Beyond Pilots: Keeping Rural Wireless Networks Alive

14 Pages·2008·0.43 MB·English

Beyond Pilots: Keeping Rural Wireless Networks Alive
Sonesh Surana∗ RabinPatra∗ Sergiu Nedevschi∗ ManuelRamos†
LakshminarayananSubramanian‡ Yahel Ben-David§ EricBrewer∗ ¶
Abstract Researchers (ourselvesincluded) tend to focus on the
sexypartsofadeployment,suchashigherperformance
Very few computer systems that have been deployedin
orahighlyvisiblepilot.However,realimpactrequiresa
rural developing regions manage to stay operationally
sustainedpresence,andthusoperationalchallengesmust
sustainable overthe long term; most systems do not go
be viewed as a first-class research topic. Analogous to
beyondthepilotphase.Thereasonsforthisfailurevary:
researchonhighavailability,wemustunderstandtheac-
components fail often due to poor power quality, fault
tualcausesofoperationalproblemsandtakeabroadsys-
diagnosis is hard to achieve in the absence of local ex-
temicviewtoaddresstheseproblemswell.
pertiseandreliableconnectivityforremoteexperts,and
In this paper, we describe our experiences over the
faultpredictionisnon-existent.Anysolutionaddressing
last three years in deploying and maintaining two ru-
theseissuesmustbeextremelylow-costforruralviabil-
ralwirelesssystemsbasedonpoint-to-pointWiFilinks.
ity.
OurpriorworkonWiFi-basedLongDistanceNetworks
We take a broad systemic view of the problem, doc-
(WiLDNet) [26] developed a low-cost high-bandwidth
ument the operational challenges in detail, and present
long-distance solution, and it has since been deployed
low-cost and sustainable solutions for several aspects
successfully in several developing regions. We present
of the system including monitoring, power, backchan-
real-world validation of the links, but the primary con-
nels, recovery mechanisms, and software. Our work in
tributionhereis the explorationof the operationalchal-
the last three years has led to the deploymentand scal-
lengesoftworuralnetworks:atelemedicinenetworkat
ing of two rural wireless networks: (1) the Aravind
the Aravind Eye Hospital [3] in southern India and the
telemedicine network in southern India supports video-
AirJaldi[1]communitynetworkinnorthernIndia.
conferencing for 3000 rural patients per month, and is
We have had to overcome major challenges in both
targeting500,000patientexaminationsperyear,and(2)
networks: (1) components fail easily due to low qual-
the AirJaldi network in nothern India provides Internet
ity power, (2) fault diagnosis is hard because of non-
accessandVoIPservicesto10,000ruralusers.
expertlocalstaffandlimitedconnectivityforremoteex-
perts, and (3) remoteness of node locations makes fre-
1 Introduction
quent maintenance difficult; thus fault anticipation be-
comes critical. All of these problems can be fixed by
The penetration of computer systems in the rural de-
having higher operating budgets that can afford highly
velopingworld has been abysmally low. Severalefforts
trained staff, stable power sources, and robust high-end
aroundtheworld thathavetriedtodeploylow-costcom-
equipment.Buttherealchallengeistofindsolutionsthat
puters,kiosksandothertypesofsystemshavestruggled
are sustainable and low-cost at all levels of the system.
toremainviable,andalmostnoneareabletoremainop-
To this end, our main contributions are (1) document-
erational over the long haul. The reasons for these fail-
ingandcategorizingtheunderlyingcausesoffailurefor
uresvary,butatthecoreisanunder-appreciationofthe
thebenefitofresearchersundertakingruraldeployments
manyobstaclesthatlimitthetransitionfromasuccessful
in the future, and (2) developing low-cost solutions for
pilottoatrulysustainablesystem.Inadditiontofinancial
thesefailures.
obstacles,theseincludeproblemswithpowerandequip-
Inovercomingthesechallenges wehavelearnedthree
ment, environmental issues (e.g. heat, dust, lightning),
importantlessonsthatweargueapplytoITdevelopment
and an ongoing need for trained local staff, as trained
projects more broadly. First, designers must build sys-
staffmoveontobetterjobs.
temsthatreducetheneedforhighlytrainedstaff.Second,
∗UniversityofCalifornia,Berkeley simple redesign of standard components can go a long
†UniversityofthePhilippines way in enabling maintenanceat lower costs. And third,
‡NewYorkUniversity therealcostofpowerisnotthegridcost,butisthecostof
§AirJaldi,Dharamsala,India overcoming poor power quality problems. By applying
¶IntelResearch,Berkeley these lessons to severalaspectsof oursystem including
monitoring,power,backchannels,recoverymechanisms,
10
and deployed software, we have made real progress in WiLDNet Aravind WiFi Aravind WiLDNet
keepingtheseruralnetworksalive. ps) 8 Emulator Testbed WiFi Testbed WiLDNet
Mb Venezuela WiFi Venezuela WiLDNet
The Aravind network now uses WiLDNet to inter- ut ( 6
connect rural vision centers with their main hospitals p
h 4
for patient-doctor video-conferencing. Currently 9 vi- ug
o
sion centerscater to 3000patients per month. Thusfar, hr 2
T
30,000ruralpatientshavebeenexaminedand3000have 0
had significant vision improvement. As all vision cen- 10 20 50 100 200 300 400
Distance (km)
tersarenowrunningwithnooperationalassistancefrom
ourteam,thehospitalconsidersthisnetworksustainable Figure 1: Comparison of TCP throughput for WiLDNet
andistargetingatotalof50centersinthenext2years. (squares)andstandardWiFiMAC(triangles)fromlinksinAr-
Similarly,AirJaldiisalsofinanciallysustainableandcur- avind,Venezuela,Ghana(the65kmlink),andourlocaltestbed
rentlyprovidesInternetaccessandVoIPservicestoover intheBayArea.MosturbanlinksinAravindhadupto5–10%
10,000usersinruralmountainousterrain. loss, and so WiLDNet did not show substantial improvement
In thenextsection we validatethesufficiencyofreal- overstandardWiFi.However,WiLDNet’sadvantageincreases
worldWiLDperformance,andoutlinethechallengesto with distance. Each measurement is for a TCP flow of 60s,
802.11bPHY,11Mbps.
