This paper explains how many models of wireless links used in research are inaccurate and how to really model the losses, variable delays and retransmissions that can occur. The paper is very important because it shows problems with many evaluations of new proposals and tells researchers how to get wireless channels right. However, this is also my main gripe with the paper: it only explains how to make better models, or at least models that are better in particular scenarios. It does not answer the question of "when is a model good enough to be accurate within X%", nor does it propose other means of evaluating the protocols, like how to set up a realistic experiment. As long as you only use modeling, all you can do in a paper like this is prove past work "wrong", but not prove any work "right", because the models could still be faulty. It's also unfortunate that the paper doesn't try to quantify how far off one can be with the simplifying assumptions in some of the work it criticizes - e.g. assuming that uplink and downlink have the same capacity may be okay if acks are very tiny compared to data packets.
My other gripe with the paper is that it contains a lot of maybes and speculation about future technology. This is good for warning researchers that some of the problems they are trying to solve (delay, reordering, etc) might go away by themselves, but it seems far to premature to discount work dealing with these problems just in case a better link layer protocol is invented. We also know that software can be much easier to change than hardware, so it makes a lot of sense to work on transport algorithms rather than worrying that a link layer complication you are trying to solve will be irrelevant in a few years. Maybe this wasn't clear at the time of the paper, if Wifi and cellular networking were not pervasive; but today, it's pretty obvious that 802.11 is not going away soon, and neither is edge / 3G / etc, so we need to design protocols that work well with them.
I was a little surprised at the paragraph that said current WLANs don't reorder packets. It actually makes sense, since you're supposed to hear the link-level ACK right away, so you can keep trying to send a packet if you don't hear it. But in that case, why was so much of the work in the previous paper (comparing transports over wireless) concerned with reordering? Did none of those papers use any evaluation beyond simulation?
Subscribe to:
Post Comments (Atom)
2 comments:
I agree that the lessons are not particularly constructive. What's missing is some generalization about what the faulty models have in common and how the authors' methodology for categorizing and justifying new models avoid these problems. Instead, we are merely presented with two lists: bad models, and their models.
I thought it was pretty good to have a nice reference on 'things undreamed of in your philosophy' of networks.
I'd definitely have preferred data, rather than speculation.
Post a Comment