►
From YouTube: IETF100-PANRG-20171116-1330
Description
PANRG meeting session at IETF100
2017/11/16 1330
https://datatracker.ietf.org/meeting/100/proceedings/
A
A
This
is
an
IRT
F
research
group,
the
IRT
F
intellectual
property
rules
parallel
those
of
the
IETF
anything
you
say
here
is
essentially
an
IETF
contribution.
You
should
be
aware
of
rules
about
disclosure
of
IPR
that
may
apply
to
anything.
You
say
here
so
our
agenda
today
we're
gonna
do
a
little
bit
of.
A
A
What
are
the
questions
that
we're
trying
to
ask
and
answer
here,
and
then
we
have
a
set
of
invited
presentations
from
Chris
Pappas,
talking
about
methods
for
path,
dissemination
and
path,
aware
networks
and
from
Laurel
and
choix
talking
about
interfaces
for
path
selection
to
the
first
is
okay,
so
you
have
a
network
that
knows
something
about
its
pounds.
It's
going
to
take
information
about
those
paths
and
put
them
down
to
the
endpoints.
They
can
do
something
with
them.
This
is
sort
of
mechanisms,
a
discussion
and
then,
once
the
endpoints
have
this
information.
A
What
can
they
do
with
it?
And
that's
more
about
their
all's
going
to
be
talking
about,
as
we
pointed
out,
one
of
the
reasons
that
we're
as
a
'print
I'm.
One
of
the
reasons
that
we're
interested
in
this
area
and
research
is
that
there's
a
lot
of
work
going
on
sort
of
in
and
around
the
IETF
that
begins
to
give
us
the
ability
to
see
a
future
where
we
have
endpoints
that
are
aware
of
their
paths,
and
one
of
those
bits
of
technology
is
provisioning
domains.
A
So
Eric
thinker
will
be
talking
about
that,
and
then
we
have
a
C
being
around
en
masse.
Talking
about
Alto
tingrong,
Zhao
I'm
messing
up
all
of
these
names.
I
apologize
talking
about
a
particular
application
of
path,
awareness
specifically
to
congestion
control.
Since
this
is
a
proposed
research
group,
we
have
a
lot
of
time
at
the
end
for
sort
of
like
general
discussion.
A
One
of
the
things
that
we're
looking
for
here
is
interest
in
this
topic:
sort
of
at
the
interface
between
research
and
the
IETF
sort
of
what
of
what
should
we
be
looking
at
as
chairs
for
the
agenda
for
next
time?
Is
there
interest
in
working
on
you
know
any
documents
talking
about
frameworks
or
looking
at
things.
They've
already
been
done
within
the
research
groups,
are
gonna
have
a
relatively
hopefully
a
relatively
long.
An
interesting
discussion
about
that.
So
I
think.
C
A
A
Owe
him
like
three
or
four
beers
anyway?
So
it's
good,
so
we've
identified
a
common
theme.
This
list
of
working
groups,
boss,
and
so
us,
and
so
on,
is
not
meant
to
be
exhaustive.
So
if
we
missed
out
on
yours,
please
don't
feel
bad
of
sort
of
path,
awareness
and
a
lot
of
research
sort
of
going
on
around
the
edges.
The
IETF
that
the
obvious
one
here
is
NP
TCP
as
I
understand
it.
There
was
just
a
lunch
meeting
of
people
trying
to
figure
out
how
to
do.
A
Rebinding
in
a
multipath
friendly
way
for
quick,
so
that's
not
what
quic
is
working
on
multipath
they're,
not
working
on
multipath
right
now,
but
that
is
a
future
development,
hybrid
access
approaches,
path,
control,
approaches,
dynamic
interface
and
transport
selection
and
then
path,
signaling
works.
So
the
fact
that
there's
a
lot
of
this
work
and
nothing
like
an
architecture
or
a
framework
around
how
you
would
put
it
together,
you
know
architectures
and
frameworks
not
really
what
the
IETF
does,
but
this
did
seem
like
a
theme
that
a
new
research
group
should
explore.
A
So
for
those
of
you
who
have
no
idea
what
I'm
talking
about,
because
you
were
over
in
ideas.
The
last
time
we
had
that
that
was
an
unfortunate
conflict
I
as
I
understand.
We
have
an
unfortunate
conflict
this
time
with
a
routing
working
group.
So
at
some
point
we're
gonna
have
to
actually
get
the
routing
people
in
here
is.
A
This
is
sort
of
what
the
internet
looks
like
right
now.
You
have,
you
know
an
endpoint
and
that's
not
working
at
all
and
another
endpoint
and
you
send
a
packet
to
the
to
one
of
the
gateway
router.
And
that
knows
how
to
afford
it,
based
on
some
control
plan,
information
that
is,
you
know,
distributed
or
centralized,
usually
distributed,
and
but
you
can
think
of
it,
you
can
abstract
it
away,
is
just
sort
of
the
control
plane
tells
the
14
plane
what
to
do
in
the
Ford
and
does
it
the
endpoints.
A
The
control
plane
in
some
way
is
pushed
all
the
way
to
the
to
the
edges.
You
have
endpoints
with
transport
protocol
stocks
that
are
path
aware
at
each
end
of
the
network
and
can
cooperate
with
the
routers
in
order
to
select
path.
So
here
you
see,
you
have
path
aware
stacks
on
the
colors.
Don't
look
great
there,
so
you
have
a
path
or
stack
on
either
side
of
that,
and
they
can
select
multiple
paths
through
the
network
that
are
joint
or
disjointed
various
points.
A
They
have
enough
information
about
those
paths
to
be
able
to
make
informed
choices
about
which
path
they
want
to
send
stuff
down.
So
yeah
again,
what's
interesting
about
path,
aware
networking.
Obviously
the
routers
are
aware
of
the
paths
that
the
traffic
is
taking.
This
is
I
mean
this.
We
solved
this.
It's
BGP
right.
Well,
not
really
we're
interested
in
extending
this
to
the
edge,
so
discovery
of
has
explicit
Association
of
properties
to
pass
by
endpoints
and
endpoint
selection
of
paths.
A
Since
we
noticed
this
is
happening
anyway
and
things
like
MP,
TCP
and
other
multiple
interface
work,
we
figured
that
we
should
be
explicit
about
it.
This
research
group
has
one
document
that
I
wrote
craft
Trammell,
panna,
RG
questions.
This
is
essentially
just
a
list
of
open
questions.
This
is
not
anything
from
the
chair.
This
is
just
I'm
interested
in
this
topic,
and
this
is
what
I
think
the
open
questions
on
that
are.
A
Please
have
a
look
at
it
and
let
us
know
on
the
pan
RG
list.
If
the
questions
there
aligned
with
what
you
think
the
questions
in
this
space
are.
So
we
also
intend
that
kind
of,
as
an
introductory
document
to
this
work
for
people
coming
in
from
outside
the
research
group.
So
if
you
are
coming
in
from
outside
the
research
group-
and
you
try
to
read
that
document
and
you
are
completely
confused
and
have
no
idea
what
I'm
talking
about
there-
that's
also
useful
feedback.
A
It
might
be
useful
to
have
a
catalog
of
failed
transport
layer,
signaling
approaches
just
a
you
know,
a
quick
document
that
says:
don't
do
this
that
didn't
work.
Don't
do
this
that
didn't
work,
but
also
with
some
information
about,
like
don't
do
this
that
didn't
work
and
here's.
Why
right
like
not
just
complaining
about
stuff
I,
understand
like
to
complain
a
lot
me
IETF,
but
sort
of
looking
at
ways
for
it
like
constraining
the
space
for
path,
aware
approaches
so
that
we
can
make
a
little
bit
of
progress
here.
A
One
question
that
I'd
like
to
ask
in
the
in
the
discussion
phase
today
is
if
there
are
any
other
documents
that
people
have
either
energy
to
write
or
think
that
other
people
should
have
energy
to
write
and
we'll
try
to
find,
as
jurors
will
try
to
find.
You
know,
match
ideas
and
authors
to
to
things
that
we
can
do
here.
A
We
are
considering
proposing
a
workshop
on
path,
aware
networking
for
sigcomm
2018
in
budapest.
If
this
were
accepted,
this
would
replace
our
fall
2018
to
fall,
2018
panarchy
meeting
but
fall
2018
meeting
I
think
right
now
is
tentatively
scheduled
for
Asia
I
hope
we'll
narrow
that
down
at
some
point
before
we
have
to
buy
plane
tickets.
So
this
would
would
essentially
replace
that
meeting
and
we'd
have
it
in
in
Budapest
instead.
A
A
F
A
A
Okay,
so
thank
you
so
yeah.
Let
the
minute
show
that
Lars
corrected
me
because
they
don't
know
what
I'm
talking
about.
G
Well
sector
does
a
treasurer
for
the
sick.
If
you
put
in
the
rupture
proposal
and
in
the
requirements
that
you
want
from
the
venue
you
put
that
in
that
they
will
budget
for
it.
