►
From YouTube: IETF103-TUTORIAL-TRAFFICENGINEERING-20181104-1230
Description
Traffic Engineering in the IETF
Traffic Engineering is now so important as a glue for the Internet that it's worth stepping back to consider what it is, how it works, and what work the IETF is currently engaged in.
This tutorial introduces the concept of Traffic Engineering (TE), describes the basic techniques, runs through the key TE technologies that have been developed in the IETF, and provides pointers to current IETF efforts to enhance Traffic Engineering. Presented at IETF 103, the talk describes recent architectural models such as the Architecture for Application Based Network Operation (ABNO), PCE as a Central Controller (PCE-CC), and a Framework for Abstraction and Control of TE Networks (ACTN), and it mentions new approaches to Traffic Engineering delivered by Segment Routing (SR) and Deterministic Networking (DetNet). It also preview the work on YANG models for the control of TE resources and virtual networks.
Slides at: https://datatracker.ietf.org/meeting/103/materials/slides-103-edu-sessk-an-ietf-traffic-engineering-overview-01
A
And
by
way
of
proof-
and
this
is
us
so
me-
I
was
routing
ad
for
for,
while
a
while
and
I've
co-chaired
a
number
of
working
groups
and
I'm
really
a
circuits
which
traffic
engineering
person
deep
down
inside
I've
just
become
the
independent
stream
editor
for
my
sins.
How
many
out-
and
this
is
brilliant-
this
you've
got
to
guess
the
age
of
hammy
on
illness
in
this
photo
and
compare
and
contrast
with
him
now
this
this
was
this
was
his
graduation
some
twenty
years
ago,.
A
A
A
Oh
No
traffic
engineering
working
group
chair
has
just
come
in
at
the
back,
so
now
we're
really
in
trouble
yeah
we're
trying
to
introduce
and
spend
quite
a
lot
of
time
on
some
of
the
new
techniques
and
architectures
for
for
traffic
engineering.
As
far
as
references
go,
there
are
just
way
too
many
to
list.
What
we've
tried
to
do
is
pick
out
on
many
of
the
slides,
a
key
RFC
or
draft
that
you
should
start
with.
If
an
technology
interests
you
and
then
at
the
end
we
have
a
list
of
relevant
working
groups.
A
A
2702
is
actually
a
requirements
RFC
for
traffic
engineering
in
an
MPLS
network,
but
what
it
does
is
set
the
scene
by
saying
this
is
the
application
of
technology
to
measurement
modeling,
characterization
and
control
of
traffic
in
the
Internet
to
achieve
specific
performance
objectives,
and
it
does.
It
then
goes
on
to
not
say
what
specific
performance
objectives
they
might
be,
but,
in
summary,
the
purpose,
then,
is
to
allow
you
to
provide
more
reliable
traffic
delivery
offer
advanced
services.
A
If
you're
a
network
operator,
you
want
to
offer
advanced
services
to
your
customers,
you
want
to
make
better
use
of
your
network
resources,
so
deployed
resources,
cost
money,
you
don't
want
them
wasted
or
tied
up.
You
want
to
be
able
to
get
as
good
a
use
out
of
your
network
as
possible.
You
want
to
survive
network
outages,
but
also
planned
maintenance
and
a
part
of
all
of
that
is
making
the
network
predictable.
So
you
can
guess
what's
going
to
happen
to
to
your
customers
traffic
when
there's
a
change
in
the
network.
A
A
First
typo,
you
need
to
know
traffic
patterns,
demands
and
matrices.
You've
got
to
know
what
traffic
your
network
is
going
to
carry.
If
you're
going
to
engineer
that
traffic
and
you
have
to
know
what
sort
of
errors,
what
likelihood
of
errors
and
what
plans
you
have
for
maintenance
once
you
know
all
that,
then
you
have
to
control
how
the
traffic
is
placed
on
the
network
and
to
do
that.
You
configure
the
network,
the
traffic
and
you
actually
configure
the
management
tools
themselves.
A
What
sort
of
networks
are
we
talking
about?
Well,
this
is
the
IETF,
so
essentially
we're
talking
about
the
delivery
of
IP
packets
turns
out
that
the
principles
of
traffic
engineering
apply
to
any
network
that
delivers
commodity,
so
I've
seen
similar
or
identical
algorithms
applied
to
electricity
networks
to
the
management
of
cars,
on
road
networks,
to
the
delivery
of
water
and
the
inverse
delivery
of
sewage
and
so
on
and
a
sewage
it's
fun,
because
I
used
to
have
a
company
that
made
a
sort
of
traffic
engineering
tool
using
machine
learning.
We
built
it
for
IGF
protocols.
A
What
that
tells
you
about
the
ITF
I,
don't
know
here
we're
going
to
focus
on
layer,
3
and
below
so
that's
IP,
MPLS,
Ethernet,
and
then
all
the
transport
technologies
that
sit
underneath.
