►
From YouTube: IETF101-MPTCP-20180322-1330
Description
MPTCP meeting session at IETF101
2018/03/22 1330
https://datatracker.ietf.org/meeting/101/proceedings/
A
B
B
B
B
B
Rejigging
of
the
milestones
the
most
recent
update
that
was
due
to
happen
now,
the
sending
it
to
the
is
geez
I.
Guess
we're
not
quite
at
that
stage,
but
we're
hopefully
fairly
close
to
it.
We
talked
about
I'm,
just
in
the
next
set
of
bullets,
just
some
discussion
summarizing
to
some
discussion.
We
had
at
previous
idf's
that
we
agreed
we
needed
to
have
implementation
of
the
of
the
Biss
or
the
revisions
that
are
in
the
the
changes
that
are
in
the
best
before
in
order
to
push
it
forward.
Given
its
on
the
standards
track.
B
I,
my
belief
is
so
our
belief
is
that
we
now
have
one
implementation
of
everything
new,
that's
in
the
Biss,
except
for
the
sha-256
part,
and
our
belief
is
also
as
far
as
we
know.
There
is
no
prospect
at
the
moment
of
further
implementation
in
the
near
term,
at
least
on
that
public
linux
implementation,
by
which
I
mean
it
doesn't
look
like
people
will
have
time
to
implement
the
sha-256
stuff
in
the
near
term,
and
our
belief
is
also
that
all
kind
of
current
suggestions
have
been
included
in
the
in
the
best
document.
B
Any
any
changes,
etc.
Binning
included
always
been
agreed
that
they
won't
be
included.
So
the
proposal
justice
is
just
sort
of
to
warm
you
up.
The
proposal
will
we're
going
to
be
making
in
the
discussion
session
in
I.
Don't
know
20
minutes
time
or
something
is
that
we
working
group
last
call
the
document.
Now
we
send
a
request
to
the
security
ATS
to
find
an
early
reviewer
and
there
you
go
it's
under
standards,
track
comments.
Justifications
will
be
appreciated,
of
that
and/or
comments,
etc.
B
So
there
are
two
other
items
on
the
on
our
on
our
Charter.
One
is
about
multipath
TCP
proxy
activity,
so
the
development
on
that
that's
happened
recently
is
that
TCP
M
working
group
has
now
added
a
working
group
item
on
what's
called
their
converter.
So
that's
that's
good
news
and
I
think
that
that
means
there
by
this.
This
topic
in
our
working
group
has
essentially
migrated
across
to
there.
B
The
other
topics
that
we
have
on
on
our
chart
at
the
moment
is
about
an
enhanced
API,
an
informational
document,
so
we've
we've
got
to
talk
later
to
discuss
about
some
and
and
some
enhanced
API
work
has
been
done.
There's
no
working
group
document
for
this
item
at
the
moment,
so
upon
some
point,
we'll
need
to
discuss
about
how
we
progress
with
that.
B
D
Hello,
so
just
a
very
short
update
on
the
implementations
of
MP
TCP
next
slide
in
terms
of
the
Linux
mpg
implementation.
We
have
few
two
weeks
ago,
I
think
released
the
latest
empty
speed
version,
0.94
release.
It
is
released
based
on
the
latest
Linux
kernel,
V
4.14,
the
one
was
the
long-term
support.
We
are
now
tracking,
always
the
long-term
support.
Kernels.
That
way
we
can
guarantee
support
for
those
for
a
longer
time.
D
D
We
still
had
a
total
of
72
changes
coming
from
six
different
developers
of
five
different
companies,
so
there's
a
community
around
it
when
people
trying
it
out
and
whenever
something
doesn't
work,
they
pop
up
end
up
trying
to
fix
it,
there's
also
the
bigger
effort
getting
more
steam
in
terms
of
up
streaming
it
to
the
official
Linux
kernel.
We
have
a
community
that
is
building
up
around.
D
There
was
people
from
different
companies
from
Tessa
res
Intel
Oracle
and
some
others
who
are
also
at
least
in
observing
this
effort,
and
this
effort
is
getting
more
steam
and
we
are
trying
to
design
the
Linux
MPGs
pimenton
implementation
so
that
we
can
upstream
it
more
directly
what
this
might
involve,
that
we
will
probably
have
to
remove
quite
some
features
out
of
the
Linux
MPT
speed,
V
0.94
implementation.
But
if
that
gives
us
a
path
towards
up
streaming,
I
think
that's
acceptable
later.
We
can
always
pull
in
those
features
back.
D
So
that's
for
the
line,
exam
participe
implementation.
So
we
also
Olivia's
team
at
the
UCL.
They
organized
an
empty
speech
a
curtain
a
few
weeks
ago,
and
the
goal
was
to
leverage
multiple
TCP
on
iOS
11
and,
at
the
same
time
deploy
corresponding
Linux
servers.
So
the
goal
was
to
pick
us
and
iOS
applications
find
out
which
connections
those
applications
are
doing,
which
of
those
connections
would
actually
benefit
from
MGP
CP
from
one
of
those
different
modes
that
we
are
exposing
in
iOS
and
then
basically
modify
those
applications.
D
D
Go
to
the
next
slide.
So
what
what
happened
at
the
hackathon
is
the
developers
we
we
chose
quite
a
few
iOS
applications.
There
was
a
web
browser
and
called
grave
a
radio
station.
Some
net
me
some
networking
measurement
tools
like
net
meter
and
some
other
custom
applications
like
program,
P,
multipath
tester,
which
counter
is
going
to
present
a
bit
later
today
and
in
terms
of
the
lessons
that
we
learned
is
that
those
applications
they
are
indeed
creating
connections
that
would
benefit
from
MP
TCP,
for
example,
the
radio
station
the
radio
station.
D
It
receives
a
list
of
possible
reservations
and
then
it
connects
to
an
URL
and
the
URL
just
simply
starts
streaming.
The
traffic
streaming
the
radio
and
for
a
certain
reason
that
we
don't
know
about,
but
for
some
reason
to
stream
it
in
support
to
basically
easily
handover
from
one
interface
to
the
other.
So
as
soon
as
we
use
this
application
and
turn
off,
Wi-Fi
did
stream
broke
so
MB
TCP
was
basically
a
perfect
candidate
for
these
kind
of
connections.
D
One
problem
is
that
the
app
and
the
server
off,
not
under
control
of
the
same
person
of
the
same
team,
for
example,
for
the
radio
station.
This
dream
was
coming
from
some
third
party
from
the
radio
station
itself.
The
web
browser.
He
was
doing
basically
searched
suggestions.
When
you
are
typing
in
in
the
search
bar.
There
was
a
connection,
it
would
have
benefited
from
MPTP
as
well,
because
it
is
a
more
interactive
connection.
D
But
even
then,
if
the
search
queries
are
going
to
the
search
providers
like
Google
or
Bing
and
those
are
not
under
the
control
of
the
application
developer,
so
that's
one
problem
and
then
the
other
problem
is
also
that
lots
of
applications
are
using
third-party
libraries.
They
are
not
using
the
libraries
that
Apple
is
exposing
that
allowed
to
use
MP
TCP
or
they
use
even
higher
layer
api's,
for
example,
for
radio
streaming.
They
use
ap
eyes
that
are
sitting
at
a
very
high
lake
at
a
very
high
layer
where
they
cannot
choose
multiple
TCP
support.