If
they
can
get
it
I
mean
they
might
say
that
it's
not
possible
on,
and
then
you
will
need
to
decide
whether
you
still
want
to
hold
a
meeting.
But
if
you
put
it
in
the
you
need
to
put
it
in
a
proposal
because
otherwise
it
won't
be
in
the
budget
and
rain
won't
happen.
Yeah.
F
Spencer
Dawkins
I'm
actually
got
up
and
walked
slowly
as
the
mic.
So
this
is
about
the
previous
slider
to
back
that,
where
you
were
talking
about
what
yeah
the
catalog
of
failed
transport
layer,
signaling
approaches
that
seemed
like
there.
That
seemed
like
a
really
useful
thing
to
do
and
if
it
is
helpful
for
me
to
collect
those
experiences,
I
can
do
you
know.
F
A
F
A
You
and
I
think
that's
all
I
have
for
my
chair,
slides
so
yeah.
So
if
you're
interested
in
helping
out
with
the
workshop,
we
are
even
just
looking
for
people
who
are
interested
in
the
area
who
would
be
willing
to
review
papers
in
this
space.
One
of
the
things
that
you
need
for
a
good
workshop
proposal
is
a
read
of
a
list
of
people
for
the
TPC.
So
please
do
send
that
to
I
guess.
Pan
orgy
chairs,
at
IRT,
F
dot
org
should
work
Alison.
A
Do
you
know
if
we
have
chair
aliases
in
the
data
cracker
yeah?
We
have
here
at
clearly
yeah,
oh
yeah.
Okay.
We've
used
that
okay,
good
yeah
nevermind
for
good,
so
yeah.
Please
send
up
an
orgy
chairs
at
IRT
org
interest
in
serving
there.
So
thank
you.
Next
up
is
Chris.
Papas
and
I
will
switch
you
over
to
keynote.
E
Okay,
it's
good
afternoon,
everyone
and
thank
you
for
coming
to
my
presentation
about
dissemination
of
paths
in
path
aware
networking,
so
I
will
start
by
the
motivation
of
the
part
of
our
networking
research
group
and
stealing
from
the
energy
agenda
slides
from
ITF
99.
One
of
the
questions
that
is
Richard
she
wants
to
answer
is
how
does
our
awareness
extent
to
the
edge
and
by
edge
meaning
the
endpoints?
So
in
order
to
answer
this
question,
there
are
many
other
sub
questions
to
be
answered.
E
Another
question
to
be
answered
is
how
do
endpoints
discover
path
properties
and
how
do
they
associate
this
properties
to
the
paths
and
this
has
to
do
with
dynamic
properties,
meaning
that
they
change
often
such
as
congestion
or
utilization
stages
of
some
links
on
the
paths
and
more
static
properties
such
as
capacitive
links
on
the
path
or
derived
properties
about
a
set
of
paths
at
source
path.
Disjointness
one
more
problem
is
how
to
endpoint
select
paths
and
how
then
do
they
inform
the
network
about
the
paths
that
they
want
to
use.
E
So
the
focus
of
this
research
group
is
the
endpoint
and
how
to
give
more
path,
control
to
the
endpoints,
meaning
more
path,
information
and
more
path
choice.
But
in
order
to
talk
about
how
path
control
looks
at
the
endpoints,
we
will
have
to
talk
about
how
the
network
assists
the
endpoints
in
realizing
this
higher
level
of
path
control.
So
by
the
term
path,
dissemination
I
will
refer
to
through
three
things
or
three
steps.
The
first
one,
the
first
one
is
part
construction,
and
this
includes
the
steps
of
exploring
the
topology
and
then
creating
paths.
E
On
top
of
that
apology
and
this
step
can
be
performed
by
the
network
alone
or
by
the
network
and
the
endpoints
together.
The
second
step
is
that
the
endpoint
will
have
to
select
or
choose
among
these
constructed
paths
and
in
the
end,
the
networks
will
have
to
inform
them
at
the
endpoints
will
have
been
formed,
the
network
about
which
parts
to
use
something
that
I
will
refer
to
a
spot
representation.
E
So
to
give
you
an
overview
about
how
about
dissemination
looks
inside
the
first
step
is
path
creation,
and
this
falls
down
to
topology
exploration,
which
is
done
through
a
process
called
beaconing,
and
then
we
have
path
segment
construction,
namely
creating
up
segments
and
down
segments
which
I
will
explain
soon
assume
how
they
work.
Then
this
path
segments
are
registered
and
this
whole
process
in
science
is
performed
by
the
network.
E
E
So
going
to
the
first
step,
the
poet
exploration
happens
through
the
process
called
beckoning
and
this
process,
the
beginning
process,
starts
at
core
a
SS
and
that
term
choruses
is
not
really
well
defined,
but
intuitively
consider
this
that
this
Ras
is
that
exists
high
in
the
era
T
of
the
interconnection
market.
So
they
have
mostly
peering
relationships
and
customers.
E
E
So
one
interesting
aspect
is
that
this
process
of
beckoning
can
be
used
as
a
transport
medium
in
order
to
disseminate
path
properties,
and
it
is
more
appropriate
for
disseminating
more
static
path
properties.
So
in
this
beacon
messages,
each
a
s
on
a
path
can
input
some
information,
such
as
MTU
capacities
of
links
on
the
paths,
some
partial
path,
policy,
supported
services
and
supported
crypto
protocols.
E
Then
we
have
the
path,
segment,
construction
process
and
this
beacons
as
they
have
traversed
the
topology.
They
now
contain
some
path
segments
that
can
be
used
by
autonomous
systems
in
order
to
reach
the
quarry
ASIS
that
started
this
beginning
process.
So
there
are
up
segments
which
are
spot
segments
that
can
be
used
by
non
quarry
Isis
to
reach
quarry
axis.
E
E
Segment
that
may
be
used
in
order
to
breach
the
up
and
down
segments
and
I
will
show
an
example
shortly,
so
this
up-and-down
segments
will
be
registered
and
looked
up
and
more
specifically,
a
enas
will
select
a
subset
of
these
path
segments
to
be
used
by
the
local
endpoints
to
reach
the
core
and
another
subset
will
be
selected
so
that
other
endpoints
can
reach
the
a
s.
Now
this
path
segments
will
be
registered
with
a
special
lookup
service
that
is
operated
by
path
servers,
and
there
are
two
types
of
this
service.
E
First,
we
have
path
servers
operated
by
choruses,
and
this
is
where
these
down
segments
are
initially
registered
with,
and
they
will
serve
these
down
segments
to
remote
end
points.
Then
there
are
path,
servers
that
are
running
at
non-core
aassists
and
they
serve
a
dual
purpose.
First,
they
serve
up
segments
to
local
endpoints,
and
then
they
resolve
and
cast
responses
of
how
to
reach
other
remote
end
points.
E
So
then
we
come
to
the
second
step,
which
is
to
combine
the
paths
in
order
to
create
a
full
end-to-end
paths
and
then
select
among
them,
and
here
we
have
the
lookup
of
path
segments,
and
at
this
step
we
have
two
translations
that
take
place.
The
first
installation
is
that
the
end
point
will
contact
a
name
server
with
with
a
name
in
order
to
obtain
the
AES
in
which
that
server
resides
in
after
this
translation.
E
You
can
see
in
the
core
to
bridge
between
the
up
and
down
segments
or
alternately,
for
example,
if
an
up
segment
and
the
down
segment
advertise
or
contain
the
same
peering
link,
a
shorter
end-to-end
path
can
be
constructed.
In
the
end,
there
will
be
multiple
end-to-end
paths
and
the
end
point
can
select
one
or
more
of
them
for
its
communication,
this
two
part
representation
and
how
the
endpoint
will
inform
the
network
about
the
selected
path.
E
A
difference
is
that
in
Neera,
the
endpoints
are
part
of
this
path,
vector
component
that
is
used
to
discover
the
up
graph.
Now
part
of
these
paths
will
be
selected
and
then
they
will
registered
with
our
name
to
route
lookup
service,
so
that
other
endpoints
can
look
them
up
in
order
to
construct
end-to-end
paths
about
path.
E
Selection,
the
sender
will
choose
a
path
from
its
up
graph
and
then
it
will
combine
that
up
path
with
looked
up
paths
in
order
to
construct
this
final
end-to-end
path
and
Pat
representation
in
ear
is
different
in
that
it
is
implemented
through
source
destination.
Addresses,
for
example,
ipv6
addresses,
and
this
addresses,
so
the
source
address
will
encode
an
app
path
and
the
destination
address
will
encode
the
down
path
and
allocation
of
others
is
in
error.
E
So
padlet
routing
is
a
more
generic
routing
architecture
and
it
defines
two
basic
concepts.
The
first
one
is
V
notes
or
virtual
notes,
which
are
created
by
na
s
and
V
notes
represent
routing
structure
in
the
network
of
an
AAS,
for
example,
there
could
be
one
we
note
per
outer
or
one
we
note
per
outer
interface.
