We consider a 6-node sensornetwork shown inFigure 2.7. It can be easily seen that
φ 1,0 = φ 3,0 = φ 5,0 = 1
, node0
being the common destination for all the packets generated in the12
Itis,hence,shownthattheproposeddistributedalgorithmprovidestheoptimal delaythatis achievable
basedontracdynamics.
network. The other available routes could also be usedfor nodes1, 3, and 5. But, we were
moreinterestedinclearlydemonstratignthetracsplit androutingprobabilities. Therouting
algorithm thus has to nd appropriate valueof
φ 2,5
andφ 4,3
in order thatthe trac ow inthe networkcorrespondsto aWardrop equilibrium.
3 4
0 1
5 2
λ1 λ4
λ2 φ25
φ43
φ41
φ21
λ5 λ3
Figure2.7: Network Simulated for Stability
Apart froma demonstration of theconvergence of theproposed algorithm, we will seein
this section that the data sampling rates that a network can support using the Open
archi-tecture is very small. This is essentially because of the stability constraints on the channel
access rates. On the other hand, the Closed systemcan support higher datasampling rates
because of the fact that itis essentiallyself-regulating, guaranteed to be stable while
main-taining large data sampling rates; this is because a node generates a new packet only if it
has no other packet inthe queue. This however does not mean that the Closed system can
supportarbitrary datasamplingrates.
We have implemented the Open and Closed system as an application layer module in
TinyOS [52]. TinyOS is an open-source operating system designed for embedded WSNs. It
features acomponent-based architecture whichenables rapid innovation andimplementation
while minimizing code size asrequired bythe severememory constraints inherent inWSNs.
The sensor network model under consideration is shown in Figure 2.8. The sensor nodes
sample the data at a predened rate,
λ 0 i s.
The sampled data is sent to the MAC queue forboth open and closed system according to the explanation given earlier inSection 2.4. The
transmitqueueofnode
i
canhaveatmostonepacketinthetransmitqueuethatwasgeneratedat node
i
. It can however have multiple packets in thetransmit queue to be forwarded, i.e.,those packets that were initially generated at some other node, and have arrived at node
i
to be forwarded to some other node. Therefore, we need not implement two-queues at the
MAC layer for sensor nodes for prioritizing trac. At simulation start up, the nodes learn
the network topology and built routes toward the fusion center (sink, node
0
). The fusioncenter isalso asensornode which has
0
samplingrate. Thislearning process,which dependsonthe network topology for the given network inFigure 2.7, can take up to
50 − 70 seconds
for larger topologies. The routing layeris initiated withthe minimum-hop routing, which is
updated during the network lifetime according to the algorithm proposed in Section 2.6. In
this section, we present the numerical results once the neighbors are discovered and routes
areestablishedtoward thefusion center. We have utilized theTOSSIMsimulator of TinyOS
to validate our proposals. All simulation run for
1000 seconds
. The results presentedin thissection arethe averageoverseveralsimulation runs.
Routing Routing Routing
ROUTING
MAC
PHY APPLICATION
Source Destination
(Fusion Center)
Forwarders
...
...
...
Figure 2.8: Sensor network architecture.
→
represents the ow of packets from the sourceto the destination. The forwarding sensor network receives a packet and queues into the
forwarding queueat the MAClayer. The routinglayerdoesnot buer theforwarding trac.
2.7.1 Open System Stability
InFigures. 2.9and2.10weplot,againsttheslotnumber,theaveragedelaysonthefourroutes
2 → 5 → 0
,2 → 1 → 0
,4 → 3 → 0
,and4 → 1 → 0
for the open system. Thedata samplingrates weresetat
λ 1 = λ 2 = λ 3 = λ 4 = λ 5 = 0.2
. Notethatthedatasampling ratesaresmall.We were forced to select small data rates in order to guarantee stability of thenodes inthe
network. Thechannelaccessrates weresetto
α i ≤ 0.2
fori = 1, ..., 5
.Observations
1. The delays on routes
2 → 5 → 0
and2 → 1 → 0
arevery close to each other, with avery fastconvergence. Similarlyfor routes
4 → 3 → 0
and4 → 1 → 0
. Thisshows thatthealgorithm succeedsinachievinga Wardrop equilibrium.
