►
From YouTube: IETF115-BIER-20221110-1530
Description
BIER meeting session at IETF115
2022/11/10 1530
https://datatracker.ietf.org/meeting/115/proceedings/
B
A
B
A
A
A
A
Yeah
you
got
to
put
your
face
to
the
mic.
I
can't
hear
you
when
you're
that
far
away
we'll.
A
You
just
faded
out:
man
fell
off
the
chair,
okay,
yeah,
I
I
need
a
I,
need
a
button,
a
release
button
if
I
pass
out
and
drop
the
button.
You'll
know
there's
just
questions
already
about
remote
control
of
the
slides
I
mean
the
PDFs
are
up
there.
I
know,
there's
two
different
buttons
on
allowing
either
they
control
it
or
you
control
it.
A
It's
not
obvious
which
is
which,
but
either
way
they
don't
do
remote
slide
and
screen.
Sharing
you've
got
the
slides.
They
either
hit
the
button
or
you
hit
the
button
for
them.
Correct.
C
C
B
A
B
B
B
B
Okay,
so
welcome
starting
beer
session.
We
have
semi-conscious
chair
online,
so
Greg
will
be
managing
the
queue.
We
have
a
projectile
song
that
came
back.
B
We
have
one:
one
of
the
fathers
came
back
a
pretty
good
agenda
today,
a
couple
of
interesting
things:
yeah,
let's
start
can.
A
I
add
a
note:
I
want
to
add
a
note
about
the
queue.
The
queue
is
a
composite
of
people
online
and
in
the
room
and
to
manage
the
queue
properly.
Everyone
should
be
using
the
tool,
so
even
if
you're
in
the
room
get
on
meet
Echo
and
as
you
come
into
the
mic,
hit
the
queue
button,
so
we
can
manage
the
queue
order
based
on
who
hit
it
first,
both
in
the
room
and
online.
B
B
Much
much
simpler
here
in
the
room,
there
is
a
note-taking
tool
right.
When
you
did
you
manage
to
Darling
that
your
eyes
doesn't
have
Wi-Fi,
so
that
sucks,
so
we
only
have
one
laptop
with
material.
The
other
one
has
only
the
Q
all
right.
How
do
we
not
taking
now
in
this
degraded
environment,
old-fashioned
sucks,
give
it
another
shot?
What
you
get
in
that
you
know.
B
All
right
there's
one
which
is
completely
open,
yeah
Legacy,
nobody
does
Legacy
except
me
and
the
eyes.
So
that's
why
it
works
all
right.
So,
let's
see
what
the
ice
gets,
the
note
taking
all
right.
So,
let's
start
so
I'll
be
flipping
those
things
and
then
we'll
see
how
the
stuff
works.
New,
tooling,
fine,
the
usual
note!
Well,
the
GJ
plus
everybody.
Please
wear
a
mask.
B
B
And
if
it's
not,
it
will
be
next
to
it,
if
I'm
sure
all
right.
So
everybody
please
signs
in
tooling,
changed
again.
Hoorah,
okay
and
please
mute
your
microphone
because
otherwise
you
know
we'll
have
looking
effects.
B
You
know
the
agenda
agenda.
Is
the
meat
Echo
I,
don't
even
know
whether
we're
running
meet
Echo
anymore
foreign
willing
to
do
the
meat
Echo
stuff,
I
thought.
So
no
so
we
have
no
taking
I
will
have
the
recording.
A
A
It
will
be
recorded,
but
what's
set
at
the
mic,
won't
be
so
if
someone
can
type
them
in.
That's
the
point
of
having
a
note
taker
to
put.
B
A
B
We're
not
having
any
technical
issues,
otherwise
you
know
what
to
report
them
good,
so
agenda,
a
bunch
of
updates
on
the
previous
stuff,
a
very
interesting
conversation
on
Ephesian
P4
based
beer
implementation
looks
actually
fascinating
on
the
Tofino
and
then
tallest
is
showing
something
towards
recursive
beat
masks
which
looked
like
a
pretty
crowded
pressure.
We
see
what
comes
out
of
it,
then
probably
an
frr
update
and
T
extension
by
Sandy.
If
we
find
time,
okay,
so
what's
the
status,
so
the
T
stuff
is
an
RFC.
Congratulations!
Yeah!
A
B
Room
Applause
room
Applause.
We
don't
get
that
many
of
those
we
need
to
know
cheap
wins.
The
beer
bar
IPA
is
also
RFC
now,
so
we
have
a
clean
architecture.
You
know
either
run
to
it
igp
or
otherwise
do
our
own
things
if
necessary,
Club.
A
B
Oh
thanks
for
Jeffrey
to
actually
hurting
that
one
along
for
quite
a
long
time.
All
right
now
comes
the
public
castigate.
The
ospv3
stuff
is
being
not
taken.
Care
of
I
was
twice
on
people
I
was
promised,
updates.
I
updates
not
delivered
one
of
the
code
that
is
sitting
here
and
looked
properly,
bedraggled
right.
So,
gentlemen,
I
give
you
last
warning
and
two
weeks
and
otherwise
I
start
to
go.
B
Look
for
another
Shepherd,
oh
I'm
sheparding,
but
the
stuff
as
far
as
he's
not
addressed
so
I
give
you
two
weeks
for
version
that
takes
care
of
the
stuff.
Otherwise
you
know
you
lose
the
opportunity
for
Eternal,
Fame
and
I
will
go
and
drag
another
author
and
make
him
editor
and
basically
start.
You
know
to
get
the
stuff
delivered
so
once
again,
last
call,
but
in
a
different
sense
a
bunch
of
things
are
moving
forward.
Nothing
particular
here:
Bunch
waiting
for
Shepherd.
B
Yeah,
nothing
very
interesting
here
unless
somebody
wants
to
come.
Yeah
Andrew.
B
B
Yeah
to
the
mic,
please
systems
yeah
right.
E
Yeah
Andrew
here
just
I
I,
don't
know
whether
this
has
been
circulated
to
all
the
working
groups,
but
we've
got
a
policy
in
routing
area.
Now
that
before
drafts
go
to
the
area
directors
after
last
call,
they
must
have
gone
through
an
early
review
through
the
directorate.
B
All
right
so
so
we
need
a
bunch
of
Shepherds,
for
you
know
earlier
reviews,
so
people
can
clean
the
stuff
up
before
we
push
it
further
least
extensive,
frr,
surely
interesting
everything
else.
Maticus
is
a
service.
Probably
interesting.
Te,
no
opinion
the
beer.
F
Okay,
Greg
mersky
Erickson
on
Saturday
the
pain
in
the
butt;
no.
F
Okay,
I'm
glad
to
so
dear
pink
beer
ping.
In
the
last
call,
we
received
Shepherd's
review.
Thank
you
for
the
shepherd
worked
under
addressing
I,
submitted
my
proposal
to
the
pen
holder
and
never
heard
since
so
the
draft
expired.
F
His
first
offer
right
so
I
would
appreciate,
if
chairs,
can
think
of
assigning
the
editor
to
the
draft,
so
we
can
move
on
because
the
problem,
the
problem
is
not.
