►
From YouTube: IETF104-BABEL-20190328-0900
Description
BABEL meeting session at IETF104
2019/03/28 0900
https://datatracker.ietf.org/meeting/104/proceedings/
A
B
D
A
Okay,
okay,
so
we'll
go
ahead
and
get
started
here.
This
is
the
table
or
Babel
working
group,
depending
on
what
you
feel
like
I'm,
not
Elise
lake,
with
whoa
a
co-chair
either
co-chair
of
Russ
white
he's
not
here.
He
we
originally
an
English
scheduled
yesterday
when
they,
so
he
was
planning
to
be
here,
but
he
made
arrangements
to
leave
either
this
morning
or
last
night,
I'm,
not
sure.
So
when
we
got
rescheduled
for
today,
he
was
not
here
so.
C
A
Note
well
probably
you've
seen
this
before
or
would
you
read
it
we're
operating
under
the
IETF
IPR
rules?
If
you
make
any
contributions,
including
saying
anything
during
a
meeting,
you
need
to
declare
any
IPR,
you
know
about
that.
Hasn't
already
been
declared
and
there's
a
bunch
of
other
stuff.
Here
too,
you
should
look
at
so.
Why
wasn't
this.
A
F
A
F
F
A
F
A
Hitting
the
down
arrow
right,
left
and
right,
oh
right,
somehow
that
doesn't
seem
logical
to
me,
but
anyway
moving
right
along
agenda.
We
have
an
agenda,
so
we
should
bash
on
it.
If
anybody
wants
to
make
any
change,
the
plan
is
talked
a
bit
about
our
status
of
milestones
and
Jerez,
which
is
I,
think
there's
good
news
there
and
then
presentations
on
changes
to
available
bubble,
h,
mac,
r,
TT,
based
routing
information
model
and
the.
H
A
A
Okay,
no
requests
for
changes
so
we're
now
that's
right.
There
document
status,
so
we
actually
have
a
nifty
kind
of
like
sort
of
cohesive
set
of
four
documents
with
publication
requested
the
applicability,
the
two
security
crafts
DTLS
and
HVAC
and
RFC
6120
six
bits.
So
it
seems
hopefully
that
package
will
advance.
A
We
do
have
three
working
groups,
including
wrapped
IETF
babble
yang
model,
which
it's
correct
here.
The
and
the
source
specific
end
information
model
drafts
which
are
in
working
group
last
call
are
there
related
drafts.
There's
the
home
net
babble
profile,
which
is
in
the
RFC
editor's
q
in
miss
ref
state
means
it's
waiting
on
normative
references
and
the
to
normative
references.
It's
waiting
on
our
RFC
6120,
six
bits
and
the
source
specific
draft,
and
then
there's
the
doughlas
fable,
RTT
extension
graft,
and
this
slide
also
shows
the
previous
set
of
three
experimental
grafts.
A
So
we
just
moved
the
milestone
of
submission
of
RFC
6120,
six
bits
to
the
IES
g
RT
to
our
ad,
to
be
more
specific
from
the
to
be
done
to
the
current
ones.
Were
you
know
kind
of
behind
relative
to
the
actual
dates
on
the
current
milestones?
But
that's
we
are
making
progress,
I
believe
on
all
fronts
here.
A
The
management
draft,
the
last
one
is
is
intended
to
me
refer
to
the
yang
draft.
In
my
opinion,
so
that's
kind
of
the
status
of
things
I
think
I'll
wait
till
after
the
presentations
or
there
were
a
couple
of
working
group.
Actions
concerning
adoption
and
last
call
that
are
I
said,
would
go
through
this
meeting
without
I.
Think
I'll
do
things
related
to
those
at
the
end
of
the
meetings
or
so
any
questions
or
comments
on
that
I
guess
not.
Then
we
will
proceed
with
the
agenda
and
with
Julia's
presentations.
B
A
A
Okay
and
you
would
you
like
a
control
device,
I
can
do
it
for
you.
I
So
hello,
I'm,
Julia
scrub,
Oh
check
and
this
thing
over
there
is
my
name
and
if
you're
ever
wondering
why
I
have
so
many
strange
consonants.
That's
because
it's
written
in
the
Polish
spelling
system,
which
dates
from
the
Middle
Ages.
Now
the
checks
over
here
used
to
have
a
very
similar
writing
system.
But
then,
in
the
early
17th
century
there
was
a
great
reformer
Covey
on
who's
who
redesigned
the
writing
system
to
make
it
sensible.
I
So
if
you
ever
look
at
my
name,
please
use
it
as
a
cautionary
tale
about
what
happens
when
you
come
to
ITF
with
an
already
deployed
protocol
and
are
too
concerned
that
ever
backwards-compatibility
next,
please,
okay,
so
last
summer,
together
with
things
online
and
the
Veronica
kodiak,
we
set
about
to
design
a
authentication
method
in
for
Babel.
That
would
be
as
simple
and
as
comprehensible
as
possible.
So
I
know
nothing
about
cryptography
and
the
design
criterion
was
something
that
I
could
understand.
I
So
it
was
based
on
some
work
that
had
been
done
by
Denis
of
Shenko
in
RFC,
seven,
two
nine
eight
and
we
had
a
lot
of
in
good
and
in
particular,
so
I've
tried
to
be
exhaustive
in
the
acknowledgements,
but
so
here
I
will
only
mention
Marcus
Stenberg,
who
gave
us
some
very
important
ideas,
because
some
properties-
and
we
ended
up.
You
know
we
did
the
opposite
of
what
happens.
I
Usually,
usually
you
have
ambitious
plans
and
you
end
up
realizing
it's
difficult,
and
so
you
scale
them
down,
and
in
this
case,
actually
we
were
expecting
to
do
and
that
turned
out
pretty
well.
We
got
a
protocol
that
protects
both
unicast
and
multicast.
So
remember
that
in
Babel
you
have
both
unicast
and
multicast
packets.
Depending
on
the
pendulum
well
various
details
of
the
protocol.
It's
not
vulnerable
to
replay
it's
reasonably
easy
to
implement
it
been
actually
reimplemented
by
Toki
and
the
properties
that
make
it
easy
to
implement
are
that
we
don't
require
the
global
clock.
I
There
is
no
global
counter
in
the
network
and
there
is
no
persistent
storage.
So
barbara
was
telling
me
at
the
time.
Well,
why
are
you
trying
to
avoid
persistent
storage?
There's?
No
such
thing
as
a
Rooter
with
no
persistent
storage?