Then
a
palette
is
a
sequence
of
such
V
notes,
along
which
na
s
is
willing
to
forward
packets.
E
So,
with
respect
to
the
apologi
exploration
and
construction,
the
routers
will
announce
their
v
nodes
to
their
neighbors
and
then
the
neighbors
can
combine
these
v
nodes
in
arbitrary
sequences
and
they
can
also
combine
paths
let's
and
stitching
together
in
order
to
create
longer
paths.
Let's
part
vector
is
then
used
in
order
to
disseminate
these
putlitz
and
v
notes
to
the
neighbors,
but
the
spot
vector
component
in
path
with
routing
does
not
enforce
any
policy.
E
At
the
same
time,
there
are
many
challenges
that
remain
unanswered,
so
there
is
a
lot
of
work
to
be
done
in
the
area,
for
example,
how
to
convey
dynamic
path
properties.
How
often
do
we
need
to
disseminate
such
information?
How
much
information
is
needed,
for
example,
for
obtaining
accurate
congestion,
information
or
latency
information.
E
Another
issue
is
how
should
the
end
point
to
look
the
end
point
to
lookup
service
interface
be
implemented
and
one
more
interesting
issues
how
to
disseminate
or
to
convey
this
part
information
higher
up
in
the
networking
stock
closer
to
the
application
and
I
think
some
of
the
subsequent
talks.
We'll
talk
will
describe
some
of
these
issues.
So
thank
you
for
your
attention
and
we'll
be
happy
to
answer
questions.
H
Hello,
Tommy
Polly
Apple,
so
thank
you
for
sharing
this
very
cool
to
see
what
I
think
is
most
interesting
is
how
you've
broken
down
the
various
categories
and
shown
the
comparisons
across
them.
I.
Imagine
for
a
lot
of
us,
the
kind
of
Scion,
architectures
and
stuff
it
seems
you
know
quite
far
off,
but
I
think
these
principles
when
we
look
at
other
solutions
like
what
we'll
see
you
later
on
with
like
PVD
and
stuff,
seem
to
fall
into
very
similar
categories
of
how
these
could
be
represented.
E
I
talked
mostly
about
the
networking
parts
or
you're
talking
more
about
this
interface
between
the
network
and
the
application.
So
when
talking
about
part
awareness,
there
are
multiple
things,
so
you
could
have
like
a
fully
path
aware,
endpoint
or
you
could
convey
some
partial
information.
So
if
a
first
step
would
be,
for
example,
just
to
give
information
to
the
end
point
about
the
available
parts
without
giving
an
option
to
select
parsimony.
So,
for
example,
I
don't
know
about
these
jointness
of
these
parts.
E
So
it's
really
a
bit
hard
to
imagine
from
based
on
today's
internet
architecture,
but
the
first
step
would
be
to
giving
information
about
the
current
status
of
the
past
that
could
be
used
by
the
end
points
that
could
be
probabilistic
information
because
it's
really
hard
to
debug
your
it
information,
but
ideally
it
the
ideal
goal.
I
think
would
be
to
go
and
if
the
actual
actual
control
to
the
end
point
so
that
they
can
combine
these
parts
in
the
way
they
want.
E
E
I
E
So
in
the
beginning
process,
if
there
is
a
peering
link,
is
peering
link
would
be
announced
down
to
the
ASIS
that
are
below
something
that
would
not
that
doesn't
happen
today.
At
least
BGP
provides
this
privacy,
but
on
the
other
hand,
there
are
approaches
that
enable
you
to
determine
what
the
relationship
is
among
to
a
SS.
E
If
you
look
at
a
few
routing
tables
and
see
how
they
look
from
different
vantage
points,
so
this
information
is
partly
disclosed
today,
but
I
agree
with
you
that
this
relationship
become
more
explicit
through
such
a
process
because
they
are
clearly
put
into
the
beacons
as
they
traverse
the
topology.
Today,.
E
I
E
But
I
mean
BGP
I
mean
referring
to
the
typical
BGP
policy
is
not
like
pgps
the
dissemination
method,
but
through
the
typical
gower
expert
of
progressing
this,
how
these
policies
are
implemented
through
BGP,
so
indeed
the
providers
don't
want
to
do
it
and
that's
why
they
just
announced
based
on
the
garlic
sort
model.
So
you
cannot
determine
this
information
today,
just
by
looking
at
one
table.
J
E
So
what
I
described
clearly
neglects
this
issue
about
representing
multi-part
information
inside
the
Naas
and
I
would
say
that
one
problem
is
scalability,
so,
inter
domain
routing
gives
us
a
good
abstraction
to
scale
this
Multi
multi
path.
Support.
One
problem
of
this
is
that
you
cannot
get
accurate,
disjoint
ins
information,
so
you
can
have
partly
disjoint
ins
information
and,
let's
say
the
more
disjoint
some
parts
are
at
the
inter
domain
level.
You
can
assume
that
there
will
be
more
descendants
at
the
intra
domain
level
if
they
are
purely
designed,
for
example,
at
the
inter
domain.
K
K
A
D
Aaron
Faulk
I
worked
on
a
project
called
Farah
which
was
sort
of
a
precursor
to
deira,
which
I
guess
an
unfortunate
selection
of
names.
That's
clever
computer
scientists
for
you
and
the
way
that,
when
we
were
working
on
Farah
and
I,
think
that
near
is
the
same.
Although
it's
been
a
while,
since
I've
read
the
paper,
it
was
a
somewhat
anachronistic
idea
of
allowing
end
users
to
select
their
long
distance
providers
when
they
use
the
intern,
and
the
whole
idea
was
to
give
some
way
of
Anna
to
provide
hints
to
the
path
that
you
wanted.
D
D
And
so
the
the
point
being
that
it
was
really
about
providing
some
hints
so
that
a
end
user
could
express
what
they
wanted,
which
kind
of
goes
to
stuff.
That's
going
on
in
taps
and
and
then
this
is
sort
of
a
mechanism
for
how
you
communicate
that
back
and
forth.
So
not
a
question
just
trying
to
add
something
to
this
guy's
cool.
A
Thanks
Alison,
do
you
get
anything
on
jabber,
all
right,
I'm,
if
you,
if
you're
still
on
and
listening
Donna,
I'm,
sorry
that
we
couldn't
get
your
question
and
if
you
could
send
it
to
the
list,
I
think
Chris
is
on
the
list
and
we
need
to
neither
session
there.
I'm
sorry
that
we
couldn't
get
that
working
for
you
all
right.
So
next
is
Laura.
M
M
This
topic
overlaps
a
bit.
What's
going
on
in
the
taps,
working
group
in
particular
I'm,
going
to
talk
about
the
socket
API
and
efforts
to
replace
it
with
the
post,
sockets,
API
and
I
will
try
to
link
this
to
research
perspectives
and
some
research
we
have
been
doing,
and
that
goes
in
the
same
direction
so
using
multiple
copies
of
the
same
components
to
improve
real
reliability,
availability,
performance
or
capacity
of
computer
systems.
It's
very
old
idea:
multiprocessor
systems,
multi-core
CPUs,
multi,
disk
storage,
I,
will
a
fairly
common
places
today.
M
But
what
about
narrow
paths
so
I
think
it's
widely
accepted
that
multi
path.
Communication
can
help
to
improve
reliability
if
one
path
goes
down
and
B
TCP,
for
example,
can
how
to
effortlessly
and
quickly
switch
path.
It
helps
with
mobility.
If
client
disconnects
from
a
Wi-Fi
network,
for
example,
he
can
gracefully
switch
to
LTE,
for
example,
but
this
is
not
entirely
multi
path.
It's
just
switching
quickly
from
a
single
path
to
another,
and
then
multiple
communication
also
helps
with
the
total
amount
of
bandwidth
but
I,
say:
I.
M
It
would
be
preferable
to
use
a
single,
very
good
network
path
and
multiple
paths
and
some
I
think
that
this
perfect
that
could
be,
for
example,
an
optical
fiber
link,
I
think
there's
no
such
thing
as
a
perfect
network
path,
whereas
radio
waves
travel
close
to
the
speed
of
light
in
vacuum.
On
the
surface
of
the
earth,
information
only
travels
about
2/3
of
that
speed
in
in
fiber,
so
latency
is
significantly
increased
when
using
fiber
and
so
wireless
has
this
latency
advantage,
but
it
also
comes
with
other
drawbacks.
M
It
has
also
the
mobility
advantage,
obviously,
but
it
has
drawbacks
in
terms
of
Gerry
run
all
kinds
of
security
implications.
So
I,
don't
think,
there's
a
perfect
network
path
and
also
are
arguably
in
today,
in
today's
Internet
there's
a
fairly
limited
amount
of
network
paths
between
one
source
and
one
destination.
But
we
don't
know
what
tomorrow's
networks
be
made
of.