The
problem
is
as
additional
as
you
listed
BFD
draft
here,
so
that
ping,
the
BFD
and
path
MTU
Discovery,
depend
on
progress
of
BRP.
B
A
We've
reached
out
to
nahendra
already
we
got
some
response,
but
nothing
really
progressed.
I
think
is
fair
enough
that
we
just
move
the
pen
yep.
B
Yeah
we'll
give
it
some
time
and
then
we
move
the
pan.
You
know
we
get
the
nature,
so
we
move
this
stuff
forward.
Okay,
anything
else
on
the
drop
shot
standing
on
the
status.
G
Yeah
Jeffrey
from
juniper
I
think
there
is
one
draft
missing
here,
probably
because
it's
expired
and
it's
my
fault,
it's
the
beer
proxy
draft.
It's
I!
We
believe
it's
ready
for
last
call,
but
I
will
refresh
it
and
then
refresh.
H
I
I
So
with
that
said,
if,
if
I
knew,
the
last
call
was
as
easy
as
going
to
the
microphone
I
wouldn't
face
anybody's
time
with
going
over
this
draft
again,
but
I
think
we
try
to
do
three
last
calls
on
this
and
I
think
it
went
to
the
working
group
too.
I
B
I
B
Show
of
hands
people
need
to
refresh.
Then
we
run
five
minutes.
Nobody
gets
hurt.
Okay,
one
two
yeah,
please
run
a
refresh.
I
Thing
is
about
I'll,
make
it
very
Snappy,
so
the
problem
we
are
trying
to
solve
is
that
some
providers
that
they
are
trying
to
go
next
slide.
Please
actually
some
providers
that
are
trying
to
go
to
the
next
to
the
next
technology,
which
is
beer
for
multicast.
They
want
to
replace
some
parts
of
their
Network,
so
they
buy
brand
new
hardware,
that
is
beer
capable
and
they
want
to
replace
one
portion
of
the
network
as
an
example.
I
In
this
case
the
network
is
being
multicast,
ldp,
mldp
and
the
small
island
within
that
network
is
going
to
get
upgraded
and
to
beer
and
with
this
draft
basically
what
they
can
do
is
they
can
I,
wouldn't
use
the
term
as
Stitch.
They
can
signal
the
mldp
to
the
Beer
core
from
one
Island
mldp
Ireland
to
another
mldp
Island
that
is
connected
through
this
vehicle.
I
So
so,
basically,
the
way
it's
done
is
like
PM
signaling.
All
that
happens
is
that
when
mldp
effects
arrive
at
a
beer,
Edge
router,
we
grabbed
that
mldp
FEC
and
the
mlbp
signaling.
If
you
will
and
we
use
RFC
7060
and
we
shoot
it
from
the.
Let
me
get
my
terminology
right:
the
the
down
a
stream
B
router,
which
is
Ingress
view,
router,
I,
guess
to
the
upper
stream
B
router
and
the
Upstream
B
router
keeps
track
of
all
these
Downstream
routers.
I
I
I
As
at
this
point,
and
then
your
data
path
starts,
data
path
starts
from
The
Source
via
mldp
VIA,
ldp
mpls
in
cap,
and
when
it
goes
to
the
to
the
Beer
domain,
we
have
a
list
of
the
Downstream
routers
that
are
interested
into
that
tree
and
we
generate
a
beer
header
that
has
the
bits
of
all
the
downstream
routers
in
the
beer
header
and
we
shoot
the
packet
all
the
way
through
the
beer
core
to
the
downstream
ldp
Island.
I
B
Okay,
last
call
request:
should
we
do
show
of
hands
or
take
it
to
the
list
crack.
B
G
All
right,
so
it's
this
is
update
on
this
bgp
extension
for
beer
signaling
on
behalf
of
the
co-authors.
Next
slide,
please
there's
a
bit
of
history
on
this
draft.
It
had
passed
working
group
last
call
before
and
I
was
the
shepherd
I
did
the
review.
I
had
some
comments,
then
the
main
authors
had
lost
Dean,
and
so
they
couldn't
address
the
comments.
G
And
additionally,
it
was
only
intended
to
for
this
bgpdc
scenario
with
bgp
being
the
underlay,
and
it
is
assumed
that
all
the
routers
would
support
beer,
and
if
there
is
one
router
or
that
does
not
support
beer,
then
the
Ingress
replication
will
happen
all
the
way
to
the
bfvrs.
G
G
So
after
considering
all
the
factors
factors
like
the
nobody
was
actually
working
on
comments,
and
then
we
have
this
problem,
so
we
decided
to
do
a
revamp,
so
I
joined
as
the
editor.
G
We
moved
to
move
the
solution
from
the
markets
as
a
service
strapped
into
this
this
one
too,
so
that
you
can
handle
generic
situations
and
not
just
for
these
pgp
DC
and
where
all
the
routers
support
beer
next
slide.
Please.
G
G
G
If
you
hadn't
known
that
bfr2
can
be
used,
then
you
could
have
just
tunnel
to
the
bfr2
only
with
the
new
approach,
all
the
bfrs
via
router,
when
they
re-advertise
the
beer
prefixes,
they
will
updates
the
beer
information
and
they
will
also
update
a
sub
a
next
hub,
sub
tlv
in
the
brtov
and
to
themselves.
So
this
has
two
purposes.
G
The
main
purpose
is
that
in
this
example
bfr2,
when
the
it
really
advertised
the
bfvr123
prefixes,
it
updates
next
hub
to
itself
so
now
in
bfr1,
guess
that
this
is
the
next
half
is
bfr
too,
and
it
has
the
encapsulation
information
and
associated
with
bfr
too.
So
it
knows
that
for
all
those
bfvr
one,
two
three
they
can
just
be
reached
from
bfr
too,
so
that
handles
the
non-bf
Rockets
very
well,
and
another
side
benefit.
Is
that
all
those
prefix
beer
prefixes?
G
So
next
next
slide.
Please.
G
So
here
bfaer123
originated,
there
are
an
RI
for
their
peer
prefixes
and
then
it
is
assumed
that
everyone
just
re-advertise
them
everywhere,
but
there
may
be
situations
where
you
do
not
want
to
re-advertise
the
Beer
prefixes
everywhere.
For
example,
bfr2
can
say
that
I'm
not
going
to
re-read
the
price,
those
bfvr
one,
two
three
prefixes
instead
bfr2
will
just
advertise
its
own
beer
prefixes,
but
it
uses
sub
POV
that
is
defined
in
the
other
draft
beer
prefix
either.
G
Excuse
me
re-advertisement
that
in
that
sub
new
sub
gov,
you
will
just
list
all
the
bfr
IDS
reachable
from
this
bfr.
So
that's
what
I
was
talking
about
on
that
slide
later,
let's
go
back
to
that
slide.
You
know
yeah
so
summary
of
the
all
the
changes
in
this
revision.
We
extended
the
scope
for
general
or
applicability.
G
G
So
with
these
changes
we
need
to
do
another
round
of
working
group
last
call.
We
also
need
a
new
shepher.