operationalsustainability.Section3providessomeback-
ground for the Aravind and AirJaldi networks. In Sec-
tion4,we documentmanyofourexperienceswithsys- andPlatilloninVenezuela.Tothebestofourknowledge,
temfailures,andtheninSection5presentthedesignof this is currently the longest distance at which a stable
alllevelsofoursystemthataddresstheseissues.Related high-throughputWiFilinkhasbeenachievedwithoutac-
workisdiscussedinSection6,andinSection7wesum- tive amplification or custom antenna design. Each site
marizethreeimportantlessonsforruraldeployments. used a 2.4GHz 30-dBireflector gridantennawith 5.3◦
beam-widthanda400mWUbiquitiSR2radiocardwith
theAtherosAR5213chipset.
2 Motivation
Figure 1 presents results from running WiLDNet on
reallinksfromourvariousdeploymentsinAravind(In-
Inthissection,weconfirmhigh-throughputperformance
dia),Venezuela,Ghana,andourlocaltestbedintheBay
of WiLDNet links in real-world deployments, and then
Area.WematchtheperformanceofWiLDNetoverem-
outline the operationalchallengesthat remain obstacles
ulated links and greatly exceed the performance of the
tosustainedimpact.
standardWiFiMACprotocolatlongdistances.
2.1 Real-WorldLink Performance Thus we find that we are no longer limited by per-
formanceoverlongdistancesinruralnetworks.Instead,
Existing work [16, 26, 29, 33, 34] on rural networking
based on our experiencesin deployingand maintaining
hasfocusedonmakingWiFi-basedlong-distancepoint-
networks in the two rural regions of India for the last
to-point links feasible. The primary goal has been high
threeyears,wearguethatoperationalchallengesarenow
performance,typicallyexpressedashighthroughputand
theprimaryobstacletosuccessfuldeployments.
lowpacketloss.Inpriorwork,wehavestudiedchannel-
inducedandprotocol-inducedlossesinlong-distanceset- 2.2 Challenges inRural areas
tings[33], andhaveaddressedthese problemsbycreat-
Addressingthesechallengesrequireslookingatalllevels
ingWiLDNet:aTDMA-basedMACwithadaptiveloss-
of the system, starting from the power supply and base
recoverymechanisms[26].Wehaveshowna2–5foldin-
hardware,upthroughthesoftwareanduserinterface,all
creaseinTCP/UDPthroughput(alongwithsignificantly
the way to training and remote management. Although
reducedlossrates)incomparisontothebestthroughput
remotemanagement,reliablepowerandtrainingofstaff
achievablebythestandard802.11MAC.Wehadshown
ishardingeneral,theseproblemsareexacerbatedinrural
these improvementson real medium-distance links and
areasforseveralspecificreasons[35].
emulatedlong-distancelinks.
First, local staff tend to start with limited knowledge
In this paper we confirm the emulated results with
about wireless networking and IT systems. This limits
data fromseveralreallong-distancelinksin developing
their diagnostic capabilities and results in inadvertent
regions. Working with Ermanno Pietrosemoli of Fun-
misuse and misconfiguration of equipment. Thus man-
dacio´n Escuela Latinoamericano de Redes (EsLaRed),
agement tools need to help with diagnosis and must be
we were able to achievea total of 6 Mbpsbidirectional
educational in nature. The effectiveness of training is
TCPthroughput(3Mbpseachwaysimultaneously)over
limited by the high turnover of IT staff, so education
asingle-hop382kmWiLDNetlinkbetweenPicoAguila
mustbeanongoingprocess.
Second,thechancesofhardwarefailuresarehigherbe-
causeofpoorpowerqualityandharshenvironments(e.g.
exposuretolightning,heat,humidity,ordust).Although
we donothaveconclusivedata aboutthe failurerate of
equipmentforpowerreasonsinruralareas,wehavelost
far more routers and adapters for power reasons in ru-
ralIndiathanwehavelostinourBayAreatestbed.This
callsforasolutionthatprovidesstableandhighquality
powertoequipmentinthefield.
Third, many locationswith wireless nodes, especially
relays,are quite remote,and thereforeitis importantto
avoidunnecessaryvisitstoremotelocations.Weneedto
enable preventive maintenance during scheduled visits.
Forexample,evidenceofagradualdegradationinsignal
strength at a remote router could indicate that a cable
needstobereplacedorantennasneedtoberealignedin
thecourseofanormalvisit. Figure 2: Aravind Telemedicine Network. Theni hospital is
Fourth, the wireless deploymentmay often notbe ac- connectedto5visioncenters.Theothernodesareallrelays.
cessibleremotelyorthroughtheInternet.Thefailureofa
singlelinkmightmakepartsofthenetworkunreachable,
evenifthenodesthemselvesarefunctional.Thismakes furtherexpansionto50clinicsaround5hospitalsisbe-
itveryhardforremoteexpertsorevenlocaladministra- ingplannedtoprovide500,000annualeyeexaminations.
torstoresolveorevendiagnosetheproblem. Hardware: The wireless nodes are 266 MHz x86 sin-
gleboardcomputers.Theseroutershaveupto3Atheros
3 Background 802.11 a/b/g radio cards (200–400 mW). The longer
links use 24dBi directional antennas. The routers con-
Over the last three years we have deployed two rural sumeabout4.5Wwhenidleandonly9.5Wwhentrans-
wireless networks in India. One is at the Aravind Eye mittingatfullbandwidthfrom2radios;7Wistheaver-
HospitalinsouthIndiawherewelinkdoctorsatthecen- agepowerconsumptionforanode.Theyrunastripped-
trallylocatedThenihospitaltovillageclinics,knownas down version of Linux 2.4.26 stored on a 512 MB CF
vision centers, via point-to-point WiLD links. Patients card, and include our software for WiLDNet, monitor-
video-conferenceoverthelinkswiththedoctorsforcon- ing,logging,andremotemanagement.
sultations.TheotherisinDharamsalainnorthIndiaand Theroutersareplacedinsmallandlightweightwater-
is called the AirJaldi network. This network is primar- proofenclosures,andaremountedexternally,closetothe
ily a mesh with a few long distance directional links antennas, to minimize signal losses. They are powered
that provides VoIP and Internet access to local organi- via power-over-ethernet (PoE); a single ethernet cable
zations.Bothnetworkshavefacedlargelysimilaropera- fromthegroundtotherouterissufficient.We useunin-
tionalchallenges,butwithsomeimportantdifferences. terruptiblepowersupplies(UPS)toprovidecleanpower,
althoughwediscusssolarpowerinSection5.2.