2. Notethe highvalueofdelayonroutes
2 → 1 → 0
and4 → 1 → 0
even forthemoderate(or, very small) loadon thesystem.
3. The delays on dierent routes are sensitive to the channel access probabilities. Thus,
there is a need for carefully tuning the channel access probabilities. In Figure 2.9 and
2.10, we also see the convergence to a load-balanced regime (equal delays on all the
possible routes froma particular source).
0 1 2 3 4 5
0 100 200 300 400 500 600 700 800 900 1000
Estimated Delay (msec)
Time -> (Seconds)
Delay on Route 2->5->0 Delay on Route 2->1->0
Figure 2.9: Delays incurred on routes
2 → 5 → 0
,2 → 1 → 0
for Open System. Whereλ 1 = λ 2 = λ 3 = λ 4 = λ 5 = 0.2
2.7.2 Closed System Stability
Simulation results for the closed system are presented in Figure 2.11 and 2.12. The data
sampling rates were set at
λ 1 = λ 2 = λ 3 = λ 4 = λ 5 = 0.2
. Nodes were expected to adapttheirchannel accessprobabilities basedon theoptimaltrac split usedby node
2
and4
.Observations
1. The delays on routes
2 → 5 → 0
and2 → 1 → 0
are very close to each other, witha fast convergence. This shows that the algorithm succeeds in achieving a Wardrop
equilibrium.
2. For routes
4 → 3 → 0
and4 → 1 → 0
, the delays are also close to each other, with afast convergence. This shows that the algorithm is successful in achieving a Wardrop
equilibrium (equaldelays on allthe possibleroutes from aparticular source).
3. Note the small value of delay on routes
2 → 5 → 0
and4 → 3 → 0
even for moderate(or, very small) load on the system. This is to be compared with the corresponding
values shown underthe results for open system where the delays on these routes were
higher even thoughthe averagedatasamplingrates weresignicantly smaller. Thus, in
comparison withtheopen system,theclosed systemprovides better performance.
Wesimulateanother6-nodesensornetworkshowninFigure2.13todemonstratetheresults
onrouting. Theonlydierencewiththerstnetworkisthatwehaveadierentroutingsetup
but its logical representation is the same. It is easily seen that
φ 1,0 = φ 3,0 = φ 5,0 = 1
, node0
being the common destination for all the packets generated in the network. The routing0 1 2 3 4 5
0 100 200 300 400 500 600 700 800 900 1000
Estimated Delay (msec)
Time -> (Seconds)
Delay on Route 4->3->0 Delay on Route 2->1->0
Figure 2.10: Delays incurred on routes
4 → 3 → 0
,4 → 1 → 0
for Open System. Whereλ 1 = λ 2 = λ 3 = λ 4 = λ 5 = 0.2
algorithm thus has to nd appropriate valueof
φ 2,5
andφ 4,3
in order thatthe trac ow inthe networkcorrespondsto aWardrop equilibrium.
2.7.3 Open System Routing
InFigure2.14and2.15weplot, againsttheslotnumber,theaveragedelaysonthefourroutes
3 → 2 → 0
,3 → 1 → 0
,5 → 4 → 0
, and5 → 1 → 0
for the open system. The datasampling rates were set at
λ i ≤ 0.2
, fori = 1, ..., 7
. Note that the data sampling rates aresmall. We were forced to selectsmall data rates in order to guarantee stability of the nodes
inthenetwork. Thechannelaccessrates weresetto
α i ≤ 0.2
fori = 1, ..., 7
.Observations fromOpen System
1. The delays on routes
3 → 1 → 0
and3 → 2 → 0
arevery close to each other, with avery fastconvergence. Similarlyfor routes
5 → 4 → 0
and5 → 1 → 0
. Thisshows thatthealgorithm succeedsinachievinga Wardrop equilibrium.
2. Note the high value of delay on routes
3 → 1 → 0
and3 → 2 → 0
even for moderate(or, very small) loadon thesystem.