I
was
the
chefer
before,
but
now
I'm
joined
as
an
editor,
oh
by
the
way
Cindy
joined
as
a
contributor.
She
helped
a
lot
in,
especially
on
that
proxy
range.
Sub.
Tlb
is
a
support.
B
Questions
on
the
queue
nothing.
In
my
opinion,
this
needs
to
be
bounds
of
IDR
as
well.
Very
significant
change
of
procedure.
Okay
and
now,
because
you
carry
the
stuff
transitive
through
I-
would
like
to
have
these
guys
look
in
terms
of
like
Atomic
attributes,
Aggregates
blah
blah
right,
you're
triggering
a
couple
of
procedures,
so
I
will
fire
something
towards
the
IDR,
and
then
we
see
whether
anything
pops
up
on
the
list
and
then
find
Visual
work,
group
plus
call
again
and
Shepard,
please
if
anyone
volunteers
to
do
that
stuff.
B
B
Okay
done
next,
one
I
think
Jeffrey.
You
know
prefix,
okay,.
E
G
So
this
is
the
updates
on
that
ear.
Prefix
redistributes
traps
on
behalf
of
the
quarters
next
slide.
Please.
G
Just
a
quick
review,
we
have
three
igp
domains
in
this
topology,
even
though
they
are
running
three
different
igp
instances
that
beer
subdomain
itself
is
actually
just
a
single
one.
They
cover
all
those
domains,
questions.
How
do
we
registering
the
beer
information
between
them?
That's
one
question.
G
Another
question
is
that
if
we
are
using
pgplu
because
with
pcpr
you,
the
the
prefix
for
Edge
routers
are
typically
not
distributed
into
igp,
they
only
advertised
by
pgp
among
those
border
routers.
So
in
that
case,
how
are
going
to
do
peer
forwarding.
C
G
So
the
the
approach
there
is
the
to
have
the
Border
routers
advertise
their
own
beer
prefixes
with
that
same
proxy
wrench
subtlb
I
mentioned
earlier
to
list
the
external
bfr
IDs
that
are
reachable
about
them.
So
that's
the
basic
idea
on
how
to
handle
the
pgplu
case.
That's
basically
the
what
this
draft
is
about.
It's
been
there
quite
for
quite
some
time.
I
will
not
get
into
the
details.
G
A
We
got
last
call
show
of
hands
in
the
room.
B
A
The
list
we
do
it
in
the
room
and
I
we
have
a
tool
that
does
that
this
was
which
draft
again
it
vanished.
What
was
the
title
of
the
draft.
A
A
B
So,
let's
also
do
a
show
of
hand
for
the
the
bgp
last
call,
but
I
would
wait
until
we
get
this
IDR
feedback
and
then
I
would
only
go
for
the
last
call.
Maybe
next
meeting.
B
A
We
have
someone
with
the
mic
hope
he
vanished.
A
B
Back
to
the
agenda,
okay,
so
it
looks
like
the
bgp
is
going
last
call
as
well,
while
I
make
the
IDR
basically
checking
with
the
IDR,
what
you
call
it
integrated
part
of
the
last
call.
Okay.
B
A
B
H
H
Okay,
can
I
start
okay,
perfect,
hello,
everyone.
My
name
is
Stephanie
I'm,
a
PhD
student
at
the
University
of
tubing
and
today
I'm
going
to
talk
about
an
efficient
before
based
beam
Plantation
on
Tofino,
and
this
is
a
joint
work
with
my
colleagues,
Daniel
melling
and
Michelle
Metz
Next
Step,
please,
okay,
I'm,
going
to
start
my
talk
with
a
short
recap.
H
Next
slide,
please!
Okay!
So
at
ITF
108
we
already
presented
the
first
implementation
of
beer
and
bfr0
in
before
for
the
Intel
Tofino,
and
there
we
used
an
iterative
processing
approach
where
in
each
iteration
one
next
top
was
served.
So
we
received
a
beer
packet
in
the
first
pipeline
iteration.
We
forward
the
packet
to
the
first
next
top
of
the
pack
of
the
gear
package.
Then
we
cloned
the
packet
recirculated
in
the
second
iteration,
the
second
Next
Top
is
served
and
so
on
next
slide.
Please!
H
This
approach
has
one
huge
drawback,
because
we
make
heavy
use
of
recirculation
and
recirculation
requires
capacity
on
the
switch.
So,
for
example,
when
we
get
100
gigabit
per
second
multicast
traffic
with
tires
and
desktops,
then
this
results
in
400
400,
Giga
per
second
recirculation
traffic.
That
needs
to
be
processed
on
the
switch
and
when
we
look
at
the
chart
here
on
the
left
side
and
when
we
get
100,
gigabit
per
second
traffic
have
only
a
singular,
select,
recirculation
port
and
want
to
send
the
packet
to
three
Next
Tops
This
Is,
the
green
bar.
H
Then
we
only
achieve
around
40
gigabit
per
second,
because
the
pipeline
is
overloaded
due
to
recirculation,
so
we
have
heavy
packet
loss.
We
were
able
to
solve
this
issue
by
adding
dedicated
recirculation
parts
to
increase
the
recirculation
capacity,
so
with
three
recirculation
Parts
on
the
right
side
of
the
plot,
we
are
able
to
serve
up
to
four
Next
Tops
at
line
rate
with
100
gigaby
per
second.
However,
these
dedicated
recirculation
parts
are
physical
parts
of
the
device,
so
these
cannot
be
used
for
any
other
traffic
than
recirculated
beer
traffic.
Next
slide,
please.
H
Well,
this
idea
has
some
challenges
in
P4
and
on
the
internet
Tofino,
because
with
beer
bit
strings
with
at
least
256
bits,
it's
quite
difficult
to
map
this
bit
string
to
the
set
of
a
packet's.
Next
stop
in
a
single
table
like
look
up
when
we
do
not
have
specialized
Hardware
as
when
we
just
have
an
Intel
Tofino,
and
a
second
challenge
is
that
we
only
have
a
limited
number
of
configurable
static,
multicast
groups.
H
So
we
cannot
simply
have
any
combination
of
possible
next
stops
as
multicast
group,
because
it's
just
too
much
so
the
question
is
now:
how
can
we
do
this
more
efficiently
next
time?
Please
next
slide,
please
just
as
a
very
small
recap:
P4
is
a
highlighted
programming
language
to
describe
data
planes
and
there
we
have
a
compiler
that
Maps
the
P4
program
onto
programmable
pipelines
of
targets,
for
example.
In
our
case,
the
inter
tufino
next
slide,
please.
H
So
our
approach
for
more
efficient
B
implementation
is
now
the
following.
We
have
a
switch
with
a
certain
number
of
ports,
for
example
32,
and
now
we
divide
all
parts
of
that
switch
in
case
so-called
configured
Port
clusters,
for
example,
the
first
10
parts
are
the
first
configure
Port
cluster
here
with
the
yellow
border.
Another
10
parts
are
the
second
here
with
the
blue
border
and
the
remaining
12
ports
are
another
configure
Port
cluster.