3.1 The AravindNetwork
Applications: The primary application is video-
The Aravind network at Theni consists of five vision
conferencing. We currently use software from Marrat-
centers connected to the main hospital in Theni (Fig-
ech [22]. Although most sessions are between doctors
ure 2). The network has total of 11 wireless routers (6
and patients, we also use the video conferencing for
endpoints,5relaynodes)anduses9point-to-pointlinks.
remote training of staff at vision centers. Typical
The linksrangefromjust1 km (Theni- Vijerani)to 15
throughput on the links ranges between 5–7 Mbps
km (Vijerani- Andipatti).Six of the wireless nodesare
with channel loss less than 2%. But 256 Kbps in each
installed on towers, heights of which range from 24–
direction is sufficient for very good quality video
42 m; the others use short poles on rooftops or exist-
conferencing.Ournetworkisthusoverprovisioned,and
ingtallstructures,suchasthechimneyofapowerplant
wealsousethenetworktotransmit4-5MB-sizedretinal
on the premises of a textile factory. Recently, Aravind
images. The hospital has a VSAT link to the Internet,
has expanded this model to their hospitals in Madurai
butmostapplicationsrequireonlyintranetaccesswithin
and Tirunelveli where they have added two vision cen-
thenetwork(exceptforremotemanagement).
ters. The network is currently financially viable and a
ThisbandwidthissufficientforapplicationssuchasIn-
ternet access and VoIP that cater primarily to the needs
oftheTibetancommunity-in-exilesurroundingDharam-
sala, namely schools, hospitals, monasteries and other
non-profitorganizations.AirJaldionlyprovidesconnec-
tivitytofixedinstallationsanddoesnotofferwirelessac-
cesstoroamingusersormobiledevices.Acost-sharing
modelisusedamongallnetworksubscriberstorecover
the operational costs. The network is currently finan-
ciallysustainableandisgrowingrapidly.
4 Operational Experiences
Figure 3: AirJaldi Network. There are 8 long distance links
withdirectionalantennaswith10endpoints. We have experienced several operational challenges in
both networks that have lead to significant downtimes,
3.2 The AirJaldiNetwork
increased maintenance costs, and lower performance
TheAirJaldinetworkprovidesInternetaccessandVoIP (e.g., increased packet loss). Initially we were involved
telephonyservicestoabout10,000userswithinaradius inallaspectsofnetworkplanning,configuration,deploy-
of 70 km in rural mountainousterrain characterized by ment,andmaintenanceofthenetworks.Ourspecificend
extremeweather.Thenetworkhas8longdistancedirec- goalhasbeentoultimatelytransferresponsibilitytoour
tionallinks rangingfrom10 km to 41 km with 10 end- ruralpartners,primarilytoensurelocalbuy-inandlong-
points(Figure3).Inaddition,thenetworkalsohasover termoperationalsustainability.Thisprocesshasnotbeen
ahundredlow-costmodifiedconsumeraccesspointsthat easy.Ourinitialapproachwastomonitorthesenetworks
use a wide variety of outdoor antennas. Three of the over the Internet and to provide some support for local
nodesaresolar-poweredrelaystationsatremoteelevated management, sometimes administering the network di-
placeswithclimbabletowers.Allotherantennasarein- rectly(bypassingthelocalstaffwheneverrequired).But
stalled on low-cost masts less than 5 m in height; the enablingremotemanagementhasbeenmorechallenging
masts are typically water pipes on the rooftops of sub- than expected because of severe connectivity problems
scribers. (Section5.3).
This aspect, combined with the desire to enable local
Hardware: Most of the routers are modified consumer
operational sustainability, has led us to design the sys-
devices, either Linksys WRT54GL or units from Buf-
tem with more emphasis on support for local manage-
falo Technologies, and cost less than US$50. They are
ment, a particularly challenging problem given limited
housed inside locally designed and built weatherproof
local experience. One way in which we have ensured
enclosures,andaremountedexternallytominimizesig-
thateducationremainsanongoingprocessisbycreating
nal losses. The antennas, power supplies and batteries
athree-tiermanagementhierarchy,inwhichlocalITven-
are allmanufacturedlocallyin India.The routerboards
dors(calledintegrators)withsomeexpertiseinnetwork-
arebuiltarounda200MHzMIPSprocessorwith16MB
ing were hired to form a mid-level of support between
of RAM, 4 MB of on-board flash memory, and a low
local staff and ourselves. With this tiered approach, the
powerBroadcom802.11b/gradio.WerunOpenWRTon
ruralstaff has graduallylearned to handle many issues;
these routers, and use open source software for mesh
the IT vendorsstill handle some, most notably installa-
routing, encryption, authentication, QoS, remote man-
tion,whileourrolehasreducedfromoperationalrespon-
agementandlogging.Forlongdistancelinksandremote
sibility to just shipping equipment. In the last year we
relaystationsweuseslightlyhigher-enddevicessuchas
have notinstalled any linksourselveseven thoughboth
the PCEngines WRAP boards, MikroTik routerboards,
networks have grown. We review this transition in our
andUbiquitiLS2s,allwithAtheros-basedradios.
conclusion.