M
So
there
are
a
number
of
projects
you
provide
or
borne
Internet
connectivity
through
balloons,
drones
or
lor
bit
satellites
and,
as
we
saw
in
the
previous
talk
future
internet
architectures,
like
Zion
in
particular,
will
provide
more
path.
Control
to
an
hosts
and
paths.
Control
means
that
and
hosts
may
use
multiple
paths
simultaneously.
M
So
when
I
say
I
work
on
the
path
communication-
and
here
o
notes,
paths
TCP-
and
yes,
this
is
part
of
the
story
not
about
TCP-
is
great
because
it
hides
all
the
multipath
communication
specificities
from
the
application.
So
it
means
that
an
existing
application
can
use
multiple
TCP
almost
out
of
the
box,
but
it
also
means
that
multipath
to
speed
is
fairly
restricted
in
what
it
can
do
to
best
fulfill
each
applications.
Objective.
M
So
I
also
hear
sometimes
should
we
optimize
for
throughput
or
latency.
She
does
the
application
care
about
throughput
all
agency
and
I
think
in
most
cases
it's
both
so
I'm
more.
Of
course,
we
cannot
optimize,
it's
not
straightforward
to
optimize
for
both
through
boots
and
latency,
but
I
think
in
most
cases,
both
matter.
So
a
most
a
more
meaningful
question
is
that
the
application
need
a
reliable
transporter
and
are
inaudible
one.
M
So
we're
browsing,
for
example,
legacy
matters,
but
the
critical
requirements
is
that
contents
be
delivered
completely
incorrectly
eventually,
and
then
we
have
applications
that
refer
to
trade.
This
complete
reliability
for
for
a
very
low
latency.
So
for
something
like
a
simple
request
and
response
protocol.
M
Multiple
communication
might
not
be
the
most
interesting.
The
best
strategy
would
be
to
send
along
the
best
path,
assuming
we
know
what
that
is,
or
along
the
best
paths,
but
I
think
first,
one
very
interesting
scenario
was
real-time
communication,
because
some
applications
have
very
high
throughput
requirements
and
they
have
also
very
strict
latency
requirements.
M
So
applications
have
been
using
the
classic
Berkeley
Circuit's
API
for
decades,
and
arguably
it
was
so
successful
because
it
was
so
simple,
but
we
also
think
it
might
be
too
simple
for
modern
applications
and
networks.
It
offers
mainly
to
abstractions
suck
stream,
so
a
TCP
stream
and
so
Jairam
for
UDP
data
grams.
M
But
the
reality
is
that
the
network
is
not
freely
file
and
datagrams
do
not
offer
much
functionality,
they're
mainly
glorified
IP
packets
and
then
there's
a
third
abstraction,
such
htech
packets,
which
is
mainly
bound
to
the
HTTP
stream
control
transmission
protocol,
which
is
not
not
very
common
and
not
extremely
deployable
in
today's
internet
and
it's
multipath,
but
not
really.
It's
mainly
for
failover.
M
So
here's
a
simple
example
of
where
classic
models.
For
short,
so
here
we
have
a
source,
a
simple
application
that
generates
a
constant
rate
I
concerned
at
a
rate
of
10
megabits
per
second,
and
the
message
is
a
lifetime
of
1
seconds.
For
some
reason
it
might
be
a
live
stream
or
multiplayer
gaming
or
something-
and
we
have
two
paths
with
very
contrasting
characteristics.
On
the
left-hand
side,
we
have
high
bandwidth,
high
delay,
high
loss
path
and,
on
the
right
hand,
side.
M
We
have
a
low
bandwidth
flow
delay,
low
loss
path,
and
so
it's
fairly
easy
here
that
the
optimal
strategy
is
to
send
all
data
initially
on
the
left
on
the
left
hand,
side
path
and
then
because
it
has
a
high
loss.
All
the
packets
that
are
lost
can
be
retransmitted
on
the
very
reliable
path
on
the
right
hand,
side.
M
So
really
here,
what
makes
us
achieve
our
goal
is
path.
Diversity
is
the
fact
that
we
have
so
diverse
paths,
so
we
addressed
this
problem
a
generalization
of
this
problem
and
these
different
paths
each
bearing
possibly
different
characteristics
in
a
paper
that
we
published
recently
at
DSM,
and
so
we
formulated
it
as
a
linear,
optimization-
and
here
we
have
a
similar
example
with
you,
simple
paths,
and
we
can
see
the
result
of
using
each
path
and
using
a
multi
path
and
optimal
respect
strategy.
M
So
the
objective
is
really
to
have
a
simple
but
but
more
powerful
API.
So
the
main
question
now
is:
what
should
be
the
knobs
and
dials?
What
should
the
application
tell
the
transport
layer
so
that,
so
that
the
transport
layer
can
make
appropriate
path?
I
gets
you
two
paths
assignment
decisions
and
also
what
should
be
the
dials?
What
should
the
application
be
able
to
learn
about
the
transmission
of
its
messages?
Should
it
just
receive
the
messages
when
they're
ready
or
should,
should
it
be
able
to
you?
M
So
I
wouldn't
mainly
talk
about
the
the
properties
that
the
application
can
set
for
the
messages.
The
first
one
is
one
I've
already
talked
about
a
lifetime
or
deadline
for
a
partial,
read
ability.
The
second
one
is
a
priority
or
more
exactly
niceness,
which
Express,
which
messages
should
be
delivered.
First,
then,
there's
dependence
this
fine.
If
other
messages
should
be
delivered
beforehand.
M
Adam
patents
specifies
whether
it
is
safe
to
use
and
in
situations
that
may
cause
message
to
be
delivered
more
than
once,
and
immediacy
is
just
to
say:
do
not
wait
to
combine
this
message
with
other
messages
or
parts
thereof
so
again
about
a
lifetime.
I
think
TCP
versus
UDP
is
often
a
false
dichotomy.
There's
something
that
lies
in
between
and
that's
partial
reliability,
so
packets
with
a
lifetime
can
be
reference,
method
may
be
retransmitted,
but
only
for
a
limited
period.
M
I
think
this
is
a
meaningful
property
because,
as
I
said,
we
cannot
optimize
for
both
latency
and
throughput
together,
it's
typically
a
trade-off,
so
that's
one
way
to
specify
strict
latency
conditions
and
there
exists
transport
protocols
that
use
this
concept
of
partial
reliability.
There's
a
personal
reliability
extension
to
the
SSB
protocol
and
there's
a
variant
of
this
pickle
deadline
of
water,
2030,
TP.
M
Then
there's
niceness,
which
is
just
the
inverse
of
priority,
and
so
it
goes
together
nicely
with
lifetime
because
as
both
lifetime
and
niceness
increase,
priority
decreases.
So
niceness
is
sort
of
the
equivalent
of
lifetime
for
reliable
transport
and
by
default
it's
zero,
meaning
that
by
default
messages
should
be
delivered
with
the
highest
possible
priority.
But
then
it
can
be
reduced
arbitrarily
and.
M
M
Then,
aside
from
message
properties,
an
application
may
want
to
set
policies
specifying
which
features
of
the
transport
protocols
it
prefers
or
which
paths
or
interfaces
it
prefers
to
use
for
our
cost
performance
or
security
or
privacy
reasons.
There
are
multiple
domains
to
do
that.
It
can
be
at
the
application
level
or
the
get
me
a
user
policy
or
system
policy.
So
the
first
example
is
that
Wi-Fi
might
be
preferred
to
LTE
for
cost
reasons.
M
So
so
that's
the
bad
news,
but
the
good
news,
I
think
is-
is
that
the
assuming
that
path
characteristics
are
fairly
stable.
The
the
problem
is
not
necessarily
solved
for
each
packet
and
then
the
other
piece
of
good
news
is
that
not
finding
the
optimal
solution
does
not
mean
that
we
cannot
find
very
good
solutions
with
your
e6
and
simplifications
and
finally,
I
think
that's
more
expressive.
Api's
such
as
the
post
second
CPI
could
drive
a
new
multi
path,
fish
search
and
a
new
kind
of
missed
path,
research,
Thanks,
so
yeah.
A
Hi
Brian
Trammell,
not
as
chair
so
I,
have
two
quick
questions
actually
go
back
to
your
last
slide.
It
might
be
a
little
I
could
see
it
from
over
there.
It
might
be
a
little
bit
hard
to
see
but
yeah,
that's
linear
number
of
pounds
in
the
bottom
and
logarithmic
time
for
computations
so
yeah.
A
Clearly
we
can't
do
this
for
every
packet
unless
we
have
yeah
more
than
two
pads
and
it
might,
there
might
be
an
optimization
loop
that
you
can
do
here
where
you
could
say:
okay,
well,
I'm
gonna
limit
the
number
of
paths
just
arbitrarily
in
order
bit,
but
that
okay,
that's
scary,
but
yeah.
That's
why
it's
research!
Let
me
you
know:
post
Achatz,
co-author
taps
participant
pat
on.
Let
me
see
if
I
can.
