50
High Capacity Wireless Networks through Collaboration and Intelligent Information Storage Leandros Tassiulas Yale University IEEE PerCom 2018, Athens Greece March 21, 2018

High Capacity Wireless Networks through Collaboration and

  • Upload
    others

  • View
    1

  • Download
    0

Embed Size (px)

Citation preview

HighCapacityWirelessNetworksthroughCollaborationandIntelligent

InformationStorage

Leandros TassiulasYaleUniversity

IEEEPerCom 2018,AthensGreeceMarch21,2018

Outline

• Introduction

• Intelligentinformationstoragesolutionsoptimizedtoaccountforthe:1. hierarchicalstructureofthenetwork

2. application-levelrequirements(videostreaming)3. mobilitybehavioroftheend-users

4. economicaspects(storageownedbythird-partyentities)

• Conclusion

2

Explosivemobiledatatrafficgrowth

• 8-foldincreaseinglobalmobiledatatraffic in2017-2023.

– Closeto95%oftraffic willcomefromsmartphones.

• Networkcapacity needstoscaleaccordinglytopreventcongestionanddegradationinuserservicequality.

Source:Ericsson.com (Nov.2017)

3

ExaBytes

perm

onth

Needfornovelnetworkcapacityexpansionmethods

• Severalnetworkexpansioninitiatives:– Developinghighfrequencytechnologye.g.,60GHz(5G).– Releasingmoreunregulatedspectrum.

– Borrowingunusedspectrum(cognitiveradionetworks).– Increasingwirelesslinkefficiencybits/hertz/meter.

– Offloadingmobiledatatocomplementarynetworks(WiFi).

• Wirelessoperators’nightmare: costsnotmatchedbyrevenues:– E.g.,ChinaMobilereported115.1% increaseinmobiletraffic,but

10.2% profitreduction(Proc.1st IEEE5GSummit’15).

4

Information/content-centric solutions:why?

1. Asynchronouscontentreuse:– 75%ofmobiledatatrafficwill bedueto

recurringrequests forpopular videocontent.

2. Highpredictabilityofcontentdemand:– “What”,“when”and“where” thevideo

requestswill appear(machine learning,…).

5

video

filesharing

webbrowsing

softwaredownload

audio

socialnetwork

other

• Storing/cachingpopularvideosatthewirelessedge:– Increasenetworkcapacity,reducenetworkscosts(tradeexpensive

bandwidthforcheapstorage)&improveuserqualityofexperience.

Maintechnicalquestions

1. Where andhowmuchstorageresourcetodeploy?i. Corenetwork(packet/servinggateways).ii. RAN(macro/femto-cellBSs,fogservers).iii. Mobileuserequipment(smartphones).

2. What videostocache?forhowlong?✘ Video clipsmaybecomeviralwithin

ashorttime& thendisappear.

3. How toroutevideostouserswithinthenetwork?

• Interrelated decisions,albeitatdifferenttimescales.6

Twofamiliesof“whattocache”approaches

• MainusageinInformation-CentricNetworking(ICN).– Explicitlynamed contentchunks/packets.– ExtensionoftheP2Pparadigm.– Minimize traditionalmetric(e.g.retrieval latency).

Caching

OfflineOnline

ContentisopportunisticallyStoredbythenodesalongthepathtothecontentsourcee.g.,LRU,LFUmethods.

7

Twofamiliesof“whattocache”techniques(cont.)

• UsageinCDNs,VoD,wirelessnetworks(recenttrend).– FacilitatedbyrecentadvancesonSDN/virtualization technologies.– Optimize variousperformancemetrics (contentdelivery delay,

bandwidth consumption, video streamingQoE,etc.).

Caching

OfflineOnline

Contentisproactively storedbasedonthepredicteddemand.

8

Recenttrend:cachinginwirelessnetworks• CommercialsystemssupportingBS/WiFi APcaching.

• Newchallengesinheterogeneouswirelessnetworks:– Severalcandidatecachelocations(core,macro-,femto-cell,WiFi APs).