D
A
Of
course,
the
funk
question
so
about
that
zero
ninety-four
Persian,
so
it
would
be
some
gap
between
bits
per
channel.
0.94
am
I,
correct.
Sorry,
so
I
mean
it's
not
fully
implemented
everything
in
this
draft.
In
my
yes,
it
is
not
the
best
draft.
It's
not
at
all
the
best.
Basically
I
mean
it's
based
on
68
24.
Yes,.
D
D
B
Thanks,
so
this
is
now
the
kind
of
open
mic
slot
for
anybody
else,
who's
working
on
implementations
and
can
share
anything
about
it.
What
they're
working
on
we
sometimes
get
somebody
who
comes
assist
something,
and
sometimes
we
don't
so
I'll,
just
give
you
another
few
seconds:
okay
right,
let's
move
on
so
Alan
I
think
you're
up
next.
F
F
Hi
everyone
just
an
update
on
where
we
are
with
version
10.
Next,
like
this,
not
model
has
changed,
although
stuff
has
been
added
which
makes
the
change
for
the
last
few
revisions
of
emojis
Corrections.
Here
we've
got
two
significant
changes:
we've
got
the
sour
for
quiet
I,
don't
really
want
to
say
it's
adding
a
new
feature
as
such.
It's
the
addition
of
what
happens
if
you
send
a
fast
close
on
the
reset
packet
and
and
also
some
considerations
on
TCP
fast
open
an
exploit.
F
F
F
So
it
eliminates
any
concerns
around
resetting
all
sub
flows
at
the
TCP
level
when,
if
one
of
those
didn't
get
dead
and
the
receiver
may
have
retained
state
by
adding
this
option
to
the
reset,
that
means
that
to
receiver
know
is
that
this
was
any
connection,
never
reset
and
it
can
just
drop
immediately.
So
that's
what
we've
done
in
draft
is
proposed
by
Christoph.
It
makes
a
lot
of
sense
and
it
was
filling
a
gap
that
wasn't
previously
documented.
F
So
that's
fine
moving
on,
we
have
interactions
with
TCP
fast
open,
so
you've
got
to
thank
Quentin
for
writing
and
the
original
draft
I
think
on
this
and
their
way,
but
it
went
in
Gregory
know
so
several
you
see
Levin
people.
Thank
you
for
writing
this,
but
we've
trimmed
down
the
key
points
for
inclusion
in
a
draft
now.
F
The
goal
of
T
to
be
first
open
is
to
save
one
round-trip
time
before
being
able
to
send
data
and
you
exchange
cookies
and
TCP
options.
In
order
to
achieve
that,
so
there's
there's
two
sort
of
considerations
there.
How
old
does
it
interact
with
MP
TCP
option
space
and
when
is
it
appropriate
to
use
it?
F
So
if
we
go
on
to
the
next
slide,
what
we?
What
we've
said,
we've
made
the
two,
the
two
key
observations
in
this:
the
first
one
being
the
options
based
one
so
in
the
CIN.
Yes,
that's
definitely
enough
space
in
this
in
AK,
it's
a
little
bit
tight.
The
maximum
TFO
cookie
length
with
only
be
seven
bytes,
but
if
you're
using
fewer
of
your
typical
options
or
you
might
want
to
consider
which
of
your
fewer
options
you
use.
F
F
The
additional
observation
on
this
we
made
is
that
because
a
middle
box
could
alter
that
initial
TFO
data,
but
there's
no
mapping
there's
no
check
some
space
in
the
sin.
You
only
discover
that
you'll
would
potentially
only
discover
that
one
more
round
trip
time
further
down
the
line.
Then
you
got
note
here:
TFO
benefit
at
all.
F
If
you
wait
for
that
option
to
verify
the
data,
so
the
alternative
solution
to
this
and
the
one
that
we
believe
to
be
the
right
one
to
recommend
is
that
the
TFO
data
just
stays
outside
the
data
level
secret
space
completely,
and
it's
not
until
the
first
data
non
TFO
data
starts
flowing.
Do
we
start
doing
data
sequence
mapping?
That's
fine,
because
there's
no
there's
no
additional
sub
flows
in
you,
so
there's
no
concern
about
getting
out
of
sync
in
any
of
that
sense.
This
is
just
the
first
initial
bit
of
data.
F
F
We
are
explicit
the
only
talking
on
the
first
sub
flow
for
DFO,
because
I
didn't
we
don't
see
there
being
any
bet.
Well,
first
start
there's
not
enough
option
space
on
the
additional
sub
flows,
and
even
if
there
was
so
isn't
really
any
benefit
to
doing
TFO
on
the
later
sub
flows,
because
the
timeliness
benefit
isn't
really.
There.
You've
already
got
two
sub
flow
set
up
the
first
upload,
so
limiting
scope
here
to
purely
the
first
sub
play
and
that's
the
changes
in
the
document.
F
D
F
G
F
Another
set
of
eyeballs
on
this
would
be
greatly
appreciated
because,
when
I,
when
we
read
the
first
text
very
person
in
the
draft
having
converted
it
from
the
thing
everything
made
sense,
then
when
I
was
actually
writing
this
slide
yesterday,
I
was
rereading
it
going.
This
is
quite
hard
to
follow
and
have,
and
then
just
discuss
was
Christoph
and
it
all
became
clear,
so
I
admit
we're
going
to
need
to
reword
a
couple
of
bits.
But
if
you'd
like
to
double
check
and
send
feedback,
that
would
be
very
much
appreciated.
Thank
you.
F
A
F
A
F
E
A
F
A
A
F
At
the
moment,
the
way
this
is
all
written
is
there
all
the
messages
and
everything
are
required
for
a
full
implementation,
because,
after
all,
you
as
a
nothing
plantation
would
want
to
assume
that
if
you
send,
for
example,
a
fast
close,
it
would
be
accepted.
You
know
we've,
that's
that's
why
we
do
standards,
I
it
could
be
worth
looking
at,
but
I'm
not
I'm,
not
too
optimistic
there
very
much
that
isn't
already
should
and
must
separated
here.
H
Hi,
my
name
is
mo
from
Hobby
I'm
I
have
no
no
questions
on
the
updates.
Actually,
but
I
have
a
general
clarification
question
if
it
may
ask
okay,
okay,
so
in
section
2.7
notable
features
third
bullet
says
to
meet
the
threats
of
identifier
identified
in
RFC
six
one,
eight
one,
the
following
steps
have
been
taken:
care
keys
are
sent
in
the
clear
in
MP,
MP,
capable
message
and
those
are
used
in
basically
four
the
MP
joint
message
for
hitchhiking
right.
H
F
Hijacking
sub
flows
that
the
original
amputees
to
be
designed
and
the
one
we're
carrying
forward,
obviously
into
the
base,
because
it's
the
only
refinement
is
that
to
you
we
are.
We
are
going,
we're
always
aiming
to
be
no
worse
as
long
as
we're
no
worse
than
TTP,
and
sometimes
we
can
be
better
than
that
so
goal.
So
we
have
the
idea
of
yes,
sir,
the
key
to
in
a
clearer
to
start,
but
then
that's
the
same
as
if
you
had
intercepted
any
TCP
connection,
and
you
can
do
sequence
number
guessing.
F
H
B
B
So
assume
you
know
no
more.
Questions
on
these
particular
changes.
We're
now
moving
on
to
basically
having
a
slot
for
discussing
whether
people
happy
for
this
to
be
working
group
last
called
mouth.
So,
as
we
said,
that's
our
in
the
kind
of
the
default
is
that
will
happen
immediately.