Applications:TheInternetuplinkofAirJaldiconsistsof Although we were prepared to expect problems such
5 ADSL lines ranging from 144 Kbps to 2 Mbps for a aspoorconnectivity,poweroutages,andmisunderstand-
totalofabout7Mbpsdownlinkand1Mbpsuplinkband- ingsaroundproperusageequipmentusage,theactualex-
width.Thelongestlink fromTCV to Ashapuri(41km) tentoftheseproblemshasbeenverysurprising,requiring
achievesathroughputofabout4–5Mbpsat2–5%packet asignificantcustomdesignofthesystematalllevelsto
loss,whilethelinkfromTCVtoGopalpur(21km)only addresstheseissueseffectively.Asaresult,thereduced
getsabout500–700Kbpsat10–15%lossduetotheab- downtimesandlowermaintenancecostshaveresultedin
senceofclearlineofsight. bothnetworksbeingsustainableenoughtopayfortheir
Veerampattinam Failures from Bad Power Quality: We have experi-
500 to 1000
Villianur enced a wide range of failures from bad power. First,
400 to 500
300 to 400 spikesandsurgeshavedamagedourpowersuppliesand
-300 to -400 router boards. In the AirJaldi network, we have lost at
-400 to -500
least 50 power supplies, about 30 ethernet ports and 5
-500 to -1000
boards to power surges, while in the Aravind network,
0 100 200 300 400
wehavelost4boards,atleast5powersuppliesandsome
Figure4:Histogramofpowerspikesfromtworuralvillages. ethernetportsaswell.
Thebins(yaxis)arethesizerangeofthespikeinvolts,while Second, voltage sags have caused brown outs. Low
thexaxisisthecount.Negativebinsimplyreversedpolarity.
voltagesleave routersin a wedgedstate, unableto boot
completely. The on-board hardware watchdog, whose
jobistoreboottherouter,isalsooftenrendereduseless
ownequipmentandtowers.Beforemovingontothede-
becauseofthelowvoltages,thusleavingtherouterina
signofoursystem,wefirstdocumentthreemajorfactors
hungstateindefinitely.Third,fluctuatingvoltagescause
foroperationaloutages;eachfactorisaresultofacom-
frequentreboots,whichcorruptandoccasionallydamage
binationofthechallengespresentedinSection2.2.
theCFcardsthroughwritesduringthereboots.
4.1 Components AreMore Inclined toFail Asatypicalexample,therouteratSBSinAravindre-
bootedatleast1700timesinaperiodof12months(Fig-
Operating conditions at Aravind and AirJaldi have
ure 5), roughly 5 times per day, going up to 10 times
greatly contributed to a substantial decrease in the ro-
forsomedays.Incontrast,anotherrouteratAravindde-
bustness of system components that would otherwise
ployed on top of chimneyof a power plantfrom where
workquitereliably.Onemajorculprithasbeenthelack
itderivesreasonablystablepowerhasshownuptimesfor
ofstableandqualitypower.Althoughissuessuchasfre-
severalmonthsatastretch.Inpractice,wehaveobserved
quent power outages in rural areas are well known, we
that routers with more frequent reboots are more likely
weresurprisedbythedegreeofpowerqualityproblems
to get their flash memory corrupted over time. We had
in rural villages even when power is available. Before
atleast3suchcasesatnodesco-locatedwiththevision
addressing the power issues (Section 5.2), not a single
centers(Figure5),whichexperiencedmorerebootssince
daywentbywithoutfailuresrelatedtolowpowerqual-
staffattheselocationsshutdownandbootuptherouters
ity in either network. Any effort that is focused on ru-
everyday.Finally,frequentlyfluctuatingvoltagealsopre-
ral deployments must necessarily fix the power issues.
ventsoptimalchargingofthebatterybackupandhalves
Thereforewe describe the quality of rural power in de-
itsoveralllifetime.
tail,particularlybecauseithasnotbeenpreviouslydoc-
Lackofqualitypowerincreasesnotonlydowntimebut
umented.
also maintenance costs. Traveling to remote relay loca-
LowPowerQuality:Figure4showsdataonspikesfrom tionsjusttorebootthenodeorreplacetheflashmemory
a power logger placed in two differentrural villages in is expensive and sometimes has taken us several days,
southern India for 6 weeks. We group the spikes based especiallyinDharamsalawheretheterrainisrough.
on their magnitude in volts; negative voltage means
OtherPower-relatedProblems:InDharamsala,oneof
the polarity was reversed. We see many spikes above
thestormiestlocationsinIndia,lightningstrikeshaveof-
500V,oftenwithreversedpolarity,andsomeevenreach-
ten damaged our radios. We have learned the hard way
ing 1000V!Clearly such spikes can damage equipment
thatwheneverwedeployedamixofomnianddirectional
(burned power supplies), and has affected us greatly.
antennas,theradiosconnectedtotheomniantennaswere
We havealsoseenextendedsagsbelow70Vandswells
muchmorelikelytogetdamagedduringlightningstorms
above350V(normalvoltagein Indiais 220-240V).Al-
comparedtotheradiosconnectedtodirectionalantennas.
though the off-the-shelf power supplies we use func-
It turned out that omni-directional antennas attract
tionwellatawiderangeofinputvoltages(80V–240V),
lightning more as they are usually mounted on top of
theyarenotimmunetosuchwidelyrangingfluctuations.
mastsandhaveasharpertip,whiledirectionalantennas
Also,locationsfarawayfromtransformersaresubjectto
aretypicallymountedbelowthemaximumheightofthe
morefrequentandextremepowerfluctuations.Ourfirst
mast. To mitigate this problem, we install omni anten-
approachwastouseUPSandbatterybackups.However,
nas about 50 cm below the top of the mast. However,
affordable UPS systems are only of the “standby” type
thiscreatesdeadzonesbehindthemastwherethesignal
where they let grid power flow throughuntouched;this
fromtheantennaisblocked.Toreducethesedeadzones,
passes the spikes and surges through to the equipment
we sometimes use an arm to extend the omni antenna
except during grid outages when the battery starts dis-
away from the mast. After loweringthe omniantennas,
chargingandisexpectedtoprovidestablepower.
wehavenotlostanyradiosduringstorms.
thingseemedfineexceptthatwecouldnotcommunicate
Periyakulam withtheremoteend.Wehadnoothernetworkaccessto
Bodi the remote host so local staff kept physically checking
Chinamanoor
Andipatti the remote end, but did not (ourselves included) think
Ambasam of checking the roof at Theni. The resulting downtime
SBS
lasted for two months until we flew there and saw the
Laxmipuram
Vijerani problem!