Well,
the
problem
is
that
persistent
storage
is
very
difficult
to
do
right.
Okay,
depending
on
the
hardware
you're
on
you
might
have
a
hard
disk.
You
might
have
an
SSD,
you
might
have
some
proprietary
NVRAM
and
trying
to
access
it.
I
Transactionally
is
going
to
be
very
difficult,
so
I've
preferred
to
sweep
the
problem
out
under
the
rug
by
not
assuming
any
persistent
storage.
It
says
algorithm
agility.
There
are
a
few
parameters
you
can
change
depending
on
what
your
needs
are
next,
please,
and
what
we
did
was
that
we
did
everything
on
the
mailing
list,
because
I
didn't
see
a
competent,
so
the
two
ladies,
our
cryptography
student,
they
understand
cryptography,
they
don't
under.
They
didn't
necessarily
understand
route
routing
protocols
at
the
time
as
to
myself.
I
I
know
apps
and
you
absolutely
nothing
about
cryptography,
and
so
we
did
everything
in
the
open
on
the
mailing
list,
except
for
two
things
that
were
done
more
recently
and
that
I'm
going
to
tell
me
about
which
were
mentioned
on
the
mailing
list.
But
didn't
really
have
a
lot
of
discussion.
I
hope
they
will
be
non-controversial.
I
We
changed
the
mandatory
to
implement
nack
algorithms.
So
now
what
it
says
is
that
you
must
implement
sha-256
and
you
should
implement
lake
2s
and
that
there
is
an
additional
requirement.
You
must
discard
your
neighbour
cryptographic
state
after
a
bounded
amount
of
time,
so
the
original
RFC
seven
to
nine
by
that
done
by
then
had
to
mandatory
to
implement
algorithms.
They
were
Shaolin
and
ripe
in
the
160.
I
So
in
the
latest
version
of
the
draft
we
say
that
you
must
implement
sha-256,
you
should
implement
like
2s,
and
the
intent
is
that
those
are
most
must
accept
that
you've
written
should
in
the
draft
in
case
you
have
some
very
good
reason
not
to
implement
like
2's,
if
you
have
no
good
reason
to
implement
like
2's,
please
implement
both
and
we
implement
both
in
both
babble,
D
and
Bert.
Next,
please
look
I'll,
tell
you
the
whole
truth.
I
The
hash
function
just
doesn't
matter
so,
first
of
all,
we're
only
using
the
hash
function
for
authentication.
So
if
there
is
a
first
preimage
attack,
that's
something
difficult,
I'm,
not
sure
what
it
is.
That
I'm
told
it's
something
difficult,
but
then
we
are
threatened,
but
we
are
not
threatened
by
collision
attacks.
If
somebody
manages
to
do
a
collision
attack,
then
we
are
not
threatened
by
that,
so
any
hash
function.
You
may
conceive
for
its
strong
enough.
I
According
to
my
computations
and
I'm,
not
a
cryptographer,
but
according
to
my
back
of
the
envelope
computations
md5
would
be
strong
enough.
Md
4
is
too
weak.
Ok,
so
anything
you
use
reasonably.
If
you
have
a
reasonable,
reaching
interval
will
be
strong
enough
and
you're
only
hashing
the
control
packets.
So
unless
you
run
in
Babel
on
an
abacus,
then
any
hash
function
will
be
fast
enough.
So
really
it
doesn't
matter.
Sha
256
is
clearly
overkill
for
what
we
are
doing,
but
even
that
will
be
fast
enough.
So
it's
fine.
Next,
please.
I
However,
some
people
are
concerned
about
slow
software.
Routers,
like
the
things
you
get
in
your
home,
with
no
hardware
assists
for
sure,
200
56
and
token
they
suggested
Blake
to
s.
I've
done
my
homework.
I've
tried
to
find
out
what
is
what
is
the
world
like
to
us?
Actually
I
asked
Paulo
Xavier
to
come
I'm
grateful
to
do
my
homework
for
me
and
what
Paul
have
been
telling
me
is
that
it's
blazingly
fast
and
software.
I
It
has
128
bit
lengths,
it
integrates
keying,
so
you
don't
need
to
hash
over
the
result
of
hashing,
because
the
King
is
integrating
the
protocol.
You
don't
need
the
H
Mack
construction
and
the
issue
is
it,
which
is
the
reason
why
it
was
rejected
as
shastri
is
that
it's
based
on
very
similar
principles
as
char
256,
and
so
if
sha
256
is
broke
and
most
probably
like
2's
will
be
broken
too,
which
should
not
be
a
problem
for
us,
because
even
if
it's
broken
it
will
still
be
strong
enough
for
our
needs.
Next,
please,
ok!
I
So
that's
the
first
thing:
the
change
in
the
mandatory
to
implement
hash
algorithm.
The
other
change
is
that
we've
added
some
protection
against
delayed
packets,
so
babel
h,
mac
is
invulnerable
to
replay.
If
a
packet
has
been
used
once
you
will
not
manage
to
use
the
same
packet
again,
it's
also
involved
to
reordering,
but
it
is
not
what
it
is
vulnerable
to
is
packets
being
delayed.
So
if
a
packet
is
in
the
right
order,
it
can
be
delayed
by
arbitrary
amounts
of
time.
I
Now,
if
the
attacker
is
at
a
network
switch,
it
might
be
able
to
block
packets
and
then
send
them
later.
We
don't
know
any
attack
that
could
be
against
Babel.
That
could
be
using
delayed
packets,
but
the
notion
that
there
are
packets
that
have
been
sent
identified
and
they're
replayed
and
they
are
intercepted
and
then
sent
say
2
hours
later,
is
something
that
makes
us
nervous.
So
we've
added
next,
please
so
one
property
of
the
protocol
is
that
a
node
may
at
any
time
discard
its
perkier
cryptographic.
State.
I
Ok,
so
you
can
say:
ok
forget
anything
about
the
exchange
I've
made
with
a
given
fear,
and
then
there
is
a
mechanoid
called
the
challenge
mechanism
that
will
create
a
new
cryptographic
state
at
the
cost
of
one
RTT
of
lost
packets,
and
the
node
can
do
that
at
any
time.
And
what
we
added
is
that
now
a
node
must
we
discard
its
cryptographic
state
after
a
bound
at
time
with
no
correctly
authenticated
packet?
Okay,
so
we
don't
give
any
figures
for
the
bound.