So
it's
all
well,
it's
nearly
all
some
kind
of
circuit,
switching,
whether
it's
connection
oriented
or
connectionless,
but
IP
is
a
little
bit
different
on
top
all
right
computer
science,
networking
101
shortest
path.
First,
it's
what
our
iGPS
do!
A
Along
comes
constraint,
based
shortest
path.
First,
and
all
this
is
doing
is
saying.
Not
only
am
I
interested
in
the
cost
of
links
but
I'm
interested
in
additional
qualities
of
those
links,
and
it
may
be
things
like
bandwidth
availability,
that's
the
most
common
one
and
from
a
computational
point
of
view,
it's
very
easy.
You
can
prune
the
graph
before
you
even
start,
so
you
take
out
all
the
links
that
don't
match
your
criteria
and
then
you
run
SPF,
it's
easy.
A
Those
computes
can
be
end-to-end
and
that's
also
easy.
You're
you're
just
running
a
limiting
factor
on
SPF,
so
it
all
turns
out
to
be
really
simple
in
in
this
picture.
You
might
want
to
compute
a
path
that
uses
only
red
links
to
get
in
to
end,
and
that
would
be
fine,
but
you
would
still
find
that
if
you
ran
your
network
saying
everybody's
interested
in
only
redlynx,
then
the
traffic
would
tend
to
converge.
A
So
this
is
a
standard
picture
from
traffic
engineering
tutorials.
It's
it's
a
fish.
We
overlay
on
the
fish.
The
reason
why
it's
a
fish,
if
it's,
if
it's
sort
of
fish,
shaped
graph,
we
overlay
on
it.
Some
nodes
and
links
and
the
links
have
bandwidth
limitations,
so
the
green
ones
are
10
megabyte,
maybe
bit
the
the
blue
ones
are
7
and
the
yellow
ones
are
5
and
then
we
place
2
demands
on
the
network.
A
One
of
them
is
a
7
made
demand
and
one's
a
5
mega
demand
and
they're
both
coming
in
and
they're
trying
to
get
to
the
fishy.
As
knows
your
fish
have
noses
anyone
know.
Nobody
knows
your
fish
have
noses.
Yeah,
ok,
virtualized,
no
nose.
Yes,
so
suppose
the
5
make
demand
comes
first
and
we
and
it's
rooted
on
shortest
path.
A
First,
well,
it's
gonna
go
on
the
and
the
blue
link
accom
help,
and
it's
going
to
pick
the
green
links
to
go
through
and
when
that
happens,
the
7
make
demand
turns
up
and
it
can't
be
placed
clearly
the
network
can
support
both
demands,
but
not
if
they're
they
arrive
in
the
wrong
order
and
they
just
handled
a
shortest
path.
First,
so
what
traffic
engineering
is
all
about
is
understanding
the
set
of
demands
you're
going
to
have
understanding
your
network
and
causing
those
demands
to
be
placed
in
the
right
way
through
the
network.
A
A
So
to
solve
that
problem,
then
you
need
to
know
the
topology.
You
need
to
know
the
status
of
the
links.
Are
they
up
down
paused
ready
to
come
down
for
maintenance?
What's
the
bandwidth
of
the
link,
but,
more
importantly,
what
bandwidth
is
available
and
it's
not
been
taken
up
by
other
services,
already
other
constraints
that
you
might
be
interested
in
and
and
those
are
increasingly
delay
type
constraints.
A
As
we
start
to
deliver
5g
services,
but
for
optical
networks
they
may
be
the
ability
to
convert
or
not
convert
wavelengths
and
things
like
shared
risk
link
groups
which
are
the
preponderance
or
the
the
likelihood
of
one
link
failing
at
the
same
time
as
another
link
fails.
So
if
you
know
all
of
those
things
and
then
you
know
the
demands
that
are
going
to
be
placed
on
the
network
and
any
relationship
between
the
demands,
for
example,
one
may
be
providing
a
backup
service
to
another.
You
don't
want
to
send
them
down
the
same
link.
A
So
we
need
tools
to
do
this.
We
need
network
management,
the
network
management
can
see
the
network
and
it's
it
can
see
the
demands
and
it's
responsible
for
working
out
how
to
place
the
traffic
and
issuing
instructions
to
the
network
to
configure
the
network
so
that
the
traffic
will
go
the
right
way.
A
We
need
the
ability
to
configure
the
network
devices
and
that
can
be
direct
configuration
or
C
protocols
and
then,
as
the
packets
go
through
the
network,
we
may
need
to
mark
them
so
that
we
can
tell
which
packet
is
meant
to
go,
which
way
through
our
network
and
essentially
the
rest
of
this
presentation
is
talking
about
different
techniques
for
doing
those
different
things.
