Modeling Wireless Links for Transport Protocols

Andrei Gurtov, Sally Floyd, "Modeling Wireless Links for Transport Protocols," ACM SIGCOMM Computer Communications Review, Volume 34, Number 2, (April 2004).

This paper reviews various models for cellular, WLAN and satellite links that are used in the design of transport protocols over wireless links and attempts to strike a balance between realism, generality and detail. The purpose of the models discussed in this paper was to evaluate the effect of link-level mechanisms on end-to-end transport protocols and higher layers.

The authors started off by highlighting the following essential aspect of models:

• Type of Wireless links: Cellular, Wireless LANs or Satellite? Each of these links have different bandwidth and link latencies and must be modelled accordingly.
• Topologies: It deals with questions like whether a single wireless link is located at the end or in the middle of a path, or multiple wireless links constitute the path. Further, the number and location of wireless links in the path greatly affects the performance and is an important modeling parameter.
• Wireless Traffic: It deals with questions like whether the traffic is one way or two way? What is the intrinsic nature of data being transferred over wireless links (web browsing data, GPRS traffic etc.)?
• Performance Metrics: This deals with the basic questions of what defines a good performance? Various performance metrics such as throughput, delay, fairness, dynamics or goodput may or may not be taken into account.

 Further, the authors presented ways to model specific link characteristics as follows:

• Error Losses and Corruption: Errors may result due to the fact that link layer only does 3 re-transmissions per data block. Further, they may be due to handovers or mobility of nodes. 
• Model: These can be modelled by dropping packets on a per-packet, per-bit basis or having a time-based loss probability. Bursty traffic should be modelled by Gilbert model with erroneous and error-free states. Another important point of distinction should be whether errors are due to data corruption or due to handovers.
• Delay Variation: This refers to the link delay (or delay spike) in the network. Apart from the inherent nature of the link, delay spikes can result from link-layer recovery, handovers, or scheduling. 
• Model: These can be modelled simply by suspending the data transmission on a simulated link. 
• Packet Reordering: This refers to reordering of the sequence in which the packets are sent. While being beneficial sometimes (in case of multiple wireless links in path), they can also incorrectly trigger packet retrasmissions and consequently congestion control responses for TCP.
• Model: These can be modeled by swapping 2 packets in a queue at any given time or delaying one packet and letting others pass it.
• On-Demand Resource Allocation: Wireless links often allocate channels based on availability of user traffic. This introduces allocation delays which must be accurately modeled.
• Model: On demand allocation delay can be modeled by introducing an additional delay when a packet arrives to a queue that has been empty longer than the channel hold time. 
• Bandwidth Variation:  Bandwidth variations may occur when a wireless scheduler assigns a high speed channel to a user for a limited time. 
• Model: This can be modeled by changing the bandwidth of a link using periods of low and high bandwidth or havning a sine pattern variation. 
• Asymmetry in Bandwidth and Latency: There is always assymetry in uplink and downlink directions mostly in Cellular link and none in WLAN links. Satellite links have significant asymmetry in bandwidth and  moderate asymetry in latency.
• Model: This can be modeled simply by having different parameters for bandwidth and latency for uplink and downlinks. 
  

Further, there are also issues of modeling queue management (accomodate bursty traffic while controlling queueing delay) and effects of mobility.  Queue management in cellular and WLAN links can be done by using a drop-tail buffer with configurable maximum size in packets while mobility is modeled by adequately defining as what the the protocol designer considers as mobility (intersystem Vs intrasystem) and subsequently modeling it thorugh mobility models such as linear, ping pong etc. Further, the authors highlight that there is always an ongoing tussle between the merits of changing wireless links or transport protocols. Due to this, there is not really a clear asnwer as to which layer should be implementing features such as bit errors, reordering or delay variations.

Critique

I found this paper to be more of like '101 ways on how not to model your wireless networks'. Majority of the paper consisted of identifying various flaws in existing models which may or may not accurately depict the shortcomings of previous models. It may very well could have been the case that the model worked well  in a specific scenario thereby highlighting the broader point that the model's authors would have tried to make. This reminded me of something that Prof. Joe Hellerstrein said in our systems class that many systems designers make a mistake of exposing so many gears to the user that in the end he has no idea as to which gear should he turn to do a specific job! I think that having such a generic model may create such a scenario with so many parameters to manipulate that in the end it may result in making the task of modeling much more difficult for those researchers who do not have a complete and exhaustive  knowledge of the system (which may be good or bad...).