Applying Machine Learning to
Intent-based Networking and
NFV Scaling Strategies
Diego R. Lopez, Telefonica
Glen Lockhart, Openet

The CogNet Project
• CogNet is building a network management solution based on machine-learning
• Relying on SDN and the NFV architecture framework
• Focused on 5G use cases
• Committed to produce and demonstrate a general framework for smart
management
• Able to support different application scenarios
• Monitoring and SLA enforcement
• Traffic identification and security
• Resource management
• Reliability enhancement
• E2E service management
• <Your favourite scenario here>
• And for sure
• Relying on open source
• Willing to contribute back to the community
• Seeking for opportunities of further engagement

The CogNet Architecture
•A machine learning module
(the CSE) connected to two
streams
• Data, inbound
• Policies, outbound
•Supporting an integrated
SDN/NFV model
• Following the SDN/NFV
integration models considered
by ETSI and ONF

The CogNet Implementation
• CogNet is committed to make its
first demonstration in September
• The team is working in a first
implementation release
• Based on open source
• Monasca, Kafka and InfluxDB for
data collection
• PANDAS, Spark MLlib, Hive and
Hadoop for the machine learning
process
• Drools as policy engine
• ODL as tenant controller
• OSM for orchestration
• OPNFV Brahmaputra as NFVI
• And open data models
• JSON for (most) interfaces
• SUPA for policy definitions

A Dialectical Approach to Intent
• Thesis: Translate intent into action
• Understanding intent statements
• Mapping onto technologies
• Considering current state
• A controlled but flexible language
• Antithesis: Address environment constraints
• Provided by network management
• Evolved from general principles defined by humans
• Equivalent but not necessarily equal language space
• The archetypal SLA enforcement
• Synthesis: Conflict resolution
• Among action requests
• And with management constraints
• Hast to be initiated before translation
• Understandable by humans for auditing purposes

Intent in the CogNet Architecture
• Intent can be expressed as
different kinds of user input
• At least two essential
computational intelligence
technologies
• Linguistic interfaces
• Machine learning mechanisms
• Identify application points and
support composition
• Horizontal: federation
• Vertical: recursion

Applying Intent Dialectics
• The generic 5G slicing
• Slices defined by intent expressions
• Matched against operator infrastructure intent declaration
• Compatible with other slice definitions
• Supporting recursion, applying slice user intent expressions
• Supporting federation, across network and slice providers
• Personalized services
• Composing intent expressions
• User, service provider, slice operator, infrastructure…
• Support for personal auditing
• Privacy preservation according to personal requests
• Reduced service TTM
• Applying business rules as intent
• Domain-specific languages to capture customer requirements
• Bridging the gap between support systems and network management
• The BOSS concept
• Computational intelligence is here to help
• And the CogNet project is paving its way

Applying Machine Learning to
Intent-based Networking and
NFV Scaling Strategies
Diego R. Lopez, Telefonica
Glen Lockhart, Openet

Introduction
•As more network functions are
virtualized and deployed in the cloud we
want to be able to scale VNF clusters
automatically to improve:
•System availability
•Operating costs

Cost Savings
0
5
10
15
20
25
30
35
JAN-15
FEB-15
MAR-15
APR-15
MAY-15
JUN-15
JUL-15
AUG-15
SEP-15
OCT-15
NOV-15
DEC-15
JAN-16
FEB-16
MAR-16
APR-16
MAY-16
JUN-16
JUL-16
AUG-16
SEP-16
OCT-16
NOV-16
DEC-16
Planned vs Actual

Hadoop Cluster
YARN
High Level Architecture
Apache
Spark
VIM
(Open Stack)
Metrics
Source
(Ceilometer)
VNF-M
MANO Layer
Scaling Events
Collection
Agent
Apache
Kafka

Hybrid Model
Reactive
Scaling
Predictive
Scaling
Voting
Engine
RPS Metrics VNF
Orchestrator

Algorithm Selection
• We used Random Forest algorithm provided in Apache
Spark MLlib to train our model

Selecting Training Data
• Biased towards more recent data to take trending performance into
account.
• Take all data points from last week
• Take from the 5th minute of the hour from the week before that
• Take from the 15th minute of the hour from the week before that
• Take from the 30th minute of the hour from the week before that

Training Data - Code
JavaRDD<List<Object>> metricsRDD= inputFileRDD.map(new
InputParser(fields));
….
String query = "SELECT datetime, rps FROM dataTable WHERE
(minOfDay%5) = 0";
DataFrame week1= /* do SparkSQL query against metricsRDD */
….
DataFrame biasedData = week0.unionAll(week1)
.unionAll(week2)
.unionAll(week3)
.orderBy("datetime");

Training the model - Code
JavaRDD<LabeledPoint> biasedData= getTrainingData();
RandomForestModel model =
RandomForest.trainRegressor(inputLabeledPointRDD,
new HashMap<>(),
100,
"auto",
"variance",
5,
32,
12345);

Predicting against the Model - Code
public Tuple2<Double, Double> predictModel(RandomForestModel
model, double[] data) {
Vector v = Vectors.dense(Arrays.copyOfRange(data, 1, data.length));
Tuple2<Double, Double> predictionAndLabel =
new Tuple2<>(data[0], model.predict(v));
return predictionAndLabel;
}

Purely Reactive versus Hybrid Model

Summary
• Predictive auto-scaling delivers clear advantages over a
reactive model, providing VNF’s with the resources they
need, before they need them, ensuring that your service
can provide timely responses 24/7