You
don't
always
end
up
using
all
of
those
tools.
At
the
same
time,.
A
A
These
are
the
existing
IETF
tools
that
we
will
run
through,
and
then
we
have
another
section
talking
about
the
future
ones,
so
we've
got
a
slide
or
so
on.
Each
of
these-
and
this
is
this-
is
where
references
to
our
RCS
and
draft
start
coming
in
there'll,
be
little
yellow
boxes
down
the
left-hand
side.
A
To
give
you
your
pointers,
source
routing,
then
source
routing
is
encoding
information
in
a
packet
to
say
which
path
through
a
network
a
packet
should
take
it's
essentially
a
series
of
hop
identifiers,
either
link
or
node
addresses,
saying
right
now,
you've
done
that
hop.
This
is
the
next
hop
to
do,
and
that
can
be
a
strict
path.
A
strict
path
is
a
full
list.
There's
never
any
ambiguity
whenever
you
have
a
packet
in
your
hand,
you
can
see.
This
is
the
next
thing.
I
must
do.
It
requires
no
information
to
be
stored
in
the
network.
A
All
of
that
information
is
placed
in
the
packet
itself
and
the
example.
Best
example
of
this
is
ipv4.
Strict
source
routing
option,
which
enables
you
to
put
in
a
list
of
I,
believe
9
IP
addresses,
and
this
is
instructive.
Just
as
a
side
issue
when
ipv4
was
designed,
it
was
clearly
expected
that
the
diameter
of
the
Internet
would
never
be
more
than
nine
hops.
A
So
that
was
a
good
prediction.
Wasn't
ipv4
strict
source
routing
option
is
generally
not
used.
Okay,
it
exists.
It's
not
used
loose
paths,
a
loose
path,
breaks
that
strict
path
down
and
says.
Sometimes
I
just
want
to
get
all
the
way
over
there.
I
don't
need
to
fill
in
the
gaps.
I'll
just
say
next
hop,
maybe
get
to
them
the
edge
of
the
next
network,
so
that
information
is
still
in
the
packet
header,
but
it
relies
on
the
network
itself
to
fill
in
the
gaps
using
normal,
shortest
path.
First
routing.
A
So
it's
it's
a
combination
of
tools
then-
and
there
are
two
examples
of
that:
ipv4
has
got
the
loose
source,
routing
option
and
ipv6
has
got
a
routing
header
extension
and
apart
from
segment
routing
which
we'll
come
to
later,
neither
of
these
options
is
actually
used
in
the
wild.
So
it's
a
great
tool.
It's
just
not
used.
A
A
So
the
metric
tweaking
approach
says
well:
if
we
just
go
in
there
and
change
a
couple
of
metrics
in
the
whole
network,
we
can
cause
the
demands
to
go
differently
through
the
network,
so
this
is
just
touched,
these
two
metrics
and
now
the
shortest
path.
First
I
can
see
you
all
busily
checking
my
mats
and
I
think
it's
right.
They
get
routed
differently.
None
of
the
links
is
overloaded.
A
A
10,000
demands
on
a
big
network.
This
becomes
an
extremely
interesting
computational
problem.
You
can
get
some
very
exciting,
unexpected
consequences
you
go
in
and
tweak
one
demand
and
all
the
traffic's
suddenly
goes
a
different
way.
You
can
flap
the
network
as
you
install
these
new
metrics
and,
if
you're
doing
that
in
a
live
network,
yeah
I
needed
to
change
both
of
these
but
I
changed
one
first
and
all
the
traffic
went
the
other
way
and
I've
just
made
it
even
worse
and
then
I
change
the
other
one
and
the
traffic
comes
back
again.
A
A
A
All
right,
you
can
look
at
one
topology
one
network
topology
and
make
multiple
copies
of
logical
copies
of
it
and
effectively
mentally
partition
the
network
resources
into
multiple
slices
instances
of
that
and
build
what
we
call
colored
graphs.
So
what
the
figure
does
is
it
says,
look
there's
a
a
topology
there
and
for
every
link
we
will
give
it
a
red
metric
and
the
blue
metric
art
from
my
notice
that
one,
which
is
a
colorblind
blue
metric
second
type.
Oh
I
thought
you
reviewed
these.
A
So
what
it
means
is
when
you've
got
some
traffic
to
inject,
you
can
make
a
decision
I'm
going
to
call
this
blue
traffic
and
it
will
get
routed
through
the
network
on
the
blue
graph
or
I'm,
going
to
call
this
red
traffic,
and
it
will
go
through
the
network
on
the
red
graph,
how
you
actually
make
that
decision.
Well,
that's
a
local
policy
choice.
A
Igps,
allow
you
to
advertise
the
different
metrics
with
their
colors
and
off.
