►
From YouTube: Using Kubernetes to Make Cellular Data Plans Cheaper... Chandrasekhar Reddy Dodda & Amar Padmanabhan
Description
Don’t miss out! Join us at our upcoming events: EnvoyCon Virtual on October 15 and KubeCon + CloudNativeCon North America 2020 Virtual from November 17-20. Learn more at https://kubecon.io. The conferences feature presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects.
Using Kubernetes to Make Cellular Data Plans Cheaper for 50 Million Users - Chandrasekhar Reddy Dodda, Mirantis & Amar Padmanabhan, Facebook Connectivity
A year ago, the CNCF Telecom User Group envisioned NFVi architecture where K8s replaces OpenStack as the virtual infrastructure manager. Today the first real-world use case of this has emerged, at an APAC telco with 50 million mobile subscribers. The speaker will discuss the use case and technical implementation, where pure K8s based NFV infrastructure is used in production to run an open source evolved packet core. The system is designed to offload cellular data to WiFi and uses K8s to orchestrate all components of the packet core, including containerd pods and Virtlet pods running VMs. The talk will explain how Genie provides connectivity between various pods with multiple, different network interfaces per pod. A demo will show how K8s powered NFVI makes it possible to seamlessly authenticate and roam between WiFi and LTE access points with no handover lag.
https://sched.co/ZepN
A
A
So
today,
we're
going
to
cover
a
do.
A
deep
dive
on
on
a
project
called
magma
and
then
chandra,
who
I'm
co-presenting
with,
will
also
do
a
demo
on
a
wi-fi
use
case,
which
is
one
of
the
manifestations
of
the
converge
core
project
called
magma
little
bit
about
myself,
I'm
amar
padmanabhan,
I'm
an
engineer
on
the
magma
project.
I
work
on
in
the
facebook
connectivity
team
and
then
with
me.
Co-Presenting
is
chandra,
he's
a
senior
devops
engineer
and
we
closely
work
together
on
a
bunch
of
our
deployments
and
chandra
is
with
us.
A
So
at
a
high
level.
What
does
it
mean
to
offer
connectivity
to
the
next?
50
million
models
have
actually
saturated
the
way
that
connectivity
is
being
offered,
and
so
the
problem
of
connecting
the
next
50
million
onto
a
faster
internet
is
a
problem
of
heterogeneity.
This
is
heterogeneity
in
access.
This
means,
like
people
are
getting
on
the
internet
over
wi-fi
getting
on
to
the
internet
or
lte,
and
now
you
know,
are
getting
onto
the
internet
over
5g
as
well.
A
A
What
this
meant
was
that
you
know
the
network
had
these
high,
like
you
know,
multi
100,
000
appliances
that
were
dedicated
and
offering
you
know
specific
services
that
were
in
the
network,
and
these
were
choke
point
devices.
So
any
traffic
that
you
know
a
user
who's
trying
to
access
facebook
is
is,
is
generating
is
having
to
get
routed
through
these
sort
of
these
choke
point
devices.
A
So
this
naturally
forces
the
topology
the
equivalent
is
like
you
know,
if
you're
in
san
francisco
and
you
have
to
direct
all
traffic,
like
you
know,
car
traffic
through
a
you
know,
traffic
light
right.
So
this
is
suboptimal
from
a
topological
standpoint.
The
other
thing
is
given
that
the
data
demands
are
increasing.
The
the
the
nature
of
traffic
that
these
middle
box
devices
need
to
support
has
to
become
much
faster
and
they
need
to
do
so,
offering
rich
policies.
So
this
naturally
resulted
in.
A
You
know
these
artists
becoming
really
expensive
fast
forward
a
few
years,
and
this
picture
is
from
our
data
center.
A
couple
of
years
ago,
we
sort
of
moved
into
a
sort
of
a
decomposed
model
of
the
network,
where
there
is
a
fabric
which
is
only
responsible
for
moving
packets
really
fast,
and
then
there
is
an
edge
which
is
typically
x86,
based
where
all
your
services,
like
your
load,
balancers
and
firewalls
sort
of
got
implemented.
