>> The 1.4% is a bit misleading. As 6.2 discusses, the transport
>> protocol in the paper achieves 94.8% of the layer 3 protocol's
>> throughput (when considering the 1/3rd effect you get from multihop).
Phil - this is exactly the sort of thing that drives me nuts about
papers in this community. First off, section 6.2 discusses details of
single-hop transmission using a 14.4kbps radio, and claims 33%
efficiency. The actual deployment is multi-hop, and uses 250kbps
radios, and achieves 1.4%.
The "1/3rd effect" that you allude to is a constraint imposed by
single-channel operation. If you use a channel-hopping protocol the
theoretical limit is 1, not 1/3.
Try telling an end user that his measured 1.4% throughput is misleading,
and that he should really take into account all of the artificial
constraints that you have come up with to explain it.
ksjp
Philip Levis wrote:
On Aug 10, 2007, at 10:47 AM, Kris Pister wrote:
JP - great presentation. I hope that it was well received. I'm
very supportive of everything in that document, with one exception.
There is one assumption that I believe is potentially fatal to the
entire 6lowpan/rsn effort in ietf. On slide 11 you state "Research
has focused on near-optimal solutions to specific problems", and then
in bold "IP is maximizing interoperability, not aiming at finding a
local optimum ;-)". I agree with the second statement, but the first
one is flat out wrong.
Very few people in sensor network research seem to care about
near-optimal solutions. Consider one of the best-received papers
[Kim07] at the recent IPSN conference. Note that the authors are
leaders in the field, this is a respected conference, one of the
better papers, and this is the dissertation work culminating four
years of research on this project. The paper presents the best
academic work to date on reliable multi-hop data collection, MP2P.
The payload goodput in this network was 1.4% of the channel
bandwidth, at 100% radio duty cycle. This is not "near-optimal" in
any sense that I know.
If we start with 1.4%, and then degrade that so that we can maximize
interoperability, we'll be left with something that no one will use.
We have an opportunity to do this right - let's not fall victim to
the Zigbee-esque assumption that all we need to do is define
*something* and adoption will be automatic.
ksjp
[Kim07] Kim, Pakzad, Culler, Demmel, Fenves, Glaser, Turon, "Health
Monitoring of Civil Infrastructures Using Wireless Sensor Networks",
IPSN07.
You're right, Kris, but it's a bit deeper than that. There's a big
debate in the research community right now about layered vs.
cross-layer design, and there are more of the latter than the former.
The protocol briefly mentioned in the paper you cite has a full paper
about it that will be appearing in SenSys. It's clear that a
cross-layer design could lead to higher performance (hence JP's
comment about optimal point solutions), and some reviewers made this
point. But you can't easily compose cross-layer protocols into larger
systems. hence the morass many research systems enter when they try to
incorporate protocols X, Y, and Z together. So the community is slowly
shifting from "how good can we get if we throw out all layering" and
moving towards "what have we learned from that and how can we apply it
to the layers?" That shift is one reason why R2LN's timing is very
auspicious.
The 1.4% is a bit misleading. As 6.2 discusses, the transport protocol
in the paper achieves 94.8% of the layer 3 protocol's throughput (when
considering the 1/3rd effect you get from multihop). The big
inefficiencies are the layer 3 and 2 protocols the particular
implementation builds on, neither of which were optimized for
throughput. It would be great to see how well the e2e transport
protocol works on top of a layer 2 and a layer 3 with this metric in
mind.
Phil
_______________________________________________
6lowpan mailing list
[email protected]
https://www1.ietf.org/mailman/listinfo/6lowpan