H
These
configure
Port
clusters
can
be
overlapping,
so
they
do
not
have
to
be
disjoint
and,
of
course,
ports
within
a
configured
podcast
that
do
not
have
to
be
physically
next
to
each
other.
This
is
just
for
Simplicity
and
all
configured
part
classes
together
have
to
cover
all
parts
of
the
switch,
and
now
we
want
to
determine
which
configure
Port
clusters
require
a
packet
copy.
So
we
get
a
b,
a
packet
that,
for
example,
should
be
sent
through
the
first
part
of
the
switch
and
through
the
last
one.
H
So
we
need
the
first
configured
part
cluster
and
the
third
configure
part
cluster,
and
this
may
be
done
again,
possibly
iteratively.
However,
only
with
a
few
iterations
like
one
to
four
and
not
like
1
to
32,
as
in
our
previous
implementation,
and
when
we
now
know
that,
for
example,
C1
needs
a
packet
copy.
We
can
use
a
suitable
static,
multicast
group
to
replicate
the
packet
to
all
required
Next
Tops
within
this
a
configure
Port
cluster,
but
only
towards
the
required
Next
Tops.
H
And
with
this
approach
we
need
a
maximum
number
of
K
minus
1
recirculations
for
K
configure
plot
clusters.
So,
for
example,
with
the
32
Port
switch
and
three
configured
Port
clusters
for
size,
11,
11
and
10,
then
we
need
at
most
two
recirculations
independent
on
the
number
of
next
stops
and
in
our
last
implementation
this
was
next
house
minus
one.
So
this
is
a
drastic
reduction
in
the
number
of
risk
conditions.
Next
slide.
H
Please,
and
the
question
is
now
okay:
how
can
we
determine
which
configure
Port
clusters
need
a
package
and
to
that
end
we
look
at
all
combinations
of
these
configure
Port
clusters
here
called
SI
with
our
example,
with
three
configure
Port
clusters.
We
have
seven
combinations,
a
single
configured,
part
clusters,
all
combinations
of
two
and
all
combinations
of
three,
and
then
we
look
at
the
so-called
combined
fbm,
which
indicates
all
beef
ERS
that
are
reachable
through
the
subset.
H
H
So
all
parts
together
build
also
the
fbm
of
the
configure
Port
cluster
and
when
we
now
have
a
subset
of
configured
Port
cluster,
we
just
bitwise
or
the
individual
fpms
of
the
configure
plug
testers
and
when
we
now
receive
a
b,
a
bit
string
and
which
is
here
shown
in
the
example
with
beef
ERS
that
are
reachable
through
the
first
configured
Port
cluster.
Here
in
Orange
and,
for
example,
the
second
configure
Port
cluster
here
in
blue,
then
we
apply
a
ternary
match.
H
Operation
ternary
match
is
simply
a
bit
wise
and
operation,
and
we
do
it
with
the
complement
of
the
combined
fbm
and
we
Define
a
match
when
the
result
is
zero.
What
does
this
mean?
This
means
that
the
complement
of
the
cspn
has
a
zero
at
every
bit
position
where
the
BRB
string
has
has
an
activated
bit,
and
this
means
that
the
combined
fpm
has
a
one
at
every
bit
position
where
the
build
string
has
bought
an
activated
bit,
which
simply
means
that
the
cfbm
serves
all
pvrs
of
video
package.
H
In
this
example.
This
would
be
the
case
for
S4
and
S7,
because
these
two
contain
the
first
and
second
configure
part
cluster.
However,
S7
also
contains
the
third.
We
do
not
want
to
send
super
application
to
the
third
content
import
cluster
because
it
does
need
a
packet.
H
H
Then
we
recirculate
and
the
second
step
we
can
work
with
the
remaining
control
clusters
that
need
a
packet
copy.
Next
slide,
please,
and
now
that
we
know
which
configure
thought
clusters
need
a
packet
copy.
We
only
have
to
determine
an
appropriate
multicast
group
so
that
we
can
replicate
the
packet
to
all
Next
Tops
within
this
configure
Port
cluster
in
one
shot
without
research
collection.
How
do
we
do
that?
First,
we
configure
all
combinations
of
paths
within
configure
Port
clusters
as
static
multicast
groups.
H
We
do
that
to
gain
some
robustness
against
changes
in
the
magic
history,
for
example,
when
in
DVR
joins
police,
and
then
we
apply
the
same
logic
as
before.
We
have
a
combined
fbm
of
the
static
multicast
Group,
which
indicates
all
the
vrs
that
are
reachable
through
this
group,
or
rather
through
the
parts
that
are
part
of
this
static.
Multicast
group.
Again
we
apply
the
ternary
match
operation
on
the
pivot
string
and
the
complement
of
the
csvm
when
the
match
is
zero.
H
We
need
around
5120
static,
multicast
groups,
and
this
is
well
feasible
on
a
device
like
the
Intel
Tofino.
We
implemented
this
approach.
It
runs
at
100,
Cubit
per
second
per
part
on
the
Intel
toofino
next
slide.
Please,
this
approach
can
be
further
optimized
as
follows:
next
slide,
please
we
have
certain
assumptions
about
multicast
traffic.
We
believe
that
multicast
traffic
is
not
completely
random.
We
mean
by
that
that
there
is
some
correlation
on
eager
spots.
H
The
second
assumption
is
that
we
only
have
a
limited
number
of
static
multicast
groups,
so
we
have
to
comply
to
yourself
maximum
number
that
we
want
to
use,
and
our
optimization
idea
is
now
the
following:
we
sample
beer,
packets,
locally
and
based
on
the
beer
bit
string.
We
know
which
parts
have
to
be
used
locally
on
the
device
to
forward
the
package.
This
information
is
stored
in
a
graph
structure,
and
then
we
can
apply
clustering
methods
to
choose
appropriate
configured.
H
Port
clusters,
such
you
know
in
a
way
such
that
all
of
them
together
cover
all
parts
of
the
switch
and
that
part
that's
statistically
and
occur
more
likely
to
get
them
via
package
have
to
be
used
together,
are
in
the
same
configured
part
cluster,
so
that
we
can
use
a
single
multicast
group
to
serve
all
of
them
and
in
the
end,
this
means
that
we
need
less
configure
Port
clusters
per
via
packet
on
average,
and
this
reduces
the
required
number
of
recirculations
next
slide.
H
Please
here's
a
small
example
and
when
we
get
via
traffic
with
around
4.5
Next
Tops
on
average,
in
our
old
implementation,
this
required
3.5
recirculation
leverage,
which
is
the
Red
Bar
in
the
graphic
here
and
in
our
new
approach
with
the
optimization.
H
When
we
have
highly
correlated
multicast
traffic,
which
means
that
there's
a
high
correlation
between
equals
parts
and
beer
packets,
then
we
need
almost
zero
rest
circulations
with
around
1000
static,
multicast
groups,
that's
the
bar
on
the
right
and
when
we
have
less
correlated
multicast
traffic,
which
means
that
there's
some
Randomness
within
the
use
of
eager
spots
for
beer
packets,
then
we
still
need
around
a
scenery
circulation
with
1000
static,
multicast
groups
instead
of
the
3.5
recirculations.