Chimney
Theni2 Independent power supply Packet Loss due to Interference: In the AirJaldi net-
Theni1 Dependent power supply work, a decrease in VoIPperformancewas reportedfor
aparticularlinkatveryregularintervals.Howeverwith-
0 500 1000 1500 2000
outanyadditionalinformationtodiagnosetheproblem,
Number of reboots
noactioncouldbetaken andthisbehaviorpersistedfor
threemonths.Finally,aftersomedetailedmonitoringby
Figure5:NumberofrebootsestimatedpernodeintheAravind us (and not the rural staff), we saw a regular pattern of
network for about one year of operation. Nodes with power packetloss between8am to 9ameverydayexceptSun-
supplies dependent on the vision center are turned on or off
days.ButscanningthechannelsshowednoexternalWiFi
everyday.Nodeswithindependentpowersuppliesaretypically
interference.We were finally able to attribute the prob-
relaynodesorhospitalnodes.
lemtoapoorlyinstalledwaterpumpthatwasactinglike
apowerfulsparkgenerator,interferingwithwirelesssig-
nalsinthevicinity.Withoutpacketlossinformation,both
4.2 FaultDiagnosisisDifficult
the rural staff and we would have had a lot of trouble
Accuratediagnosisoftheproblemcangreatlyreducere- solvingthisproblem.
sponse time andthusdowntime.The mostcommonde-
SignalStrengthDecrease:IntheTheni-Ambasamlink
scriptionofafaultbyourruralpartnersisthatthe“linkis
intheAravindnetwork(Figure2),wenoticedadropin
down.”There are a wide varietyof reasonsfor network
signal strength of about 10 dB that persisted for about
outages and it is not always easy to diagnose the root
a month.Withoutfurtherinformationitwas hardto tell
cause. The lack of appropriate tools for inexperienced
whethertheantennasweremisaligned,orthepigtailcon-
staff, combinedwith unreliableconnectivitywhich hin-
nectorsweredamaged,ortheradiocardswerenolonger
dersdetailedmonitoring,preventsaccuratediagnosis.
workingwell.Intheend,severaldifferentattemptswere
Forexample,aremotehostmightberunningproperly,
made by local staff over multiple trips; the radio cards,
yet is unreachable when an intermediate wireless link
theconnectorsandeventheantennaswerereplaced,and
goesdown.Thenon-functionallinkmakesitimpossible
thesignalstrengthbumpedbackupwithoutitbeingfully
toquerytheremotehostfordiagnosis.Infact,therehave
clearwhatfinallyhelped!
been many instances where rural staff have traveled to
theremotesitewithgreatdifficultyonlytorealizethatit Network Partition:We experiencednetwork partitions
was a regular power shutdown from the grid (in which many times, but for several different reasons. For ex-
case nothing could be done anyway), or that it was a ample, at Aravind, staff misconfigured the routing and
software problem which could have been fixed if there added static routes while dynamic routing was already
wereanalternatebackchanneltotherouter.Accuratedi- enabled.Thiscreatedaroutinglooppartitioningthenet-
agnosisofsuchproblemscansaveconsiderabletimeand work. In another instance of operator error, the default
effort,andpreventunnecessarytravel.Furthermore,our gateway of one of the routers was wrongly configured.
ownabilitytohelpthelocalstaffbylogginginremotely Therewerealsoafewinstanceswhenoperatorschanged
to diagnose the problem is limited by connectivity. For the IP addresses of the endpoints of a link incorrectly,
instance,weusetheVSATlinkatTheni(intheAravind such that the link was non-functional even though it
network)toaidthelocalstaffinmonitoringandmanag- showed up as being associated. And as mentioned ear-
ingthenetwork,buttheVSATbackchannelhasworked lier,theconstructionoftheelevatorshaftleftthenetwork
foronly65%ofthetimeinthelastoneyear. partitionedfortwomonths.
Sometimes local misunderstandingsof equipment us- “Fixing” by users: A recurring problem is that well-
age make it even harder to diagnose problems. For ex- meaningruralstaffoftenattempttofixproblemslocally
ample,asshowninFigure6,anelevatorshaftwascon- when the actualrootcause is notlocal. For example,at
structed right in front of the directional antenna at Ar- AirJaldi we have seen that when an upstream ISP goes
avindThenihospital,completelyobstructingthe line of down,ruralstafftendtochangelocalsettingsinthehope
sight to the remote end. Whenever we remotely logged of fixing the problem. These attempts typically create
in to the Theni end of the link from Berkeley, every- new problems, such as misconfiguration, and in a few
Problemdescription System
Aspects
ComponentFailures
Unreliablepowersupply P
BadpowercausingburntboardsandPoEs P
CFcardcorruption:diskfullerrors M,P,S
Omniantennasdamagedbylightning P
FaultDiagnosis
Packetlossfrominterference M
Decreaseinsignalstrength M
Networkpartitions M,B
Selffixingbyusers S
Routingmisconfigurationbyusers M,B,S
Failedremoteupgrade B,R
Remoterebootafterroutercrash B,R,S
Figure 6: TheTheni toVijeranilinkintheAravindnetwork Spyware,viruseseatingbandwidth M,S
was completely obstructed by a newly constructed elevator AnticipatingFaultsishard
shaft. This problem was not resolved until we visited Theni Findingbatteryuptime/status M,B,P
after2months. PredictCFdiskreplacement M
Table1:ListofsometypesoffaultsthatweseeninbothAr-
caseshaveevenresultedindamagetoequipment.Inall
avindandAirJaldi.Foreachfault,weindicatewhichaspectsof
these cases, the network remained non-functional (but
thesystem,aswehavedesignedit,helpmitigatethefault.The
now for a different reason) even after the ISP resumed differentaspectsareMonitoring(M),Power(P),Backchannel
normalconnectivity.Thusweneedmechanismstoindi- (B),IndependentRecoveryMechanisms(R)andSoftware(S).