This
meeting
ends
so,
just
as
a
reminder
is
on
the
on
the
standards
track,
we
made
some
changes
compared
to
the
bits
that
mean
is
it's.
B
A
version
number
has
been
bumped
so
if
it
will
obsolete
682
4,
so
the
is
your
chance
if
anybody's
unhappy
or
ask
concerns
about
lost,
calling
now's
your
chance.
Otherwise
we
will
assume
that
we
are
happy
and
that's
what
we'll
do
pretty
immediately,
because
I
think
I
mean
you
know,
that's
we're
not
actually
expanding
on
that
we've
got
now
had
a
lot
of
implementation
deployment.
Experience
of
multipath
TCP
we've
had
a
lot
of
you
know
lot
of
interest.
It's
been
used
very
widely.
B
I
Yeah
I'm
the
culprit
for
that
and
we
tried
counter
tried
to
implement
it
and
there
are
still
some
issues
from
an
implementation.
You
point,
and
since
this
is
designed
to
be
an
extensible
part
of
MB
TCP,
my
suggestion
would
be
to
try
to
remove
that
from
the
core
text.
Looking
at
the
discussion
with
Yoshi
for
me
before,
which
part
is
mandatory
and
which
part
is
not
and
try
to
write,
maybe
an
experimental
RFC
that
describes
this
expertise,
additional
possibility
of
having
experimental
and
participation
which
is
outside
the
main
document
that.
I
F
F
F
J
F
B
L
Hello,
so
I'm
going
to
give
a
quick
updates
on
the
circuit
API
written
by
a
measurement
Olivier
and
I
just
did
a
small
updates
for
segmenting.
So
next
slide
please.
So
there
is
only
small
change
which
is
I
added:
the
support
for
segment
routine
v6,
so
I'm
PGP
and
leverage
its
and
so
basically
what
this
segment
routing
this
is
so
floating.
So
it
allows
you
to
steal
packets
at
the
ipv6
level,
at
the
IP
level,
to
do
that.
The
ipv6
data
plane
relies
on
something
called.
L
The
SRH
will
change
the
segment
low
together
and
we
think
that
MPTP
could
leverage
this
to
establish
the
progress
with
this
victim
of
silica
stuff
like
this.
So
this
is
the
Suresh
I
was
talking
about.
So
this
is
a
segment
hooding
error.
This
error
is
inserted
just
after
the
ipv6
error
and
what's
important
inside
is
you
have
a
segment
list
so
with
this
you
can
specify
a
list
of
segments
so,
for
instance,
a
list
of
middle
boxes
where
the
traffic
is
supposed
to
go.
So
you
could
insert
this
on.
L
Multipass
TCP
sip
float
to
decide
where
the
sub
flows
are
going
next
slide.
Please.
So
this
is
a
quick
example
of
the
codes
ported
to
the
API
in
this
one
I
add
one
of
the
segments.
It's
the
to
do
to
do
one
segment.
So
if
you
do
that,
your
subfloor
will
go
to
this
IP
address.
So
this
could
open
some
ideas
or
what
we
can
do
next
light.
Please.
So
for
the
use
cases
we
have
disjoint
paths.
So
this
is
one
of
the
idea
you
could
establish
several
sub
flow
on
several
disjoint
paths
for
moments.
L
You
don't
know
really.
If
the
path
is
oh,
no,
not
even
if
you
have
a
3G
or
Wi-Fi
or
seven
Wi-Fi,
you
don't
know
all
close
or
distance.
The
paths
are
with
this.
You
can
make
it
happen.
You
have
several
SRH
you
and
when
you
establish
a
step
flow
you'll
know
for
a
fact:
that's
they
will
go
to
different
of
same
path.
It's
not
the
case
for
the
moment.
For
a
moment,
you
just
assume
we
pop
the
sub
flow.
We
try
to
understand
if
it
works
well
or
not,
and
then
we
use
it
here.
L
You
could
make
assumption
like.
When
you
have
a
transmission,
you
could
decide
to
retransmitted
on
pallets
clearly
different
from
the
initial
one,
so
I
we,
it
was
discussed
in
panel
g2
about
state.
Let's
go!
If
you
want
to
state:
let's
go,
you
can
make
multiple
CCP
what
could
have
survey
six
and
in
PT
CP
you
can
make
it
work
in
a
state.
Let's
go
so
we
think
there
are
things
to
do
with
that.
We
are
experimenting
to
see
what's
possible
to
do
so.
L
The
we
updated
the
API
modify
the
code
to
make
to
make
it
possible,
so
you
can
just
establish
several
sub
flow
with
their
own
ACH.
So
several
so
flow.
That's
you
know,
that's
they
go
to
a
different
path.
If
you
want
to
test
it
feel
free
to
contact
me,
and
if
you
have
ideas
about
use
case
of
things,
we
could
do
with
that.
Just
tell
me
and
that's
it.
H
L
The
old
paper
about
segment
routine
v6
to
describe
use
case.
You
could
ask
a
controller
to
give
you
a
specific
path,
respecting
certain
condition
like
I
want
this
to
good.
So
for
the
moment
the
application
is
just
an
ABI
right,
so
the
API
just
it
just
allows
you
to
say.
Oh
I
want
to
establish
a
stop
flow
that
will
go
by
this
path
and
another
stop
flow.
That
will
go
if
you
had
this
path.
So
that's
a
bit
of
those
of
the
scope
for
the
ABI.
It's
still
interesting
for
them.
Ptc
piece
of
to
discuss.
L
That's
the
question
these
questions
move
about
so
soothing
right.
It's
how
do
handles
knows
whether
it's
supposed
to
go
you
can
use.
You
could
use
a
controller,
for
instance.
Well,
you
asked
him.
Oh
I
want
this
kind
of
specification
for
my
path.
Please
give
me
buff
in
the
network
that
will
respect
that
and
then
the
controller
sent
you
in
a
cell
H
and
you
as
an
application.
You
just
said
yesterday
under
step
flow.
You
don't
show
a
try.
I
mean
you
can.
L
If
you
have
a
smart
application,
you
could
try
to
figure
that
about
yourself,
but
I
think
it
is
the
different
problem.
It's
not
really
related
to
MPTP.
What's
interesting
for
MP
TCP
is
for
fast
management
or
scheduling
of
things
like
that.
No,
you
know
for
a
fact.
That's
do
so
flow
are
different
omelets.
So
that's
what
we.
M
Shrinking
the
belly
Cisco
along
the
same
lines,
I
think
you
know
I'm
trying
to
understand
I'm
not
objecting
to
this,
but
I'm,
trying
to
understand
the
motivation
for
this
right.
So
in
MPDC,
when
you
bind
your
particular
sub
flow,
but
it's
not
IP
address
that
it's
that
itself
is
indicating
a
specific
path
are
using,
let's
say:
I
have
the
ues
endpoint
us
two
interfaces
or
finality
if
I
bind
to
a
Wi-Fi
address.
That
is
a
path
all
the
way
to
the
server
right.
L
It's
not
because
you
have
two
interfaces:
two
different
I
mean
you
have
to
like.
You
have
a
Wi-Fi
and
3G
yeah,
so
those
are
two
addresses
to
a
different
path:
yeah,
but
all
different
other
paths,
for
instance
with
that
in
Belgium
what
proxy
muses
provides
a
DSL,
so
the
Wi-Fi
and
also
the
TG.