A
A
O
Okay,
my
name
is
Rick
link
and
I'm.
Speaking
in
the
name
of
many
authors
in
at
least
two
of
them
are
in
the
room,
so
he
and
Tamia
here,
maybe
others
we're
presenting
at
the
inter
area,
something
called
corrosion,
a
domain
that
can
have
some
intersection
with
the
purpose
of
this
research
group,
for
instance,
nowaday
networks
are
becoming
much
more
complex
than
they
were
before.
Now.
O
If
you
look
about
the
PC,
which
is
there,
it's
Capel
physically
two
three
networks
it
can
be
associated
through
the
phone
doing
tethering
over
Bluetooth
can
be
connected
to
the
Wi-Fi
network
of
the
enterprise
and
can
be
a
wired
connection
to
the
enterprise
as
well.
The
enterprise
itself
can
have
two
service
provider.
Giving
two
different
paths,
including
connectivity
to
the
Senate,
is
a
VPN,
so
at
least
three
interface
here
and
at
least
three
four
paths
actually
to
the
Internet
one
here:
the
phone
service
provider,
one
switch
provider
to
and
the
VPN.
O
O
Selecting
one
pass
is
correct,
but
in
some
case
you
want
to
be
to
have
a
symmetric
path,
so
we
need
to
find
a
way
in
a
PV
six,
because
in
a
PV
six
we
don't.
If
not
you
don't
want
to,
if
not
because
it's
be
a
stupid
idea,
a
similar
stuff,
a
problem
that
has
been
raised
in
the
on
networking
group
at
the
last
tiny
TF
by
Ted
lemon.
If
you
have
to
interface
or
to
ipv6
prefix-
and
you
select
your
resolve-
let's
say
IETF
doc.
O
By
using
the
eighty
interface,
you
will
get
the
CloudFlare
CDN,
the
closer's
to
your
LTE
provider.
In
my
case
it
will
be
something
in
Brussels
in
Belgium,
oh
and
now,
I
I
want
to
go
to
this
ITF
dot.
Orc
address
VR
my
Wi-Fi
here,
so
we
need
to
travel
from
Singapore
to
Brussels,
even
if
CloudFlare
add
a
pop
in
a
cache
here.
So
we
have
multiple
aspect
here
out
to
selector
physically
a
pass
one
address.
O
We
can
use
it
the
same
another
way.
Even
if
you
have
a
single
interface
in
a
pv6,
we
can
have
multiple
addresses
per
interface
from
different
prefixes,
in
this
case,
coming
through
the
same
router,
but
could
have
been
coming
from
five
six,
seven,
ten
routers
on
the
same
layer
to
domain
now
by
selecting
the
right
source
address,
then
the
network
itself
can
route
the
packet
through
traffic
engineering
s
a
v6,
normal
rsvp-te
or
whatever,
or
selecting
basically
at
LTE
or
satellite
or
whatever
link.
O
O
A
provisioning
domain
is
basically
the
consistent
set
of
information.
Then
you
can
use
to
connect
to
the
internet,
mainly
when
I
pee
address
on
interface,
one
DNS,
recursive
server,
maybe
Elena's
search
list
next
up
and
a
few
things
Langley
meant
to
you.
Maybe
now,
until
now,
most
of
the
network
I
mean
computer
in
the
80s
I
get
only
one
connection.
So
by
default
it
was
a
single
PVD,
no
choice
for
the
application.
Now
I've
seen
a
previous
slide.
O
You
may
have
a
you,
must
have
multiple
choice,
so
you
must
have
multiple
pvd's
and
we
may
want
to
name
them,
and
this
is
part
of
this
draft
to
put
a
name
to
a
PVD.
So
an
application,
for
instance,
can
select
one
specific
PVD
and
we'll
get
a
specific
set
of
consistent
information
and
in
the
same
shot
we
are
extending
the
notion
of
PVD,
which
was
just
layer
tree
information
the
to
do
tree
information
to
go
to
the
application
layer
by
doing
a
two-stage.
O
O
Because
it's
two-stage
so
in
the
array
you
receive
the
layer,
2
layer,
2
information
and
then
based
on
the
fully
qualified
domain
name,
you
use
HTTP
over
TLS
and
fetch
this
well-known
URL,
the
container
JSON
file
that
contains
some
metadata
or
application
data.
So
there's
an
example
here
of
this
JSON
file.
You
can
have
a
name
which
is
more
readable
than
a
fully
qualified
domain
name.
O
It
could
be
even
a
localized
name
put
either
in
French
or
they
can
be
in
Japanese,
Korean
or
whatever
to
help
the
end-user
to
make
a
decision,
but
we
can
also
put
there
some
characteristic
of
the
network
until
now,
we
only
put
them
first
up
so,
let's
say
between
the
CPE
and
the
access
network,
because,
typically
you
know
the
maximum
bitrate
and
the
minimum
latency.
If
it's
a
satellite
line,
you
know
it's
capable
of,
but
ten
or
hundred
of
milliseconds
there.
O
But
there
are
other
example:
I,
don't
know
whether
you
can
read
on
the
bottom
of
the
slide,
but
we
can
say,
for
instance,
as
a
captive
portal.
It's
also
important
for
the
application.
If
you
are
even
IOT
device,
that
means
to
abreu
to
upload
its
results
somewhere
of
usually
can
traverse
a
normal
captive
portal.
So
there's
no
point
of
trying
to
use
this
address.
Is
it
meteorite,
yes
or
no
meaning?
Is
it
free
or
not?
O
Okay
and
a
few
things
like
this,
so
this
could
be
an
opportunity
as
well
for
this
within
my
research
group,
sorry
to
propose
information
to
be
put
there
and
we
will
be
all
ears.
Of
course.
Currently
we
were
thinking
just
an
example,
not
sure
that
will
be
finalized
in
the
PVD
when
it
becomes
a
standard.
The
cost,
for
instance,
in
the
network
access
for
now
is
little
bit
too
complex
to
under
it,
but
for
an
x-ray
vision
or
whatever
an
extension,
as
I
said
what
about
the
bandwidth
between
the
CPE
and
the
access
network?
O
Typically,
it's
known
because
if
you
are
using
VDSL,
you
know
the
which
speed
the
videos
and
modem
is
synchronize
captive,
portal
wall
garden
as
well.
All
of
us
will
take
the
plane,
I
guess
to
fly
back
home
tomorrow.
The
day
after
in
most
of
the
airline's,
you
get
a
free
Wi-Fi
only
to
go
to
the
VOD
right.
The
video
on
demand,
that's
free!
If
you
want
to
go
to
the
real
internet
me
to
pay
so
there's
a
specific
characteristic
as
well.
For
this
thing,
I
mean
a
plane.
O
You
don't
have
a
choice
anyway,
so
maybe
less
useful,
so
we've
been
working
on
this
for
about
a
year.
Other
authors,
by
the
way
from
Apple
there
are
people
from
Google
interested
and
Microsoft
is
interested
as
well.
So
there
is
a
big
momentum
behind
it,
so
at
Cisco
I'm
working
we
have
public
domain
implementation.
The
O'reilly
is
there
with
partially
working
al,
an
external,
including
a
user
space
demon,
because
that's
where
we
need
to
get
the
application,
the
metadata
with
a
washout
detector.
O
We
have
modified
array,
DVD
and
OD
HCP
for
open,
wrt
and
ledee
to
work,
and
we,
of
course
we
have
implemented
this
information
sending
into
the
router.
But
your
shipping,
of
course,
may
be
a
specific
thing
to
say:
we
are
working
as
well
in
the
project
calling
meet
h,
2020
project
from
the
european
union,
where
basically,
they
want
to
implement
a
firmware
like
a
post,
socket
API
right,
and
we
have
some
meetings
as
well
the
brian
on
this
six
months
ago.
Something
like
that
to
cannot
synchronize
the
effort
where
this
api,
you
can
ask.
O
The
application
can
ask
some
properties,
are
requests
low,
latency,
request
security,
for
instance,
and
based
on
some
characteristic
of
the
network
and
some
parts
of
the
network,
information
up
on
the
network,
trying
multiple
transport,
UDP,
TCP,
httpd,
TLS
or
whatever,
or
quick,
and
find
the
best
compromise.
Of
course,
this
is
by
default
done
on
information
coming
only
from
the
open
operating
system
on
a
kernel,
but
with
PVD
we
can
provide
some
additional
information
coming
from
the
network,
so
it
was
a
good
intersection
here.
So
it
is
basically
what
I
wanted
to
say
today.
O
So
there
is
a
way
there
is
a
need
as
well,
for
application
to
select
the
path,
maybe
not
the
path
in
the
sense
you
mean
it,
but
I
think
pretty
close
was
one
information
we
can
receive
by
PVD.
It's
also
the
segment
routing
header
to
be
inserted
in
all
the
packet.
If
you
select
this
service,
so
then
we
select
a
real
path
on
the
network.