I
We
say
you
must
ensure
that
there
is
a
bound
and
the
time
on
which
you
keep
cryptographic
state
for
somebody
who's,
not
speaking
to
you,
okay,
and
we
are
not
specifying
the
mechanism,
but
we
do
give
implementation
advice
next,
please
so
we've
made
it
to
non-controversial
changes
to
the
protocol
and
like
this
random.
In
this
talk,
thank
you
for
attention.
G
G
I
G
I
Okay,
hello,
I'm,
Julia
scrub
attract.
Did
they
mention
the
spelling
of
my
name
next
so
I'd
like
to
tell
you
you
know:
can
you
grow
really
old?
You
start
taking
the
things
you
did
when
you
were
younger
and
trying
to
recycle
them,
and
so
today,
I'm
going
to
tell
you
about
something
we
did
four
years
ago
and
that
I've
just
asked
a
for
adoption
by
this
working
group
now
and
that's
joint
work
with
Betty
Jean
glaze.
I
The
problem
is
that
the
public
ipv6
infrastructure
is
not
reliable
enough,
and
if
you
have
enough
nodes,
then
the
probability
that
anyone
can
speak
to
anyone
else
is
basically
zero,
and
so
they
came
up
to
me
and
I
said
well.
Why
didn't
you
put
a
bunch
of
tunnels
and
run
some
dynamic
routing
protocol
over
the
tunnels?
Okay,
so
we
designed
over
a
restaurant
napkin,
an
algorithm
that
would
allow
us
to
have
enough
tunnels
so
that
the
network
remains
connected,
and
then
they
told
me,
okay,
that's
what
routing
protocol
shall
we
use?
I
I
We
were
trying
to
solve
an
actual
problem
that
real
people
have
and
in
this
topology
what
happens
when
the
leading
marks
a
link
breaks.
The
tunnel
has
broken
for
some
reason,
because
some
ISP
have
decided
that
it's
cheaper
to
break
connectivity
and
violate
their
SLA
rather
than
pay
for
transit,
and
you
have
no
connectivity
between
liliane
masse.
Well.
Babbles
is
two
hops
between
through
Paris
and
two
hops,
through
Tokyo
and
so
non-deterministically,
one
half
of
the
time
it
will
decide
to
route
the
package
from
limb
to
Marseilles
through
Tokyo.
I
We
have
nothing
against
Tokyo,
but
clearly
routing
packets
from
Linton
are
safe
through
Tokyo
is
a
bad
idea.
Next,
please
so
my
first
suggestion
the
world's:
why
don't
you
put
a
GPS
in
every
router
and
then
we
can
use
the
geographical
distance
for
breaking
ties
in
our
routing
and
they
said
that's
out
of
the
question,
that's
out
of
the
question
and
what
they
told
me
was
that
the
problem
was
not
the
price
of
the
GPS.
The
problem
is
the
human
being.
I
Who
is
going
to
make
sure
that
all
of
those
GTS
are
running
at
any
given
time
so
you're
introducing
too
much
failure
modes,
you're
introducing
unusual
things
in
the
data
center?
They
said
that's
out
of
the
question,
so
we
decided
to
measure
the
two-way
delay
the
rtt
and
derive
a
metric
from
that.
So
before
somebody
comes
up
and
tells
me
about
TCP
Vegas
or
led
that
or
bbr
or
is
it
B
RR
v
BR?
Okay,
all
of
those
things
that
measure
one-way
delay.
I
I
The
goal
of
the
routing
protocol
is
to
select
the
best
routes,
so
you
are
not
to
accurately
model
reality
as
long
as
you
have
managed
to
find
the
best
route
you've
won,
even
if
your
modeling
of
reality
is
somewhat
coarse
and
RTT
appears
to
be
good
enough,
at
least
for
this
application,
and
the
problem
is
that
the
natural
way
to
measure
RTT
basically
pink,
requires
asymmetric,
there's
a
client
and
a
server
and
synchronous
interaction.
You
have
to
answer
the
ping
as
fast
as
possible.
Now
babble
is
a
symmetric
protocol
incident,
a
synchronous
protocol.
I
Nowhere
in
Bayville
does
it
say
you
have
to
reply
straight
away.
There
are
times
where
it
says
you
must
send
a
packet
in
a
timely
manner,
but
at
no
point
do
you
have
to
react
in
an
immediate
and
the
other
problem
is
that
using
RTT
as
input
the
routing
metric
causes
a
negative
feedback
loop
which
I'm
going
to
describe,
which
might
lead
to
persistent
oscillations.
Next,
please
so
RT
t
RT
T
is
on
the
diagram.
I
I
Actually,
there's
no
big
deal
with
doing
symmetric
asymmetric
stuff
within
babble,
but
the
other
problem
is
that
when
the
server
replies
the
thing
it
must
answer
Punk
as
quickly
as
possible
and
the
delay
you
have
between
the
ping
and
the
pound
is
what
is
going
to
kill
your
accuracy
and
now
because
in
babble
you
are
allowed
to
reply
later.
You
might
have
buffered
some
stuff
to
be
sent
later,
and
here
you
have
tricky
issues
of
do
I,
flash
everything
right
now
and
cause
congestion.
Do
I
rear
their
packets
and
cause
pathology
due
to
packet,
reordering?
I
I
It
contains
three
time
stamps
which
are
teo,
which
is
copied
from
the
previous
packet
in
the
other
direction,
which
is
called
the
origin
timestamp
TR,
which
is
the
time
at
which
we
received
the
previous
packet,
which
is
there
called
the
reference
timestamp,
and
it
also
contains
the
transmit
timestamp,
which
is
the
time
we're
sending
this
packet.
Okay,
I'm
now
be
careful,
tio
and
T
are,
according
to
the
left,
peers,
clock
and
TTT
r
TT
are
according
to
the
right,
Pierce
clock,
and
we
don't
assume
that
the
clocks
are
synchronized.
I
That's
asynchronous
that
doesn't
require
clocks
to
be
synchronized,
and
the
reasons,
on
the
other
hand,
are
something
you
must
do.
You
must
compute
the
timestamp
at
which
you
send
the
packet.
The
TT
as
late
as
possible,
just
before
you
send
the
packet
not
on
your
buffer
and
same
thing
as
soon
as
you
receive
a
packet
before
you
even
parse
it
you
timestamp
it
and
that's
what
your
accuracy
is
going
to
depend
on
next,
please.