A
So
what
is
an
edge?
Server
is
good
for
it's
fundamentally
distributed,
so
it
makes
a
good
candidate
for
x86,
and
this
is
where
a
lot
of
the
containerization
has
happened
in
terms
of
networking
workloads
software.
Only
policy
enforcement
points
allow
for
rapid
iteration,
and
that's
why
all
the
web
scale
companies
are
sort
of
you
know
fully
on
board
with
this
model.
A
So
if
you
draw
the
parallel
in
today's
lte
or
gsm
architecture,
there
are
these
choke
point
devices.
Some
of
you
are
not
familiar
with
these
acronyms,
but
you
know
they're,
typically
like
packet
gateways
and
mmes.
These
are
all
like
large
boxes
that
are
pretty
centralized,
so
they
suffer
from
the
same
exact
problem
that
the
data
center
had.
So
our
argument,
or
our
thesis
here,
is
that
you
know
given
the
learnings
that
we've
had
from
the
data
center
and
the
industry
as
a
whole.
A
On
the
networking
side,
we
think
that
modularizing,
the
cellular
network
is
actually
the
the
way
forward
for
enabling
the
heterogeneity
that
we
need.
So
the
first
and
key
point
is
distribute
policy
enforcement
point.
This
lets
you
build
the
ideal
network
topology
that
is
required
for
your
network
say
you
have
like
you
know
if
you
have,
if
you're
deploying
in
like
rural
peru
or
something
you
may
need
a
different
network.
Topology
from
that
of,
like
a
very
dense,
like
stadium
and
so
being
able
to
adapt
for
either
of
them
is
pretty
critical.
A
Moving
policy
enforcement
to
software
has
a
strategic
advantage,
and
this
sort
of
can
spur
innovation
and
keeping
the
core
network
simple
and
just
purely
about
moving
ip
packets
makes
it
much
cheaper
to
sort
of
achieve.
So
you
know
bringing
down
the
cost
enabling
the
right,
topology
and
innovation
is
sort
of
the
central
thesis
of
what
we're
proposing
now.
Given
this
framing
focusing
on
operationalizing
this
distributed
network
is
the
key
key
sort
of
computer
science
problem
ahead
of
us,
and
this
is
where
you
know.
A
A
So,
finally,
what
the
problem
has
now
translated
to
is
the
high-level
problem
that
we're
trying
to
build
with
the
orchestrator
is
managing
a
fleet
of
heterogeneous
edge
deployed
service
meshes
for
core
convergence.
What
does
this
actually
mean
like
an
orchestrator
is
a
a
sort
of
an
abstraction
that
tries
to
declaratively
configure
an
edge
fleet
from
a
centralized
management.
What
was
historically
like
a
single
device
like
a
network
device
like
a
packet
gateway,
is
now
like
50
or
100
packet
gateways
that
are
distributed
across
like
geographically
disparate
locations
and
so
being
able
to
orchestrate.
A
All
of
these
things
from
a
central
location
is
a
key
concept
that
we
think
is
going
to
keep
the
opex
still
manageable,
regardless
of
the
size
of
the
deployment.
The
second
principle
that
we
really
believe
in
and
a
lot
of
it
came
from
even
you
know.
The
way
that
kubernetes
itself
is
designed
is
that
you
sort
of
build
on
level
triggers.
So
you
always
push
state
and
not
procedures.
A
So
the
orchestrator
has
two
sort
of
large
components.
There
is
an
edge
control
plane.
This
one
is
one
instance
per
managed
device,
so
this
is
the
most
shrunk
form
factor
where
we
run
the
edge
control
plane
on
a
single
node,
which
is
like
a
2u
box,
and
it
applies.
