26
Data Driven Networking Sachin Katti Platform Lab Review, Feb 2017
Data Driven Networking
Sachin KattiPlatform Lab Review, Feb 2017
Canwelearnthecontrolplaneofthenetwork?
Control&OrchestrationPlaneautomaticallysynthesizedbylearning fromtelemetrydata
IndustryStandardNorthboundApplication Interfaces
IndustryStandardSouthbound InfrastructureInterfaces
Real-timeNetworkTelemetry
OperatorPolicies,Intent&SLA
EdgeCloud
RAN
InfrastructureControlPoints
VIRTUALREALITY4KVIDEODRONEATC
OperatorPolicies,Intent&SLA
OperatorPolicies&Intents
ApplicationswithAccesstoNetworkControl
Result:Dynamicperuser/flowcontrolimplementedtomeetthepolicyanddeliverontheSLA
Youspecifywhatyouwant,andthenetworkfiguresouthowtodeliverit!
DataDrivenControlLoop
ServiceKPI
Prediction
WillCurrentControlsMeetKPI?
RecommendCurrentControl
Settings
Y
N
What-IfAdjustControl
Knobs
WillAdjustedControlsMeetKPI?
N
Y RecommendAdjustedControl
Settings
ApplyControls
MeetsOperatorPolicies?
N
Y
OperatorPoliciesIntent&SLAs
Real-timeData
NetworkTelemetry
Twolearningapproaches:
ClassicalMachineLearning DeepReinforcementLearning
NeuralNetworkReinforcementLearning
(Prediction &What-if)
RewardFunction
Requiresextensivedomainknowledge Largelyblind, trulyapproximatesintentbasedsystemdesign
ThisTalk:ApplyingthisFrameworktoCommercialMobileNetworks
ServiceKPI
Prediction
WillCurrentControlsMeetKPI?
RecommendCurrentControl
Settings
Y
N
What-IfAdjustControl
Knobs
WillAdjustedControlsMeetKPI?
N
Y RecommendAdjustedControl
Settings
ApplyControls
MeetsOperatorPolicies?
N
Y
OperatorPoliciesIntent&SLAs
Real-timeData
NetworkTelemetry
Controlplaneapplication:LoadBalancinginmobilenetworks
Operatorpolicy:Optimizeuserthroughputforthe20%worstaffectedusers
Data:PerflowRecordsfromthemobilenetwork
Mobile traffic demand is skyrocketing
2000 2005 2010 2015 2020 2025
3.7 EB/month
30.6 EB/month
0.4 EB/month
< 1 PB/month
300 EB/mon
th?
Demanding content
New, smarter devicesPhones Smartphones Tablets Wearables
Audio, Text Web Video Virtual Reality
< 10 GB/month
Courtesy: Cisco VNI, 2015-2020
How will mobile networks meet this surging traffic demand?
Densificationwillbringthenextwaveofcapacityscalinginmobileaccessnetworks
Spectral efficiency
Densification
Spectrum
Densification has challenges
Fordensificationtoscale,networksneedtoproactivelymanageload,interferenceandmobilitytooptimizeuserexperience
Load (demand) and other factors (e.g interference, mobility) become highly dynamic
à need to be proactively managed
E.g: Cells can offload users to less-congested neighbors to balance load,
need to know expected user experience
Why is this challenging?
Existing approachesAggregate cell-level counters
blind to dynamism across time and users;reactive
Need to predict user performance every ~1sfor the best load distribution
ChallengeUser performance is hard to predictComplex function of many, potentially unknown and/or dynamic, network variables
Needtopredictuserexperiencesecondsinadvance,hardinpracticalnetworkswithmanyunknown/dynamicvariables
To packet core
13:00 14:00 15:00 16:00
05
1015
TIME
LOA
D (i
n M
B p
er s
econ
d)
13:00 14:00 15:00 16:00
05
1015
TIME
LOA
D (i
n M
B p
er s
econ
d)
0 10 20 30 40
01
23
4
USER #
LOAD
(in
MB)
Dedicatedfiber/micro
wave
PublicInternet
What do we need today to enable wide-scale densification?
Can we use a data driven approach to tackle this challenge?
without making invasive changes to existing network infrastructure without deploying expensive network infrastructurewithout modifying user devices/cellular standards
Predict user performance every secondto manage load dynamically
Coordinate across cells in realtimeto balance load
ForeCData-driven control
in dense mobile networks
How to use realtime analytics & learning for network control?
Network effects
Network state
Network actions
change causes
Networksrequiretheabilitytomonitorandforecastnetworkeffectsperuser/cell,persecond,potentiallybasedonnetworkactions
MONITORRight now, what is the
avg throughput of user/cell?resource utilization of user/cell?contention faced by user/cell?
FORECASTIn the next 1s, what will be the:
avg throughput of user/cell?resource utilization of user/cell?contention faced by user/cell?
FORECAST IMPACTIn the next 1s, what if:
handover users?admit/reject new users?
increase/decrease tx power?
throughputresource utilization
contention
handover useradmit/reject userinc/dec tx power
cell: bandwidth, #users, demand etc.user: serving cell, link quality etc.
ForeC: Analytics on ‘after-the-fact’ logs to monitor & forecast network effects
ForeC
user session logs
queries responses
Network control
Cells already expose usage stats per user session (~10s)
One report per event of the session, logged after the
eventsetup report, release report, traffic report,
radio measurements report, mobility report etc.
ForeC has been tested in diverse deployments
Football Stadium Metro Area
How does ForeC work?The internals of ForeC
What can ForeC do?and how well?