If
you
look
at
the
path
they
are
not
so
different,
I
mean
they
could
really
well.
The
access
network
is
a
bit
different,
but
you
end
up
with
pretty
fast
on
the
same
route
or
so
those
paths
are
not
really
so.
L
These
joints
they
are
they're
close
to
each
other,
so
in
the
congestion
is
inside
the
network,
I
mean
after
when
the
paths
are
joined.
You
don't
know
that
I
mean
at
the
transport
layer.
For
the
moment
you
have
no
idea
about.
What's
under
you,
just
co2
difference
at
the
phase
two
paths.
So
so,
let's
say
you
have
three
or
four
addresses
and
you
have
to
do
as
a
injection,
because
one
of
the
second
one
of
the
Sigmund
has
been
lost,
you
need
to
they
inject
its.
You
have
to
choose
the
interface
totally
injected.
L
For
the
moment,
what
we
are
doing
in
the
scheduler,
it's
just
trying
to
take
the
best
one,
but
you
don't
really
know
if
one
path
would
would
be
a
better
fit
to
do
it
or
not.
I
could
say
it's
a
backup
of
something
like
this.
Here
you
have
more
insight
inside
the
transport
protocol
to
know.
Oh,
this
path
is
probably
best
because
it's
totally
disjoint
from
the
two
other
ones.
That's
the
main
idea.
Is
you
get
more
insight
at
the
transport
layer?
Yeah,
it's
a
cosplayer
violation,
but
you
have
more
insight
about.
M
I
think
insertion
is
fine,
I
think
the
question
is
an
endpoint
like
a
horse.
Doing
that
is
not
completely
getting
the
picture.
Well,
you
know
it.
Can
you
know?
Typically
it's
a
default
right.
I
use
an
IP
address
like
weifare,
it's
a
default
route
for
me.
I,
never
know
the
path.
If
you
say
it's
a
router,
maybe
I
can
understand.
But
if
it's
a
house
doing
it
yeah.
M
Another
question
is,
like
you
know,
I
think
I'm.
Now
that
you
know
your
your
opening
of
this
right,
what
is
the
implication
to
the
proximal
that
mean
only
we
are
doing
like
you
know,
will
I
be
able
to
say
that
you
know
you
know,
steer
it
to
the
same
particular
node
and
there
I
activate
a
sit
as
something
then
that
becomes
my
proxy.
Will
I
be
able
to
do
that.
I'm
just
curious
yeah.
L
I
Yeah,
just
just
one
additional
comment
is
that
there
was
a
presentation
from
Daniel
Bowie
from
Bell
Canada
on
when
LJ
on
Tuesday,
showing
that
you
want
it
well
visibility.
We
wanted
to
use
ipv6
a
Marathi
inside
this
core
network
and
he
wanted
to
expose
information
about
the
network
in
the
in
2d
and
O's
that
you
and
O's
can
make
in
form
this
year
and
in
the
presentation
he
refers
one
of
our
papers
that
will
be
presented
next
week
at
saucer.
That
shows
how
to
expose
the
information
by
using
the
DNS.
I
E
Sorry
good
afternoon,
everyone
so,
firstly,
this
is
my
first
IETF
little
overwhelmed
with
all
these
guys
around
whom
I've
been
always
looking
at
all
right.
So
can
we
just
skip
to
the
next
slide?
So
this
is
also
about
the
sake
API,
which
already
Fabian
has
spoken
about,
and
you
know
the
first
thing
is
I
will
talk
about
the
small
motivation.
Maybe
you
can
go
to
the
next
okay.
Thank
you.
E
So
we
were
like
experimenting
with
drones
and
we
realize
that
what
was
happening
is
a
drone
is
flying
and
your
control
signal
is
not
reaching
to
it
because
of
some
reasons
it
was
just
going
going
going
you
set
down,
but
it
doesn't
turn
it
right
to
the
wall
in
front
of
it.
So
we
realize
that
okay,
something
is
not
right.
Maybe
the
link
is
getting
too
much
congested
or
maybe
something
else
is
happening.
E
So
the
idea
of
using
multipath
NCP
emerged
in
a
mine
and
we
thought
okay,
let's
send
the
control
signal
on
a
different
path.
All
together,
then
your
standard
user
data
goes
in.
So
that's
how
that's
the
basic
motivation
of
this
entire
work.
Okay,
so
obviously
you
want
your
control
signals
to
reach
properly
uninterrupted
without
any
delay
and
at
the
highest
priority,
so
it
should
be
given
a
dedicated
path,
even
if
you
have
two
interfaces
but
by
multipath
TCP
in
itself
doesn't
do
that
because
it
balances
your
data
between
the
two
interfaces.
E
Okay,
so
we
thought,
let's
do
something.
Let
us
prioritize
one
of
the
paths
for
a
control
data
and
to
say
other
paths
will
prioritize
only
for
the
user
data,
but
in
in
the
event,
if
a
path
fails,
you
automatically
fall
back
to
the
other
path,
that's
the
basic
property
of
multipath
TCP
and
we
were
running
it
over
robot
operating
system,
so
I'm
very
sure,
but
everybody's
come
aware
of
that.
E
So
while
we
were
trying
to
do
that,
we
wanted
to
have
this
control
sitting
at
the
application
layer
so
that
you
can
prioritize
these
paths
using
your
Ross
interfaces,
not
right
directly
doing
it
in
the
kernel
tweaking
it
around
or
doing
something.
So
we
came
across
couple
of
socket
API,
which
were
already
existing.
One
was
MP
TCP
info,
which
was
introduced
in
0.92
when
we
started
working,
it
was
not
there
in
91,
but
anyway,
the
problem
here.
What
we
realized
is
the
subfloor
information
was
not
getting
available
at
the
application
layer.
E
For
me
to
play
with
the
another
was
the
husband's
draft
which
Heath
Fabian
just
presented,
so
that's
the
one
which
fit
suited
to
our
requirement,
at
least
from
the
sub
flow
ID
being
available.
So
we
took
up
the
idea
of
that
from
there
and
then
we
built
our
socket
API
at
the
top
of
it.
So
can
we
move
to
the
next
one?
So
what
we
did
is
we
built
a
socket
API
to
control
the
priority
of
these
parts?
E
One
of
the
paths
we
make
as
a
primary
path
using
the
low
prior,
that
is,
for
the
control
data,
which
is
being
exposed
by
your
Ross
and
for
the
user
data,
which
is
the
other
link.
We
make
the
other
path
as
the
primary
path.
But
what
we
realize
here
say
when
your
system
is
flying
or
one
of
the
sub
flow
goes
down
or
disconnected
because
of
any
particular
reasons,
and
when
it
comes
back,
it
loses
its
priority.
E
It
doesn't
know
what
was
you
what
priority
you
have
set
up
a
priori,
so
that
was
the
biggest
problem
we
face.
So
there
was
another
modification
which
we
proposed:
that's
the
next
one,
ok,
so
what
we
suggest
here
is:
can
we
have
two
data
structure
being
maintained
inside
the
amputee
CP
kernel,
which
maintains
the
list
of
your
sub
flows,
which
you
have
set
up
as
a
primary
path,
or
you
know
the
primary
path
for
a
particular
session
and
and
we're
back
a
path
for
put
that
particular
session.
E
B
E
A
pre
profit
or
yeah,
it's
still
a
test
system.
It
is
still
in
the
laboratories
I
kept,
so
we
have
not
really
deployed
it
on
a
drone.
