On Jan 8, 2007, at 9:29 AM, Kris Pister wrote:
(Reassembly at the forwarder also means different fragments cannot
choose different paths; maybe a more theoretical consideration.)
This is absolutely not a theoretical consideration. We call these
networks and headers "mesh" for a reason: there's an intuitive
sense that multiple routes provide better performance. This
intuition is now supported by substantial empirical evidence (in
addition to theory and simulation): implementations that say "mesh"
and deliver quasi-static trees are not reliable. Zigbee did a
disservice to the community by defining "mesh" to be "multi-hop" in
the 2003 standard.
Right -- the big question is the rate at which the routes can adapt.
Multiple routes do not in and of themselves necessarily provide
better robustness; unlike in the wired case, nodes are physically
correlated and so can fail together. I'm personally not convinced
that one of fine-grained packet next-hop diversity or rapid route
failover is inherently superior to the other.
I'd argue that the lack of reliability we've seen in quasi-static
trees has been due to a mismatch between route adaptation and data
traffic, with no backpressure. That is, if your route changes at most
every 8 seconds, you can buffer 10 packets, the sender drops on layer
2 acknowledgment, and there is no way to slow down your reception
rate of 5 packets per second, then a route failure will lead to
packet loss. If you look at the current layer in TinyOS 2.0, for
example, it delivers two nines, with the packet losses being due to
false positives on layer 2 acks. I know that this is well below what
Dust can deliver, but to be fair, it uses an ad-hoc rather than
centralized approach and has been written by a few grad students
spread across the country rather than an experienced development team.
Phil makes a good case for single-hop reassembly in his 12/14
treatise, and his arguments are valid from a tinyOS and/or Zigbee
perspective. F, X, and L will be large, and if link-level
acknowledgements are not used, then the energy cost of multi-hop
reassembly will be prohibitive. Indeed, the probability of success
decreases so rapidly with increasing F and X that the question
"does it work at all?" takes precedence over "is it energy
efficient?".
There are two international sensor networking standards in the
works right now (HART and SP100) which will both include mesh
routing and link-level acknowledgements. Both will support energy-
efficient fragmentation and end-to-end reassembly with large F, X,
and L. Both are based on existing products built on top of 802.15.4.
I think that option 2 below is probably still fine. Giving people
this level of flexibility is still useful as we explore this
comparatively young space, but let's do it informed by the pre-
existing solutions to similar problems.
Does saying headers can appear in either order break from IPv6?
Phil
_______________________________________________
6lowpan mailing list
[email protected]
https://www1.ietf.org/mailman/listinfo/6lowpan
