Minimising Perceived Latency in AudioConferencing Systems over ApplicationLevel Multicast

1
Minimising Perceived Latency in
Audio-Conferencing Systems over Application-Level
Multicast

• By
• Nick Blundell and Laurent Mathy
• n.blundell, laurent_at_comp.lancs.ac.uk
• Nicks website http//www.nickblundell.org.uk

2
Introduction

• First, we will look at the problems of group
  communication over the Internet, in the setting
  of audio-conferencing systems.
• Next, we will consider patterns of conversation
  observed in the fields of conversation and
  discourse analysis.
• Then, we will introduce our conversation-adaptive
  approach to routing over application-level
  multicast (ALM), building on the observed
  conversation patterns and on user perception.
• Then, we will discuss simulation results of the
  proposed ALM routing protocol and future
  directions of the work.

3
The Problem
?
?
?
?
?
The Internet
?
?
?
?
?
?
?

• How do our audio-conference participants achieve
  group communication?

4
Group Communication Method Network Multicast
The Internet

• Efficient duplication of packets performed by the
  network.
• Unicast latency observed by all participants.
• Sadly, not widely deployed in the Internet
  (typically only academic networks are peered
  through tunnels with IP multicast).

5
Group Communication Method Multiple Unicast
The Internet

• A naïve approach to ALM, placing high stress on
  the transmitting nodes network.
• All participants observe unicast latency that
  is, until the network becomes saturated and
  packets get dropped.

6
Group Communication Method Conference Server

• A transmitting node sends a packet to a
  conference server, which replicates the packet to
  the other members.
• This effectively moves the problem of multiple
  unicast from the transmitting node to the server.
• For Internet audio conferencing, where should the
  server be placed to fairly serve participant
  nodes?

7
Group Communication Method ALM

• Nodes arrange themselves into one or more overlay
  trees, mapping the overlay as closely as possible
  to the underlying network.
• Packet duplication is then shared among the
  participants, who flood data packets over the
  tree through their children.
• Inevitable imbalance in overlay stretch between
  participant node pairs.
• Distributed organisation leads to good
  scalability.

8
Observations of Conversation Patterns

• Conversation is typically structured into related
  participant turns termed adjacency pairs (i.e.
  questions and answers) (Sacks, 1992).
• Adjacency pairs lead to localisation in
  conversation (i.e. a small number of present
  participants exchange turns over a temporal
  interval of time) (Sacks, 1992).
• Such localisation results in a strong correlation
  between the previous few participants to take
  turns and the next participant to take a turn,
  making the next speaker highly predictable.

Our analysis of next-speaker predictability.
9
User Perception of Latency

• As latency in the communication channel
  increases, participants become confused,
  miss-interpreting high latency as extended pause
  in their interlocutors speech (Mehta, 2001).
• Generally accepted levels of user tolerance of
  latency are (Mehta, 2001)

10
User Perception of Latency
When a pair of participants are in discussion.
When a participant only listens to another
participant.
11
Hypothesis Dynamic Application-Adaptive Routing

• Our hypothesis
• If the overlay latency is minimised between nodes
  that host currently-talking participants ---
  those participants who are most sensitive to
  latency --- and is allowed to increase for those
  participants who are currently listening only
• participants will perceive an unchanging, low
  latency, comparable to network-level multicast.
• participants will become less sensitive to
  scaling of the ALM group than they otherwise
  would be for standard, non-adaptive tree flooding.

12
Introducing ALNAC Application-Level Network
Audio-Conferencing Routing Protocol

• Rather than nodes simple flooding packets down
  the overlay tree through their children,
  transmitting nodes send audio-frame packets
  directly to a number of previously-speaking
  participant nodes.
• As a result, a nodes children may be deprived
  from directly receiving the audio frames and so
  are supplied, eventually, through a process of
  per-packet delegation.
• We also ensure that, for basic operation, no
  loops are formed in the delivery paths of
  packets.
• Lets look at an example

13
Introducing ALNAC Application-Level Network
Audio-Conferencing Routing Protocol
14
Introducing ALNAC Application-Level Network
Audio-Conferencing Routing Protocol
15
Introducing ALNAC Application-Level Network
Audio-Conferencing Routing Protocol
16
A Closer Look at Delegation
17
Simulation Results
42-speaking participants in a 200-participant
audio-conference
18
Simulation Results
42-speaking participants in a 200-participant
audio-conference
19
Conclusions

• We have looked at existing methods for group
  communication, used for audio conferencing in the
  Internet.
• We have argued the benefit of adaptive routing
  techniques for Internet audio conferencing,
  leveraging on patterns observed in conversation
  and on a users perception of latency in the
  audio channel.
• We have simulated, and shown the adaptation and
  impact, of the proposed ALM routing protocol.

20
Further Work

• We are currently implementing an ALM proxy that
  will allow existing VoIP applications to use the
  adaptive protocol, along with other ALM overlay
  and routing protocols for subjective comparison
  of perceived audio quality.
• From subjective, user experiments, we can
  determine the true gain in scalability of the
  ALNAC routing protocol over existing,
  non-adaptive routing techniques.
• Following on from this work, we plan to consider
  the user of distributed audio mixers to further
  constrain bandwidth usage of the ALNAC protocol
  to support low-bandwidth clients.

21
Questions
22
References

• H. Sacks. Lectures in Conversation. Blackwell,
  Oxford, UK, 1992
• Princy C. Metha and Sanjay Udani. Overview of
  VoIP, technical report MS-CIS-01-31, University
  of Pennsylvania, Feb 2001.

23
Our Characters