Yet
the
reason
being,
as
I
said,
the
commercial
drones
are
really
not
exposed
there
colonel
to
tweak
it
around,
but
we
have
tried
it
on.
They
are
PI
boats
and
placed
it
over
the
drone
and
test
it
with
that.
So
we
are
working
with
some
vendors.
If
you
know
they
can
really
give
us
that
flexibility,
so
that's
very.
It
stands.
A
This
is
your
C
from
flora,
so
I
have
a
one
question.
So
yeah
I
understand
your
motivation,
but
so
when
you
establish
one
subfloor
and
then
is
some
for
some
reasons,
a
subfloor
is
down,
and
then
you
want
to
reconnect.
But
in
this
case,
can
we
get
same
for
top
roots
before
the
same
previous
one?
So
you
might
add,
your
IP,
others
might
be
changed
well,
for
solids
might
be
changed
and
then,
in
this
case,
which
kind
with
file,
we
can
be
done
if
we
can
apprise
same
priority
or
not.
Okay,.
E
So
here
I
have
a
basic
assumption
that
your
source
and
the
destination
IP
hasn't
changed,
because
that
list
which
have
maintained
is
on
the
basis
of
that
source
and
the
IP
destinations,
so
I'm
using
will
obtain
one
source
and
one
IP
destination.
That
is
one
specific
flow.
So
if
you're
establishing
multiple
sub
flow
is
that,
yes,
there
is
a
conflict,
so
the
all
the
subclause
which
will
be
established
between
the
same
source
and
same
destination,
we'll
come
up
with
the
same
variety
as.
B
I
I
Since
the
last
idea,
a
draft
was
presented
in
ITF,
100
and
attendees,
showed
some
interest
on
the
topic
and
pointed
to
related
work
from
this
group
and
also
from
research,
especially
handling
of
multiple
IP
addresses
per
network
interfaces
by
itself
was
not
seen
as
a
problem
because
it's
common
with
ipv6,
and
so
we
concluded
when
we
needed
to
clarify
the
problem.
So,
for
example,
those
multiple
IP
addresses
are
in
5g
provided
by
different
anchors.
So
what
does
that
mean?
So
we
wanted
to
explore
that.
I
So
in
this
new
draft
presented
today,
we
take
a
step
back
and
try
to
do
a
broader
analysis
of
we
needs
to
to
be
done.
So,
especially,
we
take
a
look
at
new
behavior
in
5g
that
can
impact
on
PCP.
So,
for
that
we
found
the
session
and
service
continuity
feature
is,
can
introduce
some
changes
that
are
related
to
my
PCP
and
also
another
part
of
the
draft
describes.
Our
MP
TCP
could
complement
5g.
So
basically,
that's
the
dual
connectivity.
I
The
dual
connectivity
feature
possible
outcomes
that
we
envision
for
this
work
that
could
be
a
set
of
guidelines
for
implementing
a
PCP
over
5g,
f,
IG
stack
and
also
possibly
derive
some
requirements
for
NPT
TP
for
4G
and,
of
course,
wait.
We
need
to
analyze
them
to
see
if
anything
either
is
new
or
if
everything
can
be
done
with
the
current
state
of
PCP
and
4G
yeah.
Thank
you.
So
we
start
with
session
and
service
continuity.
I
Ssc
addresses
various
continuity
requirements
of
applications,
so
it
is
based
on
the
distributed
mobility
system.
Then
it
has
multiple
modes.
So
the
mode
one
is
fixed.
Anchor
we'll
go
through
the
details
in
the
next
slide
right.
Well,
two
is
decimated
anchors
plus
break
before
make,
and
what
three
is
determined
estimated
anchors
plus
make
before
break
so
FFG
device
select
an
appropriate
SSE
mode
for
every
application,
and
it
is
based
on
local
policy,
so
otherwise
the
network
could
also
set
it
so
before
really
going
forward
with
the
the
analysis.
I
I
won't
just
wanted
to
speak
a
little
bit
about
the
PDU
session,
because
that's
but
on
topic
to
make
the
link
between
the
5g
and
then
beat
cpworld.
So
the
video
session
in
effigy
that's
a
unit
of
network
service.
It
can
hold
multiple
cue
as
flows.
It
is
connected
to
a
single
data
network
and
through
one
or
more
anchors-
and
it
is,
it-
is
associated
with
a
single
non-modifiable
session
continuity
mode
and
also
a
single
non-modifiable,
Network
slice
and
type,
and
we
focus
on
the
IP
types.
I
Basically,
it
goes
over
a
single
radio
access
type
at
any
given
time
and
it
correspond
to.
It
will
correspond
to
a
network
interface
on
the
device,
but
what
we
think
to
be
probably
probably
the
case
it
would
be
an
implementation
implementation
choice
for
the
5g
stack
developers,
but
that
that's
the
one
that
makes
sense
to
us
based
on
previous
previous
versions
of
produce.
So
he
releases
different
applications
running
on
the
device
may
use
different
video
sessions
right
so
because
different
applications
will
have
different
needs
in
term
of
data
network,
SSE
mode
or
slice
to
use.
I
So
in
a
sassy
mode,
one
Virgie
behaves
like
this
fixed
next,
a
network
anchor
provides
a
fixed
IP
address
to
the
device,
and
the
network
may
additionally
add
or
remove
additional
network,
and
course
or
additional
IP
addresses
dynamically
on
the
same
network
interface,
and
so
in
this
mode.
Mpcp
should
always
keep
using
the
initial
IP
address,
because
the
only
one
which
is
guaranteed
to
be
available
over
time
in
this
mode,
but
when
new
IP
addresses
are
made
available
to
the
application,
it
makes
sense
to
use
them
because
likely
that
would
be
a
very
short
path.
I
Probably
so
in
assessing
what
o5g
FFG
behavior
is
is
the
following.
You
have
an
IP
address
that
changes
when
the
device
leaves
a
network
anchors
service
area
and
it
will
change
to
another
IP
address,
and
this
is
a
break
before
make
IP
address
change
so
in
this
mode,
and
pet/ct
should
use
some
right
before
my
behavior
similar
to
what
exists
now,
except
that
it
this
time
it
could
occur
even
only
on
cellular
and
similarly
to
the
existing
record.
I
For
my
behavior,
the
application
should
not
request
a
specific
network
interface
in
this
case,
for
example,
when
opening
a
socket
for
the
same
reason
as
in
the
current
mode
and
finally,
MPTP
should
maintain
a
valid
backup
IP
address
by
using
this
same
vb
break
before
mag
behavior
in
mode
3.
Now
the
IP
address
changes
when
leaving
a
network
anchors
service
area,
and
this
is
an
egg
before
break
on
the
same
or
new
network
interface.
I
Both
case
can
occur
when
the
lifetime
of
the
old
IP
address
is
communicated
by
the
network
to
the
device
using
a
writing
advertisement
so
in
this,
and
PCP
should
always
create
and
use
new
sub
flows
when
a
new
IP
address
becomes
available,
and
it
should
stop.
I
mean
to
this
application
to
a
particular
application
yeah.
It
should
stop
using
older
sub
flows,
gracefully
and
should
do
that.
You
know
in
a
way
that
help
for
performance
and
also
recycle
helps
recycling
resources
faster.
So
it
should
wait
a
little
bit
for
the
slow
start
to
finish.
