The Reality and Mythology of QoS and H.323

1
The Reality and Mythology of QoS and H.323

• pschopis_at_itecohio.org
• pcalyam_at_oar.net

2
Overview

• H. 323 bounds testing
• QoS models
• Implications of applying models
• Engineering to need.

3
Test Motivation

• Abilene is trying to provide DiffServ EF
• Is H.323 suitable candidate for APS?
• DiffServ lacks hard bounds, it is totally
  probabilistic.
• What really would help? What are the performance
  bounds?

4
Video Artifacts

• Spatial Augmentation Video artifacts are added
  to the picture. Objects appear that are not in
  the captured video such as video tiles.
•  Spatial Depreciation Parts of the picture or
  objects in the picture are missing.
•  Temporal Distortion Over time the flow of an
  event is distorted by missing data, in mild cases
  resulting in an inter-frame jerkiness. In more
  severe cases resulting in video freezing.

5
Video Artifacts

•  Audio Augmentation Audio artifacts added to
  audio stream such as pops, clicks and hiss.
•  Audio Depreciation Parts of the audio are
  missing.

6
Scope of H.323 Bounds Testing

• What network conditions can be mapped to
  certain qualities of video.
• It can be highly subjective.
• We did not desire to engage in a Cognitive
  Science experiment.
• Needed simple reproducible test procedure.

7
Test Procedure

• Still office scene, count the number of defects
  over a 60 second sample.
• Motion in scene and count the number of seconds
  needed to recover.
• Tested in a variety of setups
• Point-to-point
• MCU
• Cascaded MCUs
• Isolated Latency, Loss and Jitter

8
Network Emulator

• Operating System Linux Mandrake 7.2 Kernel
  recompiled and optimized for the device to be a
  router.
• CPU Pentium III 733Mhz
• Memory 256 MB.
• Motherboard Asus CUSLC2-C AGP4X
• NICS Intel Etherpro 10/100.
• Emulator Software Nistnet 2.1.0

9
Used to test H.323

• Verified Nistnet system prior to test.
• Tested platform with SmartBits.
• All parameters were met with in a /- 1 msec
• (Actual resolution .5msec)
• With SmartBits we could verify switches etc. to
  further validate our findings. Worst case is
  total accuracy within /- 3msec.

10
Point-to-Point tests

• Latency does not matter. (holds true for all
  scenarios)

11
(No Transcript)
12
(No Transcript)
13
(No Transcript)
14
(No Transcript)
15
(No Transcript)
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21
End-to-end Delay Components
SENDER SIDE
NETWORK
RECEIVER SIDE

• Compression
• Delay
• Transmission
• Delay
• Electronic
• Delay

Propagation Delay Processing Delay Queuing
Delay
Resynchronization Delay Decompression Delay Pre
sentation Delay
22
Delay Values

• Transmission Delay Electronic Delay
• Modem delay 40ms
• Transmission delay 10 chars over 56Kbps
• 80/56000bps 1.4ms
• Switch Propagation Delay lt2ms
• Presentation Delay 17ms

23
Encode and Decode Latency
SWITCH
END POINT 1
END POINT 2
MIC I/P
AUDIO O/P
MCU
MCU
METRONOME (PULSE GENERATOR)
A
B
OSCILLOSCOPE
SCOPE I/P A METRONOME I/P SCOPE I/P B ENDPOINT
2 AUDIO O/P
24
Oscilloscope Waveforms
25
Experiment and Results

• Dialing Speeds 256K, 384K, 512K, 768K
• Metronome setting 113
• Propagation delay Switch delay 0
• Encode Decode delay 240ms
• (independent of dialing speed)
• Delay through MCU 120ms to 200ms
• (delay increasing with dialing speed)

26
Network Requirements

• Latency users may find annoying but the it does
  not break the protocol.
• Loss Can tolerate some loss, must be below 1
  in p-2-p and 0.75 in MCU
• Jitter Very jitter intolerant. For 30 Fps must
  be lower than 33 msec. Seems very intolerant in
  cascaded MCU scenario.

27
Network Calculus 101

• All functions are cumulative distribution
  functions, i.e. wide-sense increasing.
• Uses min-plus Algebra.
• Uses classes of primitive functions to describe
  various network behaviors
• Employs convolution and deconvolution with
  primitives to arrive at meaningful conclusions.

28
Models

• IntServ Has the necessary per flow state but is
  not here yet.
• It also probably has many unforeseen maintenance
  and administrative issues. (see next section).
• Experience from ATM SVCs suggests many
  scalability issues. Possible solutions include
  MPLS or Policy routing.

29
Models

• Any E-2-E solution has scalability problem in the
  sense that in packet switched networks the
  solution vector is more than number of hops and
  delay etc.
• x-gt lt Ax-gta-gt
• In other words it is also a function of topology.
  (More in DiffServ).
• .Source Network Calculus A Theory of
  Deterministic Queuing Systems for the Internet
• by Jean-Yeves Le Boudec Patrick Thriran,
  Springer-Verlog, Berlin Heidelberg, 2001.

30
Models

• DiffServ lacks the per flow state necessary for
  tight performance bounds because..
• ß1(t) ß(t)- a2(t) Where ß is the
  rate-latency function. ßR,T(t) Rt-T i.e.
  Service Curve.
• b1 b1 r1T r1(b2r2T/R-r2) Where b is a
  component of the Affine function ? r,b(t) brt
  if tgt0.
• Source Network Calculus A Theory of
  Deterministic Queuing Systems for the Internet
• by Jean-Yeves Le Boudec Patrick Thriran,
  Springer-Verlog, Berlin Heidelberg, 2001.

31
Models

• V 0.564 for bounded delay so when v0
• converges to V the latency bound explodes to
  infinity. For vl Si?m ri/Cl. Where
• v link utilization, iflow, r rate and C
  service rate.
• Source Network Calculus A Theory of
  Deterministic Queuing Systems for the Internet
• by Jean-Yeves Le Boudec Patrick Thriran,
  Springer-Verlog, Berlin Heidelberg, 2001.

32
Engineering to the need

• What realistically can we do?
• It depends on ones network.
• Appropriate queuing for congested links for maybe
  a single to only a few flows.
• Packet shaping on receiver with a Greedy Packet
  Shaper.
• GPS will not increase latency or buffering
  requirements if and only if network was
  previously lossless.