You
go
I've
seen
two
principal
use.
Cases
for
this.
One
is
providing
resilience
so
you're
able
to
partition
your
network
and
say
if
there's
a
failure
on
any
link.
I
know
that
the
other
color
network,
the
other
color
graph,
will
be
up
anyway.
A
A
It
functions
by
having
control
packets
that
fly
around
between
protocol
instances
and
are
used
to
describe
the
traffic
and
to
Grahame
their
network
to
expect
that
traffic,
those
control,
packets,
follow
the
the
flows
or
they
follow
the
same
path
that
the
flow
will
will
follow.
So
they
ideally
come
before
the
traffic,
but
they
could
actually
happen
once
the
traffic
is
already
flowing.
A
They
have
the
same
IP
addressing
so
they
follow
this
exactly
the
same
SPF
path
and
what
they
do
is
they
go
out
words
describing
the
flow
saying
this.
There
is
traffic
coming
here
and
it
needs
this
amount
of
bandwidth
and
then
the
packets
are
the
control.
Packets
are
fired
back
down
the
network,
saying
yeah.
Okay,
we
have
a
destination,
we
understand.
What's
going
on
now,
reserve
some
bandwidths
and
buffers
or
whatever
in
the
devices
so
that
the
the
packets
can
be
processed.
A
So,
what's
going
on
here
is
there's
an
overhead
of
control
message,
exchange
their
state
needed
in
the
network
to
actually
reserve
resources
and
to
understand
the
relationship
between
the
control
messages,
but
it's
a
highly
adaptive
to
changes
in
the
SPF.
So
if
a
link
goes
down
and
all
the
traffic
is
routed
a
different
way,
the
control
packets
are
also
routed
the
different
way.
So
they
everything
the
network
quickly
learns
about
the
changes.
It
also
handles
merging
of
flows,
which
is
not
uncommon.
There
are
lots
of
traffic's
coming
from
many
places
going
to
the
same
destination.
A
A
So
the
mpls
data
plane
has
its
origins
in
a
belief
that
we
couldn't
do
sufficient
IP
address
lookup
at
line
speed,
which
has
turned
out
to
be
a
false
assumption,
we're
perfectly
happy
to
forward
IP
traffic
ridiculous
line
speeds,
because
the
silicon
got
much
better,
but
along
the
way,
some
other
uses
for
MPLS
came
along
that
were
were
pretty
handy.
It's
based
on
a
little
header,
that's
inserted
on
top
of
the
IP
packet
and
constructs
a
forwarding
state
based
on
a
label
lookup.
A
So
at
every
hop
you
look
up
the
label,
and
that
tells
you
one
interface
to
send
the
packet
out
of
and
put
a
different
label
on
and
then
along.
It
goes
through
the
network,
it's
creating
a
connection-oriented
packet,
switching
technology,
and
that
turns
out
to
work
really
well
for
traffic
engineering.
What
MPLS
traffic
engineering
does
is
use
a
constraint
based
shortest
path.
A
First,
where
the
constraint
is
basically
bandwidth
find
me
the
shortest
path
with
sufficient
bandwidth
for
this
flow,
and
it
allows
you
to
place
traffic
in
quite
a
sophisticated
way
in
your
network
to
reduce
the
cost
to
make
more
efficient
use
of
the
bandwidth.
So
this
is
a
capex
saving
to
ensure
that
you
get
the
the
right
policies,
the
right
behavior
for
certain
types
of
of
traffic.
So
again
it's
really
nice
for
emergency
service
traffic.
You
can
find
a
really
short
path
and
make
sure
it's
got
higher
bandwidth
reserved
ready
for
it.
A
A
The
ultimate
goal
it
turns
out
for
impe
LST
is
its
cost
saving
and
you
can
look
at
cost
saving
two
ways
it,
but
principally
it's
it's
squeezing
more
out
of
a
network
right.
The
service
provider
has
spent
money
deploying
equipment.
They
want
to
get
as
much
traffic
running
through
that
network
as
predictably
as
possible.
A
There
some
protocols
involved
here
what
was
done
was
to
take
the
existing
iGPS,
which
are
advertising
and
reporting
on
the
quality
of
the
links
in
the
network
and
to
enhance
them
with
a
little
bit
of
extra
information
to
report
on
the
information
and
status
of
the
to
you.
Qualities
of
the
link
and
again
bandwidth
is
the
top
of
that
list,
so
the
IGP
on
its
own
would
say,
link
status
and
metric
the
two
year.
Extensions
gives
you,
the
bandwidth
and
a
te
metric
on
top
of
that.
A
There's
a
signaling
protocol
required
to
go
and
program
the
nodes
in
the
network
and
what
we
did
was
we
took
a
RSVP
and
made
some
additions
to
it.