I
Maybe
and
well,
then
it
should
release
the
old
IP
address
before
it
actually
released
automatically
and
then
mptp
should
not
close
sub
flows
using
the
latest
IP
address,
since
this
IP
address
is
the
only
one
that
will
remain
after
a
transition
period
and
finally,
it
should
maintain
a
valid
valid
backup
IP
address
using
the
latest
as
a
single
three
IP
address
for
that
application.
Okay,
so
now
the
other
part
of
the
drafts,
that's
about
general
connectivity
in
5g,
that's
from
a
research
project
and
basically
the
5g
behavior
associated
with
dual
connectivity.
I
That
is
the
following.
So
this
he
relies
on
resilient
to
path
over
radio
access
and
individual
QoS
flows
will
use
one
path
or
another,
and
so
typically,
your
connectivity
is
not
visible
outside
of
the
radio
access
technology
layer.
So
the
proposal
made
in
this
draft,
the
dual
connectivity
feature
of
algae
could
leverage
mptp
to
support
a
range
of
features,
no
robustness,
reliability
and
bandwidth
aggregation
without
duplicating
this.
In
the
5g
stack
to
use
MPTP
for
dual
connectivity,
we
identified
roughly
four
now
two
requirements.
You
know
there
should
be
multiple
these
here.
I
I
The
dual
connectivity
aspect:
integrate
feedback-
and
maybe
you
know
in
the
draft
intersection,
identify
clearly
the
requirements
for
MP
TCP
and
if
there
are
any
four
5g
as
well
and
of
course
we
were
looking
if
anyone
think
of
any
additional
5g
feature
that
could
potentially
impact
MP
GDP
so
that
we
could
gather
that
in
one
in
one
draft
yeah.
So
basically,
you
just
wanted
to
add.
I
N
I
F
I
That's
possibility,
it
depends
also
on
how
it's
you
know
the
what
happens
here,
a
type
of
discussion
here,
but
we
have
people
in
our
company
who
are
going
to
3G
and
I
think
that's
a
possibility
that
to
try
to
make
especially
we
we
propose.
You
know
some.
We
say
that
that
could
be
some
impact
on
MPCP,
but
we
also
for
the
second
part
that
would
have
some
impact
on
5g.
So,
of
course,
in
this
case
we
might
need
to
present
that
in
the
treaty.
M
Oh,
thank
you
nabela
Cisco,
so
I'm
trying
to
understand
this
proposal
so
I
understand
address,
has
different
properties,
but
essentially,
if
you
forget
that
Phi
G
terminology
forget
the
end
of
the
day,
there
are
dresses
which
are
shortly
with
mobility
property
versus
with
order
alright.
So
somehow,
if
I
understand
correctly,
you
want
this
to
be
exposed
to
the
MPD
should
be
layer.
Is
that
the
objective.
I
The
way
5g
handles
mobility
is
different
and
you
see
like
we
have
these
three
modes
before
you
had
only
one
right
and
and
so
before
it
was
I,
think
more
similar
to
mode
one
now,
and
so
now
it's
kind
of
exploded,
it's
more
distributed.
So
there
isn't.
You
know
it
can
be
leveraged
to
to
get
a
better
behavior,
better
performance
from
MPCP.
M
M
The
idea
is
to
essentially
include
some
additional
information
about
the
type
of
address,
in
other
words,
when
a
host
gets
an
address,
it
know
what
is
the
type
of
address
right
now.
3Gpp
is
all
of
these
modes.
Will
map
to
the
IETF
specified
coloring
or,
if
I
call
it
as
a
coloring
or
whatever
properties,
that
all
of
these
three
modes
will
have
one-to-one
mapping
to
the
to
the
modes
that
we
are
defining.
We
will
be
defining
in
DMM
work
Europe.
M
So
that
is
one
aspect,
so
that
does
a
some
part
of
your
concern
in
the
sense
that
addresses
have
different
properties
with
our
mobility
non
mobility.
That
from
when
you
obtain
an
address
that
that
those
property
will
be
will
come
as
part
of
it
in
the
address
assignment
procedures,
maybe
dhcpv6,
maybe
neighbor
discovery
or
I
query
to
control
protocol.
That
is
one
aspect
now.
The
second
aspect,
probably
what
is
missing,
is
how
do
we
expose
that
to
the
socket
layer
that
also
Indiaman
working
group,
somebody
is
specified
extensions.
It's
a
working
group
document.
M
It's
already
plans
last
call
there.
We
expose
that
to
the
higher
applications
at
application
layer.
If
you
use
a
standard,
socket
interface
with
that,
you
know
exactly
what
is
the
type
of
address
so
now,
if
I
add
those
two
things
now
I'm
wondering
like,
in
fact,
if
an
application
is
in
using
MPDC
be
with
those
socket
extensions,
is
they
still
a
problem?
Is
what
I'm
thinking?
So
maybe
you
know
and
read
the
draft.
M
You
know
I
just
looked
at
the
INA
section,
it
doesn't
say
anything:
no
actions
are
required,
maybe
you're
specifying
certain
behavior
I
do
not
know,
but
I
strongly
suggest.
Look
at
the
work
in
DMM
working
group,
maybe
maybe
there's
a
gap.
Maybe
there's
not
no
gap,
but
I
I
feel
it's
addressed
to
some
mood
to
most
part.
H
Thanks
for
the
information,
actually,
the
3gpp
intelligent
statement
for
TMM
working
group
for
the
prefix
map,
coloring
right.
So
that's
what
you're
talking
about
so
I
looked
into
that
draft
and
it
requirements
and
also
the
draft
he
is
talking
about.
So
this
draft
is
all
about
how
these
requirements
are
mapped.
Mpt,
CP,
if
they're
anything
they
said,
you'll
need
to
be
done
in
MPD
CP
to
address
sse
more
three.
That's
what
it's
looking
into.
H
So
it
may
be
possible
that
nothing
significant
need
to
be
added
in
MPT,
MPT
CP,
but
it's
good
to
have
the
requirements
put
it
in
one
place,
because
3g
people
are
tracking
all
the
drafts
related
to
Phi,
G
and
they'll.
Watch
and
they'll
come
back
to
this.
If
they,
if
they
need,
if
they
want
something
need
to
be
added.
Yeah.
M
Thanks
for
the
clarification,
I
think
I'm.
Ok
with
this,
but
I'm
trying
on
at
this
point,
the
requirement
is
not
clear
to
me:
I
understand
you
know:
Phi
G
has
different
things,
but
already
there's
work
happening.
This
I
think
if
this
on
top
of
that
is
a
new,
if
you
can
clearly
articulate
what
that
is,
maybe
if
it
is
impacts,
NPD
CP,
it's
fine,
but
otherwise.
At
this
point,
I
don't
see
that
impact.
So,
but
thank.
B
Just
to
follow
up
on
that
some
quite
understood
in
terms
of
this
DMM
work,
that's
been
done
so
far
woody.
Would
you
be
able
to
use
some
beaches
to
be
able
to
use
multipath
TCP
to
do
this?
Dear
memorability,
that's
being
done
with
Phi
G
or
so
our
change
is
needed
on
your
your
site
about
to
use
this,
or
can
you
use
all
of
the
small
tip
of
TCP
stuff
that
Zoe
has
been
working
on
anyway,
yeah.
M
So
the
way
I
said
filipenko,
if
you
are
already
exposing
these
three
classes
of
you,
know
the
modes
into
the
socket
layer.