Now
about
approach
next
slide,
please!
H
So
to
conclude
this
talk.
We
presented
the
mechanism
to
forward
beer
traffic
with
only
a
few
iterations
instead
of
the
number
of
next
talks,
and
to
that
end
we
choose
so-called
configure
podcasters
and
then
apply
an
iterative
processing
where,
in
each
iteration
we
select
the
configure
Port
class
that
it
needs
a
packet
copy,
and
then
we
send
a
packet
to
all
required
Parts
within
this
configure
Port
cluster
in
one
shot,
the
Second
Step
may
be
achieved
through
predefined
internal
multicast
groups,
as
in
the
case
with
Intel
Tofino.
However,
these
multicast
groups
are
static.
H
They
are
not
Dynamic,
so
we
do
not
need
any
Dynamic
state,
and
this
may
also
be
done
differently
depending
on
the
technology.
So
when
there's
another
replication
mechanism,
that's
more
efficient,
then
multicast
groups.
This
can
be
also
used
with
this
mechanism.
We
have
a
maximum
number
of
iterations,
which
is
bounded
by
the
maximum
number
of
block
clusters,
for
example
three
instead
of
32,
and
we
can
even
further
optimize
this
approach
by
packet,
sampling
and
then
applying
machine
learning
and
to
get
appropriate
configured
podcasters.
H
B
B
Would
have
yeah
Greg.
B
So
I
have
one
question.
The
rest
was
pretty
clear
when
I
look
at
the
picture.
H
Because
maybe
there
are
certain
multicast
groups
or
certain
number
of
peer
packets
that
have
a
subset
in
the
yellow
cluster
configurable
part
cluster.
But
there
are
also
some
beer
packets
that
need
parts
of
the
second
one
and,
as
we
configure
all
possible
configuration
combinations
of
Parts,
when
we
have
disjoint
configure
Park
clusters,
we
need
at
least
one
recirculation
when
we
need
to
have
a
packet
that
needs
parts
of
both
clusters
and
when
we
can
do
overlapping
clusters,
then
there's
the
chance
that
we
need
no
rest
circulation
at
all.
B
G
The
mic
go
ahead,
Jeffrey
from
juniper,
so
if
I
understand
this
correctly,
the
each
cluster
is
kind
of
like
a
line
card.
It's
on
some
routers.
G
Another
thing
is
the
so
inside
each
cluster.
It
seems
that
your
you
may
be
doing
like
a
string,
P
string,
lookup
on
the
entirety
or
just
on
these
bits
for
that
cluster.
Only
and
both
ideas
have
been
documented
or
in
some
in
some
scenarios.
G
In
some
companies
internal
implementations,
we
can
discuss
further
offline
and
additionally,
I
guess
it's
really
depending
on
the
implementation.
I
know.
There's
some
implementation
do
not
require
a
recirculation
for
each
iteration
of
that
lookup
that
that
bit
in
the
bit
string,
so
indeed
I
guess
for
for
some
Hardware
each
iteration.
What
that
does
require
re-segration-
and
in
that
case
this
optimization
will
help.
H
Yeah
definitely
so
we
are
a
bit
Limited
in
the
number
of
table,
lookups
that
we
can
do
before
and
on
the
Tofino,
and
so
we
do
not
have
a
way
to
make
30
lookups
in
one
iteration,
for
example.
So
we
have
to
do
these
recirculations,
and
this
is
a
way
to
optimize
how
we
do
these
three
circulations,
but
I'm
really
interested
in
more
information
about
the
line
card,
capabilities
of
existing
B
implementations.
H
J
Yeah
I
just
had
a
quick
question.
You
know
on
the
so
P4
I
think
one
of
the
major
advantages
of
P4
does
help
with
the
pipeline
programming
and
staging
the
ability
to
eliminate
recirculation
that
happens
on
the
on
the
Asic.
J
You
know
by
doing
just
the
recirculate,
you
know
the
programming
with
P4.
Are
you
able
to
eliminate
the
pipelining
or
eliminate
recirculation
I
think
have
you
have
you
tried
without
having
to
deal
with
the
randomness
I?
Think
like
you
had
to
in
in
your
one
of
your
last
slides,
you
mentioned
that
you're
working
on
how
to
the
optimization.
J
So
if
you,
without
the
optimization
and
if
you
just
strictly
did
everything
just
with
P4
and
and
and
with
the
pipeline
programming,
were
you
able
to
get
to
a
point
where
you
can
eliminate
recirculation
without
having
the
optimization
I.
H
Think
well,
yeah.
Thank
you.
Thank
you
for
the
question.
Of
course,
you
could
do
it
all
without
recirculation
by
just
copying
the
packet
to
each
egress
Port
that
has
a
potential
receiver
and
then
dropping
the
packet
in
the
egress
when
the
receiver
is
not
needed.
But
in
this
way
you
will
block
the
response.
So
when
you
get
100
Gig
and
each
Port
has
a
receiver
at
the
end
and
we
make
a
packet
copy
to
each
Port,
then
the
whole
switch
is
full
and
we
cannot
have
any
other
traffic.
H
B
H
It's
not
a
multicast
address,
it's
an
internal
way
of
telling
the
Tofino
to
do
a
packet
replication
to
certain
parts,
and
basically
we
tell
it.
The
control
plane
give
us
a
multicast
group
for
these
parts.
Then
we
get
a
number
and,
based
on
this
number,
we
can
later
tell
the
Tofino
to
replicate
the
packet
to
this
port
and
we
have
up
to
2
to
the
power
of
16
multicast
groups
available
on
the
Tofino
and
on
our
example
on
one
of
the
slides.
With
these
three
configurable
podcasters,
we
used
five
thousand
of
them
so.
B
Yeah,
okay,.
H
A
I
have
a
question
so
when
you
say
internal
multicast
group,
these
are
actually
group
addresses,
like
ipv4
multicast
addressing
is
built
into
the
structure
of
Tofino.
Or
what
do
you
mean?
No.
H
B
K
All
right
so
I
had
this
earlier
version.
I
think
two
years
ago,
so
be
back
here
now
next
slide
coming.
K
L
So
we
continue
to
think
that
traffic
engineering
is
really
very
important
right.
So
we've
got
brt
out
and
we're
getting
back
at
that
net
to
all
the
good
things
where
we
could
actually
for
for
multicast
use,
brte
and
I'm
not
going
to
to
read
through
the
long
list,
I
I
think
most
people,
hopefully
should
know
them
and
then
through
working
through
brte
I,
think
we
got
to
the
point
where
I
was
happy.
L
If
it
was
working
in
in
small
scale,
and
then
everybody
was
asking
me
about
large
scale
and
to
make
it
larger
and
everything.
And
so
then
you
know
the
ideas
it
came
around
on
how
what
the
root
cause
of
the
problem
is,
and
it's
really
what
I
would
call
now
the
flat
bit
strings
right.