catewhenalinkishavingproblemsattheremoteend,so Theinformationonfaultshasbeencollatedfromlogsandinci-
astopreventlocalattemptsatrepair. dentreportsmaintainedbythelocaladministratorsandremote
Thegeneralthemeisthatnomatterwhatthefault,ifthe expertsrespectively.
link appearsto be down with no additionalinformation
orconnectivityintothewirelessnode,itishardforeven Predicting Battery Lifetime: Battery life is limited by
experiencedadministratorstoresolvetheproblem. the number of deep cycle operations that are permit-
ted. This lifetime degrades sharply because of fluctuat-
4.3 Anticipating Faults isHard
ingvoltagesseeninourdeploymentsthatdonotcharge
Some of the node locations in our networks, especially the batteryoptimally,AtAravind,batteriesratedwith a
relays, are quite remote.Site maintenancevisits are ex- lifetimeoftwoyearslastforroughlythreetosixmonths.
pensive, time consuming, and require careful planning Informationaboutremainingbatterylifecanalsoenable
around the availability of staff, tools, and other spare preventionofcatastrophicfailures.
equipment.Therefore,visitsaregenerallyscheduledwell
Predicting Disk Failure: We have observed that with
inadvance,typicallyonceeverysixmonths.Inthissce-
frequent reboots over time, the disk partition used to
nario, it is especially important to be able to anticipate
store system logs accumulates bad ext2 blocks. Unless
failuresso thatthey can be addressedduringthe sched-
we run fsck periodically to recover the bad blocks, the
uled visits, or if a catastrophic failure is expected, then
partition becomes completely unusable very soon. We
a convincingcase canbemadeforanunscheduledvisit
have also seen thatmanyflash disks show hardwareer-
for timely action. But without an appropriate monitor-
rors,anditis importantto keeptrackofdisk errorsand
ing and reporting system that includes backchannels, it
replacethembeforetheycauserouterstocompletelyfail.
isdifficulttoprepareforimpendingfaults.
Battery Uptime: At both Aravind and AirJaldi we use
5 System Architecture Design
batterybackups.Lossofgridpoweratthenodescauses
their batteries to start discharging It is generally not
In this section, we present five aspects of our system:
knownwhenthebatterieswillfinallyrunout.Ifthisin-
monitoring,power,backchannels,independentrecovery
formationissomehowprovidedtothestaff,theycanpre-
mechanisms, and software. Each has been designed to
ventdowntimeofthelinkbytakingcorrectivemeasures
specifically address our goals of increasing component
suchasreplacementofthethebatteryintime.Suchfeed-
robustness,enablingfaultdiagnosis,andsupportingfault
backwouldalsosuggestiftheproblemwereregional(as
prediction. For each aspect, wherever appropriate, we
otherrouterswouldalsosufferlossofgridpower)orsite-
also discuss tradeoffs affecting our design choices. Ta-
specificsuchasacircuitbreakertrip.
ble 1 indicates which aspects of our system design are
important for reducing the impact of some of the com- Scope Type MeasuredParameter
monfaultspresentedintheprevioussection. Passive CPU,diskandmemoryutilization,in-
Node
terrupts, voltage, temperature, reboot
5.1 Monitoring
logs (number & cause), kernel mes-
Allaspectsofsystemmanagementrequiresomelevelof sages,solarcontrollerperiodicdata
monitoring.DuringtheinitialdeploymentatAravind,we Active disksanitycheck
facedtwomainchallengesindesigningamonitoringsys- Link Passive traffic:, traffic volume(#bytes, pack-
ets)
tem. First, the Aravind network at Theni only allowed
wireless: signal strength, noise level,
ustoinitiateconnectionsfromwithinthenetwork.Sec-
#controlpackets,#retransmissions,#
ond,localstaffwasnotfamiliarwithLinuxorwithcon-
droppedpackets
figuration of standard monitoring software such as Na-
interference:#ofstationsoverheard&
gios[10].
packet count from each, # corrupted
This led us to build a push-based monitoring mecha- packets
nismthatwecall“PhoneHome”inwhicheachwireless Active liveness, packet loss, maximum link
router pushes status updates upstream to our US-based bandwidth
server.Wechosethismethodoverthegeneralpull-based Passive routechanges,pairwisetrafficvolume
System
architectureinwhichadaemonrunningonalocalserver &type
pollsallthe routers.Thepull-basedapproachwouldre- Active pairwise end-to-end delay & max
throughput
quireconstantmaintenanceviare-configurationofa lo-
calservereverytimeanewrouterwouldbeaddedtothe
network. In contrast, the push-based approach enabled Table2:ParameterscollectedbyPhoneHome.
us to configurethe routersonlyonce,at installation, by
andremoteservers.
specifyingtheHTTPproxytobeused.
PhoneHomeprovedtobehelpfulinunderstandingfail-
The Aravind network features two remote connectiv-
ures, diagnosing and predicting many faults. First, it
ityoptions,bothofwhichareslowandunreliable(Sec-
helpedmaintainnetworkreachabilityinformation,alert-
tion 5.3): (1) a direct CDMA network connection on a
ingthelocalstaffwhenthenetworkwasdownandaction
laptopatthecentralhospitalnode,and(2)aVSATcon-
neededtobetakentorecover.Earlier,onlyaphonecall
nectiontoanotherhospital,whichhasaDSLconnection
fromaruralcliniccouldalertthelocaladministrator,and
to the Internet. PhoneHome is installed on each of the
dependingontheawarenessofthestaffattheruralclinic,
wirelessrouters.Alltheroutersperiodicallypostvarious
thiscallwouldnotalwayshappen.
parameters to our US server website. Server-side dae-
Second, kernel logs transferred using PhoneHome
monsanalyzethisdataandplotvisualtrends.
helpedusdiagnoseseveralinterestingproblems.Forex-
Wecollectnodeandlink-levelinformationandend-to-
ample, in certain instances routers configured with two
end measurements. The comprehensivelist of the mea-
network interfaces reported only one interface as being
sured parametersis presentedin Table 2. Most of these
active.Pairingthisinformationwithpowerdata,wereal-
parameterscan be measuredpassively,withoutinterfer-
izedthatalowvoltagesupplycanpreventtworadioin-
ing with normal network operation. However, several
terfacesfromfunctioningsimultaneously.Inanotherin-
of these measurements, such as maximum link or path
stance, kernel logs and system messages allowed us to
throughput, require active testing. Some of these tests
examineflashdiskerrormessagesandpredictwhendisk
can be performed periodically (e.g. pinging every net-
partitionsneededrepartitioningorreplacement.
workhost),andsomeofthemaredoneondemand(e.g.