– Wireline/wirelessbackhaullinkswithdifferentcapabilities.

– Otherwirelessaspects(mobility,interference,unreliability).

March2014LinksyssmartWiFi routerswithexternalstoragedevices

9

Goingbeyondstateoftheart

• Wedesignnovelcachingschemesthatareoptimizedtoaccountforthe:1. hierarchicalstructureofthenetwork

2. application-levelrequirements(videostreaming)3. mobilitybehavioroftheend-users4. economicaspects(storageownedbythird-partyentities)

10

Goingbeyondstateoftheart

• Wedesignnovelcachingschemesthatareoptimizedtoaccountforthe:1. hierarchicalstructureofthenetwork

2. application-levelrequirements(videostreaming)3. mobilitybehavioroftheend-users4. economicaspects(storageownedbythird-partyentities)

11

Whyhierarchiesareimportant?

• Radioaccessnetwork(RAN)andMobilePacketCoretendtohavehierarchicalstructure.

• Cachescanbeinstalledatvariouslevels.

• Userrequestsforcontentfilesareroutedupwardsuntiltheyreachacacheorserverwiththesevideos.

• Technicalquestion7:– Whichfilestoplaceineachcachesoastomaximize“cachehits”?

12

SGW SGW

PGW

Internet

100GB

200GB

300GB

200GB

100GB100GB 100GB

servers

7K. Poularakis, L. Tassiulas, “On the Complexity of Optimal Content Placement in HierarchicalCaching Networks”, IEEE Transactions on Communications, 2016.

Formulatingthehierarchicalcachingproblem• Combinatorialproblem(NP-Hard).

• Integervariable𝒙𝒏𝒇 ∈ {𝟎, 𝟏} forplacingvideo filefatcachen.

• Constraint1:eachcachencanstoreuptoanumberoffiles.• ∑ 𝒙𝒏𝒇∀𝒇𝒊𝒍𝒆𝒇 ≤ 𝑪𝒏

• Constraint2: eachusercanaccessonlythecachesonhispathtoservers(no‘U-turn’isallowed).

• Maximizecachehits:∑ ∑ 𝝀𝒖𝒇𝟏{∑ 𝒙𝒏𝒇4𝟏}𝒄𝒂𝒄𝒉𝒆𝒏∈𝒑𝒂𝒕𝒉(𝒖)𝒇𝒊𝒍𝒆𝒇

∀𝒖𝒔𝒆𝒓𝒖

13Anticipateddemandofuser1λ1f

servers

Anticipateddemandofuser2λ2f

𝑪𝟏

𝑪𝟐

𝑪𝟑

𝑪𝟒

𝑪𝟓

Hitifatleastonecacheonthepathofuseruhascachedfilef

Indicatorfunction1{c}=1ifconditioncistrue;otherwise0.

Atractablespecialcase

• Whencachesareinstalledonasinglehierarchypath,theconstraintmatrixistotallyunimodular.Therefore,solvingthelinearrelaxedproblem(e.g.,usingSIMPLEX)yieldstheoptimal integralsolution.

• Example:

14

servers

𝑪𝟏

𝑪𝟐

𝑪𝟑

TotallyUnimodular Constraintmatrix(thedeterminantofeverysquare

submatrixis+1,-1or0).

Cachesonlyontheredpath

Cachesizeconstraints

Eachfileiscachedinat

mostonecache

Anapproximationalgorithmforthe2-levelcase

15

• Observation: Foragivenfileplacementattherootcache,theoptimalpolicyplacesateachleaf thelocallymostpopularfilesnotincludedintherootcache.– Wecanexpress theobjective (cachehits)asa

functionofthecachedcontentattherootonly.– Noneedtooptimizeallthecachesatthesame

time.Instead,optimizeonlytheroot!

• 1.582-approximationalgorithm(bestinliterature):– Iterativelyplacesthefiletotherootcachethatincreases theobjective

themostuntiltherootcacheisfull.Ateachiteration itadjuststhecachedcontentattheleavesaccordingly.

