[SensorNetArch] Hotnets submission

[Rodrigo Fonseca]

Hi,

I'd like to share with you our HotNets submission that has to do with
link estimation. This is joint work with Omprakash Gnawali from USC,
Kyle Jamieson from MIT, and Phil Levis from Stanford.

The 4B paper proposes a link estimator that uses 4 very narrow
interfaces (1 bit each) with the link layer, the network layer, and
the physical layer, and uses the information to greatly improve the
performance of collection. Our avg cost per delivered packet is 44%
better than that of the original CTP, avg depth in the tree is 55%
better.  Against MultiHopLQI, we are 29% better in cost and 11% in
path lengths, and we got 99.97% of the packets delivered.

By using information from acks of sent data packets (ACK Bit), we
don't need to do bidirectional probe-based link estimation
(unidirectional suffices). This has a huge impact, because you don't
constrain the in-degree of the tree to the size of the receiver's
neighbor table.

By using hardware receive quality indication (WHITE Bit), we are able
to quickly assess if a link is worth considering for inclusion in the
link table, before wasting resources trying to estimate quality.

We couple this with information from the network layer ( COMPARE Bit)
to see if this incoming link is part of a better path than some we
already have, and insert it if so.

Lastly, we also use information from the network layer (PIN Bit) to
see if some neighbor should not be evicted from the neighbor table
(for example, if the neighbor is a root of the tree).

These interfaces are not limited to collection, we believe they can be
used in other protocols as well.

Comments are welcome.

Cheers,
Rodrigo
-------------- next part --------------
A non-text attachment was scrubbed...
Name: submit.pdf
Type: application/pdf
Size: 458170 bytes
Desc: not available
Url : http://mail.millennium.berkeley.edu/pipermail/sensornetarch/attachments/20070928/cba7f26b/submit-0001.pdf