I
I
Perhaps
it
should
have
been
simply
a
TLV
rather
than
a
sub
TLD
of
hello
and
the
other
property
and
the
other
other
to
timestamp
the
origin
and
reference
those
are
copied
from
the
latest
packet
you
received
from
the
peer,
so
they
are
specific
to
a
given
peer
and
they
are
sold
as
sub
charities
of
the
ihe
govt.
And
originally
we
started
with
a
10,
millisecond
granularity,
and
then
dave
told
us
what
he
thought
about
it.
You
know
Dave,
you
know
how
he
tells
you
that
he
disagrees
like
seven
times
in
a
row.
I
Okay,
no
secrets
here
are
32
bits
rolls
around
every
36
years,
or
is
it
72
I'm
not
sure
about
the
sign?
But
okay,
no
particular
issues
here.
Next,
please,
okay,
so
Mills
algorithm
is
going
to
give
you
some
RTT
samples
and
that's
a
noisy
signal
and
the
signal
that
is
not
immediately
useful
for
a
routine
metric.
Remember
we
are
doing
a
routing
protocol
here.
I
G
I
Well,
I
suggest
I'd
suggest
we
tackle
this
problem
when
we
have
it
yeah
sure
yeah
sure
I
was
completely
off,
so
we
are
assuming
here
that
you
send
at
least
a
hello
every
hour.
Okay,
so
you
get
a
bunch
some
noisy
signal,
but
your
goal
is
to
select
the
routes.
Okay
and
you
need
to
do
quite
a
bit
of
processing,
so
the
RTT
sample
that
you
get
or
a
noisy
signal.
You
do
some
basic
smoothing
over
there.
Then
that
gives
you
the
smoothed
RTT
and
the
RTT
is
a
time
in
milliseconds.
I
That's
not
a
metric.
In
order
to
get
a
metric,
you
have
to
map
it
to
the
abstract
units
that
we
use
for
our
metrics.
So
you
take
your
link,
cost
the
one,
the
normal
link
cost,
which
mode
is
the
96
that
you
have
with
normal
battle
and
you
compute
a
penalty
from
the
smooth
RTG
and
this
penalty
is
not
a
linear
map
from
the
RTT
of
how
you
add
the
penalty
to
the
link
cost
from
that.
You
derive
a
metric,
and
here
that's
important.
There
is
a
third
mechanism
which
is
hysteresis.
I
I
You
just
want
a
smooth
signal,
but
we
did
not
do
any
real
evaluation
of
what
happens
when
you
change
your
smoothing
algorithm.
So
we've
just
stolen
the
algorithm
from
TCP
the
one
that's
used
for
computing,
the
RTT
in
TCP
and
the
constant
alpha
where
TCP
says,
should
be
between
zero
point.
Eight
and
zero
point:
nine:
we
chose
a
zero
point.
Eight
three,
six
and
first
person
who
works
at
Y
I'll
be
very
impressed.
If
you
work
out
why
you
won't
work
at
F
now,
you
will
need
at
least
an
envelope
next
thesis.
I
You
want
the
answer.
It's
easy
to
compute
with
integer
arithmetic,
where
in
doing
Taylor's
is
some
stuff,
that's
the
whole
reef.
Okay!
Now
so
now
we
have
a
smooth
RTT
signal,
but
using
RTT
for
computing,
geometric
colors,
a
feedback
loop.
You
know
when
you
were
a
kid
on
the
highway
and
it
was
really
hot
in
the
back
of
the
car
and
the
highway
was
really
congested
and
then
the
radio
would
say
well,
there's
nobody
on
the
other
Highway
and
then
everyone
would
move
to
the
other
congested.
I
So
the
radio
will
say:
hey,
there's
nothing.
Nobody
on
the
first
Highway
and
everyone
would
move
back
was
going
back
on
here
is
that
there
is
a
negative
feedback
loop,
the
more
you
push
traffic
over
a
route,
the
more
its
RTT
increases
and
therefore
the
less
desirable
it
becomes,
and
therefore
the
less
traffic
you
push
over
it
now.
Usually
an
egg
Luke,
a
negative
feedback
loop
is
what
you
like,
but
converges
because
the
more
you
do
something
the
less
you
do
it.
I
But
here
when
we're
in
a
discrete
space,
either
we
go
to
the
left
or
we
go
to
the
right.
The
fixed
point
doesn't
actually
exist,
so
you
end
up
oscillating
now,
Babel
doesn't
care
even
in
the
presence
of
oscillations.
Babel
will
push
packets
towards
the
destination
according
to
Luke
free
pass,
there's
not
much.
That
can
be
wrong,
but
if
you
also
like
between
routes,
you
might
reorder
packets
and
higher
layer
protocols
do
actually
care.
So
you
need
to
limit
the
frequency
of
oscillations.
I
And
the
insight
came
from
some
old
paper
from
the
1980s,
which
was
trying
to
use
RTT
and
link-state
protocols,
and
the
answer
is
you
want
to
saturate
your
metric
so
instead
of
mapping
RTT
to
a
metric
linearly,
as
you
would
expect,
what
you're
doing
is
that
you
have
a
range
below
min
RTT
at
which
you
consider
look.
This
route
is
good.
This
link
is
good.
It
doesn't
matter
how
good
it
is,
and
above
and
that's
not
important.
The
important
bit
is
that
above
max
RTT,
you
consider
this
link
is
bad.
I
And
what
you
want
is
that
this
Max
Artie
T
value
is
well
chosen
and
this
that
part
of
the
protocol
requires
manual
tuning
next,
please
now
the
saturation
bit
avoids
oscillating
between
congested
links,
but
you
might
still,
if
you're
in
the
intermediate
range,
have
two
routes
that
have
very
similar
RTT
and
they
are
oscillating
around
a
value
and
you
will
end
up
switching
turn
to
apply
hysteresis
and
there
are
many
hysteresis
algorithm.
So
I
worked
out
this
one.
So
in
this
one
you
have
in
a
which
is
the
metric.
I
That's
actually
announced,
and
there
is
a
second
metric
which
I
called
the
smoothed
metric,
which
is
basically
an
exponential
average,
a
smooth
version
of
the
announced
metric
so
think
of
MA
as
being
the
short-term
metric.
It
tells
you
how
good
the
route
is
right
now
and
m/s
being
a
history
sensitive
metric.
I
It
tells
you
how
lot
good
this
route
has
been
recently
and
when
you
switch
routes,
you
only
switch
routes
if
the
new
route
is
better
than
the
old
route,
both
short-term
and
long-term,
and
this
and
there's
a
symmetry
between
better
and
worse
is
what
is
going
to
give
you
hysteresis
and
hence
more
stability.