Configuration
updates
based
on
the
local
mesh
and
life
cycle
manages
the
services
it
ships,
local
state
and
operational
history
onto
the
central
controller,
and
so
the
central
controller
has
sort
of
global
awareness
of
what
is
happening
across
a
lot
of
the
edges.
A
What
we've
sort
of
embarked
on
from
as
we're
moving
from
the
ultra
rural
to
like
the
peri-urban
sort
of
scenarios,
is
where
we
start
running
on
edge
clusters,
and
this
is
where
we
think
you
know,
there's
a
lot
of
synergy
between
us.
Moving
into
a
like
a
k
test
based
model
where
you
know,
there's
kubernetes
operators
where
we
leverage
and
still
go
with
our
declarative
approach
of
pushing
state
and
having
the
edge
kubernetes
cluster
sort
of
orchestrate
the
life
cycle
and
the
management
of
these
edge
services.
A
The
other
thing
worth
noting
is
that
it
comes
with
a
bunch
of
batteries
included,
so
there
are
multitenant
wrappers
around
graphana,
and
then
we
use
prometheus
as
our
time
series.
We
also
use
elastic
for
for
log
export
and
then
there's
a
rest,
a
basic
rest
authorization
framework
that
allows
you
to
access
apis
as
well
as
a
as
a
ui
for
interacting
with
the
system
as
a
whole.
A
A
B
B
So
I
want
to
talk
overview
of
mcp,
which
we
call
that
as
miranda's
cloud
platform,
it
is
an
enterprise
grade
open
cloud
software,
so
which
consists
of
three
components
on
the
right
side.
You
can
see
it's
a
stack
light,
which
is
our
element
tool
on
the
left
side.
It's
a
drivetrain
which
is
our
lcm,
which
is
used
for
your
day
zero
day,
one
day,
two
operations
and
the
middle
part.
It
says
infrastructure
stack
which
comes
with
a
lot
of
given
stacks
like
openstack,
kubernetes,
safe.
B
B
B
B
B
So
here
is
the
worklet
actress
or
the
flow,
how
it
is
how
it
is
connected
and
talks
between
the
components
and
when
you
differentiate
the
wattage
with
existing
communities
manifest
the
both
manifests.
The
same
kubernetes
supports
all
standard
kubernetes
objects
such
as
deployment
replica
set,
demonstrate
and
so
on.
B
B
This
is
also
on
a
hl
like
active
standby,
I'll,
be
deploying
wi-fi
access
gateway
as
a
vm,
using
what
late
on
kubernetes
next,
I
will
be
connecting
to
the
hotspot
and
access
the
internet
from
mobile
so
and
then,
which
is
not
covered
here.
So
we
deployed
access
point
for
this
demo,
but
I
am
not
going
to
show
you
how
we
deploy
it
and
how
how
we
are
going
to
configure
it,
and
also
we
are
not
going
to
see
how
the
user
provisioning
has
happening
at
hss,
pcrf
and
ocs
side.
B
So
when
we
go
into
the
demo
part,
I
just
want
to
talk
about
what
is
magma
orchestrator
very
high
level.
So
it
is
a
cloud
service
that
provides
a
simple
and
consistent
way
to
configure
and
maintain
the
wireless
network
securely
and
when
it
comes
to
the
federation
gateway,
it
acts
as
a
proxy
between
the
magma
and
your
existing
manual
core
network
and
comes
to
the
access
gateway.
B
So
next
I'll
be
moving
to
the
very
very
high
level
how
we
are
deploying
all
the
components
here
on
mcph.
So
initially
this
will
be
the
flow.
I
have
a
three
node
of
bare
metals
right
now,
so
initially
I'll
be
deploying
my
sql
ports
on
three
nodes
on
a
head
shape.
Next
I'll
be
deploying
a
magma
orchestrator
with
nms,
which
we
call
as
a
network
monitoring
system
which
will
be
the
ui
to
access
your
orchestrator
magma
orchestrator.
So
once
that
is
done,
I'll
be
deploying
federation
gateway
on
my
second
node.
