More on Wireless ...

Two more papers discussing wireless technologies, Modeling Wireless Links for Transport Protocols [MODELING] and Architecture and Evaluation of an Unplanned 802.11b Mesh Network [ROOFNET].

The authors of [MODELING] were intent on showing the dismal state of modeling and analyzing existing wireless deployments and suggesting alternative, hopefully better, characteristics to use for modeling. I found this study to be a bit dry, but I did have a few of the take-away points:

• There is a definite tension between changing wireless links versus transport protocols. I actually alluded to this in a previous post where I mentioned the hurdle researchers have in the wireless space, because it is not clear where you need to make compromises for either the wireless protocols or the transport protocols.
• TCP is not robust in the face of new link layer technologies. IN FACT, it is rather ironic that TCP is sensitive to packet reorderings! The original design goals of networks were to be robust in the face of lost links, lost packets, or packets traveling down different paths ... yet this is clearly not the case with TCP. IN ADDITION, TCP performs poorly when corrupted packets are received ... TCP has been over-optimized for link layers that are perfect (like ethernet). IN ADDITION, spurious timeouts due to link layer retransmissions are a severe problem and the transport protocols (like TCP) inflate transmit timers because they are agnostic of these link layer retransmissions (again, it seems that TCP has been over-optimized for an ethernet like link layer).  
• The packet loss==congestion signal is inadequate in wireless networks because packet loss might be due to errors ... do we need more explicit signals?
• A possible solution might be to push information about the link layers into the transport protocols (and maybe even vice-versa as well) ... but doesn't this break the nice abstraction we had before?

The authors of [ROOFNET] showed how an ad-hoc mesh network could be deployed using 802.11b technology. Interesting take-away points:

• In the authors words, "the highest-throughput routes often involve links with relatively high link-level loss rates, since a single-hop route with 40% loss can deliver more data than a two-hop route with perfect links. In addition, because the 802.11b bit-rates are at roughly power-of-two intervals, a high bit-rate with up to 50% loss is preferable to the next-lowest bit-rate." Interesting!
• Fast short hops are the best policy (as determined by their routing algorithm Srcr).
• The majority of nodes use many of their neighbors ... I actually didn't quite understand this conclusion. Does this mean that the majority of nodes have constantly fluctuating best routes to gateways?
• Multi-hop has a throughput advantage over long low-quality single-hop because it can exploit lots of short high-quality links ... and the overhead of retransmissions appears to be low enough.
• Roofnet does not use RTS/CTS ... in fact they have shown that it does not improve performance in their network!

An interesting link between the two papers was that [MODELING] claimed that out-of-order delivery or corrupted delivery is shown to be beneficial for real-time applications by decreasing delay, and this is especially true when there are multiple wireless links in the path. However, because none of the connections tested in [ROOFNET] were of the real-time nature, they did not conclude the same thing nor did they show how a new transport protocol might have excelled in their network.