Next,
please
that's
with
hysteresis
and
no
smoothing
this
datum
line
the
pity.
So
this
is
worst
case.
I
We
constructed
a
diamond
topology
in
simulation
with
huge
amounts
of
buffer
bloat,
so
buffer
bloat
is
really
bad
for
us,
because
it
means
that
the
RTT
is
very
very
widely
with
nothing
to
break
the
feedback
loop.
So,
though
those
are
four
bloated
routers
in
a
diamond
topology,
measured
over
4,500
seconds,
Michael
helped
me
because
I'll
say
something
stupid
again.
I
That
will
be
about
an
hour
and
a
half
yes,
and
what
you
can
see
is
that
you
have
oscillations
between
the
red
route
and
the
green
route.
So
you'd
have
to
look
carefully
at
the
frequency
next
leaf,
and
now
you
add,
the
you
tune.
The
metric
to
have
a
max
RTT
value
that
as
well
chosen
is
the
dashed
blue
line
around
hundred
and
something
so
what
you're
doing
is
that
the
links
are
spending
most
of
the
time
and
they
are
saturated
State.
And
at
this
point
you
pick
a
link.
I
The
RTT
of
the
link
increases
because
you
have
buffer
bloat.
So
you
switch
to
the
other
link,
but
the
RTT
of
the
rig
will
only
decrease
once
the
buffers
have
drained.
And
then
you
have
the
hysteresis
that
takes
some
time
to
take
it
into
account
and
supplied
after
saturation
and
together
between
the
fact
that
you're
waiting
for
the
buffers
to
drain
without
measuring
the
RTT,
because
you're
in
the
saturated
range
plus
the
hysteresis
bit.
You
end
up
having
oscillations
that
are
order
on
the
order
of
a
minute.
But.
J
I
G
I
I'll
put
it
differently.
Perhaps
I
wasn't
sufficiently
clear.
We
don't
want
this
to
happen.
The
feedback
loop
is
a
bad
thing
and
it
doesn't
happen
in
practice
in
practice.
You're
going
to
have
plenty
of
throughput,
okay,
you're
running
over
gigabit
links.
Okay,
nobody
has
ever
managed
to
use
a
whole
gigabit,
except
perhaps
David
I'm,
not
sure.
G
It's
great
so
the
one
that
came
default
in
the
Linux
distributions
that
I
was
using
did
this,
so
this
is
not
as
silly
as
it
might
seem.
No,
it's
not
silly!
So
basically,
we
are
assuming
that
you
run
this
right.
Okay,
pathological
was,
it
was
probably
the
name
that
did
what
I
was
using
for
not
silly
but
yeah.
This
is
this.
Can.
I
Actually
be
a
real
problem,
you
know
what
we
want
is
a
protocol
that
those
mechanism
are
not
used
when
your
network
is
working.
Fine.
What
we
don't
want
is
a
catastrophic
mode
of
failure
when
your
network
happens
to
trigger
this
thing.
Okay,
ideally,
of
course
you
don't
get
the
feedback
loop
in
the
first
place,
and
you
have
no
problem.
I
Okay
is
that
yeah?
Next,
please,
okay,
I'll,
tell
you
the
whole
truth:
we've
done
some
really
bad
science.
We
wanted
to
have
a
solution
that
worked.
We
ended
up.
We
had
only
a
few
months
because
that
was
effects
your
internship
and
we
did
not
at
the
time
work
out
a
complete
understanding
of.
What's
going
on,
we
did
an
engine
engineering
job
instead
of
doing
a
scientific
job.
I
That
is,
we
did
manage
to
get
something
that
works
beautiful
and
that
has
been
working
in
production
for
eight
years
now,
but
we
did
not
actually
study
all
the
trade-offs
and
we
don't
fully
understand
what's
going
on.
So
there
is
one
minor
issue
in
the
packet
format,
which
is
that
the
IH?
U
sub
TLD,
can
only
be
interpreted
if
the
packet
contains
a
hello.
So
remember,
there
are
three
timestamps
and
two
are
carried
by
the
HU
because
they
are
per
peer
and
one
is
carried
by
the
hello.
I
So
that
means
that,
in
order
to
interpret
the
data
in
the
HU,
you
need
to
put
a
hello
in
every
packet.
So
that
requires
some
inter
TLV
scheduling
in
Babel
and
that's
something
I
really
dislike.
So
I
wrote
it
and
written
an
alternative
here,
which
is
to
include
the
transmit
timestamp
in
each
HU,
but
that
causes
a
different
problem.
So
remember
that
you
are
time,
stamping
the
packet
as
late
as
possible.
So
what
we
do
is
that
when
we
buffer
there
hu,
the
hello,
we
put
a
and
sub
TLV
in
it.
I
At
the
very
last
moment,
we
patch
the
hello
with
the
timestamp
and
that's
fine,
because
there
is
only
one
hello
in
the
packets,
but
there
are
multiple
ih
use
and
if
you
need
to
patch
it
multiple
time
the
implementation
will
become
more
complex.
The
other
alternative
that
I
didn't
write
here.
Ins,
which
I
actually
like
more,
is
to
use
the
timestamp
on
the
hello
to
you,
the
timestamp,
in
a
dedicated
timestamp
TLV,
rather
than
attaching
it
to
the
hello,
that's
something
we
will
need
to
discuss
and
we
did
some
Mike.
I
And
what
we
did
is
that
we
did
do
evaluating
of
the
interesting
bits
like
why
saturating
is
important,
but
there
are
some
things
that
were:
we
did
not
do
our
job
at
the
end,
we
didn't
evaluate
what
happens
when
you
change
smoothing
functions,
that's
something
that
doesn't
bother
me
much,
because
the
answer
is
pretty
clear.
Nothing
much
will
happen,
because
this
new
thing
is
not
important.
I
Bad
things
happen
in
this
protocol.
If
max
RTT
is
too
large.
If
max
RTT
is
too
large,
you
get
oscillations
in
the
bad
cases
and
we
didn't
do
any
work
on
auto-tuning
max
RTT,
which
should
be
doable
and
I.
Think
that's
our
greatest
sin.
We
didn't
fully
evaluate
the
effect
of
hysteresis
and
how
the
hysteresis
should
be
tuned.
I
So
it's
obvious,
you
need
some
form
of
hysteresis,
but
whether
we've
done
the
right
thing,
that's
not
clear
at
all
next
please,
and
that
is
the
reason
why
so
what
we
have
here
is
something
that
will
be
beautiful
in
production,
we'd
like
to
get
it
standardized
and
the
packet
format
I.