So
it
was
traffic
engineering
aware,
and
that
means
that
those
RSVP
packets,
instead
of
following
shortest
path.
First,
they
actually
have
an
explicit
path
that
nailed
into
the
control
messages
so
that
that
we
can
set
up
the
flow
for
the
LSP.
B
Okay,
Jim
Paris
is
exactly
the
technique
of
where
I
started
my
ITF
and
I.
We
have
already
been
demonstrated
to
saying
that
the
are
so
apt
and
OSPF
T
is
a
works
and
they
save
the
class
to
sell
ITF.
C
is
really
good
at
doing
something
like
generalization
and
makes
the
same
philosophy
applied
everywhere.
So
for
every
potential
circuit,
switching
technology.
This
concept
is
generalized
and
to
construct
a
label
hierarchy
on
the
lambdas
on
the
label
packet
and
also
on
the
fibers.
B
So
this
one
is
extended
only
for
4
from
the
NP
RS
and
the
into
ours
is
very
typical.
We
keep
the
OSPF,
p
and
r
rsvp-te
extensions
and
all
the
same
time,
we
also
develop
the
new
protocols
called
RMP
for
the
coordination
between
the
physical
links,
and
this
is
all
distributed.
All
the
nodes,
how
the
gim
PRS
is
working.
B
And
then
pass
computation.
So
will
you
have
everything
on
the
distributed
manner?
Then
there
would
be
a
problem
on
the
vision,
because
considering
each
often
node,
you
can
only
have
restricted
vision
of
the
whole
network
rather
than
we
have
a
global
topology
information.
So
it
is
request
that
the
routing
and
path
computation
function
to
be
moved
to
a
component
or
entity
that
can
have
the
global
issue
to
provide
further
optimization.
So
PC
is
evolved
in
this
way
and
now
very
popular
and
can
be
applied
as
a
component
or
can
be
applied.
B
The
application
or
a
network
note
everywhere.
As
long
as
there
is
a
need
for
the
past
computation
function,
so
the
earliest
RFC
46:55
specify
the
functionality,
but
it
can
be
embedded
in
any
routers
in
every
router
and
to
say
the
cost
where
you
need
it,
and
the
second
RFC
in
this
page
effective,
wanted
to
provide
a
more
advanced.
The
use
case.
That
is
a
word
to
P,
o
P
use
case,
which
announced
the
PCE
domains
can
be
interacted
with
each
other.
B
B
So
there
are
the
different
kind
of
realizations
of
the
PDE
in
the
architecture
and
historically
okay.
For
noted,
when
we
call
the
historically
means
is
something
like
twenty
years
ago
and
here
when,
because
this
is
a
previous
technique.
So
so
this
slide
shows
something
that
just
three
or
five
years
ago,
maybe
so
BCE
actually
locates
at
so
wherever
there
is
a
need
for
past
computation.
The
first
the
scenario
would
be
the
RS
are.
B
The
second
scenario
would
be
a
in
ms:
they
already
have
this
kind
of
functionality
as,
but
as
they
do
not
just
to
use
the
PC
terminology
to
do
that.
But
after
that,
when
we
have
a
PCE
separated
out
as
a
dedicated
server,
there
are
also
special
configuration
like
the
T
database,
like
our
SP
database.
How
many
connections
are
in
the
whole
network?
These
are
all
understood
on
the
PC
server
and
there
are
also
different
ways
to
interact
with
other
kind
of
PCs
and
the
opposite
to
a
PC.
B
And
then
we
comes
the
topology
aggregation
because
of
different
component
have
different.
The
topology
informations
there
from
the
perspective
of
PCC
is
a
network,
is
URI
rather
simple
and
for
the
PCE.
Rather
it's
a
big
and
more
complicated
topology
information,
so
here
the
RFC
79
26
provider
very
typical
examples
for
how
the
abstraction
is
done
between
the
client
layer
network
and
the
service
layer
network.
So
you
relieve
that
when
you
look
into
the
operators
or
even
look
into
the
windows
there,
it's
very
large
networks,
but
the
representation
of
the
clients
idea
is
do
really
simple.
B
So
here,
for
example,
we
have
five
nodes
in
the
summer
Network,
but
all
the
other
client
connections
is
so
only
useful
to
say
three
to
see
three
of
them
and
you
can
just
hide
them
to
simplify
the
work
of
the
client,
so
the
abstraction
abstraction
layer
resource
can
be
either
node.
Abstraction
can
be
also
link
abstraction
and
also
just
show
that
all
three
to
the
client.
B
And
given
this
kind
of
abstraction,
there
can
be
a
few
technical
about
virtualization,
which
means
something
you
pretend
you
have,
but
you
really
have
in
another
shape.
So
previous
page
shows
an
example
on
how
to
simplify
while
the
abstraction,
but
this
one
also
show
another
kind
of
variation.