• WecanextendittoN-levelhierarchies,foranyN>2,withaslightdeteriorationoftheapproximationratio.

𝑪𝟒

root

leaves

Comparewith3state-of-the-art cachingschemes

• Max-Popularity:Eachcachestoresthemostpopularfilesbasedonitslocaldemand[noapproximationratio].

• Iterative-Greedy1,8:Startingwithallthecachesbeingempty,iterativelyplacesthefiletothecachethatimprovesmosttheobjectivefunction.Endswhenallcachesbecomefull.[2-approximationratio]

• Swapping8:Startingwitharandomcacheplacement,iterativelyswapsafileinacachewithafileoutofitifthisimprovestheobjectivefunction.Endswhenthereisnosuchswappingpossible[2-approximationratio]

16

1Golrezaei, et al. IEEE Infocom’12. 8Borst et al. IEEE Infocom’10

Evaluatinghierarchicalcaching

17

ü Upto56%server loadgains.ü Gainsincreasewiththeroot

cachesize.ü Evenhighergainsforsteeper

demanddistributions.

• Setup:a3-levelhierarchy,13caches,500files,Zipf (0.8)demand.• Measuretheserverload (totalfilerequestsminuscachehits)achievedbyeachalgorithm.

• Thefigureshowstheperformancegains (%)ofouralgorithmoverstate-of-the-artmethods.

Impactofrootcachesize

10%

56%

Goingbeyondstateoftheart

• Wedesignnovelcachingschemesthatareoptimizedtoaccountforthe:1. hierarchicalstructureofthenetwork

2. application-levelrequirements(videostreaming)3. mobilitybehavioroftheend-users4. economicaspects(storageownedbythird-partyentities)

18

Uniquechallengesofvideodelivery

• Incontrasttoothertypesofcontent,videos shouldbeavailableatvariousqualities toserveuserswithdifferentrequirements.

19

Videoencodingtechnologies&caching

• ScalableVideoCoding(SVC/H.264)createsmultiplelayers foreachvideo,whichwhencombined producedifferentqualitylevels.

– Layer1byitselfproducesquality1,layer1combinedwithlayer2producequality2,etc.

• Idea: cachelayers,notentirevideofiles.– Userscandownloadtherequiredlayersfromdifferentcaches.

– Videoplaybackisconstrained bythelayerwiththelargestdelay.

Layer1 Layer2Cannotbedecodedbyitself+ =

20

Modelinglayeredvideocaching

• Anabstractdistributedcachingarchitecture:– A setofcachesreceiverequestsforlayeredvideos.

– Cachescanexchangelayersondemand(“coordinatedmode”).

– A serverdeliversthelayersnotfoundinthecaches.

21

9K. Poularakis, G. Iosifdis, A. Argyriou, I. Koutsopoulos, L. Tassiulas, “Caching and OperatorCooperation Policies for Layered Video Content Delivery”, in Proc. IEEE Infocom, 2016.

cacheN

cache1

cache2

Remoteserver0

d10 d20 dN0

d12

𝜆CDE 𝜆FDE 𝜆GDE

Delaybetweenacacheandtheserver(sec/bits)

Predicteddemandforvideovatqualityq(requests/sec)

Layeredvideocachingproblemformulation• Optimizationvariables xnvl 𝝐 𝟎, 𝟏

• Goal:minimizeaveragedeliverydelay:

∑ 𝜆IDE 𝐦𝐚𝐱MNOPQM:CSTE

𝑑𝑒𝑙𝑎𝑦I,M∀ZNZ[PI,D\]PTD,E^NM\SOE

delayd,e = g𝑑Ih , 𝑖𝑓𝑛𝑜𝑛𝑒𝑜𝑓𝑡ℎ𝑒𝑐𝑎𝑐ℎ𝑒𝑠ℎ𝑎𝑠𝑠𝑡𝑜𝑟𝑒𝑑𝑙

𝑑II∗, 𝑖𝑓𝑐𝑎𝑐ℎ𝑒𝑛∗𝑖𝑠𝑡ℎ𝑒𝑐𝑙𝑜𝑠𝑒𝑠𝑡𝑡𝑜𝑛𝑡ℎ𝑎𝑡𝑠𝑡𝑜𝑟𝑒𝑑𝑙0, 𝑖𝑓𝑐𝑎𝑐ℎ𝑒𝑛ℎ𝑎𝑠𝑠𝑡𝑜𝑟𝑒𝑑𝑙

storing of lthlayer of video v at cache n

22

delay for downloading layer l to a local user of cache n

cachesizeconstraint: } o~exd~e∀~���� ~,∀e���� e

≤ Cd

size of layer l of video v

size of cache n

Non-linear objective

“Easy”specialcaseforonecache• Optimalsolutionstructure foraspecialcasewithN=1cache:

– Thelayer𝑙 ofavideoshouldnotbecachedunlessallpreviouslayers𝑙� ≤ 𝑙ofthisvideoarecached.

• Polynomial-timereducibletothemultiple-choiceknapsack(MCK) problem.

• Examplewith2videosand3layerspervideo.

• Atmostoneblue item,atmostonereditemintheknapsack.

– Pseudopolynomial-time optimal andFPTA algorithms.23

Knapsack size=cachesize

Layer1

Layers1+2

Layers1+2+3

Layer1

Layers1+2

Layers1+2+3

Video1 Video2

Item1

Item2

Item3

Item4

Item5

Item6

Layer-awarecooperativecaching(LCC)algorithmformanycaches• UsetheMCKsolutiontosolvethegeneralcase.

• LCCalgorithm(inputparameterF𝜖 0,1 ):1.partition eachcacheintotwoparts:• a partFforglobally(acrossallcache-nodes)popularvideolayers,• Apart(1-F)forlocallypopularvideolayers.

24

Cache1 Cache2 Cache3

F1-F

Layer-awarecooperativecaching(LCC)algorithmformanycaches• UsetheMCKsolutiontosolvethegeneralcase.

• LCCalgorithm(inputparameterF𝜖 0,1 ):2.Fill inthefirstcacheparts bysolvingaMCKproblemwithknapsack

sizeequaltothetotal cachesizeallocatedforgloballypopularvideos.

25

Cache1 Cache2 Cache3

F1-F

Layer-awarecooperativecaching(LCC)algorithmformanycaches• UsetheMCKsolutiontosolvethegeneralcase.

• LCCalgorithm(inputparameterF𝜖 0,1 ):3.foreachcache-noden:

fillinitsremaining cachespacebysolvingaMCKproblemwithknapsacksizeequaltotheremainingcachesize,consideringthelayersalreadyplacedatthepreviousstep.

2-approximationratio forthesymmetricdelayscase.26

Cache1 Cache2 Cache3

F1-F

Lowerdelayoverstate-of-the-art

• Compareouralgorithm(LCC)with:– IndependentCaching(IC):eachcacheserves its localrequests only.Caching is

performedbasedonMCKsolution.

– Iterative-Greedy: iteratively,cachesthevideo layerthatreduces theobjective function(averagedelay)themostuntilall thecachesbecomefull.

• Setup: 3caches,1TBsizeeach,10,000 videos,5layerseach,layersizesbasedonhttp://trace.eas.asu.edu,~10TBstotallayersize,1Mbpsserverrate,Zipf(0.8)videorequests,uniformacrossqualities. 27

Impactofratebetweencaches(1-10Mbps)25%delaygains

Sidebenefit:improvedstreamingperformance

• LCCisdesigned tooptimize averagedelay,notperformancemetricsrelatedtovideo streaming.

• SimulateDASHstreamingprotocol(dynamically adaptplaybackqualitybasedonnetwork load)andmeasure streamingperformance.

• Setup: asbefore,butnon-constantrate:requestsarrivedynamically,andthedeliveryrateofalayeristhecapacityofthelinkoverthependingrequests(10Mbpsserverlinkand100Mbpscachelinks).