3. Figure 2.15 shows the delay obtained by varying the channel access rates to
α i = 0.1
for
i = 1, ..., 5
, andλ 0 s
remaining the same as earlier. The estimated delays show thesensitivity to channel access probabilities. Thus, there is a need to carefully tune the
α 0 i s
. InFigure2.15,wealsoseethatconvergencetoaload-balanced regime(equaldelays on all the possible routes froma particular source) isviolated bychanging theα 0 i s
. Aswe will seelater, this is not a problem intheclosed systembecause the systemadapts
0 1 2 3 4 5
0 100 200 300 400 500 600 700 800 900 1000
Estimated Delay -> (msec)
Time -> (Seconds)
Delay on Route 2->5->0 Delay on Route 2->1->0
Figure 2.11: Delays incurred on routes
2 → 5 → 0
,2 → 1 → 0
for Closed System. Whereλ 1 = λ 2 = λ 3 = λ 4 = λ 5 = 0.2
its channelaccessprobabilities tomeet thetargettrac and thereisno needof further
tuning this parameter.
4. Thedelaysondierentroutesarealsoclose toeachother,withafastconvergence. This
is also reected in the trac split obtained by the algorithm, as in Figure 2.16we see
that node 3 uses node 2 a little less than the other available route because of smaller
delay on
3 → 1 → 0
. Similarly, node 5 also use5 → 1 → 0
more than5 → 4 → 0
becauseofsmallerdelayontheformer. Itisalsointeresting tonotethatthetracsplit
obtained inthis gureis proportional to thedelays on dierent routes inthe network,
i.e.,
φ 32
is very close to 0.5 due to a smaller dierence in estimated delays on routes3 → 1 → 0
and3 → 2 → 0
, whereas,φ 54
isnot due to relatively large dierence intheestimated delays on routes
5 → 1 → 0
,5 → 4 → 0
. Thisis Wardrop equilibriumwhere a slightly higher delay path is less usedi.e., the+ve
valueof trac on alternate routeis imposedbythealgorithmto ensure thatallthealternativesareprobedoftenenough
to cope upwitha change intrac patterns.
2.7.4 Closed System Routing
Simulation results for the closed system are presented in Figure 2.17, 2.18, 2.19, and 2.20.
Thedata sampling rates were setat
λ 1 0.1, = λ 2 = 0.2, λ 3 = 0.1, λ 4 = 0.005, λ 5 = 0.1, λ 6 = 0.1, λ 7 = 0.1
. Nodeswereexpected to adapt their channel access probabilities based on the optimaltrac splitused bynode3
and5
.Observations fromClosed System
1. The delays on routes
3 → 1 → 0
and3 → 2 → 0
are very close to each other, with0 1 2 3 4 5
0 100 200 300 400 500 600 700 800 900 1000
Estimated Delay -> (msec)
Time -> (Seconds)
Delay on Route 4->3->0 Delay on Route 4->1->0
Figure 2.12: Delays incurred on routes
4 → 3 → 0
,4 → 1 → 0
for Closed System. Whereλ 1 = λ 2 = λ 3 = λ 4 = λ 5 = 0.2
a fast convergence. This shows that the algorithm succeeds in achieving a Wardrop
equilibrium.
2. Forroutes
5 → 1 → 0
and5 → 4 → 0
,thedelaysarealsoclose toeachother,withafastconvergence. This is also reected inthe trac split obtained by the algorithm, asin
Figure2.19we seethatnode5uses node1 formost ofitstrac,thusobtainingsmaller
delay. Thisis again Wardrop equilibrium where thehigher delaypath isnot used(the
small
+ve
valueoftraconroute5 → 4 → 0
isimposedbythealgorithm toensurethatall thealternativesareprobed oftenenough to copeup withachange inthenetwork).
3. Note the small value of delay on routes
3 → 1 → 0
and3 → 2 → 0
even for moderate(or, very small) load on the system. This is to be compared with the corresponding
values shown underthe results for open system where the delays on these routes were
higher even thoughthe averagedatasamplingrates weresignicantly smaller. Thus, in
comparison withtheopen system,theclosed systemprovides better performance.