A
A
A
V,
actually
I
think
I
should
probably
take
the
rest
of
these
questions
offline,
because
these
these
people
don't
just
need
to
see
us,
have
a
really
long
conversation
about
this.
But
you
know
when
I
have
but
I
have
an
architectural
question
about
about
pvd's,
so
the
information
that's
there
right
now
is
basically
architectural
II
limited
to
the
first
hop
right.
It's
just
about
your
access
and
then
anything
that
was
sort
of
as
part
of
the.
A
H
Tommy
Paulie
Apple,
so
co-authoring
this
as
well,
but
I
would
like
to
just
solicit.
You
know
feedback
from
anyone
else,
especially
the
presentations
we
saw
earlier
on
how
they
could
see
this
type
of
thing
fitting
in
I,
do
think
there
is
overlap
of
especially
how
we
could
imagine
applications
viewing
the
network,
viewing
things
to
preview
T's
to
Brian's
point
I!
Think
it's
key.
How
do
you?
H
Let
application
see
the
pvd's
right
now
we
are
focused
on
one
hop,
because
that's
all
we
can
really
do
but
to
some
degree
the
fact
that
a
router
is
telling
you
that
it
has
multiple
up
links
is
actually
multiple
how
hops
behind
you.
So
your
first
hop
is
telling
you
what
it
knows
about
the
next
one's
and
those
are
represented
as
separate
DVDs.
If,
for
whatever
reason,
your
router
was
able
to
learn
about
more
fancy
things
and
it
could
actually
control
your
routes
through
a
Scion
architecture.
H
If
applications
were
aware
of
pvd's,
those
could
just
show
up,
as
maybe
you
know,
25
different
pvd's,
and
maybe
the
scalability
problem
there.
But
hopefully
you
know
each
of
these
would
be
a
self-contained
version
of
information,
because,
beyond
the
extended
information,
one
of
the
really
key
things
of
the
PVD
is
just
the
ability
to
say
there
is
something
different
here.
There
are
two
things:
even
if
you
don't
know
what
they
are,
you
can
try
probing
on
them
and
just
use
them
differently.
Yeah,
but
Vic.
You
see.
A
P
So
the
IETF
protocol
is
a
transport
protocol
that
exposes
a
network
centric
abstraction
of
the
network.
So
the
goal
is
to
guide
the
application
to
for
a
selection
of
their
endpoints,
so
as
to
trade
network,
cost
efficiency
with
equal
or
better
application
performance.
So
this
way
to
to
this
end,
the
auto
exposes
via
restful.
Api
is
information
services
that
are
the
outer
network
map,
which
is
a
set
of
network
locations
that
not
network
location
groupings
with
provider
defined
identifiers
and
enumeration
of
the
related
endpoints.
So
the
provider
identifier
may
be
expressed
in
natural
languages.
P
So
it's
an
indirect
and
network
agnostic
way
to
aggregate
Network
endpoints
that
shares
some
common
property
a
characteristic.
So
the
second
information
service
is
the
outer
coast
map,
so
that
exposes
cows
among
the
P
IDs
and
the
end
points
of
the
map.
So
again,
these
costs
do
not
need
to
be
the
real
values,
because
that
will
never
happen.
Everybody
knows
it,
but
typically
will
be
an
abstraction
on
real
costs
that
give
some
hint
on
the
network
performance.
P
P
P
We
may
say
that
is
actually
looking
at
how
to
connect
so
where,
as
the
base
protocol
is
looking
at,
where
to
connect.
So
the
auto
working
group
is
specifying
and
working
on
a
number
of
extensions
to
address
this
aspect,
so
they
mostly
built
on
an
extension
that
is
defining
abstractions
of
network
traffic
engineering
like
or
network
performance
like
metrics.
P
That
can
be,
of
course,
abstracted.
So,
let's
go
quickly
through
those
extension,
so
you
have
one
extension
that
is
called
auto
path
vector
so
the
auto
path,
vector
abstraction
details,
some
in
within
a
path
in
an
end-to-end
path.
It
details
some
parts
of
the
path
that
it
the
network,
considers
as
significant,
and
it
describes
them
as
like
abstracted
network
elements
which
is
a
set
of
one
or
more
switches
or
links
or
any
part
of
the
network,
and
the
a
in
the
ane
properties
may
be
exposed
by
the
network
in
a
separate
property
map.
P
So
here,
for
example,
so
you
have
now
some
detail
on
the
path
and
the
application
when
it
gets
the
fact,
for
example,
that
from
this
point
to
this
point,
the
routing,
although
the
path
goes
through,
these
particular
steps
and
the
path
here
goes
to
some
other
particular
steps.
It
can
figure
out
whether,
for
example,
they
share
a
bottleneck
and
so
and
figure
out
what
performance
or
what
capacity
it
will
get.
P
So
the
second,
how
to
connect
a
kind
of
extension
here
assumes
that
there
are
several
paths
available
to
the
application
in
Putnam
lis
path
that
relate
to
the
access
type.
So
here,
for
example,
the
the
application
knows
that
it
can
connect
via
Wi-Fi
or
cellular.
So
it
requests
a
path
performance
given
its
access,
type
and
SLA,
and
here
it
gets,
for
example,
two
values.
P
If
I
ask
cell
cell
and
SLA
of
three
and
Wi-Fi
and
SLA
three
so
then
here
I
can
get
two
values
and
figure
out
how
to
connect,
given
that
you
do
not
have
necessarily
a
better
quality
with
Wi-Fi
and
now
with
40.
You
often
have
better
you're
better
off
and
if
there
is
any
special
on
the
weekend
better
of
using
a
cellular
connection,
the
other
another
extension
there
is
called
multi
cast
with
constraints.
P
So
a
multi
cast
with
multicast
I
can
request,
for
example,
routing
cost
and
bandwidth
capacity
on
paths
that
I
would
like
to
have
with
either
lower
routing
or
lower
cost
and
lower
bandwidth
or
higher
cost.
But
then
better
bandwidth,
so
I
will
ask
all
those
to
matrix
and
the
auto
will
return
values
only
on
those
paths
that
are
meeting
the
constraints.
P
So
this
way,
it's
kind
of
selecting
telling
me
do
not
use
that
that
that
one
path
so
besides
other
how
to
connect,
if
it
may
be
of
interest
to
the
panel
G
group
to
have
service
and
information
service
that
helped
applications
to
decide
when
to
connect-
and
this
is
done
with
Auto
cost
calendars.
And
so
here
you
have
an
array
of
of
time
dependent,
cost
values,
and
here
you
have
attributes
that
help
you
to
interpret
those
values
in
the
array.
So,
and
here
are
other
references
and
thank
you
for
your
attention.
Thank.
A
Q
Q
The
world
is
easy
and
explicit
congestion
notification
now
currently
seeing
is
the
only
a
standard
way
hosted,
use
the
advanced
properties,
feedback
from
network
and
order
for
congestion
control
either
case.
Could
superintendency
into
the
center
network
will
offer
information
continue
a
more
reach
and
I
created
past
properties.
Hang
our
research
for
networks.
It
is
founded
at
some
dynamic
networking.
Information
from
a
bottleneck
device
on
the
past
is
very
useful
to
the
our
congestion
control
lack
of
ease
the
properties
such
as
ingress
rate,
egress,
read
performance,
feedback
of
phone
bottom
naked
note.
Q
Q
Radio
lingo
stated
such
as
a
radio
recorded
a
spectrum
you
tonight
edition
at
her
arte,
Penn,
Station
or
Wi-Fi
device
people
to
adapt
the
variable
wireless
network
very
well,
as
we
go
to
our
shows
in
Figure
2.
The
red
line
is
the
capacity
of
the
wireless
network
under
the
president
non-islamic
our
method,
we
can
see
the
bragging
I
could
adapt
to
the
red
line
very
well.
Q
The
other
kana
then
how
to
a
distributed
the
possible
of
this
easy
I
used
to
invent
the
beats
for
information
distribution
as
the
ITF
standard,
but
for
rich
and
greater
properties.
There's
no
ascend
in
a
way.
Now
we
only
could
use
the
TCP
and
IP
options
for
distribution,
but
this
method
is
limited
by
the
option:
dance
and
has
no
problems.
That
means
the
middle
box
as
well.
So
we
may
prefer
a
standard
embedment
way
to
disputed
these
problems
between
paths
and
a
host
specific
sub
layer
between
TCP
and
IP.
Q
Maybe
it's
better
like
layer,
3.5,
the
more
of
passed
awareness,
but
before
passed,
awareness
hoster
has
no
idea
about
of
the
pass
after
pass
awareness
hosted,
Mendoza
passed
them
could
select
and
write
the
past
and
I
used
pass
the
more
efficient
than
before.
So
could
we
are
to
more
about
to
the
past?
Could
we
control
the
past
properties?
We
have
least
idea
because
we
found
the
service
like
knife
VIII,
which
leaves
automated
properties
like
a
very
high
up
enemies
and
a
known
tendency.