28

Playbacktimedistributionacrossvideostallsandqualities(Q1toQ5)

ü 4xfewervideoplaybackstalls

ü Higherplaybackquality

Goingbeyondstateoftheart

• Wedesignnovelcachingschemesthatareoptimizedtoaccountforthe:1. hierarchicalstructureofthenetwork

2. application-levelrequirements(videostreaming)3. mobilitybehavioroftheend-users4. economicaspects(storageownedbythird-partyentities)

29

Whymobilityisimportant?

• Ultra-densedeploymentoffemto-cells isconsideredtobeanintegralpartof5Gwirelessnetworks.– Densificationwillresultinfrequenthand-offs betweenfemto-cellsas

theusersmove(evenbywalking– withoutvehicles).– Hand-offsmayoccurbeforedatadownloadisfinished.

• Mobilitycandrasticallyaffecttheefficiencyofthecachingpolicies11.

30

11K.Poularakis,L.Tassiulas,“Code,CacheandDeliverontheMove:ANovelCachingParadigminHyper-DenseSmall-cellNetworks”,IEEE Trans.onMobileComputing,2017.

Femto-cells(frequencyf2)

Makemobilityafriendnotanenemy• Traditionalcachingmethods arebasedonpredictionsofthe

contentdemand.– Themorepopularafileis,themorecopiesofitarecachedinthenetwork.

• Mobility-awarecaching isanewideawhichexploitsboththedemandandmobility behavioroftheusers:– Popularfilepackets arespread acrossdifferentfemto-cellcaches thatarelikely

tobeencounteredoneaftertheother bytheusers.

Traditional Mobility-aware

31(Cachethemostpopularfiles– red&blue) (Cache filepacketsacrossusertrajectory)

ü Mobileusers mayencountermanyfemto-cellsandcandownloadmorefilesthanstaticusers.

ü 3timesmorefilesintheexample.

Modelingmobility-awarecachingproblem• Markovmobilitymodel: usersmovetodifferentlocationsfrom

slottoslot inapredictivemanner.pl qll’

• Timedeadlineofdslots forausertodownloadtherequestedfilebytheencounteredfemto-cellcaches;otherwiseredirectedtothemacro-cell.

-wesaythatausertakesawalkw (sequence ofdlocations)withprobability rw

• BandwidthconstraintBn :maxamountofdatadownloadedinaslotbyfemto-celln. 32

probability that a user is at location l

probability that a user moves from location l to l’

Variablesusedintheoptimization• WeuseNetworkCoding toformulatethecachingproblemina

moretractablemanner.xnf 𝛜 [0,1] (continuous insteadofintegervalues)

– Afilecanbedecoded iftheamountofencodeddataused ismoreorequaltothesizeoftheoriginal(un-coded) file(MaximumDistanceSeparablecode).

• Amobileuserrequestingfilefwilldownloadbytheencounteredfemto-celln:

𝐲𝐧𝐟𝟏 = min xd�, Bd fileportionduringthe1st contact

𝐲𝐧𝐟𝐤 = min xd� − ∑ 𝑦I�

S��CS�C , Bd portionduringthekth contact,

k=2,3,..,d.

portion of encoded data of file f cached at femto-cell n

33

Auxiliary

varia

bles

Optim

izatio

nvaria

bles

MaximumDistanceSeparable(MDS)Code

• Giventwopositiveintegerskandn>k:an(n,k)MDScode separatesafileintokpackets.Subsequently,theseareencodedintonpackets(ofthesamesize)suchthatanykoutofthesensufficetorecovertheoriginalfile.

• Example(4,2):– Any2outofthe4columnblockssuffice torecover theoriginalfile

data(A1,A2,B1,B2).

34

Problemformulation

• Objective:minimizeprobabilityofmacro-cellservice:

} 𝑟� } 𝜆��∀�\MP�∀�NM��

1{∑ O��� �C∀��� ¡� � �¢£¡¤¤�¥¤¤� }

• NP-Hardeventoapproximatewithinanyconstantfactor(reductionfromindependentsetproblem).

predicteddemand

35

predictedmobility

condition for service by macro-cell

(the total amount of data downloaded by femto-cell

caches is not enough)

Adistributedcachingalgorithm

• Weproposetominimizeanupperbound oftheobjectivefunction.– Thiscouldbeaneasier problem.– Thisway,weindirectlyminimizetheinitialobjective function.