4. Figure 2.20 shows that the algorithm is also able to adapt the channel access rates in
a distributed fashion. It can be checked that the values of
α 0 i s
converged-to by the algorithm indeedarejustenough to serve thetrac oered to thedierent nodes.Intheaboveanalysis,wehaveonlypresentedresultsontheperformanceofourdistributed
routingalgorithm forboth open andclosed systems. We, now, considera randomly deployed
sensornetworkwith50 sensornodes. There isonly corner-sink which istherepresentative of
datacollectioninthe network. Sensorstransmit theirreadingsinamultihop fashion towards
this sink. The sampling rate of all the nodes is a random variable uniformly distributed
between 0and 0.2. Thechannelaccessrateofallthenodesissetaccordingtotheir sampling
0 1
φ31
φ32 φ51
φ54
λ1
2 3
4 5
λ2 λ3
λ4 λ5
Figure 2.13: Network Simulated for Routing
rates. Thesimulationrunsfor3000sand thesamplingvector ischanged every100sto seethe
impact of trac pattern change on the performance of closed system. Figure 2.21 displays
individual node delays over time w.r.t change in the sampling rate. It can be easily seen,
for each node in the network, that the delay does not change much due to a change in the
trac pattern overtime. The coupling intheclosed systemautomatically regulates thedelay
by adapting its own sampling process to the change in network dynamics. The last degree
of freedom, i.e., the channel access rates are also adapted using the proposed optimization
criteria. TheCDF oftheestimateddelayinthenetwork ispresentedinFigure2.22.
2.7.5 Closed System with Two Transmit Queues
We,again, considerthe6-nodesensornetwork showninFigure2.13. We considerthis simple
network toclearly demonstrate the stability region in closed system withtwo transmit queues.
The transmit queue of node
i
can have multiple packets in thetransmit queue (bothQ i
,i.e,selfgenerated,and
F i
,i.e.,thosepacketsthatwereinitiallygenerated atsomeothernode,andhavearrivedatnode
i
to beforwardedtosomeothernode). Therefore,weneedtoimplementtwo-queuesattheMAClayerforsensornodesforprioritizingtrac(basedontheappropriate
weights given to
Q i
andF i
). We have implemented the Closed system with two-queues as a cross-layer (application-mac) module in TOSSIM [52]. The routing layer is initiated withthe minimum-hop routing, which is updated during the network lifetime according to the
algorithm proposedinSection 2.6. Inthis section, we presentthe simulation resultsoncethe
neighbors are discovered and routes are established toward the fusion center. All simulation
runsfor
10 8
,seconds.We present in Table 2.3, the results on stability region and throughput for sensors 1, 2,
and4 assensors3 and5 donot forward anytrac and
y i
fori = 3, 5
is setto 1.In order to demonstrate theresults on delay-and-stability together using a closed-system
with two-queues, we have implemented a 50-nodes sensor network with a common sink. In
Figure 2.23 we plot, against the slot number, the average delays for our closed-system with
0 5 10 15 20 25 30 35 40 45
0 1e+07 2e+07 3e+07 4e+07 5e+07 6e+07 7e+07 8e+07 9e+07 1e+08
Estimated Delay (s)
Time ->
Delay on Route 3 -> 1 -> 0 Delay on Route 3 -> 2 -> 0 Delay on Route 5 -> 1 -> 0 Delay on Route 5 -> 4 -> 0
Figure2.14: Delays incurred on routes
3 → 1 → 0
,3 → 2 → 0
,5 → 1 → 0
,5 → 4 → 0
foropen system.
α 1 = 0.2, α 2 = 0.15, α 3 = 0.1, α 4 = 0.2, α 5 = 0.2
,λ 1 = 0.01, λ 2 = 0.01, λ 3 = 0.04, λ 4 = 0.05, λ 5 = 0.05.
two-queues and single-queue system. The datasampling rates were set at
λ i ≤ 0.1
,∀ i
. Notethatthedatasampling rates are small. We were forcedto select smalldata rates inorder to
guaranteestabilityof thenodesinthenetwork.
Observations from the Simulations: The average delays on routes in two-queues
closed systemareverysmall comparedto single-queuesystem. Thisisduetotheappropriate
choice of weights given to both
F i
andQ i
(as discussed in Section 2.4) compared to thesingle queue system where we do not have the service dierentiation. The routing schemes
(Section2.6) allows both systems to pickthe shortest-delaypaths based on delayestimates.
Theseresults complywithour motivationthatservicedierentiation at MAClayer results in
better overall performanceof thesystemand canhelpstudy theimpact of dierentnetwork
parameters on its performance.