MOS Predictor

• Mapping from relative recognition ratio Rrel to
  MOS

10
Human Recognition Results

• Listeners are asked to transcribe what they hear
  in addition to MOS grading.
• Human recognition result curves are less smooth
  than MOS curves.

11
Research infrastructure

• After 10 years, still no low-latency audio
  research application
• existing applications (rat/vat) have erratic
  latency behavior
• no good access to audio queuing information for
  lip-sync
• no good diagnostic tools ? blame network

12
QoS is about reliability

• cant sell premium service thats unavailable one
  day a year
• auxiliary cost of failure people scheduling,
  interruption, embarrassment, ...
• consistent 5 packet loss is much better than 5
  probability that network is unavailable for
  seconds
• BGP convergence time minutes
• conjecture applications (conferencing, VoIP)
  don't migrate for lack of predictability