Think,
is
something
that
we
understand.
We
understand
the
trade-offs,
but
in
order
to
make
use
of
it,
you
have
to
use
a
bunch
of
algorithms
and
those
alder
is
are
not
fully
understood.
I
So
what
I
would
like
to
see
would
be
a
situation
which
we
managed
to
standardize
the
package
for
and
an
enormous
part
of
the
document.
Just
say:
hey
folks:
that's
how
you
measure
RTT
and
babble.
If
you
need
RTT
and
what
you
use
this
RTT
for
I'd
like
to
avoid
standardizing,
because
I
don't
think
we're
ready
to
standardize
that
and
just
have
an
informative,
appendix
saying:
look,
that's
what
we've
used
it
for
it
works
pretty!
K
David's
Knaus
e
google
thanks
joyous
first
point:
I,
really
supportive
of
this
work
and
I
do
think
it
should
happen
here.
I'm,
really
glad
that
we
have
one
like
actually
deployed
extension
like
this,
which
really
reassures
us
that
the
extension
mechanism
in
Babel
all
works.
So
that's
a
great
use
case
that
we
don't
were
shipping,
something
that
we
know
is
solid
and
now
a
question
have
you
just
out
of
curiosity?
Have
you
tried
running
this
in
babel
d
running
both
the
one
RT,
the
rtt
extension
and
h
mac,
and
how
does
that
impact
timing.
I
It
doesn't
work,
it
doesn't
work
due
to
a
limitation
of
the
implementation.
So
remember
that,
in
order
to
make
this
work,
you
need
to
have
a
hello
together
with
every
I
hu,
so
the
timestamp
will
be
ignored
if
an
I
achieve
is
spent
in
a
packet
that
doesn't
have
a
hello
now.
Remember
that
will
DTLS
the
H.
H
I
I
I
I
A
Are
there
any
other
comments,
questions
so
there's
a
call
for
working
group
adoption
that
was
it
I
was
posted
which
ends
at
this
meeting
and
there's
been
strong
support
on
the
mailing
list
and
no
opposition.
There
are
anybody
in
the
room
that's
supposed
to
adopting
this
as
a
working
group
draft
hearing,
nobody,
it
is
so
adopted
and
all
set
ask
the
authors
to
post
the
zero-zero
version
as
a
working
group
draft.
Thank.
A
J
A
M
M
M
And
we've
been
doing
it
over
on
github
in
my
github
site,
where
I've
been
trying
to
keep
editors
drafts
and
things
like
that,
I
have
not
but
I
know
I
promised
I
would
I
have
not
put
the
updated
broadband
form
data
model
there
I
just
sort
of
ran
out
of
time
with
everything
else.
I
was
doing
this
week,
but
it's
in
good
shape.
I
had
great
comments
out
of
broadband
form
as
well
so
major
changes.
M
Clearly
we
did
some
stuff
on
some
stuff
on
DTLS
I,
put
in
the
statistics
and
message
that
we
agreed
to
coming
out
of
Bangkok,
and
there
was
also
some
additional
discussion
on
the
list
that
helped
me
clarify
some
stuff
I
thought
that
was
great
and
the
other
stuff
out
of
Bangkok
I
did
okay.
G
M
I've
got
the
rest
of
the
slides,
are
about
some
issues.
I
would
really
like
to
get
some
input
on
number
one.
Do
we
have
the
link
types
registry?
Do
we
have
the
right
names
currently
in
the
registry
and
is
it
meaningful
to
describe
what
sorts
of
links
it
is
used
for,
and
is
there
any
idea
on
how
extensive
the
list
should
be
now
Julius
has
previously
said
this
is
I
think
you
really
only
want
the
main
four
and
then
we
can
leave
some
room
for
experimental
stuff
at
this
time
if
they
want
it.
M
Mahesh
and
I
have
discussed
that
in
fact,
now
that
we
understand
better,
that
Ethernet
isn't
really
just
Ethernet,
but
it
supposedly
incorporates
everything
that's
wired,
including
all
of
the
no
new
wire
protocols
like
powerline
and
mocha,
and
things
like
that.
We're
really
confused
and
wondering
if
we
should
change
it
back
to
Wired
or
Julius.
I
So
if
you
look
at
the
Babel,
the
implementation
you're
going
to
have
a
lot
of
per
length
tweaks,
are
you
going
to
use
packet
loss
for
computing?
The
metric?
Are
you
going
to
use
RTT
for
computing?
The
metric?
Are
you
going
to
use,
split
horizon
and
so
on,
that's
impossible
for
a
normal
human
being
to
chip,
and
because
of
that,
we
have
bundled
sets
of
parameters
into
names
that
you
can
give
to
the
actual
users
who
don't
want
to
understand
all
of
the
details.
G
O
I
M
M
O
M
M
Yeah
cool
I
hope
you've
got
that
on
the
the
notes.
I
read:
oh
my
god.
Well,
hopefully,
you
know
the
Medeco
recording
will
work.
Okay,
DTLS
interfaces
modeling
had
some
discussion
of
additive
interfaces.
Oh
good
Michael's
leaving
us.
P
M
Said
one
hour
go
away:
okay,
now,
Mahesh
and
I
were
discussing
and
actually
I
got
some
input
in
broadband
form
that
we
should
actually
change
the
direction
of
the
reference
right
now.
I
was
trying
to
reference
two
interfaces
from
H,
Mac
and
DTLS,
but
it's
been
suggested
that
it
might
really
solve
a
whole
lot
of
issues
if
we
change
the
direction
and
from
interfaces
reference
to
H,
Mac
and
DTLS
objects,
and
that
would
actually
match
more
closely
with
what
Julius
was
describing
on
the
list
as
to
what
you
were
doing.
M
F
I
M
I
M
That's
what
good
that's,
what
you
nodded
about
and
I
agree
with
you
completely
and
that's
what
my
colleagues
at
broadband
forum
said
too.
So
it's
like
everyone
agrees
that
Barbara
is
wrong:
okay,
okay,
but
now
the
the
other
question
I
really
have
is
about.
Are
these
other
parameters
global
or
can
they
be
associated
with
the
set
of
credentials.
I
Okay,
verification
of
receive
packets.
That's
clearly
you
want
that
to
be
per
link.
You
have
a
secure,
router
you're,
adding
a
new
link.
