Page 1
VIA:ImprovingInternetTelephonyQualityUsingPredictiveRelaySelection
Junchen Jiang,Rajdeep Das,GaneshAnanthanarayanan,PhilipA.Chou,Venkata N.Padmanabhan,VyasSekar,Esbjorn Dominique,Marcin
Goliszewski,Dalibor Kukoleca,Renat Vafin,HuiZhang
1
Page 2
Keytakeawaysinoneminute
Onefifth ofcallsusepathswithpoornetworkperformance
“ManagedOverlay”couldalleviateoverhalfofcallsonthesepaths
VIA:Data-drivenrelayingcanrealizemostManagedOverlay’spotential
Westudied430millionSkypecallsandfound…
2
Page 3
0
20
40
60
80
2004200520062007200820092010201120122013
Million
Peak#ofusersconcurrentonlineonSkype
Internettelephonyiseverywhere!
MoreappsfocusingonInternettelephony
https://blogs.skype.com/2013/04/03/thanks-for-making-skype-a-part-of-your-daily-lives-2-billion-minutes-a-day/
Rapidgrowthoverthelastdecade
https://www.statista.com/chart/1417/skype-usage/3
Page 4
Callqualityissensitivetobadnetworkperformance
4
0
0.2
0.4
0.6
0.8
1
0 200 400 600Normalize
dPo
orCallRate
RTT(ms)
0
0.2
0.4
0.6
0.8
1
0 1 2 3 4Lossrate(%)
0
0.2
0.4
0.6
0.8
1
0 5 10 15 20 25Jitter(ms)
Thresholdsofpoornetworkperformance
Datasets:Avg RTT,lossrate,jitterforeachcallof430millionSkypecallsSmallfractionhasuser-providedqualityscores
320ms 1.2% 12ms
Page 5
Manycallshavepoornetworkperformance
5
17%17%12%Onefifth ofcallshavepoornetworkperformance
Ourgoal:AlleviatepoornetworkperformanceforSkype
Page 6
Outline
• Problem: NetworkperformanceofSkypeisbad
è Opportunity: ManagedOverlayhashugepotential
• Solution: VIAforoptimalrelayselection
• Evaluation: VIAisclose-to-optimal
6
Page 7
RevisitingOverlayNetworksbyManagedOverlay
ManagedOverlayhasnewbenefits• World-wide distributedDCs asrelays• Wellconnected• Deployedbymanyproviders• Single administrativeentity
HowmuchcanInternettelephonybenefitfromit?
7
Page 8
Selectingthebestrelayoption
Keyistoselectthebestrelayoption(direct,one-relay,ormulti-relay)
Q1:Doespickingbestrelayoptionhavesignificantimpact?Q2:Ifso,howtopickthebestrelayoption?
8
Directpath
One-relayhop
Multi-relayhops
Page 9
ManagedOverlayhashugepotentialbenefit
Substantialfractionofbad-performancecallscould bealleviated
Consideranoraclethatpicksthebestrelayoptionforeachsrc-dst ASpairin24hours
9
0102030405060
RTT Loss Jitter Atleastonebad
Poorperformance:RTT>320ms;Lossrate>1.2%;Jitter>12ms
%calls“Bad”è “Good”
Page 10
Outline
• Problem: NetworkperformanceofSkypeisbad
• Opportunity: ManagedOverlayhashugepotential
è Solution: VIAforoptimalrelayselection
• Evaluation: VIAisclose-to-optimal
10
Page 11
VIA:RealizingthebenefitofManagedOverlayusingCentralizedPredictiveControl
Qualityofexistingcalls
Relayselection
VIAControlAlgorithmPredictthebestrelayoptionbased
onothercalls’performance
11
Page 12
Strawman 1:Pureprediction-based
12
Uselong-termhistorytopredictperformance
Relayoptions
Anewcall
Problem:CallperformancehasgreatinherentvarianceE.g.,predicting“nextday”using“lastweek”leadstoover30%erroronlatency.
CallhistoryQualityprediction
Page 13
Strawman 2:Pureexploration-based
13
Inashorttimewindow,explorerelayoptions,thenexploitthebestone
Relayoptions
CallsofoneASpairperday
Problem:Calldistributionishighlyskewed“#ofactualcalls>>#ofrelayoptions”NOTtrueformostASpairs
Page 14
Keyidea:Guidedexploration
Prediction-based Exploration-basedXORStrawmenVIA
14
00.20.40.60.81
0 1 2 3 4 5 6 7 8 9Prob
abilityofthe
best
relaybe
inginto
pk
k
Topk canbemoreeasilypredictedandcanbemore efficientlyexplored!
VIA’sidea:GuidedExplorationRoughpredictioncanstillidentifytopk candidates,whichcanbeexploredefficiently.
Predictingtop1istoohard
Page 15
Step1:Prediction-basedpruning
Latency
Confidenceintervalofqualityprediction
Top-kcandidates
Focusonrelayoptionswhoseconfidenceintervalsarebetterthanthoseofothers.
15
Callhistory
Upperboundof“blue”and“yellow”isbetterthanthelowerboundof“green”and“red”
Page 16
Step2:Exploringtop-kcandidates
16
Multi-armedbanditprocessUpperConfidenceBounds(UCB1)
Reward
OurproblemlookslikeMABUCB1withdomain-specifictwists
Reward
Reward
Howtomaximizerewards?
Howtominimizelatency?
UCB1:AlwayspicktheonewiththehighestUCB
Page 17
Puttingthemtogether:GuidedExplorationinaction
Callhistory
Performancemeasurements
Tomography-basedcoverageexpansion
Predictiveselectionoftopk candidates
ModifiedUCB1onthetopk candidates
Toexploremorerelays Real-timerelayselection
Runspercall
Prediction-basedPruningUpdateseveryT hours
Top-k candidatesperASpair
17
Allrelayoptions
Page 18
Moreinourpaper
• Budgetedrelaying
• Networktomography
• Granularityofprediction
• Internationalvs.domesticcalls
18
Page 19
Outline
• Problem: NetworkperformanceofSkypeisbad
• Opportunity: ManagedOverlayhashugepotential
• Solution: VIAforoptimalrelayselection
è Evaluation: VIAisclose-to-optimal
19
Page 20
VIAachievesclose-to-optimalperformance
20
Poorperformance:RTT>320ms;Lossrate>1.2%;Jitter>12ms
%calls“Bad”è “Good”
0
10
20
30
40
50
60
RTT Loss Jitter Atleastonebad
Optimal VIAPrediction-based Exploration-based
Page 21
BenefitvariesacrossASes
0
20
40
60
80
100
SourceAS1 SourceAS2 SourceAS3
Oracle VIA
Substantialimprovement
Limitedroomforimprovement
Roomforfurtherimprovement
21
%calls“Bad”è “Good”
Page 22
Conclusion
• Internettelephony isheretostay• Problem:Onefifthofcallshavebadnetworkperformance.
• Opportunity:ManagedOverlaycould significantlyreducebad-performancecalls.
• Challenges:OptimalrelayselectionPurepredictionandpureexplorationwon’twork!
• Solution:VIAcanrealizemostofthebenefitofManagedOverlay.Keyidea:Guidedexploration:
Predictivepruningofrelayoptions+Explorationovertopkcandidates22