The long, interesting tail of Indie TV

1
The long, interesting tail of Indie TV

• Daniel Cutting, Aaron Quigley, Björn Landfeldt
• CTSB Workshop, Pervasive 2006, 7th May 2006

2
Indie TV

• Producing video content is now easy and cheap
• More publishers and more niche content
• Already specialised TV channels on the web
  appealing to niche audiences
• Sail.tv, Democracy TV, YUKS TV
• Logical conclusion is a tailored channel for each
  viewer based specifically on their interests

3
Indie TV

• Indie TV has 3 components, Creators, the
  Disseminator, and Blenders
• Content is produced by creators who describe its
  audience in terms of interests
• E.g. a dramatic thriller is destined for an
  audience interested in drama and thrillers
• The Disseminator delivers the content to this
  audience
• The Blenders combine content as it arrives for
  playback
• We focused on the dissemination aspect only

4
Indie TV
5
Dissemination via implicit groups.

• Explicit groups
• Viewers named
• Pre-defined by creator or viewers need to join
• Kim, Julie

• Implicit groups
• Viewers described
• Creator defines on the fly, viewers dont need
  to join
• Drama Thriller

6
Implicit group messaging.

• Multicast messages from any source to any
  implicit group at any time in a P2P network
• Each peer described by interests, e.g. Drama,
  Sci-Fi
• Implicit groups are specified as logical
  expressions of attributes, e.g. Drama AND
  Thriller
• System delivers messages from creators to all
  viewers matching target expressions
• Iterative design process
• Theoretical, implementation, simulation,
  theoretical

7
Initial theoretical model.

• A fully distributed, structured overlay network
• Peers maintain a logical Cartesian surface (like
  CAN)
• Each peer owns part of the surface and knows
  neighbours
• Peers geometrically ROUTE to locations by passing
  from neighbour to neighbour

8
Initial theoretical model.

• Peers locations on the surface determined by
  their attributes

Benoit Action, Thriller
Kim Drama, Sci-Fi, Thriller
Julie Drama, Thriller, Romance, Action
9
Initial theoretical model.

• Can calculate all regions on the surface where
  the matching viewers must exist
• Multicast content from creators to the regions
  matching the audience description

10
Initial implementation.

• OMNeT/INET simulation of a real network
• The simulation raised some concerns we had not
  considered in the initial design
• The overlay hop between peers on the surface
  resulted in many IP hops at the network layer
  which led to extremely long end-to-end delays
• The design was adequate for large/medium implicit
  groups but required too much overhead for small
  groups

11
Revised theoretical model.

• The simulation led us to revise the model taking
  these problems into account
• To counter the latency problem, we stored
  pointers to the peers on the surface, rather than
  locate the peers there themselves
• This allowed us to have peers that were
  physically close to be close on the surface,
  regardless of their attributes
• To counter the group size problem, we introduced
  a hybrid approach
• Smaller groups used a distributed index to find
  members
• Initial model retained for large groups

12
Distributed index.

• Every peer registers at a rendezvous point (RP)
  for each of its attributes
• Every registration includes IP address and all
  attributes

13
Distributed index.

• To CAST, select one term from target
• Route CAST to its RP
• RP finds all matches and unicasts to each

14
Evaluation.

• New implementations performance was vastly
  better
• Delay was greatly reduced and within required
  limits
• Overall network peer and link stress was also
  reduced, especially when delivering content to
  small or empty groups (load was now proportional
  to group size)

15
Conclusion.

• We had an elegant theoretical model to begin with
• But, abstracted details of the system too much
• A structured overlay network has to be based upon
  physical computer network with peers, routers,
  fast and slow network links
• The possibility of highly variable group sizes
  had been similarly neglected
• Implementing the simulation brought these
  problems to the fore and allowed a quick revision
  of the theoretical model

16
Questions?