You
can
also
multiplied
your
your
that
works,
it's
a
abstraction
layer
and
allocate
them
to
different
work
networks
to
satisfy
the
requirement
for
different
users,
and
this
is
nothing
but
a
map
between
the
physical
ones
and
or
what
you
want.
But
what
you
need
is
just
a
memory.
B
How
many
you
really
have
and
don't
just
overload
your
network.
So
there
are
different
kinds
of
policies,
different
kinds
of
algorithms
working
on
how
to
efficiently
virtual
eyes
and
how
to
make
best
to
use
our
distractions.
If
you
do
it
with
a
very
abstracted
manner,
it
will
actually
sacrifice
your
efficiency
because
you
have
too
little
information.
But
if
you
have
maybe
one-to-one
mapping
and
show
the
whole
picture,
then
the
complexity
or
the
client
is
really
really
high.
So
there
is
always
a
balance
between
the
degree
of
virtualization
and
that
introduced
different
kind
of
policies.
A
Bgp
implementations
are
really
good
at
klein
policy,
so
abstraction
is
all
about
policies.
So
that's
a
good
reason.
Bgp
is
actually
quite
good
at
moving
bulk
data
around
and
the
export
of
topology
is
an
example
of
moving
bulk
data,
and
it
also
turns
out
that
the
places
in
the
network
that
you
want
to
do
the
export
from
are
typically
nodes
that
are
already
running
BGP
and,
lastly,
a
feature
of
BGP.
A
The
the
root
reflector
allows
you
to
send
the
information
to
multiple
places
without
overloading
the
network
nodes
themselves,
so
we
invented
link
state,
BGP,
GG
pls,
and
it's
a
set
of
encoding
to
take
information
that
describes
links
and
nodes
in
their
traffic
engineering
form
from
the
network
and
send
them
out
in
BGP
to
important
notes
on
that.
Bgp
is
just
a
protocol
that
could
do
it.
There's
also
work
now
to
write
yang
models
to
describe
topology.
A
A
A
So
the
first
one
is
improving
IGP
flooding,
and
this
has
been
talked
about
in
the
ITF
for
the
longest
time.
The
intention
is
to
reduce
the
processing
load
in
the
network,
as
the
iGPS
continue
to
announce
and
maintain
the
network.
State
particularly
important
is
when
there's
a
link
failure
that
information
has
to
get
propagated
to
everybody.
A
A
It's
about
Billy
building,
deterministic
paths,
predictable
paths
for
data,
it's
aimed
at
layer
to
bridged
and
layer,
three
routed,
Network
segments
quite
often
looking
at
a
concatenation
of
those.
So
you
might
start
it
in
layer.
Two
do
some
hops
there
and
then
move
into
a
layer,
3
network
and
then
pop
back
out
into
a
layer,
2
network.
A
The
intention
is
that
the
paths
provide
some
predictability
in
terms
of
latency
and
loss
and
jitter,
so
not
necessarily
controlling
those
things,
but
knowing
what
the
worst
case
might
be,
and
that
is
really
useful
in
delay.
Sensitive
networking
for
things
like
video
transfer,
especially
popular
for
for
video
editing,
Suites,
probably
quite
useful
for
augmented
reality.
Unless
you
like
falling
down
holes,
the
work
is
being
split
between
the
IETF
det
networking
group
and
I
Tripoli
802
dot,
one
their
time-sensitive
networking
working
group,
which
is
looking
at
the
layer.
Two
parts
of
this.
A
It's
still
a
little
early
to
know
where
this
is
ending
up,
but
it
looks
like
it's
doing:
data
plane,
encapsulation
for
NP,
RS
and
IP.
It's
relatively
simple,
although
not
not
completely
trivial
and
ends
up
based
on
tunnels,
with
the
MPLS
tunnels
placed
with
some
knowledge
of
network
behavior
and
well.
That
could
well
be
rsvp-te.
A
A
Segment,
routing
segment
routing
is
essentially
a
tunneling
technology,
so
the
the
payload
data
is
encapsulated
in
a
packet,
and
it's
forwarded
based
on
information
in
that
encapsulation
header.
In
a
way
that
is
very
similar
to
I
P
source
routing,
that
is
to
say,
there's
a
list
of
hops
placed
in
the
packet
header
and
those
hops
describe
the
route
that
the
path
that
the
packet
must
take
through
the
network
and
each
packet
is
then
a
free
agent.
It
goes
off
with
its
information
in
it.
It
doesn't
follow
any
programmed
path
in
a
control
plane.
A
Each
packet
contains
all
the
information
it
needs
to
get
to
the
end.
That
makes
it
traffic
engineering
technology,
where
some
entity
can
compute
a
path
that
wants
the
packet
to
go
on
program,
that
information
into
the
packet,
header
and
off
the
packet
goes
and
follows
that
path.
