CS268 Reading Review

Improving Reliable Transport and Handoff Performance in Cellular Wireless Networks

Hari Balakrishman, Srinivasan Seshan, and Randy H. Katz
Review by Feng Zhou
2/2/2003

The problem: Two problems are addressed by the paper. One is the poor performance of TCP in lossy wireless networking environment. The reason of the problem is the assumption made by TCP that packet loss is due to congestion does not hold in wireless networks, in which data corruption is common. The other problem addressed is that long latency of handoff between base stations makes TCP perform poorly or unreliably.

Key points:

  1. The authors choose to solve these problem in a way that mixes network layer and tranport layer changes. The design space is tight, because we can only expect to have control of the base stations and mobile hosts. Compared to the alternative of doing retransmission in link layer, changing the network layer and part of the transport layer can potentially have better performance, because more information is available in the upper layers. However, it is often undesirable to change the end-to-end property and "snooping" information of the upper layer, which is exact what the snoop module does. These changes incur the maintaining of more information in the network and impede future changes of end- to-end transport protocol. Nevertheless, the problems discussed are important and solutions are effective, although not perfect.
  2. The snoop module sits on base stations and monitors each TCP connection. By "understanding" the TCP traffic, it differentiates between losses due to congestion and those due to data corruption. Then it retransmits locally those packets lost due to data corruption. The authors show that this can be done effectively (20-fold throughput increase) and efficiently (using zero-copy buffering in the base station). On the other hand, the mobile hosts are modified to implement SACK and exchange SACK packets with the snoop module, to enable both of them to detect packets lost due to data corruption and fast retransmit them in both directions.
  3. The approach of letting nearby base stations buffer data for fast handoff is useful. However, using IP multicast to transfer the data is probably not a good idea. This is require all routers between the home agent and all base stations implement IP multicast, which is not very feasible. However, if we are not implementing wide-area mobility, then we can let the home agent (probably the local IP router) send data to the base stations of interest. This can be done efficiently because the base stations are within local administration and the home agent (router) probably knows about them.