An Architecture and Communication Protocol for Picocellular Networks

One-line summary: As cell size decreases, the number of handoffs and therefore handoff overhead within a wireless picocellular networks increases; a method for reducing handoff overhead and buffer space required using multicast groups and mobile trajectory prediction is described, as well as a network architecture that supports this technique.

Overview/Main Points

• Picocells: Picocells (indoor, O(10m)) offer greater frequency re-use, lower power transmitters, and higher throughput. There will also be many more handoffs between cells; overcoming the overhead associated with handoffs therefore is the critical design element in a picocellular network.
• "Dumb" Handoff: The wrong way to do handoff is to require old basestations to forward data received during handoffs for open connections to the new basestation. The number of forwarded messages can be shown to be NxMxR, where N is the number of packets that arrive at the old BS during and after a handoff, M is the number of mobile users per cell, and R is the number of cells in the network.
• "Smart" Handoff: BS's (called mobility support stations or MSSs in this story) no longer retain intelligence, but rely on a centralized authority called a supervisory host (SH). The supervisory host calculates the mobile's likely trajectory, and forms a multicast group for MSSs that the mobile is likely to handoff to in the near future. All packets are multicast to this group; MSSs that do not currently host the mobile buffer packets in anticipation of a handoff; such buffered packets are tossed when the MSSs receive an update of the mobile's ackowledged sequence numbers.
• Delayed Buffering: By keeping track of how long a mobile has spent in previous cells, the SH can guess how long the mobile is likely to remain in the current cell. The SH therefore can let MSSs in the mobiles multicast group know when they should start buffering in anticipation of the handoff. This reduces the buffer space required at the MSSs, but increases handoff delay when the SH's guess is wrong.
• Network Architecture: A connection-oriented network architecture is also described in this paper. Virtual circuits are set up between end-points. The communication is optimized for the case of two mobiles in the same network region managed by a one SH (the ACTIVE_LOCAL connection state). Communication with a mobile under a foreign SH is in the ACTIVE_REMOTE state. POINT means that a SH is forwarding packets for a mobile that moved out of its network region. ACTIVE_LOCAL destinations are easy to find since the common SH is omniscient. ACTIVE_REMOTE destinations are found by broadcasting LOCATE packets. Finally, a NASCENT connection is one in the process of being established.

Relevance

Flaws

• The network architecture described is centralized, connection oriented, and has broadcast route-finding as a common case. These are all bad properties (IMHO) for a wireless network - the handoff work in the paper is much more interesting than the network architecture work.
• The buffering approach was justified with some rather dubious analytical calculations. The queuing model used for buffering packets is known to be weak. The numbers chosen were totally unbased in reality. The probability distributions used were ad-hoc. No measurements of any real-life working implementation were presented. (Possibly because there is not an implementation in existence?)
• From the paper: "During our experiments we noted that, on occasion, packets do not necessarily get delivered to an MH in either the current cell or even the next one..." They don't talk about fading or packet loss at all in this model, nor any other known bad property of wireless links.