Third,byexaminingthepostedroutingtableandinter-
findingthethroughputachievableona particularlinkat
face parameters,we were able to diagnoseroutingmis-
agiventime).
configurationsorbadlyassignedIPaddresses.
We also use the PhoneHome mechanism for remote
Fourth,continuousmonitoringofwirelesslinkparam-
management. Every time PhoneHome connects to our
etershelpedusnarrowthescopeoftheproblemsinmany
US server, it opens a reverse SSH tunnel back into the
cases. Figure 7 shows the signal strength variation in
wirelessnode,enablinginteractiveSSHaccesstotheAr-
someofournetworklinks.Whilemajorityoftheselinks
avindmachines.AstheVSATconnectiononlyallowsac-
showfairlystablesignalstrength,someofthemshowim-
cessoveranHTTPproxy,wearerequiredtorunSSHon
portantvariationovertime.Forexample,asudden10dB
topofHTTP,andconfigurePhoneHomewiththeproxy.
signaldroponthelinkbetweenAmbasamtoTheniindi-
IncaseofadirectconnectiontotheInternet,nosuchcon-
catedsomekindofadrasticeventsuchasapossiblean-
figuration is required. Another option (employed in the
tenna misalignmentthat needed an immediate visit. On
remotemanagementofAirJaldi)istousetheOpenVPN
theotherhand,asteadydeclineinsignalstrengthonthe
software to openVPN tunnelsbetween networkrouters
Bodilinkindicatedagradualdegradationofaconnector
incidentswe haveexperiencedhavebeenrelatedto low
Chimney
40 power quality. Thus, designing to increase component
Laxmipuram
B)30 Ambasam reliability in the face of bad power is the most impor-
d Bodi tant task. We have developed two separate approaches
NR (20 to address the effects of low power quality. The first is
S
a Low Voltage Disconnect (LVD) solution, which pre-
10
vents both routers from getting wedged at low voltages
0 andalsoover-dischargeofbatteries.Thesecondisalow-
0 60 120 180 240 300 360 420 480 540
Time (days) cost power controller that supplies stable power to the
equipmentbycombininginputfromsolarpanels,batter-
Figure7:Signalstrength(shownindB)variationforalllinks. ies,andeventhegrid.
Each point isaverage of measurement over 2 days. The Am-
LowVoltageDisconnect(LVD):Over-dischargeofbat-
basamlinkshowsatemporarydropinSNRof10dBforabout
teries can reduce their lifetime significantly. Owing to
40days.WhiletheBodilinkisgraduallydegradingasitsSNR
the poor quality of grid power, all AirJaldi routers are
hasdroppedby4dBoverthelastyear,theChimneylink’sSNR
onbatterybackup.LVDcircuits,builtintobatterycharg-
hasremainedconstant.
ers, prevent over-discharge of batteries by disconnect-
ortheRFcabletotheantenna,andrequiredaneventual ingthe load(router)whenthe batteryvoltagedropsbe-
visit. lowathreshold.Asabeneficialside-effect,theyprevent
therouterfrombeingpoweredbyalow-voltagesource,
Tradeoffs:We contrastthiswith monitoringat AirJaldi
which may cause it to hang. Off-the-shelf LVDs oscil-
where we use various off-the-shelf tools such as Na-
lated frequently, bringing the load up and down, and
gios[10]andSmokePing[13]tocollectnode,link,and
eventually damaging the board and flash memory. Ev-
network level parameters. Informationis stored at a lo-
ery week, there were roughly fifty reboot incidents per
cal data server in Dharamsala and then copied to a US
routerduetohangscausedbylowvoltage.However,we
server for detailed analysis. Various graphing toolkits
designedanewLVDcircuit[24]withnooscillationand
suchasMRTG[25]areusedtovisualizetrendsandde-
better delay; since then the hangs per week per router
tectanomalies.
havereducedtonearzerointheDharamsalanetwork.
The differencein approachcomparedtoAravindisin
part due to the higher experience of the AirJaldi staff, Power Controller: We have developed a
andinpartduetothebetterconnectivitywehavetoAir- microcontroller-basedsolarpowerchargecontroller[31]
Jaldi. The advantageof having local serverspolling for that provides a stable input of 18 V to the routers and
informationisthattheycanbeconfiguredbylocalstaff intelligently manages the charging and discharging
to look for relevant problems, but such an approach is of the battery pack. It has several features such as
maximumpowerpointtracking,lowvoltagedisconnect,
beneficial only if local staff are experienced enough to
takeadvantageofthesefeatures. tricklechargingandveryimportantly,supportforremote
After three years of operation, the local Aravind staff managementvia ethernet. The setup is trivial as it sup-
(someofwhomwelostduetoturnoveraftertheygained pliespowertotherouterusingPoE.Thiscombinationis
more experience through our training) are more famil- novelforitspriceofaround$70.
iar with system configuration,and show less apprehen- We use TVS diodes to absorb spikes and surges and
sion in taking the initiative and maintaining the system a robustvoltage regulatorto get clean 18V power from
ontheirown.Therefore,we arenowbeginningto usea wide ranginginputconditions.Figure 8 shows the flow
ofcurrentthroughtheboardovera60-hourperiod.First,
pull-basedmodel.