There
are
unsecure
Reuters
and
this
link
you
want
to
transition
the
new
link.
You
don't
want
to
disable
security
and
all
the
other
links
just
because
you've
added
a
new,
insecure
link.
Okay,.
M
I
I
H
Aeternam
Donnie
I
think
we
should
try
to
answer
the
question
of
which
end
phase
refers
to
Detailers
H,
Mike
object
or
the
other
way
around
is
by
asking
the
question:
do
you
have
one
interface
that
they're
first,
the
multiple
VPLS
/a
track
objects
or
that
you
can
have
a
given
DTLS
/eric
object
being
used
by
multiple
interfaces?
M
I
think
a
single
DTLS
can
be
used
by
multiple
interfaces
and
that's
highly
desirable,
because
we
want
there
to
be
kind
of
a
default
or
to
be
able
to
use
a
single
for
everything
yeah
the,
but
the
other
one
that
Toki
did
bring
up
was.
It
would
also
be
desirable
to
have
a
single
interface,
be
able
to
point
to
multiple
objects
and
I
think
we
can
have
both
its
and
and
for
H
Mac.
It's
once
we
pull
this
global
parameter
out.
M
G
I
think
it's
hard
to
do
that
here,
but
typically
you
want
to
have
a
many-to-many,
so
an
interface
can
have
many
something,
and
this
this
thing
should
have
like
a
name
that
you
reference
it
by
they've.
This
credential
like
an
SSH
key,
if
you're
doing
that
or
so
on.
You
should
be
the
have
that
flexibility,
because
that
always
comes
up
sometimes
later
I
would
say
that
you
wouldn't
want
to
do
something
like
that.
M
M
M
So
we've
also
taking
one
step
up.
That
was
the
that
was
the
the
this
was
about
the
actual
certificate
and
key,
but
one
level
down.
We
had
these
objects
and
some
of
this
yeah,
where
we're
trying
to
figure
out
how
to
refer
to
these
H
Mac
and
DTLS
objects
that
contain
keys,
multiple
one
or
more
keys
and
Mahesh
and
I
had
disagree
and
on
where
to
specify
the
unique
key,
because
in
some
of
my
modeling
that
I
do
on
the
broadband
form
side,
we
have
auto-generated
things.
M
I
M
I
M
M
It
can
be
it's
it's
the
information
model
at
this
level.
It
doesn't
care
once
you
get
into
a
yang
data
model
or
broadband
form
data
model.
They
do
different
things
or
even,
if
you're
doing
some
other
way
of
data
modeling
it.
Those
things
are
going
to
care,
but
it's
left
up
to
the
data
modeling
mechanism
and
not
dictated
by
the
information
model.
M
Four,
we're
going
to
need.
There
is
inherent
in
the
information
model
as
currently
specified.
There
is
nothing
you
can
point
to.
That
is
a
unique
key,
and
so
what
I'm
saying
is
we're
going
to
leave
it
up
to
the
data
model,
to
figure
out
what's
the
right
mechanism
within
the
context
of
that
data
model
as
to
how
to
uniquely
identify
these
entries.
M
M
M
When
you're
looking
at
a
set
of
objects,
you
need
to
be
able
to
address
the
object
that
you
want
right.
You
know
like
if
you've
got
yeah,
and
so
that's
simply,
what
we're
saying
is.
How
do
you
address
the
object
that
you
want
and
in
all
of
our
other
things
you
know
like
when
we've
got
their
interface
lifts?
We
have
a
unique
way
of
identifying
the
interface.
M
Q
A
M
Then
we're
done
in
this
one,
it
should
be
a
quick
one,
and
this
is
something
that
kind
of
came
out
just
in
Mahesh
in
my
recent
discussion,
and
so
we
figured
we
just
toss
it
out
here
we
have
this
the
the
route,
the
calculated
metric
and
the
received
metric,
and
we
previously
had
input
that
we
have
to
have
one
of
these.
M
Okay,
so
in
this
this
is
this
is
a
personal
problem.
This
actually
didn't
come
up
with
Mahesh
and
I
just
wanted
well,
actually
it
sort
of
did
I
just
wanted
to
be
absolutely
sure
that
when
we
reference
an
interface
that
it
can
from
our
interface
list,
when
we
reference
some
other
interface,
where
this
interface
lies,
it
can
be
at
any
layer,
physical
layer,
Mac
layer,
IP
layer,
tunnel
IP.
We
don't
care
what
layer
that
interface
references.
M
So
we
have
right
now
when
I
look
at
the
implementations,
you
know
they're
referencing,
something
like
eath.
You
know
Linux
ETH,
one
which
generally
in
ETH
one
has
the
full
IP
stack
on
it.
It's
really
not
the
specific
ethernet
layer.
It's
this
thing
with
all
of
these
IP
address
stacks
on
it.
If
we
had
say
in
Linux
the
ability
to
reference
the
100
base,
T
interface,
would
that
make
sense
or
are
there
layers
that
don't
make
sense
of
interfaces?
No.
I
M
And
I
think
that
answers
my
question
because
both
yang
in
broadband
form,
we
basically
stack
our
interfaces
and
so
you've
got.
You
know.
Here's
an
ipv6
interface
in
this
ipv6
interface
by
the
way
happens
to
ride
on
this
Mac
interface,
and
this
Mac
interface
happens
to
ride
on
this
physical
interface.
And
so
you
know
your
interface
stack.
But
the
question
is
at
what
entry
point
we're
we're
in
the
stack?
Should
you
be
pointing
and
you're
saying,
I
need
to
be
pointing
for
the
ipv6
answer.
M
I've
got
we're
good,
we're
good
I've
got
my
answer.
Thank
you
and
it
was
not.
The
answer.
I
was
going
to
go
for
if
you
hadn't
stood
at
the
mic,
so
it's
a
really
good
thing.
You
got
up
there
so
we're
in
wgl
see
additional
review
would
be
really
great
in
the
meantime.
We
just
here
just
result
just
about
every
single
one
of
our
issues,
I
think
so
I'm
going
to
be
revving
the
information
model
based
on
this
and
we'll
do
the
Yang
and
the
tier,
181
and
I
think
we're
in
pretty
good
shape.
A
D
H
All
right,
the
status
of
the
document
it
is
of
a
group
document,
as
Donna
has
already
verified.
We
did
manage
to
augment
the
routing
management
interface
as
re
RFC,
which
brought
us
support
for
addition
of
Europe.
The
vrf
instance
is
something
that
is
now
available
for
us
to
also
add.