L
So
brt
was
designed
to
be
as
much
as
possible
using
the
same
hardware
for
wooding
as
beer,
which
means
one
single
flat
bit
string,
and
now
we
think
that
we
can
do
better
when
we
are
taking
into
account
that
we're
10
years
after
the
original
redesign,
our
Hardware
forwarding
planes
hopefully
can
do
better,
and
we
think
that
we
can
encode
trees
more
efficient
than
brte,
make
larger
headers
easier
to
process
and
make
it
easier
to
Auto
configure
next
slide,
and
this
is
by
the
way,
highly
experimental
right.
L
L
It
is
simply
updating
the
offset
and
the
links
to
point
to
the
downstream
neighbors
what
we
call
recursive
unit,
which
is
the
bit
string
and
then,
of
course,
all
the
downstream
neighbors
bit
strings
themselves
next
slide.
This
is,
of
course,
very
high
level.
L
L
One
two
router
B
one
for
router
c,
and
to
do
that
it
examines
only
its
own
bit
string
then
for
the
serialization
we
need
to
have
the
length
of
the
subtrees
for
B
and
C.
One
is
actually
derived
from
reading
the
ru
links.
B
L
No,
this
is,
this
is
arbitrarily
many,
so
in
this
case
we're
just
showing
two
bits,
so
it
can
be
many
bits
and
the
the
AF
field
is
an
array.
Okay,.
L
Example
with
two
so
two
AF
entries.
L
This
is
an
area
and
we're
saving
one
area
element,
because
we
can
calculate
the
length
of
the
last
recursive
unit
because
we
can
take
the
global
Ru
links
field
and
do
the
calculation
that
is
in
that
field.
So
obviously
you
know
this.
This
takes
longer
than
the
the
first
slide
deck.
L
Hopefully,
the
visuals
are
somewhat
helpful,
but
yeah
I
think
we'll
take
more
time
to
breathe
in,
but
the
the
main
point
is:
every
router
only
needs
to
look
at
its
own
bit
string
and
that
bit
string
only
needs
as
many
bits
as
that
router
has
interfaces
so
to
speak
or
host
neighbors
next
slide
right.
So
this
is
a
little
bit
different,
a
view
of
the
same
thing
right.
Obviously,
this
could
replace
the
bitstring
field
of
an
RFC
8296
packet
header
and
then
the
cost
of
the
serialization
right.
L
We
love
the
fact
that
we
only
got
every
router's
bit
string
that
we
need
not
the
bit
strings
from
any
router
where
the
tree
doesn't
go
through,
but
the
cost,
of
course,
are
exactly
these
addressing
fields
that
are
telling
us
the
length
of
the
subtrees
and
then
the
global
Ru
links,
and
are
you
offset?
So
this
is
kind
of
the
overhead
for
the
savings
that
we
get
next
slide,
and
so
we
did
some
simulation
some
calculation
for
large-scale
Network.
L
So
this
is
I,
think
Beijing
Mobile
in
in
China
with
a
couple
of
hops
right,
so
every
hop
is
individually
in
there
and
then
this
comparison
actually
not
even
against
brt,
but
against
beer
and
the
total
number
of
replication
needed
to
reach
a
certain
random
subset
of
receivers
and
then
going
through
all
possible
random
subset
of
receivers,
starting
from
a
small
number
going
to
a
very
large
number.
L
B
B
L
Colleagues
and
Beijing
have
done
right,
so
obviously
the
more
we
get
into
that
assessment.
The
more
we'll
we'll
have
to
revisit
these
numbers
and
and
look
for
different
topologies
as
well.
Yeah
next
slide.
B
L
Okay,
so
the
three
big
things
right,
so
replication,
efficiency
for
small
trees
and
large
networks
right
so
and
it's
easier
to
make
the
comparison
with
beer,
because
brte
with
all
these
different
optimizations
is
it's
very
difficult
right.
Imagine
you
have
a
network
with
10
000,
bfer
and
you've
just
got
a
bit
string
length
of
256
right
so
even
replication
to
a
small
set
of
10,
20,
30
or
40
random
destinations.
L
A
large
Network
right
across
that
Network
may
require
up
to
40
packet,
copies
in
BR
and
brte,
because
they're
all
in
different
bits
drinks
right.
So
you
have
20
256
bits,
so
that's
40
different
bit
strings.
So
even
for
small
trees,
you
may
need
a
large
number
of
copies
right,
and
so
why
do
we
care
about
small
trees
and
large
networks?
Well,
that's
what
we
did
started
in
1994
right,
which
we
called
it
Pim
sparse
mode
right.
L
We
want
to
have
efficient
delivery,
small
trees
in
large
networks,
and
that
hasn't
changed
just
because
we
want
to
go
stateless.
So
that's
why
small
trees
are
actually
quite
interesting
next
slide.
So
then
we
had
this
fun
discussion
with
Dino.
He
isn't
even
here
so
that
was
basically
I
was
trying
to
be
really
provocative
and
that's
the
red
warning
offensive
technology,
Right.
L
So
consider
really
why
don't
we
have
bit
strings
that?
Are
you
know,
10
000
bits
long
right,
what's
bad
about
it,
so
and
then
go
through
the
reason
why
it
does
or
does
not
work
right.
So
in
the
first
place,
I
think
it
would
work
perfectly
fine
from
the
benefit
side.
Okay,
it's
one
kilobyte
worth
of
bit
strings,
one
kilobyte
worth
of
data,
but
we
still
only
have
an
overhead
of
50
we're
sending
it
to
a
large
number
of
bits,
10
000
receivers,
and
we
have
an
overhead
of
two
times
right.
L
If
we
do
Ingress
replication,
it's
an
overhead
of
ten
thousand
or
right.
So
if
we
go
Pim
sparse
mode
sure
we
only
have
an
overhead
of
one,
but
then
all
you
know
the
stateful
operation.
So
why
didn't
we
like
large
headers
up
so
far?
I,
don't
think
the
MTU
is
the
issue
right,
one
thousand
byte,
more
MTU!
If
we
start
with
data
center,
I've
I've
worked
with
so
many
crazy
networks
with
really
you
know
gigantic
packet
sizes
right
I,
don't
even
think
it
is
long-term.
L
The
accessibility
of
the
header
right
we've
started
with
header
from
128
byte.
Now
we're
at
512
byte
would
say
we
can
access
one
kilobyte
of
of
header
in
the
forwarding
plane
in
10
years,
but
the
problem
that
we
have
is
we
don't
want
to
process
10
000
bits
worth
of
you
know
stuff
when
we
only
need
to
replicate
to
10
outgoing
interfaces
right.
L
If
we
only
want
to
replicate
to
10
outgoing
interfaces,
we
should
only
have
to
look
and
process
something
that
is
proportional
to
the
number
of
replications
we
really
can
do,
and
that
is
exactly
what
RBS
gives
us
right.
We're
only
looking
always
at
a
local
bit
string
and
all
the
other
header
is
just
payload
that
you
know
some
other
router
down.
The
tree
will
look
into
next
slide.