Now,
if
there's
an
application
sitting
at
the
you
know,
which
is
opening
up
sockets
already,
are
the
visibility
like
anytime,
you
bind
to
a
particular,
you
know,
IP
address
you
fetch,
you
know
getting
for
whatever,
like
you
know,
the
property
of
that
artists.
You
know
class
one
more
one,
more,
two,
more
three
right
now
now
the
I
as
the
question
back
now
now
that
that
lay
now
it's
visible
to
the
application.
M
I
So
what
I
get
is
that,
based
on
this
DMM
work,
the
MPG
city
would
be
able
to
get
in
which
mode
a
particular
interface
is
attached
to
right,
I
mean
in
CL,
and
so
additionally,
you
know.
Maybe
it's
done.
Also
I,
don't
know.
When
you
see
the
there
is
a
sort
of
mapping
like
if
a
new
interface
goes
up
its
visible
or
it
should
be
usable
by
some
applications,
but
not
others,
because
this,
the
implication
that
that
goes
up
relates
to
one
slice
or
one
sec
mode.
I
That's
that
we
that's
covered,
but
you
see
that
some
some
applications
will
not
be
able
to
use
that
particular
interface,
maybe
and
may
be
able
to
use
others.
So
there
is
a
sort
of
mapping
between
interfaces
and
applications,
which
is
known
by
the
5g
stack,
which
normally
the
application
doesn't
need
to
know,
because
the
stack
will
expose
the
the
interface
that
is
visible
by
this
application.
But
since
MPTP
is
between
the
application
and
5g
stack,
there
may
be
something
needed
a
form
for
this
type
of
mapping.
I
mean
yeah.
M
B
I
M
B
B
I
B
Exists
in
some
kind,
so
I
think
I
probably
should
be
sitting
over
there,
but
I'd
encourage
you
during
this
working
group.
Last
call
if
there's
any
changes
that
you
think
are
needed
to
the
protocol
to
help
our
stuff
get
used
by
5g.
Please
make
the
suggestions
or,
if
they're
part,
of
the
API
work,
which
is
sounded
like
there
might
be
more
part
of
that
in
this
API
discussion.
So
there
now
meaning
about,
will
continue.
Please
make
the
suggestions
and
work
with
the
API
authors.
Sure.
I
O
So
I,
all
this
is
a
quick
informational
presentation
about
some
experiments.
I
do
with
multiple
CCP
in
iOS,
so
next
slide
so
appellees,
probably
the
most
the
biggest
deployment
of
multipath
CP
students.
Initially
it
was
made
for
Siri
and
since
iOS
11
it
is
available
for
all
applications.
So
it's
nice
for
research,
point
of
view,
because
you
can
try
and
participate
with
a
lot
of
desired
vison.
That's
what
the
next
that.
So
the
question
is
all
performance:
multipass
sippie's
about
its
design
goals.
O
O
So
I
concrete
example
of
what
we
can
observe
with
this
is,
for
instance,
here
I
show
the
mobile
experiment,
whether
you
try
to
estimate
how
you
can
reach
the
Wi-Fi
access
point
whilst
and
sending
data,
and
you
can
see
some
results-
life
resource
of
of
the
experiment.
So
next
slide
about
the
methodology.
O
A
sweater
some
far
is,
if
you
have
cumulative
distribution
function
over
the
mobile
test,
where
the
Wi-Fi
was
actually
lost,
and
so
we
showed
the
maximum
applicative
delay
experienced
by
the
application
in
the
flow
direction
by
comparing
and
PTC
pn
m
p
quick.
So
we
see
that
in
most
cases
seems
that
it
works
well,
but
we
have
a
long
tail
distribution
at
the
end
where
there
are
more
interesting
cases
where
seems
difficult
to
perform
and
over
efficiently,
and
so
next
slide
in
the
load
will
observe
the
stand.
O
Behavior
next
slide,
and
so
so
currently
we
only
have
10
users.
So
it's
not
a
lettuce,
so
the
idea
is
to,
of
course
collect
more
data
about
it,
and
so
we
are
looking
for
volunteers
and
we
push
interesting
results
about
those
findings
on
multipath
we
talk,
we
push
one
about
address,
validation
of
quick
but
formal
to
recipe.
We
also
have
some
interesting
and
so
next
slide.
So
if
we
want
to
give
it
a
try,
you
can
have
the
application
here
and
since
it's
still
in
development,
feel
free
to
provide
feedback
and.
A
O
Yeah
in
that
case
so
far
so
I've
my
own
implementation
of
with
the
past
week,
of
course,
and
in
both
case
at
server-side,
you
are
using
classical
low
latency
with
latency
first
and
at
client-side.
It's
also
low
lowest
latency
first,
but
considering
also
the
fact
that
you
prefer
the
Wi-Fi
/
silica.
So
you
have
some
network
prioritization
over
the
Wi-Fi
until
you
consider
Wi-Fi
as
too
bad
and
switch
on
this
area.
A
O
Rt
t
mainly
but
I,
know
in
the
in
the
linux.
Can
implementation
for
extents
you
ever
saw
when
a
pass
is
considered,
potentially
failed
when
you
experience
a
lot
of
losses,
and
so
you
avoid
sending
up
that
path.
Until
you
see
it
works
again,
so
I
suppose
there
is
something
similar
in
iOS
and
for
multi
pass
quick.
It
used
that
so
far,
but,
of
course
the
scheduling.
This
is
a
work
to
improve.
Of
course,
the
scheduling
algorithms
are
also
the
past
management's,
and
so
it's.
O
Yeah
but
ax
one
of
the
the
stuff
is
that,
because
I
look
at
the
twist,
sometimes
my
a
multipass,
quick
implementations
tries
to
send
on
the
cellular
while
it
should
not
and
also-
and
so
it
increases
the
delays
because
sometimes
with
UF
out
of
order,
delivery
and
also
add
some
performance
issue
if
multipath
squeak
but
I'm
fixing
them.
So
it's
an
iterative
process
about
fixing
the
screening
issue
with
multipath.
P
O
P
B
Thanks,
so
next
up
is
Anthony,
he's
doing
it
remotely.
Isn't
he
so
I
can
see
your
own
meter
currently.
Q
C
B
Q
J
J
For
example,
we
when
we
look
at
the
FH
global
forty
tango
gray,
it's
in
the
sixteen
sixteen
oh
seven
letter
P
per
second,
but
they
didn't
what
they
didn't
tell
you
is
that
the
upload
speed
is
much
slower,
for
example,
in
the
LTE
one
of
the
example
of
that
you
really
have
to
look
look
at
this
more
pin
and
the
difference
between
the
upload
and
download.
In
this
case,
our
typical
our
provider
is
saying
that
the
tango
is
fight.
J
J
It
can
also
be
different
in
the
server
side,
so
we
are
struggling
it
three
and
so
basically,
what
we're
proposing
is
that
the
round-trip
time
measurement
is
not
sufficient
and
in
the
next
line,
number
four
stage.
Number
four
is
going
for
the
one
of
the
reason
for
the
asymmetry.
If
you
take
the
3gpp
Network
it
first
of
all,
it
has
to
look
for
the
control
channel
two.
When
you
do
an
uplink,
you
had
to
look
for
the
the
uplink
on
the
left
and
the
darling.
J
On
the
right
hand,
side
you
had
to
look
for
the
the
channel
was
assigned
to
allow
you
to
go
down
and
up
load
up
link
and
depending
on
whether
the
transmission
time
interval
is
10,
millisecond
or
1
millisecond
odd
it
can.