So
if
the
control
entity
is
capable
of
seeing
the
whole
topology
and
accounting
for
bandwidth,
you
can
do
a
fairly
sophisticated
traffic
engineering
resource
reservation
protocol,
the
control
plane.
All
the
signalling
is
now
not
needed
in
a
network,
so
there's
great
save
there.
B
Okay
and
given
this
kind
of
technologies
and
then
it's
time
to
talk
about
how
the
controller
is
designed
and
what
kind
of
techniques
can
be
used,
while
the
controllers
so
PC
yes,
previously
discussed,
it
includes
functionalities
on
the
past
commutation-
and
this
is
exactly
one
of
the
functionality
need
to
be
achieved
by
the
controller,
interesting
architecture.
Software-Defined
networking,
so
PC
protocol
actually
provide
a
very
mature
there.
Mature
solution
to
the
Sdn
southbound
interface.
The
solution
and
this
kind
of
control
and
computation
may
also
be
applied
on
a
single
node.
B
So
the
applications
can
be
it's
proposed
from
both
NP
RS
non
packet
and
IP
environments.
So
this
is
widely
used
as
a
controller
and
can
be
constructed
as
the
controller
hierarchies.
If
necessary
and
based
on
this
idea,
we
have
from
promoted
a
new
architecture,
namely
as
ICT
and
obstruction
control
of
T
networks
by
constructing
the
controller
of
the
high
Rockies.
B
So
this
this
is
actually
some
kind
of
state
of
art
into
the
industry
and
we
are
implementing
this
kind
of
as
the
in
controller
hierarchy
to
make
use
of
the
abstraction
techniques,
and
this
is
very
well
accepted
by
the
carriers
for
multi
domain
in
Europe
and
from
a
T
domain,
but
not
when
they're
interrupts
and
multi
domain
coordination.
So
we
separate
to
the
different
than
list
of
the
controllers
and
from
the
customer
from
the
carrier
and
this
from
the
winter.
B
So
the
customer
early
do
not
have
much
knowledge
on
the
network
and
they
only
take
care
of
a
few
boxes
located
in
their
working
area.
So
what
they
need
is
just
to
represent
their
request
to
the
carriers
by
Chris
parking
interface.
We
named
this
as
a
CMI
and
the
Perea
is
actually
is
a
critical
component
in
the
whole
architecture.
Because
to
the
above,
we
need
to
be
connected
with
a
customer
to
understand
their
demand
and
to
the
bottom.
They
need
to
interact
with
a
winner.
B
So
on
the
detail,
the
network
configuration
so
to
interface,
all
them
for
the
customer,
understanding,
virtual
service,
coordinating
for
the
different
customers
and
the
coordination
between
multi
domains
for
among
the
winners-
and
this
is
all
the
support-
it
applies-
the
abstraction
and
the
virtualization
techniques,
so
the
window
specific
ones
is
actually
the
most
complicated
ones.
It's
very
challenging
because
the
scalability
of
the
network-
and
there
is
a
lot
of
parameters-
need
to
be
configured
on
both
the
network
level
and
the
device
level.
B
So
currently,
I
think
about
most
of
the
winners
have
their
all
at
one
tutors
for
their
own
implementation.
But
people
ask
you
meet
requesting
for
a
standard
interface
to
be
to
enable
the
internal
friction
among
multiple
vendors.
This
is
done
on
the
MPI
interface
and
carioca
handle
that
are
not
you
in
their
case,
so
typical
scenarios
for
a
city
and
how
the
topology
is
discovered,
how
the
topology
is
abstracted.
The
Union-
and
this
page
shows
an
example
on
how
this
is
done.
B
B
It
will
receive
the
topology
information
free
from
three
different
domains
and
thus
teaching
them
together
to
four,
a
very
brief
topology
that
can
be
used
and
understand
easily
by
the
controller
by
the
customer
controllers,
which
is
a
DC
here,
and
everything
can
be
representing
the
ass
of
connectivity
matrix
to
have
local
information
and
all
the
capability
on
the
switching
and
the
bandwidth
information
of
the
network.
Everything
is
understandable
from
the
abstraction
topology
and
given
this
kind
of
topology
people
are
when
people
are
capable
to
to
deploy
their
connections
onto
the
networks.
B
It
is
worth
noting
that
this
can't
be
a
kind
of
technology,
specific
domains,
saying
that
we
have
different
techniques
like
IP,
like
article
like
something
together,
but
everything
is
nothing
but
kind
of
decomposition,
as
there
is
already
a
kind
of
standardized
interface
and
models,
we
can
easily
instruct
the
same,
which
domain
to
complete,
which
part
of
the
work-
and
this
will
be
very
helpful
on
the
setting
up
the
end-to-end
tunnels
and
deploys
or
connections
so
on
the
different
level
of
the
interface.
