2008-09-30 Readings - Wireless Networks Overview and Architectures

Per suggestions in class, more succinctness to help facilitate peer discussion.

MACAW: A MEDIA ACCESS PROTOCOL FOR WIRELESS LAN's

Vaduvur Bharghavan, Alan Demers, Scott Shenker, Lixia Zhang

Summary:

Talks about some pretty generic problems in wireless networks. Happened to be in the context of some network with a five letter acronym. Hidden & exposed terminal problem in vanilla CSMA. RTS & CTS interaction. Inherent unfairness and starvation of "normal" exponential back-off, pretty hacky solution (can't think any better alternative ...). Multiple stream buffers. Need for ACK, DS, RRTS messages. Multicast problems. Need for per-destination backoff. Some issues for pathological topologies unresolved.

Background:

You'd better know your basic wireless CSMA/CA stuff and maybe see some of the RTS/CTS or exposed/hidden terminal problems before. Else it's gonna be tough going like the router papers. The pictures helped though.

Discussions & Criticisms:

Review of state of the art in 802.11 stuff? How much of the paper got into the current wireless protocols?

Mobility not touched - so results valid for mobile hosts. Moving hosts? Hosts on Eurotunnel type trains? Relativistic speeds in space not allowed :)

Scalability?

If there is a shortcoming of the paper is that the simulation is not described in any depth. Thus fails the "reproduceable" criteria for experiments? Too few papers outlines their methodology in a truly reproduceable fashion anyways ...




A COMPARISON OF MECHANISMS FOR IMPROVING TCP PERFORMANCE OVER WIRELESS LINKS

Hari Balakrishnan, Venkata N. Padmanabhan, Srinivasan Seshan and Randy H. Katz

Summary:

Comparative study of several techniques for handling wireless link layer losses. Including link layer retransmit mechanisms, both TCP aware ones (snoop) and not aware ones. End-to-end TCP loss recovery, like vanilla TCP Reno, TCP New Reno, TCP SACK, SMART TCP (cummulative ACK, and seq no. of pkt that caused receiver to generate the ACK), TCP Explicit Loss Notification (TCP ELN), and TCP ELN with retransmit on every duplicate ACK. Also looked at split-connection with TCP proxy and split-connection with TCP SMART.

Bottom line - link layer TCP aware stuff with SMART is best.

Background:

Most of the required background is already provided in the paper - just need to know what the techniques are and such. The snoop paper is worth a read though. Just because it also has Randy's name on it and this paper basically says snoop rocks.

Discussions & Criticisms:

Paper claims exponential distribution error model give results that are suggestive for other error models. Why is this the case? I think one of Stoica or Shenker or Paxson's early papers provided a good justification re why exponentail arrival/loss model is fairly general, but I forgot the reason. Anyone remember why?

Another methodology question. How do we ensure we get only controlled, artificially inserted errors and not "real" bit errors coming from the wireless link?

How much of this is just a glorified measurement and comparison of the snoop protocol with other stuff? Much of the result saz basically snoop rules.

Follow up results with regard to the temporal model of wireless loss?

Changing error rate results showed each solution has critical error rate above which throughput dramatically decreases, as marked by inflection points in Fig. 12. Would be interesting to plot the vertical axis in % of achievable theoretical throughput, given the error rate, and horizontal axis extend to the left for even higher error rates. Basically vanilla TCP Reno starts sucking majorly at very low error rates.

Fundamental question re Nyquist limit - AT&T WaveLAN at 915MHz means max theoretical data rate of ~2Gbps. 802.11 is at 3.7GHz (?). Hence 7.4Gbps max theoretical data rate. So future wireless tech fundamentally limited in speed? Ethernet is moving to 10Gbps soon, and 100Gbps Ethernet in the works. Would be weird to have dramatic speed mismatch. Whole new cycle of panic -> research -> more panic -> more research? Or we solve problem with CDMA or some ingenius crap like that ...