B
L
No
here
so
Auto
configuration
right
so
I
mean
I
started,
brte
I'm,
not
really
a
fan
of
centralized
controllers
right,
I,
love,
traffic
engineering,
but
I
would
love
to
have
you
know
as
lightweight
as
possible
traffic
engineering
in
the
source
router
in
the
first
top
router
doing
explicit
Source
route
and
I
don't
want
to
have
to
configure
new
address
basis
on
every
node
and
with
this
mechanism
RBS
we
really
only
need
to
have
every
router
locally
assigning
a
number
from
1
to
20
for
all
the
20
bits
of
its
interfaces
and
don't
have
to
have
global
address
space
like
the
bif-t
ID
that
we
need
to
do
sorry,
bfr,
ID,
that
we
need
to
do
now
in
beer
and
beer
to
e
to
basically
split
up.
L
You
know
the
identifiers
of
routers
across
the
different
SI
right,
so
that
problem
goes
away
and
I
think
we
could,
with
this
mechanism,
also
get
back
to
something.
That
is
a
lot
more,
if
not
fully
auto
configuring.
So
next
slide
yeah
so
scale
right.
That
primarily
comes
because
we're
not
doing
this
hard
subdivision
into
fixed
256,
bit
streaks
or
larger
right.
So
we
are
not
wasting
any
bits
that
we're
having
in
the
bit
strings.
L
Right
so
we're
looking
for
interest
feedback
collaboration
right.
Is
this
easy
enough
packet
processing?
Obviously
you
know
we
have
the
plus.
We
have
the
minus.
We
want
to
continue
evaluating
and
improving
the
approach.
L
Scalability
comparison,
implementation,
feasibility,
such
as
MP4,
you
saw
one
of
the
co-authors
being
Michael
right,
so
we're
working
on
that
together
and
then,
of
course,
there
may
be
other
proposals
that
are
looking
beyond
the
flat
bit
strings
that
we've
done
and
that
hopefully
we're
starting
to
successfully
deploy
for
the
Next
Generation
beyond
that
right
and
then,
ultimately,
you
know
when
we
feel
that
we
have
something
that
enough
people
feel
could
work
right.
What
are
the
criteria
that
would
make
any
such
non-flat
bit-free
Solutions
acceptable
as
beer
working
group
right?
Obviously
rechargering.
L
We
had
to
do
that
all
the
time,
but
what
are
the
criteria
right?
What
are
the
bounds
of
what
beer
ultimately
would
want
to
do?
I'm
not
asking
to
answer
that
question
now,
but
I
think
it
would
be
really
good
to
have
an
answer.
You
know
within
the
next
half
a
year
or
a
year
so
that
we
know
where
to
put
what
works,
because
the
more
we
get
working
group
support
and
people
interested
I
think
the
better.
The
result
will
be.
A
Okay,
we
got
queue
filled
up
here.
I'm
just
going
to
give
you
comments.
First
from
the
chat
from
Edward
V
saying
MTU
is
I,
think
is
he
meant
smaller
problem?
Much
bigger
problem
is
how
big
header
A6
May
process
it
may
be
as
low
as
128
bytes.
L
You've
seen
we've
done
the
comparison
for
256
bits,
which
is
by
itself
less
than
128
byte.
But
of
course
we
also
have
other
header
overhead,
so
yeah.
Let's
do
the
comparison
there
as
well.
Of
course,.
G
So,
if
you're
down
from
juniper,
so
when
PRT
was
first
brought
up,
I
I
was
not
believing
and
I
was
in
doubts
because
I
didn't
see
it
would
scale.
But
when
this
RBS
was
brought
up,
I
liked
it
and
I
I
can
clearly
see
the
advantage
compared
to
brt.
So
this
is
a
good
thing.
G
There
are
some
benefits
on
the
slide
you
claimed.
I
will
need
to
understand
it
better,
but
the
basic
idea
of
using
the
recursive
structure
and
to
save
the
the
overhead
using
local
bits
a
bit
Precision
that
that's
the
key
I
I.
It
does
work
and
I
like
it
and
once
a
suggestion
in
the
name
itself
is
to
me
it's
really
just
beer
RBS,
don't
call
it
cgm2.
Oh.
B
Okay,
so
speaking
purely
as
participant,
okay,
I,
like
that
I'm
surprised
by
this
256
bits
that
result
is
so
good.
So
you
know
trust
is
good.
I.
B
See
that
verified
independently,
because
the
gains
are
impressive.
It
looks
simple
enough
to
process
in
the
hardware.
There
are,
of
course,
open
things
like
how
do
you
make
sure
that
everybody
computes
the
same
tree?
What
will
happen
while
the
tree
is
converging
like
Loop
prevention
right,
the
the
usual
discussions?
What
will
happen
there?
Nothing,
because
each
node
has
to
kind
of
see
the
same
tree
and
know
the
offset?
Who
is
the
next
note?
B
B
This
will
be
a
fundamentally
different
way
to
process
the
bit
mask.
Yeah
me
likes,
okay,
that's
what
I
can
say.
It's
I
think
it's
a
great
proposal.
If,
if
the
numbers-
and
then
you
know
these
little
things-
that
we
don't
see
to
work
out
and
we
yeah.
L
Okay,
my
biggest
challenge,
of
course,
is
so
we've
done.
You
know
an
evaluation
of
course
in
actual
Hardware,
which
is
you
know,
like
everybody
else
here,
vendor
proprietary
Hardware,
that
researchers
can't
verify.
So
that's
why
Michael
is
doing
it
on
P4
and
P4.
Has
problems
I'm,
not
sure
how
much
that
is
going
to
tell
us
right,
so
the
decisions
of
what
is
actually
feasible
everywhere,
I
think
in
our
space
is
really
a
difficult
one
to
do,
and
I
hope
that
we
can
all
try
to
do
that.
You
know
as
good
as
possible.
B
L
B
C
No
I
say
so:
yeah
I
also
find
it
a
very
Nifty
approach.
So
it's
quite
nice,
although
so
I
saw
my
comments
to
the.
C
Yeah
because
I
think
there's
some
design
choices
made
to
not
put
the
more
state
in
the
packet,
but
put
it
in
the
control
plane
where
it
doesn't
belong
so
in
my
I
do
expect
actually
the
the
header
size
to
increase.
If
you
want
to
solve
that
problem.
So
the
data
that
you're
showing
here
for
this
256
bits
might
not
be
as
good
as
it
slips
two
days.
Yes,
definitely
foreign.
A
L
C
How
about,
but
if
we
support
looking
into
this
more
for
sure,
yeah.
J
Yes,
just
a
few
comments,
so
I
I
really
like
the
idea,
you
know
it's,
it
seems
like
an
optimization.
It
kind
of
takes
some
of
what
exists
today,
I
guess
in
Optimus
ideas
of
optimization
and
it
seems
like
a
really
a
good
Improvement
and
works
with.
As
you
said,
small
trees
are
important,
Universal
Big
Trees,
but
it
really
is
a
is
a
real
nice
optimization
to
severe.
J
L
I
mean
I
described
how
it
would
fit
into
an
80
296
header
right.
So
obviously
we
only
have
power
of
two
length
of
the
bit
string
right
now,
so
we
would
need
to
do
some
padding
and
ultimately
we
could
use
all
the
different
lengths
depending
if,
if
the
hardware
does
it
right
so
yeah
it
could
be
fit
in
there.