• UseHoeffding’s inequalitytofindthisupperbound:– ConsideranergodicMarkovchainandat-step randomwalkwithtotal

weightY.Then,forany𝛅ϵ 0,1 ,itholds:

Pr 𝐘 ≤ 𝟏 − 𝛅 𝛍𝐭 ≤ c φ ¬exp{−­®¯S°F±

}

whereφ istheinitialdistribution,π isthestationarydistribution,μ istheexpectedweightofthewalk,T isthemixingtime,c isaconstant.

– Fort=dandδ = 1 − C³�,weget: Pr 𝐘 ≤ 𝟏 ≤ c φ ¬exp{−

¯]´ µ¶·�F

°F±}

36Initialobjectivefunction Upperbound

Adistributedcachingalgorithm(cont.)• Minimizeupperboundómaximizetheexpectedweight(μ) ofawalk:

• Separable functionacross femto-cells.

• Sub-problemforafemto-cell nó fractionalknapsackproblem:

Itemvalues =stationaryprobabilities:higherfractionsarecachedforfilesthatarepopular,andspreadinfemto-cellsthataremorelikelytobeencountered.

max𝜇 = } Pr[𝑦I�� ]∀¼½NMMZPMMI,�\MP�,ZTISNZS�

𝑦I��

stationaryprobability thatauserencountersfemto-cellnfor thekth timerequestingfilef

37

Knapsack size=cachesizeofn

𝑦ICC 𝑦I¾C

File1 FileF

1st slot

𝑦ICF

𝑦IC]

𝑦I¾F

𝑦I¾]

2nd slot

dth slot

Greedilyplacingitemfractionsisoptimal

.. .𝑦IFC

File2

𝑦IFF

𝑦IF].

..

..

....

Trace-drivenevaluation

Mobilitypatterns Requestpatterns• Wireless TopologyDiscoveryproject

(UniversityofCaliforniaSanDiego)• 275PDAusers,15WiFi APs/femto-cells• 11-weekperiod(year:2002)• Mobility patternsarerecordedevery20sec

• Amherstcampus,UniversityofMassachusetts (year:2008)

• Alluserrequestsarerecorded(10,000files)

Publiclyavailabledatasets 38

Trace-drivenevaluation(cont.)• Comparewithexisting(mobility-agnostic)schemes

– Max-Popularity: cachethemostpopularfileseverywhere.– Femtocaching: usersareassumedtobestaticfollowingtheinitial

distributioninthetrace.Iteratively,placesthefiletothecachethatimprovesthemosttheobjective,untilcachesarefull.

– Setup:3timeslots(1minute)deadline,8mbpsfemto-cellrate,10%ofentirefilelibrarycachesizes,40MBfilesizes.

Upto65%moremacro-celltrafficisoffloadedbyourscheme.Thegainsincreasewithcachesizesandfemto-celldensity. 39

Cachesize(%offilelibrarysize)

Probabilityofm

acro-cellservice

Numberoffemto-cells

Probabilityofm

acro-cellservice

}

Goingbeyondstateoftheart

• Wedesignnovelcachingschemesthatareoptimizedtoaccountforthe:1. hierarchicalstructureofthenetwork

2. application-levelrequirements(videostreaming)3. mobilitybehavioroftheend-users4. economicaspects(storageownedbythird-partyentities)

40

Harvestingresourcesfromresidentialusers

• Highavailabilityofwirelessbandwidth and storage inresidencestoday:– ManyresidentialusersdeploytheirownWiFi APstoserve theirownneeds.

– Theymayalsoownseveral storagedevices (externalharddisks,USBflash,etc).

• Idea:anoperatorcanharvest/lease theseresourcestoservemobileusers12.– Benefit:CAPEXcostforinfrastructure isreduced.