We
are
having
different
kind
of
protocols
so
further
above
interface.
B
The
mode
is
you
early
represented
by
the
protocol
and
for
the
negotiation
between
the
operator
and
the
winders.
There
are
multiple
ways,
including
the
piece
AB
and
the
rest
confit
on
which
is
fully
automated.
There
are
also
QF
Kaba
and
the
SNMP
used
to
for
many
many
years
and
that,
but
some
kind
of
manual
works
request
it
and
the
first
bone
interface
there
can
be
PC
protocol.
This
is
between
the
controller
and
note
and
a
Molineux
there
can
be
GMP
are
so
a
city
has
good
progress
eitf.
B
It
was
first
proposed
to
maybe
use
four
years
ago.
2014
and
directed
to
teas
working
group
in
2015
and
in
2016
we
did
develop
two
hackathon
events
and
then
promote
the
draft
sought
out
in
the
working
group
and
though
our
hackathon
also
evolved
from
transport
only
to
IP
plus
transport.
So
that
is
a
kind
of
demonstration
of
how
the
entire
operation
can
be
achieved
while
standard
models,
and
now
it's
becoming
an
RFC.
So
the
architecture
is
now
and
what
we
are
investigating
only
is
a
kind
of
young
models
to
fit
into
this
architecture.
B
B
B
That
describes
how
how
the
in
from
network
information
is
more
daughter
or
how's
it
model
it's
used
in
the
whole
architecture,
so
each
of
the
box
in
the
car
is
actually
a
module
again
module
and
they
located
in
separated
crafts,
and
the
arrows
here
shows
the
relationship
on
the
modules
so
who
the
the
arrow
points
the
pointed
one
will
augmented
to
the
the
base
model,
and
this
allows
people
to
generalize
a
model
and
augment
it
for
the
technology
specifically
use.
So
from
the
basic
categories
and
categorization.
B
It
is
classified
into
topology,
Tunnel,
service
and
types,
and
we
also
have
the
layer,
0
technology
to
layer,
three
technology
for
every
type
of
this
kind
of
models
we're
coming
to
and
as
a
summer
session.
So
T
traffic
engineering
is
really
a
valuable
tool.
You
do
not
need
to
spend
more
money
on
your
infrastructure,
but
you
get
more
from
your
network,
so
it
helps
people
to
guarantee
their
quality
of
the
delivery,
and
it
also
provides
some
kind
of
bonus
on
their
networks
by
carefully
and
efficiently
use
their
network
resource
and
to
achieve
this
goal.
B
Actually,
this
is
still
an
ongoing
work
and
there
are
really
a
lot
work
in
the
past
and
in
the
at
this
moment,
so
most
of
them
widely
deployed
and
we
are
developing
more
tools
and
techniques.
So
there
are
a
lot
of
opportunity
to
get
to
be
involved
here.
If
you
understand,
if
your
expert,
so
if
any
layer
of
the
technologies
and
want
to
bring
such
kind
of
models
highly
welcomed.
B
And
this
page
just
to
show
some
pointers
but
yeah
ITF
for
started
30
years
ago
when
I
was
born.
So
so
one
page
should
not
be
sufficient
to
list
all
the
reference,
but
the
PRC's
has
been
listed
previous
slides
and
you
can
look
into
that
one
and
figure
out
that
will
help
you
to
to
point
to
more
RFC's
and
some
interesting
working
groups
for
the
existing
technology
promotion,
including
the
t's
for
the
architecture
of
all
the
te
technologies
and
NPR
ste,
is
something
specially
designed
for
NP
RS.
B
Secondly,
is
something
specially
designed
for
optical
switching
and
the
PCE?
Is
everything
about
the
pc
protocol
so
sprint?
Working
group
c
is
now
working
on
the
segment,
routing
and
IDR
works
for
the
BGP
and
the
BGP
Ras
extension.
That
model
is
a
right
working
group
to
the
young
modeling
and
the
corresponding
protocols
and
error
error.
S
are
provides
extension
to
the
I
gps,
their
networks,
all
the
time,
sensitive,
networking
and
I
think
that's
all
right
time
for
the
questions
and
you're
welcome
to
provide
us
the
feedback.
By
doing
this
away.
A
A
C
A
A
C
A
And
I
think
one
of
her
Mayans
pictures
for
a
CTN
shows
centralized
controller
talking
down
to
a
domain
controller.
Now
whether
that
domain
controller
then
uses
individual
southbound
programming
or
uses
a
control
plane
in
that
domain,
nobody
further
up.
The
stack
knows
about
that
or
needs
to
know
about
it.
So
different
domains
can
do
different
things,
but
still
be
underneath
the
central
control
architecture.