Information Technology Reference
In-Depth Information
(a)
Line topology.
(b)
Example of a learned wake-up schedule for duty cycle
of 5%.
(c)
Average end-to-end latency for different duty cycles.
Fig. 3.
Experimental results for the line topology
line topology had remained active during the sleep period of its immediate neighbors,
its messages, together with those of its higher hop neighbors would not have been de-
livered to the sink. Even though neighboring nodes are awake at the same time (or have
synchronized), one can see that schedules are slightly shifted in time. The reason for
this desynchronicity is to reduce the overhearing of higher hop communication and to
increase throughput by compensating for propagation delays — a behavior that nodes
have learned by themselves.
Figure 3(c) displays the average end-to-end latency of the learning and the synchro-
nized nodes respectively, where error bars signify one standard deviation across
30
runs.
Since the learned wake-up schedules of our approach closely resemble the prescribed
behavior of S-MAC, the latency improvement over different duty cycles is marginal.
Nevertheless, the end-to-end latency of our learning agents is on average
2
seconds
less than under the S-MAC protocol. The reason for this improvement lies in the fact
that with S-MAC all nodes wake up at the beginning of the frame, while with our ap-
proach agents learn when it is best to wake up. Since each node periodically generates
messages at a different time within the
10
-seconds frame, the latency of S-MAC is on
average
5
seconds. Learning, however, allows flexibility in the wake-up times, such that
a node could wake up immediately after generating its messages (and all other nodes
will learn to wake nearly at the same time) and therefore reduce the queuing time of
messages for at least one node.
As evident in Figure 3(c), both approaches are inefficient at very low duty cycles.
The reason for this high latency is the fact that the active period of nodes is too short
compared to the propagation delay. Therefore, messages need to be queued for more
than one frame on average, which results in traffic congestion.
In contrast to the previous topology, our second set of experiments investigate the
performance of the network where all nodes lie on the same hop from the sink. This
setup presents agents with the opposite challenge, namely to find an active period where
no other node is awake. The latter behavior will eliminate communication interference
with neighboring nodes and will ensure proper reception of messages at the sink. Fig-
ure 4(b) displays an example of the wake-up schedule of the learning nodes for a duty