41

12K. Poularakis, G. Iosifidis, I. Pefkianakis, L. Tassiulas, Martin May, “Mobile Data Offloadingthrough Caching in Residential 802.11 Wireless Networks”, IEEE Transactions on Network andServiceManagement, 2016.

Whycanthisbeagoodidea?

• WiFi bandwidthisoftenunused formanyresidences:– weanalyzedadataset of167realresidentialusers,subscribersof

Portugaltelecom,fora4-monthsperiod.Thedatasetcontainsthenumberof(transmittedorreceived)bitsevery30-secondsforeachwireless deviceinsideeachresidence.

– wefoundtheprobabilitythatWiFiisusedbytheresidentialusertobeupto0.25(peakintheevening)=>only1outof4APsareused!

• ResidentialuserswillnotbeveryreluctanttoleasetheirWiFi bandwidth.

42

Techno-economicoptimizationframework

• Theoperatordecidestheincentivestoresidentialusern:𝐩𝐧𝐂 ≥ 0 and𝐩𝐧𝐁 ≥ 0

• Residentialusersmaximizetheirownutilities(U)andpayments.

𝑪𝒏(𝒑𝒏𝑪)Ã = 𝑎𝑟𝑔𝑚𝑎𝑥Ç∈ h,C 𝑈IÉ 1 − 𝑏 + 𝑝IÉ𝑏 (leased cacheportion)

𝑩𝒏(𝒑𝒏𝑩)Ã = 𝑎𝑟𝑔𝑚𝑎𝑥Ç∈ h,C 𝑈IÎ 1 − 𝑏 + 𝑝IÎ𝑏 (leased bandwidthportion)

• Theoperatordecidesthecachingandroutingpoliciesaswell:

– 𝐱𝐧𝐟 ∈ {0,1} and𝐲𝐧𝐟𝒌 ∈ [0,1]

perunitpriceforleasingcachespace

perunitpriceforleasingWiFi bandwidth

43

caching file f at residence n

portion of demand of mobile user k for file frouted to residence n

Jointincentive,cachingandroutingproblem

• DeterminepC,pB,x,y tominimizetheleasingcost+costforservingmobileusers:

∑ (𝐩𝐧𝐂𝑪𝒏(𝒑𝒏𝑪)Ã +𝐩𝐁𝚩𝐧(𝐩𝐧𝐁)Ã∀𝒓𝒆𝒔𝒊𝒅𝒆𝒏𝒄𝒆𝒏 ) + Q(λ,y)

• Q(λ,y)canbeanyconvex functionofthetrafficthatreachesthecellularBS.

• Tradeoff: themorefilesarecachedtoresidentialusers,themoredemandcanbeoffloadedtothem,andhencethelowertheservicingcostQ(.)becomes;howevertheleasingcostincreases.

• NP-Hardproblem.Aprimal-dualmethod isusedtoapproximatetheoptimalsolution.

44

Dataset-drivenevaluation(cont.)

• ComparePrimal-dualwithastateoftheart cachingalgorithmusingadatasetof167realresidentialusers:• “Femtocaching” algorithm:leasesthecompletecachespace.Then

iterativelyfillsincachesgreedily(eachtimestoresthefilethatreducesoperatorcostthemost).Eachrequestisroutedtothenearestcachehavingstoredtherequestedfile(andthenecessarybandwidthisleased).

19%lesscostthanFemtocaching Compensationupto9euros/monthperresidence

Impactonoperator Impactonresidentialusers

45

Cachinginothercontexts

• Capacityscaling: ifthenetworksupportscontentaccesswithZipfpopularitiesthentheuseofcachingateachnodemayimproveitsscalingbehaviorfromtheinversesquarelaw(Gupta-Kumar)upto1/N^(2/5).13S.Gitzenis, G.S.Paschos,L.Tassiulas,“AsymptoticLawsforJointContentReplication andDelivery inWirelessNetworks”,IEEETrans.onInformation Theory,vol.59,no.5,2013.