Ingeneral,webelievethatduringtheinitialphaseofa we note that power is always available to the router.
networkdeployment,minimalconfigurationpush-based Whenenoughsunlightisavailable,thesolarpanelpow-
mechanisms are more appropriate for data collection. erstherouterandchargesthebattery.Duringperiodsof
However,afterbuildingenoughlocalexpertise,themon- no sun, the battery takes over powering the router. The
itoringsystemshouldbemigratedtowardsamoreflexi- frequentswingsobservedontheleftpartofthegraphare
blepull-basedapproach. typical for a cloudy day. The graphs also demonstrate
how the batteryis continuallychargedwhen sunlightis
5.2 Power
available.Wehavemeasureda15%moreefficientpower
Powerqualityandavailabilityhasbeenourbiggestcon- drawfromthe panels,and also expectthatwe can dou-
cern at both Aravind and AirJaldi. Low-quality power blebatterylife.Usingthecontroller,wehavenotlostany
damages the networking equipment (boards and power routersfrombadpower,butithasbeenonly8monthsof
adapters)andsometimesalsobatteries.Over90%ofthe testing.
4000 NetworkBackchannel: Atthe AravindThenihospital,
wealreadyhadsomeformofbackchannelintotheTheni
networkthroughVSAT.WeusePhoneHometoopenan
SSHtunnelovertheVSATlinkthroughanHTTPproxy
s
p at the Aravind Madurai hospital. We configure Phone-
m
a 0 Home to post monitoring data to our US-based server
Milli every 3 hours and also to open a reverse SSH tunnel
throughwhichwecanlogbackinforadministrationpur-
Load Current poses. Out of the 2300 posts expected from the router
Battery Current
at Theniover143 days (2 posts every 3 days), we only
Solar Current
-4000 received 1510 of them, or about 65%. So this particu-
0 6 12 18 24 30 36 42 48 54
larbackchannelwasnotveryreliableinpractice,some-
Time (hours)
timesnotworkingforlongstretchesoftime.Asaresult,
Figure8:Currentflowover60hours.Theloadstaysevenat weusedthesolitaryhospitallaptoptoconnectdirectlyto
7W, whilethe solar panel and battery shift their relativegen- the Internetusing a 1xRTT CDMA cardto improvethe
eration over time. The battery current is negative when it is availabilityofabackchannelintothenetwork.However,
charging.
this laptop was used for several other purposes (shared
hardware is a common feature in rural areas) and was
Thecontrollerreportssolarpanel,loadandbatterysta-
mostly unavailable.Furthermore,in many instances the
tusinformationthatcanbeusedforremotediagnosisand
networkbackchannelwas notenoughasthe localwire-
somepredictionofbatteryuptimeandlifetime.Asecond
lessnetworkwoulditselfbepartitioned.
version of the controller, currently under development,
Node Backchannel: At AirJaldi, we built a node
willaddthefeaturetotakegrid-suppliedpowerasinput.
backchannelmechanismusingGPRS.InIndiaatthemo-
This has two major advantages: the same setup can be
ment, GPRS connectivity costs roughly $10 per month
usedto stabilize gridpowerlocally,andgridpowercan
forunlimiteddurationandbandwidth.WeusedaNetgear
alsobeusedtochargethebatteriesinadditiontothesolar
WGT634U router, interfaced through its USB 2.0 port
power.
with a mobile phone. The router runs PPP over GPRS
Tradeoffs:Therealcostofpowerinruralareasisnotjust
and sets up an OpenVPN tunnel to a remote server. To
the raw grid electricity costs, but the cost of overcom-
enable remote diagnosis using this link, the backchan-
ing power availability and quality issues through UPS,
nelrouterisconnectedtothemainwirelessrouterusing
battery-backups, and and chargers. The recurring costs
ethernet and optional serial consoles. The backchannel
can be quite high, and therefore solar power, although
router can also power-cycle the wireless router using a
stillexpensive,becomesmorecompetitvethanexpected
solid-staterelayconnectedtooneofitsGPIOpins.
as it can produce clean power directly. Currently we
Thisapproachhastwoadvantages.First,thecellphone
choosetousesolarforveryremotelocations.Atlessre-
networkis completelyindependentof the wireless link.
mote and critical sites, we tend to use “dumb” analog
Second, even thoughthe mobile phone is chargedfrom
chargerstoreducecostsevenfurther.
the same powersource, it hasits own battery which al-
5.3 Backchannels lowsaccessviaGPRS evenif themain powersourceis
down. However, for the Netgear router, we needed ad-
AwidevarietyofproblemsatAravindandAirJaldihave
ditional battery backup which adds to the maintenance
caused link downtimes, leaving remote nodes discon-
complexity.Oneapproachtosimplifythissetupforcon-
nected.Thefailureofasinglelinkmakespartofthenet-
soleaccesswouldbetouseaLinuxGPRSphonebutwe
work unreachablealthoughthe nodesthemselvesmight
havenottriedityet.
be functional. In many cases, if we had alternate ac-
Tradeoffs:Our experiencewith the GPRS backchannel
cesstothenodes,thefixeswouldhavebeensimplesuch
intermsofprovidingrealutilityforsystemmanagement
ascorrectingaroutermisconfiguration,orrebootingthe
hasbeenmixed.Manycommonproblemscanbesolved
router remotely. It is important to have out-of-bandac-
by alternative means in simpler ways. In cases of in-
cessorabackchanneltothenodesthatisseparatefrom
correct configuration of routers, we can imagine using
the primary wireless path to it. Backchannel access is
theGPRS backchannelto fixproblems.ButatAravind,
alsousefulincaseswherethebatteryisdischargingbut
when misconfigurationsresulted in routing outages, we
therouterisalreadydownforotherreasons.Information
used cascaded hop-by-hop logins to move through the
aboutthebatterystatusfromthechargecontrollerviathe
network,althoughthisdependedonatleasttheendpoint
backchannelwouldstillbehelpful.Wehavetriedseveral
IP addresses to be set correctly. However, we can also
approachestobackchannelsinbothnetworks.