I
haven't
quite
done
that
as
yet
so
in
the
next
Rev
I'll
take
care
of
that.
H
So
with
the
changes,
the
the
model
now
looks
something
like
this.
As
we
say:
I'm
augmenting,
routing,
83,
RFC,
83
49
in
as
per
percent,
we
have
now
H
back
and
DTLS
port,
which
we
have
added
to
the
yang
on
at
the
bottom.
Is
the
addition
for
web
and
rip
support
under
that
so
issues
so
much
like
Barbara
I
also
have
my
own
github,
which
is
tracking
all
the
issues
related
to
the
data
model.
H
So
Barbara
and
I
had
to
go
back
quite
a
bit
on
this
discussion
of
mandatory
values
and
I.
Finally,
after
several
attend
and
a
good
breakfast,
we
find
Barbara
explained
to
me
that
the
two
attributes
that
she's
trying
to
address
one
is
mandatory
to
implement,
which
is
not
what
this
is
about.
Slide
is
about
and
there's
mandatory
from
a
model
perspective
as
far
as
values
that
need
to
be
provided
for
the
protocol
to
work
correctly.
H
So
I
don't
have
a
slide
on
mandatory
to
implement,
but
just
I'll
just
say,
I'll
wave,
my
hands
and
I'll
explain
this.
Maybe
in
in
the
model,
is
that
we
need
some
capability
of
a
feature
statement
which
allows
and
implemented
to
declare
what
features
they
supported
and
I
and
an
example
of
that
feature
would
be
I
support,
DTLS
capability
or
I
support,
H
Mac,
and
once
you
declare
that
feature,
then
the
model
it's
expected
to
have
all
the
capabilities
defined.
H
If
you
don't
declare
the
feature
that
part
of
the
model
pretty
much
vanishes,
and
you
don't
have
to
worry
about
trying
to
implement
that,
so
that's
mandatory
to
implement
so
now
mandatory
as
far
as
values
that
need
to
be
provided
so
I
think
in
this
particular
case,
what
I've
highlighted
in
red
are
two
parameters
which
I
believe
are
mandatory
parameters
in
what
happens
is
that
when,
when
you
do
specify
that
something
is
mandatory,
when
the
client
tries
to
configure
the
Rueter
or
out
of
with
that
information,
if
those
valadon
exists,
the
client
won't
even
be
able
to
send
that
config
request
to
the
server.
H
H
It
also
means
that
this,
the
scope
of
mandatory
nests
is
limited
to
the
parent.
So
in
this
particular
case,
the
two
parameters
that
are
highlighted
in
red
are
mandatory
only
if
an
h
mac
is
being
specified,
because
if
it's
mac
is
not
even
being
specified,
then
it
doesn't
matter
whether
there
other
parameters
within
which,
within
that
that
are
mandatory,
the
other
clarification
was
around
read-only
attributes
in
the
information
model
in
yang.
Typically,
it
doesn't
really
make
much
sense
to
declare
something:
that's
read-only
mandatory,
so
something
like
Rooter
or
outer
ID
that
Bible
uses
is
read-only.
H
K
Think
what
Joyce
is
trying
to
say
give
its
Knaus
e
I
think
what
Joyce
is
trying
to
say
is
that
you
can
actually
have
an
implementation
of
Babel
that
does
not
have
a
router
ID.
That
is
just
a
like
simple
client
node
on
the
edge
that
doesn't
offer
any
routes,
and
so
there
it
so
that
it's
valid
for
that
to
be
non-mandatory.
K
I
It's
not
quite
clear,
so
does
it
make
sense
to
put
restrictions
on
the
implementation?
Does
the
management
model
put
restrictions
on?
Is
it
legitimate
for
the
management
model
to
put
restrictions
on
the
implementation,
so
it
is
technically
possible
not
to
have
a
Rooter
ID,
but
it
doesn't
cost
much
to
say
we
always
have
a
Rooter
ID.
I
Q
Q
K
H
K
I
M
I
What
we're
hearing
here
is
that
the
natural
data
structure
that
is
used
by
babo
says
that
you
only
need
to
have
a
ruler
ID
if
you're
originating
groups,
if
the
data
model
needs
to
say
that
implementations
that
implement
the
data
model
must
generate
ruther
ID,
even
they
don't
need
us
just
for
the
routing
part
I'm
fine.
With
that
it'll
cost
much
to
generate
64
bits
of
randomness
sure.
H
Okay,
so
that
then
answers
our
question
of
how
we
need
to
update
the
information
model
and
I'll
reflect
that
in
the
young
one
yep,
thanks
for
the
clarification,
okay,
so
I'm,
not
a
security
expert
and
I
just
go
based
on
what
is
there
and
some
of
the
be
remodeled.
So
my
effort
is
to
try
to
just
bring
attention
to
the
fact
that
there
are
other
data
models
that
we
have
to
deal
with
as
far
as
he
setting
up
keys,
key
exchanges
and
everything
related
to
that.
H
There
is
a
draft
on
the
asymmetric
key
and
certificate
support
that
is
in
the
works
in
the
net
count
working
group,
but
none
of
these,
of
course,
have
any
thing
that
supports
key
rotation.
I,
don't
know
if
that's
important,
if
that's
important,
that
I,
we
need
to
figure
out
how
we
will
support
that
and
then
I
think
this
is
the
question
I
think
Julius
brought
up
on
mailing
list
is:
how
do
you
share
keys
across
domains
as
H
Mac
uses,
symmetric
keys?
H
H
I
would
appreciate
trying
to
get
some
configuration
examples
that
I
can
use
not
only
to
validate
the
model,
but
also
showcase
to
people
who
want
to
use
them
on
how
it
we
configure.
So
if
there
are
different
implementations
of
the
model
that
have
different
ways
to
configure
I
would
be
I
would
imagine
that
would
be
very,
not
very
kosher,
but
if
it
is
the
case,
let
me
know,
but
basically
I
would
appreciate
some
examples
that
I
can
use
for
validating
the
data
model.
I
M
A
A
A
I
think
we're
making
great
progress.
We
have
a
big
package
of
documents
already
in
publication,
requested
State
and
more
documents
coming
along,
so
see
you
all
on
the
mailing
list
and
this
Montreal
with
the
next
IETF
meeting
I
think
anybody
have
anything
else.
They
want
to
bring
up
I
guess
not
if
you
did
not
sign
the
blue
sheets
they're
up
in
frontier,
please
come
and
sign
them.
Thanks
again,.