Q
A
K
K
Q
K
A
So
I
am
very
animal.
Okay,
that
wasn't
obvious,
but
now
I
put
together
a
little
draft
a
little
bit
before
this
meeting,
sending
out
to
the
list
I
think
about
a
week
ago
based
on
so
the
idea
is
basically
to
try
and
focus
all
of
the
discussion
on
the
list.
Could
you
know
maybe
something
going
about
what
I
think
the
questions
are
in
this
space
and
there
are
six.
We
have
some
questions
and
we
have
some
more
questions.
So
I'll
just
go
through
these
very
briefly.
A
So
how
are
path?
Two
properties
defined
and
represented
right,
so
you
could
call
this
a
closed
question
right.
Everything
that
does
anything
in
in
exposing
path
or
even
last
hop
information
to
the
end.
Points
must
have
its
own
vocabulary
in
order
to
say
right,
so
Alto
has
the
cost
maps
and
the
network
maps.
The
PVD
has
this
extension
thing
where
you
have
key
value
pairs,
and
you
know
some
of
the
points
in
this
key
value
pair
space
are
assigned
so,
for
example,
the
you
know
this
is
I.
A
Think
metered
is
one
of
the
things
and
that
has
a
meaning,
I,
think.
Oh,
it
seems
to
me
that
all
of
the
work
sort
of
coming
from
the
ietf
side
has
I
don't
want
to
say
arbitrary,
but
they've
kind
of
gone
for
the
most
obvious
way
to
represent
those
properties
for
their
use
case,
and
it's
not
clear
to
me
that
in
a
fully
path
or
network
that
that's
the
best
way
to
do
it
right,
like
so
Aalto
is
focused
on
a
particular
application.
Space
PVD
is
focused
on
a
particular
application
space.
A
All
of
them
are
built
in
such
a
way
that
a
common
vocabulary
could
be
built
into
them
right.
Not
they
all
have
points
of
extensibility
that
there
are
exercises
that
went.
You
could
use
them,
but
it
seems
like
it's
interesting
to
step
back
and
consider
what
those,
what
these
sort
of
vocabulary
items
are
and
how
you
know
what
they
mean
and
how
the
definition
of
that
space
implies
what
we
can
and
can't
do
with
path
or
we're
not
working
in
place
sort
of
like
whether
we
can
or
cannot
deploy
it.
A
So
we
have
a
couple
of
other
questions.
How
to
implement
discover
trustworthy
path.
Properties
Sion
represents
a
point
in
this
design
space,
so
it
is
Aalto
for
the
set
of
each
for
the
set
of
path
properties
they
care
about.
So
we
have.
We
have
sort
of
like
running
code.
Here
we
have
solutions
that
we
think
that
we
can
take
from
there.
How
can
in
point
select
has
to
use
for
traffic
in
a
way
that
he
trusted
by
the
network.
A
A
My
computer
is
telling
me
to
set
that
sit
down.
Okay,
please
confirm
your
dick
German.
Okay,
so
in
here,
like,
as
we
saw
from
from
Chris's
talk,
there's
running
code
on
these
for
both
addressing
based
approaches
and
PC
FS
based
approaches.
There
are
ways
to
do
this
right,
so
Sion
represents
again
another
point
in
this
design
space.
How
can
interface
is
the
transport,
an
application
layers,
support
the
use
of
path.
A
Probably
here
one
is
how
can
a
path
or
a
network
and
a
path
aware
Internet
be
effectively
operated
right,
so
the
the
network
operations
model
that
we
have
right
now,
as
you
tell
BGP,
where
you're
willing
to
accept
traffic
from
for
which
prefixes
and
magic
happens,
and
you
get
traffic
and
inbound
interfaces
and
you
can
change
your
announcements
and
you
have
a
effective
and
predictable
control
loop.
Well,
okay,
predictability
in
BGP
is
an
open
area,
but
you
haven't.
You
have
an
effective
control
loop
for
being
able
to
say
okay,
this
traffic's
going
to
come
here.
A
I
want
this
traffic
to
go
away,
I'm,
going
to
stop
announcing
that
prefix
on
that
on
on
that
interface
and
the
traffic
goes
where
you
tell
it
in
a
path
aware:
Network,
the
endpoints
that
are
attached
to
the
network
are
also
making
decisions
about
which
paths
are
going
to
use.
Do
something.
Since
this
is
already
happening
in
in
sort
of
lake,
in
multipath
networks,
right
now
and
multiple
connected
networks,
endpoints
are
mobile
and
they
are
shifting
traffic
from
one
interface
to
another
sort
of,
and
they
degenerate
to
point
case.
A
So
this
is
might
be
a
problem
that
we
in
the
internet
engineering
community
have
to
handle
anyway,
and
this
framework
might
give
us
a
way
to
think
of
that
in
a
way
that
we
get
a
better
handle
on
it.
But
the
operations
model
is
a
completely
open
question
and
it
would
be
interesting
maybe
to
have
some
ops
area
and
input
here
or
people
have
work
from
the
management
space
to
come
in
and
help
us
out
with
that.
The
second
one
is
really
a
layer.
A
Eight,
an
economic
layer
question:
how
can
the
incentives
of
network
operators
and
end-users
be
aligned
to
realize
that
vision
here?
Great?
So
why,
as
this
was
actually
a
question
that
was
posed
to
me
by
Martin
Thompson
as
soon
as
he
read
the
draft
like
five
minutes
after
I
post,
it
he's
like?
Why
didn't
you
consider
this
I'm
like
Thank,
You
Martin,
it's
in
github
now
so
yeah?
A
The
working
copy
of
this
in
in
github
actually
has
this
question
and
we'll
post
that
in
any
other
questions
that
we
enable
dictation
language,
German,
unbound,
Radnor,
dhalsim,
next
and
I'm-
sorry,
and
then
we'll
we'll
post
that
after
this
meeting,
if
there
are
any
other
open
questions
that
people
want
to
come
up
and
address,
I
obviously
have
fuzzy
ideas
here.
That
might
be
a
document
that
we
talk
about,
but
we
need
to
have
an
answer
for
the
question:
why
would
somebody
actually
buy
a
gateway
that
would
do
this?
Why
would
they
buy
a
router?
A
A
C
A
A
Feedback,
my
personal
view
on
that
it
does
seem
to
me
like
there
are
a
lot
of
enabling
technologies
that
work
much
better
on
v6
than
on
v4.
For
this
sort
of
thing
right,
you
could
build
like
you
could
replace
pbd
with
some
hacks
Indian
DHCP
you
could.
Oh,
are
you
gonna
design
this
for
me
Aaron
or
are
you
gonna
go
talk
to
somebody
about
the
feedback?
Okay,
perfect
yeah!
Thank
you.
A
So
I
actually
see
this
as
a
if
this
is
something
that
v6
does
much
much
better
than
v4
right
like
if
you
saw
the
the
API
talk
that
Lauren
gave
we're
actually
talking
about
having
to
rewrite
all
the
applications
anyway.
So
moving
to
v6
is
not
going
to
be
harder
than
rewriting
all
these
applications.
I,
don't
think.
R
R
So
I
could
imagine
so
that's
conceivable
I'm,
not
sure
the
works
worth
doing,
but
that
might
well
come
out
of
the
essentially
applicability
statements
that
you're
talking
about
in
the
last
couple
points
here
right
on
that
which
I'd
love
to
see
hey.
How
is
this
work
going
to
be
useful?
Being
one
of
their
early
questions
great.
E
A
So
I'll
actually
point
out
again
so
another
thing
that
just
came
to
mind
so
I
know
that
song
on
actually
does
work
with
v4
right,
so
it
basically
uses
song
on
routing
through
the
internet
and
then
you
can
have
the
last
hop
address
as
be
ipv4
addresses
and
it
just
hands
off
in
the
end
point
as2
v4.
So
this
there's
a
way
that
you
can
use
an
overlay
waiting
to
get
v4
endpoints
talking
to
v4
end
points
over
song
on,
but
that's
kind
of
a
life
extension
hack
for
v4.
As
I
see
it.
S
A
A
really
easy,
so
you're,
basically
talking
about
so
I,
see
that
as
an
intermediate
stage
right
lakes.
Are
you
build
like
the
way
that
we
often
build
these
new
technologies
as
you
go
as
you
go
local
net
to
local
net?
You
do
the
tunnel
thing
because
you
own
bits
of
the
tunnel.
That's
a
really
good
question
and
I
don't
have
an
answer,
but
but
that
should
get
written
down
in
this
document.
Actually,
so
you
know
how
how
are
these
affected
by
like
so
we're
like
all
of
these
properties?
So
you
know
things
like
laurels.
A
Work
wouldn't
work
there
right,
because
you
don't
have
enough
information
to
say
the
application
can't
go.
Okay,
this
packet
or
the
transport
layer
can
go.
This
packet
goes
this
packet
goes
here,
but
what
do
you?
You
know?
What
are
the
benefits?