B
With
help
of
what
light-
and
I
do
have
a
cni
plugins
here
and
I
have
two
plugins
attached
to
the
vm,
one
is
connected
to
your
existing
manual
core
and
the
third
node,
which
I
am
going
to
deploy
carrier
wi-fi
gateway.
Here
we
have
three
cni
plugins
one
plugin
is
connected.
One
interface
is
connected
to
the
wi-fi
access
point.
Another
one
is
connected
to
the
internet,
gateway
yeah,
so
I'll
be
playing
the
video
now
so,
which
will
see
the
all
this
stuff.
B
B
So
once
that
is
up,
there
will
be
a
bootstrapping
process
to
create
a
bunch
of
users.
Here
the
ports
are
wrapped
and
now
the
bootstrapping
process
is
happening
where
it
creates
bunch
of
users
which
will
be
used
to
bootstrap
the
magma
diplomat.
So
once
that
is
done,
we
are
going
to
deploy
here.
Orchestrator,
which
is
again
will
be
on
like
we
have
two
replica
sets
for
each
service,
so
once
the
orchestrate
is
up,
it
will
take
a
few
minutes
to
bootstrap.
B
B
B
B
So
we
are
in
the
final
stage
of
bootstrapping
process.
So
now
the
automation
script
will
do
the
user
creation
yeah.
We
are
done
with
the
bootstrapping.
So
now,
let's
get
into
the
nms
the
dashboard
where
we
will
be
creating
the
networks
for
isolating
two
gateways
different
gateways,
and
then
we
will
register
the
gateways
which
you
just
created
and
then
we
will
proceed
with
the
final
part
of
the
demo.
B
B
C
B
B
B
So
I
will
name
it
as
effect
01
for
the
demo
purpose
and
the
same
way
I
already
got
the
hardware
ua
id
from
my
from
the
vm,
which
I've
just
spent
up
previously
the
same
challenging
key.
I
got
already
from
the
vm,
so
I
will
use
the
upgrade
tray
as
the
default
here
also
yeah.
Now
both
gateways
are
created,
so
it
will
take
out
around
30
seconds
to
get
the
status
yeah.
B
So
now
I
connect
it
to
my
mobile,
so
you
can
see
there
is
no
wi-fi
right
now
at
latched,
so
what
I
will
do,
I
will
try
to
configure
it
using
a
couple
of
special
identities.
So
here
we
have
to
use
a
method,
authentication
method
as
aka
which
will
be
used.
Your
sim
will
be
used
to
authenticate.
So
so
you
can
see
that
wi-fi
is
already
latched
and
there
is
no
username
and
password
asked
so
directly.
Your
sim
will
be
doing
the
authentication
with
your
existing
mno
core
ym
magma.
B
Okay,
so
to
summarize
here
so
what
we
are
trying
to
do,
we
are
solving
for
heterogeneity.
The
first
thing
is
like
flexibility,
so
mobilizing
the
network
into
fast
fabric
and
policy
rich
edge
to
maximize
the
deployment
and
business
models.
That's
what
we
are
trying
to
solve
here.
The
next
is
hierarchical
orchestration,
so
distributing
the
edge
service
requires
hierarchical
obstacles
so
that
we
are
achieving
here
and
next
will
be
the
consistent
run
time
where
we
are
using
kubernetes
based
edge
clusters,
which
offers
a
constraints
runtime
for
each
service.
B
B
A
Thanks
everybody,
so
I
think
we
have
one
question
that
I
think
I
know
the
answer
to
the
other.
Two
I'll
probably
have
to
sync
up
offline
on
slack,
so
the
first
one
is
around.
I
haven't
seen
any
new
comments
in
the
worklet
project
since
december.
4Th
2019
is
the
project
still
alive,
I
believe
it's
still
alive,
but
for
the
integration
that
we're
exercising
here
with
magma
in
certain
components
that
require
a
custom
kernel
and.