• Capacityofthebroadcasterasurechannelwithfeedback: itisachievableifthereceiversareabletostore allreceivedbroadcasttrafficbroadcastandadynamicback-pressurestylenetworkencoderisemployedinthetransmitterbasedonfeedbackofcachedtraffic.

14M.Gatzianas, L.Georgiadis, L.Tassiulas “MultiuserBroadcastErasureChannelwithFeedbackCapacityandAlgorithms”, IEEETrans.on InformationTheory, 2013.

46

Closingremarks

• Caching popularcontenthasagreatpotentialforreducingcostsandimprovingperformance inwirelessnetworks.

• Tobetterreapthebenefitsofcaching,weproposedalgorithmsthat:– exploitthehierarchicalnetworkstructure,– applyadvancedvideoencoding technologies,

– leverageusermobility predictability,– harvest residentialuserresources.

• Futurework:– extendouralgorithmsfortime-varyingand/or(partially)unknown

contentpopularitydistributions.47

Acknowledgements

48

GeorgeIosifidisAssistant ProfessorTrinityCollegeDublin,IE

Konstantinos PoularakisPost-doctoralResearcher, YaleUniversity

References1. N. Golrezaei, K. Shanmugam, A. Dimakis, A. Molisch and G. Caire, “FemtoCaching: Wireless Video Content

Delivery through Distributed Caching Helpers”, in Proc. IEEE Infocom, 2012.2. M. Dehghan, A. Seetharam, B. Jiang, T. He, T. Salonidis, J. Kurose, D. Towsley and R. Sitaraman, “On the

Complexity of Optimal Routing and Content Caching in Heterogeneous Networks”, in Proc. IEEE Infocom,2015.

3. M. Dehghan, et al., “Jointly Optimal Routing and Caching for Arbitrary Network Topologies”, ICN, 2017..4. M.A. Maddah-Ali, et al., “Fundamental Limits of Caching”, IEEE Trans. Information Theory’14.5. Liu et al., “Mixed-timescale precoding and cache control in cached MIMO interference network”, IEEE Trans.

on Signal Processing’13.6. A. Khreishah, et al., “Joint Caching, Routing, and Channel Assignment for Collaborative Small-Cell Cellular

Networks”, IEEE JSAC’16.7. K. Poularakis, L. Tassiulas, “On the Complexity of Optimal Content Placement in Hierarchical Caching

Networks”, IEEE Transactions on Communications, 2016.8. S. Borst, V. Gupta, A. Walid, “Distributed Caching Algorithms for Content Distribution Networks”, in Proc. IEEE

Infocom, 2010.9. K. Poularakis, G. Iosifdis, A. Argyriou, I. Koutsopoulos, L. Tassiulas, “Caching and Operator Cooperation Policies

for Layered Video Content Delivery”, in Proc. IEEE Infocom, 2016.10. K. Poularakis, G. Iosifidis, A. Argyriou, L. Tassiulas, “Video Delivery over Heterogeneous Cellular Networks:

OptimizingCost and Performance”, IEEE Infocom2014.

49

References(cont.)11. K. Poularakis, L. Tassiulas, “Code, Cache and Deliver on the Move: A Novel Caching Paradigm in Hyper-Dense

Small-cell Networks”, IEEE Transactions on Mobile Computing, 2017.12. K. Poularakis, G. Iosifidis, I. Pefkianakis, L. Tassiulas, Martin May, “Mobile Data Offloading through Caching in

Residential 802.11Wireless Networks”, IEEE Transactions on Network and Service Management, 2016.13. S. Gitzenis, G. S. Paschos, L. Tassiulas, “Asymptotic Laws for Joint Content Replication and Delivery in

WirelessNetworks”, IEEE Trans. on Information Theory, vol. 59, no. 5, 2013.14. M. Gatzianas, L. Georgiadis, L. Tassiulas “Multiuser Broadcast Erasure Channel with Feedback Capacity and

Algorithms”, IEEE Trans. on InformationTheory, 2013.

50