Is
it
the
best
header
but
I,
don't
think
we
can
save
a
lot
by
by
trying
to
optimize
that
header
just
for
you
know,
saving
every
single
bit.
B
Well,
I
mean
so
there
was
a
strong,
architectural
and
practical
consideration
to
put
the
Beet
mask
at
the
very
end.
Right
so
hence
I
mean
the
power
of
two
is
artifact,
because
the
algorithms
to
look
up
look
up
so
good,
but
there
is
nothing
that
will
you
know,
put
a
crimp
into
our
whole
thing.
It
would
make
the
big
mess
different
length,
so
you
know
we'll
have
to
think
how
to
record
it
in
terms
of
the
beer
and
coating
that
we
have
today,
we
have
a
version
on
it.
B
We
can
always
bump
it
to
version
two
or
we
can
think
of.
You
know
how
we
semantically
distinguish
just
like
with
the
brke.
How
do
we
semantically
interpret
the
bit
mask
differently,
so
I?
Don't
think
or
it
actually
anything
would
prevent
us
to
go
and
have
yet
another
semantic
interpretation
of
the
bead
mask
and
all
the
existing
Hardware
right,
yeah
yeah.
L
B
Have
in
this
slightly
whatever
I
mean
the
control
plane,
we
can
go
wild
right
so
that
we
can
synchronize
who
assigns
which
semantics
you
know
to
to
what
kind
of
bitmaps
is
long
in
the
hardware.
We
don't
start
to
pile
stuff
up,
so
yeah,
I,
fundamentally
I,
don't
see
like
architect,
actually
that
wouldn't
feed
and
then
Coatings
we
have
and
the
architectural
degrees
of
freedoms
that
we
provided
yeah.
B
B
Here
we
are
so
what
else
is
on
the
tank?
We
still
have
to
be
a
frr
update,
yeah,
we
can
run
through
it
quickly,
I
mean
otherwise.
People
can
just
you
know,
Fade
Out,
foreign.
M
M
So
in
addition
to
that,
and
then
there's
some
people
mentioned
that
also
the
relationship
between
bfr
and
ipfr.
So
we
just
we
dressed
next
page.
M
M
Ipfr,
so
one
question
is
that
we
already
have
an
ipfri:
why
do
we
need
a
bfr?
This
is
because
beer
package
is
a
forward
without
IB
header
I
mean
outside
header.
So
in
this
case,
ipfr
cannot
apply
to
the
Beer
package
because
it
don't
have
to
be
a
Target
that
doesn't
have
a
IP
out
headers.
So
in
this
case,
when
one
beer,
that's
hot,
failed
or
unreachable
and
then
be
a
package
will
be
dropped
and
then
lost.
So
that's
why
we
need
bfr
so
regarding
to
beer,
LFA
and
the
IPL
fa.
M
So
for
for
IPL,
we
have
different
types
and
we
have
basic
IPL
Alpha
and
we
have
remote
one
and
we
have
a
property
independent
LFA.
So
those
are
three
types
will
also
reused,
because
we
have
a
architecture
to
support
normal
action,
which
is
just
forwarding
the
package
to
the
laptop
and
then
we
also
have
a
tunnel,
and
then
we
have
also
you
first.
So
what
is
a
covered?
So
next
page.
M
So
we
just
want
to
welcome
work
comments
and
the
suggestions.
I
Yeah,
who
made
me
goalie
Nokia
I,
mean
yeah
great
idea,
but
I'm
still
scratching
my
head.
Usually
the
deployments
that
we're
seeing
with
beer
unicast
is
enabled
so
most
of
the
deployments
they
want
to
follow
unicast,
so
I'm
still
kind
of
scratching
my
head.
Why
do
I
want
to
separate
my
beer
faster
you're
out
from
unigas.
M
I
I
D
I
Is
in
the
mercy
of
the
igp
and
what
what
we
have
realized
I
mean
our
implementation.
Is
you
just
turn
on
LFA
in
the
network
and
beer
just
follows
it,
there's
nothing
special
about
it
that
that's
how
the
input
like,
even
from
implementation
point
of
view.
We
didn't
have
to
do
marriage.
We
just
turned
on
the
LFA
under
your
igp,
and
you
just
follows
it.
M
As
long
as,
if
you
have
one
but
that's
I,
think
the
yeah
normally
is
okay,
but
sometimes
because
some
notes
from
the
support
appear.
In
that
case,
we
may
have
a
issue
if
we
ever
know
the
support
that
one,
it's
okay,
yeah
every
ID,
every
IP
LFA
is
also
p
l
of
a
and
then
we'll
watch
that
go
forward
right.
I
B
Jumping
quickly,
I
mean
we
had
kind
of
this
discussion,
not
ongoing
discussion,
but
to
keep
everybody
open-minded
right.
Yes,
today
we
deploy
the
right,
GPS
or
unicast
or
works
like
a
charm
but
fundamentally
or
it
actually
were
not
slave
to
igp
right,
we
can
run
completely
different
computation
run,
build
completely
forwarding
table
installed
via
controllers
if
we're
not
slave
to
igp
and
igpfrr
is
not
good
enough.
B
I
Well,
that
okay,
again
I
just
want
to
be
very
careful
here,
even
with
the
controller
like
if
you
have
Sr
policy
or
anything
that
comes
down
with
the
controller
again
from
what
we
see
from
the
implementation.
It
just
follows:
what
is
installed
in
the
data
path?
Right
I
mean
we
don't
see
anything
as
specific,
but
yeah
I
mean
I'll.
I
B
Okay
I
mean
this
discussion
has
been
going.
You
know
that
was
the
first
comment
given
to
the
draft
okay.
So
thanks
for
the
update
right,
something
else
chasing
me,
but
it's
okay
and
we
come
to
the
very
last
one
and
we
weigh
over
time.
Okay,.
D
Okay,
I'd
like
to
say
this
is
a
very
simple
draft-
is
used
for
prte
VP,
a
beefed
ID
at
what
husband
in
in
brt
in
calculate
circulation.
Sorry
because
the
previous
slides,
the
previous
structure
is
no
are
not
opting
optimal,
so
we,
you
would
like
to
use
a
new
advertisement
for
the
gift
ID
advertisement.
D
So
the
only
change
for
the
new
draft
is
that
we
move
the
Deep
ID
advertisement
from
the
DP
to
the
prefix.
So
we
just
only
Edward
has
the
gift.
Id
was
so.
This
can
reduce
the
advertisement
and
bytes
consume
consumption,
greatly
yeah.
So
that's
very
simple!
Oh
sorry,
please
next.
D
D
D
So
we
have
the
the
previous
structure
is
the
word
highest,
the
the
proof,
the
Edward
husband
in
previous
structure
is
and
what
has
the
gift
ID
under
BP
the
POV?
So
we
just
moved
the
operation
to
the
prefix,
so
the
Edward
husband
may
be
more
efficient,
so
the
if
the
ID
is
located
per
SD
BSL.
B
A
One
Tony's
got
to
pay
for
beer
for
everybody
tonight.