Modeling Wireless LInks for Transport Protocols

Summary:
This paper looks at a number of reasons that current models of wireless networks, and why they are inadequate. They show that modeling a duplex connection gives quite different results than one way traffic, which is the model often used in simulations.

The paper also highlights a number of features of wireless networks, how they effect the performance of transport protocol, what types of links they are present in, and how to model them. These features are:

• Error losses and corruption. These losses can make TCP think there is congestion when there isn't, leading to poor performance. While FEC and link-layer retransmission ease this somewhat, there can still be delays and re-ordering due to limited retransmit buffers at the end-points of the wireless links. This can be modeled by dropping packets, but we need to consider whether our losses are correlated or not.

• Delay Variation. Link delay can change suddenly in a wireless world due to mobility, or scheduling at the wireless end-points. If the TCP connection assumes this is due to congestion (i.e. it's using delay based congestion-control) then it will back off sending unnecessarily.
  

• Packet Reordering - This can be caused by link-layer recovery mechanisms and can cause retransmission from the end-host if it causes enough of a delay in delivery.

• On-Demand Resource Allocation - In cell networks a channel needs to be allocated if a client hasn't send any data for a threshold amount of time. This can cause delays. The authors show how to model this by introducing a delay when a packet is queued on an empty queue that has been empty longer than the threshold time.

• Bandwidth Variation - Different channels can have different bandwidths and clients might be moved from channel to channel. This can cause spurious timeouts in TCP if the RTT suddenly goes up.

• Asymmetry in Bandwidth and Latency: With an asymmetric link TCP can get congestion on the ACKs. This can easily be modeled by setting the right parameters.

The paper also looks at queue management and how drop-tail queuing can lead to delay and problems due to spurious retransmission timeouts. They argue that drop-tail modeling is okay for cell and WLAN links, but that RED is perhaps a better model for satellite links and for future WLAN links. Finally the paper looks at ways transport protocols could handle the various problems modeled in the paper.

Thoughts:
I can't say I loved this paper. It felt a bit like a laundry list of wireless link issues, which might be very useful to designers but wasn't particularly illuminating from a theoretical perspective. Some of the experiments showing the real problems with past models was interesting.

I could imagine a very interesting follow-up to this paper in which solutions to the various issues they have identified here are proposed and evaluated. This is an important starting point toward that paper.