And
then
what
are
the
the
changes
in
the
model
when
you
go
from
a
a
local
to
local
net,
as
opposed
to
endpoint
to
endpoint
or
maybe
even
local
net,
to
end
point
or
endpoint
to
local
net?
A
That's
that's
a
really
good
question,
so
I
will
go
off
and
and
and
and
think
about
that
for
this
document
and
I
invite
you
to
help
me
out.
S
That
would
probably
I
mean
there
is
no
encapsulation
for
it
for
it,
but
it
defines
what
could
be
collected,
but
that's
mostly
focused
on
how
you
can
operate
your
network,
better
feed
it
back
into
the
application.
I
guess
that
also
has
to
find
some
place
in
case.
If,
if
we
go
with
modifying
the
transport
protocol
to
carry
and
return
the
data
in
the
transport
in
quick
and
TCP
and
so
on,
that
would
probably
be
one
other
way
of
returning
the
data
for
awareness
for
the
application
of
amis.
So.
A
It
seems
to
me
that
the
IOM
stuff
could
be
used
sort
of
in
the
in
the
in
the
inter
domain
or
the
intra
domain,
but
also
in
the
inter
domain,
like
so
IOM
actually
right
now,
IOM
is
explicitly
single
domain
or
single
iom
domain,
where
that
doesn't
necessarily
overlap
with
yes,
but
usually
does
so
that
could
be
used
as
a
way
to
to
measure
dynamic
properties
in
a
way
that
could
be
then
disseminated
through
a
path
dissemination
protocol
like
something
on
yeah.
A
H
H
On
the
evil
things
that
they're
doing
right,
and
so
the
one
thing
that
that
brings
up
is
you
know,
as
we
are
trying
to
build
a
solution
for
captive
portals.
Hopefully
there
is
some
sort
of
compromise
that
is
mutually,
not
terrible.
That
allows
the
network
to
communicate
the
terrible
things
they're
doing
to
block
the
advice,
but
also
enable
the
system
to
help
show
their
ads
a
little
bit
better,
which
is
kind
of
terrible
but
important.
H
A
We'll
say
that
I
didn't
want
to
bring
this
up
at
all,
that
Aaron
kind
of
already
did
that
one
of
the
things
that
having
if
you
have
path,
property
dissemination
in
a
trustworthy
way,
so
those
are
secured
and
you
have
path
control
in
a
trustworthy
way.
So
each
participating
AAS
can
know
that
the
path
isn't
forging
anything
right
like
so
so
you're
in
a
situation
where
we've
basically
made
KC
is
wonderful,
spoofing,
detector,
obsolete
which
would
be
awesome
because
I'm
sure
she
does
want
to
have
to
do
that.
A
Then
you
actually
can
enter
into
business
relationships
with
a
SS
that
you're
not
directly
connected
to.
Is
that
a
thing
that's
enough
of
an
incentive
to
get
people
to
put
these
things
up?
Maybe
I,
don't
know
that's
C,
but
it
seems
like
we're
changing
the
business
relationships
that
you
can
build
if
we
do
this
and
that
sets
up
that.
That
takes
this
problem
from
the
complexity
of
the
captive
portal
thing,
where
it's
like.
A
It's
pretty
clear
to
see
how
you
can
analyze
this,
because
these
guys
are
trying
to
screw
you
as
hard
as
they
can
and
you're
trying
not
to
get
screwed.
You
know
you
can
kind
of
meet
in
an
equilibrium,
if
maybe
that
equilibrium
space
is
zero,
but
there
might
be
points
in
that
equilibrium
space
here
by
basically
changing
the
way
that
you
can
build
business
relationships
you
throw
this
into
a
giant.
I
mean
like
this
is
this
is
now
an
economic
problem
of
a
complexity
that
I'm
certainly
not.
H
I
think
that
yet
the
points
that
aaron
brought
up
earlier
about
the
point
of
doing
this
path,
awareness
is
potentially
just
to
create
an
open
market
yeah
of
path.
So
you
can
get
to
is
really
interesting
and
really
points
towards
having
multiple
paths
as
being
maybe
a
way
out
of
some
of
the
terrible
incentives
that
we
have
today
on
locked
paths.
A
A
H
Tommy
so
I
don't
remember
exactly
what
it
was,
but
I
think.
The
first
question
that
I
had
given
to
Chris
was
essentially
that
you
know
we're
seeing
that
a
lot
of
these
patterns
between
the
various
protocols
in
the
various
architectures,
Scion
or
pvd's
or
Alto
have
these
similar
patterns
of
how
we
do
path,
discovery
path,
communication
and
we
also
brought
up
the
point
of
api's
and
how
our
applications
interacting
with
these,
and
we
could
design
kind
of
the
perfect
API
for
each
of
yes
exactly
right.
H
So
I
think
this
is
a
really
good
table
and
when
we
look
at
the
API
is
that
we
can
design
whether
they're,
post
or
neat
stuff,
I
think
those
overlap,
but
we've
also
often
defined
the
api's
to
be
very
specific
to
the
problems
we
want
to
solve
in
that
case.
So
maybe
we
can
make
it
more
general.
But
how
do
we
evolved
this
to
make
sure
that
let's
say
today?
H
I
can
write
an
API
for
PVD
that
works
on
my
multihomed
ipv6
networks
and
can
I
use
those
same
properties
to
work
on
Scion
Network
when
it
comes
along
like
what's
the
transition
path.
For
this
are
there
is
there
a
difference
between
kind
of
soft
path,
aware
networks
in
which
I
can
get
away
with
just
knowing
a
little
bit,
and
maybe
a
world
in
which
path
awareness
is
critical
to
be
able
to
use
it
well.
So
how
does
that
work?
I
don't
know.
Do.
A
A
Cool
any
other
questions
on
this,
or
should
we
kind
of
try
and
wrap
this
up
and
figure
out
what
the
next
steps
are?
I
know:
I've
learned
a
lot.
This
was
interesting
for
me,
I,
don't
know
I
hope
it
was
interesting
for
for
for
you.
Let
me
put
my
chair
hat
back
on
then
and
say
so.
We
got
one
action
item
here.
A
So
I
can
put
this
back
over
on
the
non-functional
nightstand,
because
I
have
my
own
mic
over
there.
F
Just
you
know,
assuming
that
the
chairs
are
not
looking
at
the
mailing
list,
because
they're
looking
at
the
meeting
I
did
send
a
note
to
the
mailing
list
about
that.
Just
making
sure
people
can
find
me
and
asking
them
to
send
me
the
name
so
that
when
I
have
a
good
template,
I'm
gonna
do
my
two
first
and
see
what
you
know
see
what
information
you'd
ask
for
it.
No
circle
back
could
ask
them,
but
I
just
want
to.
Let
you
guys
know
that.
Could.
F
A
Heard
a
couple
of
bits
of
input
for
the
open
questions
draft,
so
I
will
read
that
and
send
it
to
the
list
soon.
I
actually
want
to
get
a
little
bit
of
a
sort
of
a
sense
of
who's
in
here,
as
the
chair,
who
here
has
no
idea
what's
going
on
and
is
willing
to
say
so,
okay
Aaron,
do
you
find
this
area
interesting
and
useful,
and
you
you
think
you
get
kind
of
a
handle?
You
don't
really
have
to
have
it's
not
like
a
Boff
question
where
it's
like.
Do
we
understand
the
problem?
A
Because
the
point
of
this
is
that
we
don't
understand
the
problem.
Does
this
work
interest
you
just
a
sort
of
a
show
of
hands.
I
mean:
are
you
here
for
this,
or
are
you
here
for
the
internet
and
the
and
the
air-conditioning?
A
B
A
And
yeah
we're
also
looking
for
for
suggestions
for
what
fee,
like
the
general
theme
of
that
agenda
should
be
like.
So
these
were
sort
of
the
first
two
we're
booting
the
research
group
ideas
and
we'll
try
to
keep
the
energy
going.
It's
much
much
fewer
people
in
here
than
last
time.
I
noticed
because
you're,
the
first
meeting
of
anything
is
like.
Oh,
what's
that
I'm
gonna
come
sit
down
so
hopefully
we'll
have
about
this.
A
T
Yeah
I
think
the
speaking
with
my
assassin,
with
my
RTL
I,
think
we,
you
have
to
probably
put
like
way
way
more
primary
conflicts
and
you
might
end
up
on
Friday,
but
and
just
tell
everybody
if
you
come
on
Friday
right.
But
the
other
thing
is
that
you
could
consider
having
two
sessions
and
kind
of
bridging
between
them.
If
it's
to
allow
for.
A
N
N
T
N
A
B
S
A
Yeah,
maybe
people
took
it
over
to
doe
and
they're
signing
with
Banerjee
blue
sheet
and
oh
that
would
be
cool
alright.
So
with
that,
thank
you
very
much
for
coming
to
the
second
meeting
of
the
pannier
G
and
we
will
see
you
in
London
on
the
mailing
list.