How can a network use ForeC?Load management in a stadium
How does ForeC work?The internals of ForeC
What can ForeC do?and how well?
How can a network use ForeC?Load management in Levi’s Stadium
futurepast now
EFFECTEFFECT
ACTION
How does ForeC forecast network effects?
STATE 2STATE 2STATE 1STATE 1
forecast causes predict
effect
ForeC’s approach:Forecastthecauses,predicttheeffectDecouplingsimplifiesdesignandimplementation,alsoworksquitewell!
What will be throughput of user A if handed over from cell 1 to 2
over the next 10 seconds?
forecast effect
decompose
The internals of ForeC
ANALYTICS LAYER
MODELING ENGINE
DATA LAYER: Data sanitization and organization
QUERY LAYER: Query parsing and response
FORECAST ENGINE
MONITORING ENGINE
PREDICTION ENGINE
Future effects
Future states
Currentstates
Currenteffects
Currentstates
OFFLINEPATH
ONLINEPATH
Current data
Past and current data
user session logs
queries responses
Proposed state change
How do we decompose an effect into its constituent causes?
ANALYTICS LAYER
MODELING ENGINE
DATA LAYER: Data sanitization and organization
QUERY LAYER: Query parsing and response
FORECAST ENGINE
MONITORING ENGINE
PREDICTION ENGINE
Future effect
s
Future states
Currentstates
Current
effectsCurrent
states
OFFLINEPATH
ONLINEPATH
Current data
Past and current data
user session logs
queries responses
Proposed state
change
Usedomainknowledgetoformulate/hypothesizecauses,uselearningtoolstotestsignificanceanddiscoverrelationships
Spectral efficiencybits per resource element
Throughputbits per time
Cellbandwidth
Cell contention
User&totaldemand
Resource allocation rateresource elements per time
?
?Σ Active
users
Σ Pktrate, size
Rank
MCSCQI
X
0 20 40 60 80 100
-6-4
-20
24
6
USER SESSION #0 20 40 60 80 100
-6-4
-20
24
6
USER SESSION #0 20 40 60 80 100
-6-4
-20
24
6
USER SESSION #
58% error 40% error 12% error
00:00 01:00 02:00 03:00 04:00 05:00
1520
2530
3540
00:00 01:00 02:00 03:00 04:00 05:00
1520
2530
3540
How do we forecast the dynamic causes?
ANALYTICS LAYER
MODELING ENGINE
DATA LAYER: Data sanitization and organization
QUERY LAYER: Query parsing and response
FORECAST ENGINE
MONITORING ENGINE
PREDICTION ENGINE
Future effect
s
Future states
Currentstates
Current
effectsCurrent
states
OFFLINEPATH
ONLINEPATH
Current data
Past and current data
user session logs
queries responses
Proposed state
change
Testforseasonalityandtrend,useappropriateversionofARIMAtime-seriesforecasting
Number of users on a cell
-20
2060
data
-20
2060
seasonal
-20
2060
trend
-20
2060
0 20 40 60 80 100
remainder
time
ActualForecast
4% error6% error with naïve forecasting
How does ForeC work?The internals of ForeC
What can ForeC do?and how well?
How can a network use ForeC?Load management in Levi’s Stadium
ForeC can forecast network effects per user, per second, aggregates
Primitive Examplequery Medianerror
BasicprimitivesPREDICTeffectperuserbasedon(changesin)networkstate
ThroughputofuserAif itwereoncell2insteadof1? 8.5%
FORECAST effectperuserpersecond Throughputof userAoverthenext1s? 13.0%
Extensionsto aggregatesoverspaceandtime
FORECASTeffect per cell,sector,network… Averagethroughputofcell1inthenext1s? 4.0%
FORECASTeffectper10sof seconds,minutes… Averagethroughputofcell1inthenext5s? 1.7%
Networkeffects:throughput, resourceutilization, contention…&anyfunction ofthemNetworkactions:handoveruser,admitnewuser, increase/decreasetransmitpower
How does ForeC work?The internals of ForeC
What can ForeC do?and how well?
How can a network use ForeC?Load management in a stadium
ForeC letsoperatormonitor,forecastandoptimizenetworkKPIs
[SWITCHTOTABLEAU]MONITOR– FORECAST– OPTIMIZE
Dash1:OperatorscanmonitornetworkKPIsatafineresolutioninspaceandtime(show“ACTUAL”KPIforBand700,1900,2100)[canalsodefinenewKPIstomonitor]
Dash2:OperatorscanforecasttheseKPIstomanageloadproactively(show“ACTUAL”KPIand“FORECAST”KPIforacell)
Dash3:OperatorscanchoosethebestloadmanagementtooptimizetheseKPIs(show“ACTUAL”and“OPTIMIZED”KPI)
Canwedothisunsupervised?ClassicalMachineLearning DeepReinforcementLearning
NeuralNetworkReinforcementLearning
(Prediction &What-if)
RewardFunction
• KeyChallenge:DeepRLisverydatahungry,youcannotrunbillionsofA/Btestsonalivenetwork
• Whatweneed:Network/SystemSimulatorswetrust!• Veryhardgiventhecomplexinteractionsinthesesystems
Summary&Takeaways
• Goal:Buildingaprogrammableconnectivitylayerforfutureapplications– Applicationsprogrammaticallyspecifytheirconnectivityrequirements
– Thenetworkplatformleveragesrealtimestreamingtelemetryandlearningtoautomaticallydeliverthatconnectivity
• Youspecifywhatyouwant,andthenetworkfiguresouthowtodeliverit!