J
The
FH
is,
is
about
17
or
choke
point
5,
heartbeat
second,
depending
on
the
x
times.
Our
interval
under
down
linnaeus
is
is
different.
It's
only
processing
the
incoming
data
from
the
for
the
for
the
base
station
and
it
allows
50
times
the
transmission
time
interval
and
then
it
would
transmit
the
download
data
and
then
the
you
negotiate.
So
it's
about
seven
and
half
we
second.
This
is
these.
Are
the
numbers
in
the
40
in
the
5g?
The
number
would
be
different
I
in
the
next
slide
and
how
about
the
access?
J
Besides
the
access
that
source?
Only
fought
the
theorists
I
was
only
talking
about
the
uplink
and
downlink
in
the
wireless
access
part,
and
when
you
get
to
the
network
that
with
the
routers
are
they
are
going
to
be
cues.
When,
when
we,
when
there's
a
cue
on
one
direction,
it
doesn't
mean
the
other
direction
and
also
have
a
long
view.
J
J
J
If
you
had
them
prior
to
2.2
to
15,
if
the
minimum
speed
for
the
downlink
is
for
Maccabee
per
second
and
the
uplink
is
one
per
second
and
it
is
con
distance
in
the
pot
broadband
and
it
is
based
on
a
number
of
measurement
that
it's
made
so
that
they
may
such
the
definition
and
in
two
and
five
we
define
that
the
definition
of
broadband,
so
basically
the
service
providers,
if
providing
apartment
surface,
it
has
to
provide
a
minimum.
The
other
way
of
25
megabits
per
second
download
and
three
megabit
per
second.
J
Are
there
says
that
minimum,
but
if
it
is
but
mobile,
it
is
a
little
different
and,
and
the
reference
is
down
there.
So,
basically,
besides
the
wireless,
we
also
have
been
in
the
service,
for
that
poet
in
asymmetric
are
dead
away
for
the
user.
So
in
page
seven
is
that
so
other
parts?
Thank
you.
As
far
as
motivated
CP
is
concerned.
We
are
picking
different
hops
and
one
of
the
criteria
is
to
pick
the
path.
J
So
if,
if
if
it
is
not
so
if
we,
if
the
launch
of
time
is
not
enough,
you
would
be
looking
at
the
one-way
delay
and
if
it
measure
the
one-way
delay
in
in
the
past,
we
had
to
synchronize
both
clocks
and
and
and
and
and
we
don't.
First
of
all,
we
don't
have
enough
information
to
measure.
Both
we
basically
set
nas
tell
what
time
you
send
a
packet
and
more
times
it
comes
back
and
then
measure
the
bunch
of
time.
J
If
you
had
to
measure
the
wrong
ship
delay
or
the
one-way
delay,
you
would
need
to
know
the
time
the
TV
series,
if
the
packet,
so
one
option,
is
to
synchronize
the
clock
and
any
other
acknowledgment
or
sec
K,
which
could
be
delay
that
one.
So
do
you
want
you
Tommy
will
be
even
more
in
actually
and
in
the
knowledge
Minh.
It
would
have
to
include
the
time
that
the
packet
was
received
so
that
the
receiver
can
calculate
it
and
and
then
take
the
difference.
J
So
in
one
half
pipe
half
I
on
the
left-hand
side
and
they
were
inside
different,
Popov
J
and
it
it
it
measures
to
you.
Do
you
one
way,
delay
and
then
calculate
the
difference
and
then
in
the
next
slide?
Page
nine?
Is
that
it?
If
all
we
care,
is
basically
difference?
We
we
only
want
to
look
at
the
difference
and
if
you
look
at
the
difference,
whether
you
synchronize
the
crop
are
not
synchronized
its
God,
it
really
doesn't
matter
because
it
tends
yourself.
J
So,
on
the
right
hand,
side
bottom,
you
are
getting
exactly
the
same
yourself,
whether
you
synchronize
the
clock
on
or
not
synchronize.
The
car,
although
they
D
the
measurements,
are
not
actually
because
of
the
synchronization,
but
when
but
because
you're
between
the
same
Center
and
the
same
receiver,
but
two
different
pop,
the
synchronize,
the
clock
synchronizes,
the
Crocker
retains
herself
so
you're
getting
the
same
itself,
so
they
were
so
that's
there
is
where
you're.
So
the
proposal
is
a
dancer
away,
you
to
measure
the
one-way
delay
and,
and
and
and
that
would
require
an
addition.
J
True
Trudy
to
the
stand
up
so
in
the
next
slide
is
that,
first
of
all,
you
had
to
negotiate
whether
you
are
using
you're
using
this
additional
option
or
a
subtitle
or
ever
you're
going
to
use.
It
could
be
a
subtype
of
the
multipath
type
message
message
that
you're,
adding
and
then
on
page
eleven
is
talking
about.
There
are
many
different
approaches
to
measure
the
one-way
delay
or
latency,
and
we
were
late
you'll
hear
from
your
feedback.
Whether
this
is
one
of
the
approach.
J
What
was
outlined
up
before
in
PHP
and
I
was
that
the
Cayenne
just
remembers
the
time
you
send
out
the
the
packet,
and
then
we
see
if
I've
only
had
choose
sandy
time.
The
packet
was
received,
there's
only
one
information
that
you
receive
essential
and
when
Wendy
Wendy
can
be
saved
the
time
it
received
a
packet,
it
will
calculate
the
difference
and
then
take
that
if
and
then
comparatively
featuring
the
different
hops
so
that
it
will
pick
the
best
part
to
send
a
packet
and
an
alternative
in
the
next
one.
J
J
The
kind
doesn't
need
to
EC
remember
the
time,
but
you
it
does
have
to
add
time
stem
in
the
packet
that
is
sending
compared
with
the
previous
one
and
then
again
you
would
just
compare
the
different
paths
and
then
pick
Tibet,
that's
a
better
route
and
and
and
the
another
alternative
in
the
in
the
bottom
is
that
the
D
server
well
I'm,
stand
and
serve.
I
will
calculate
the
difference
and
then
send
a
one-way
latency
back
to
D
to
the
client.
So
the
car
doesn't
need
to
do
any
computation.
J
They
only
need
to
compare
the
one-way
delay,
so
they're
posting
on,
depending
on
what
you
have,
what
you
need
and
then
in
the
next
slide.
Page
12
is
that
Oh
spit,
that
is
more
straightforward
and
it
probably
doesn't
need
and
additional
changes
in
the
standard.
Basically,
if
the
are,
if
there
is,
if
these
surface
eyes
sent
a
packet
in
the
downloads
in
the
download,
Direction
are
people
times
them,
then
the
client
knows
what
time
it
receive
it.
Then
it
would
just
calculate
the
one-way
agency.
J
B
B
B
B
R
R
Currently,
the
RTD
is
commonly
used
as
a
condition
for
data
scheduling
amongst
multiple
paths.
However,
as
just
described
by
the
Antoni,
he
sliced.
The
delays
of
the
forward
path
in
the
reverse
path
may
be
different,
especially
in
the
wireless
environment
and
also
in
the
scenario
when
the
congestion
may
not
have
heard
the
same
so
which
will
cause
the
different
cue
delays
so
we'd
better
to
consider
the
one-way
latency.
Yet
for
the
data
scheduling
next,
we
have
several
examples
and
to
show
why
the
one-way
latency
should
be
considered
for
the
data
scheduling.
R
R
Okay
in
this
picture,
if
we
consider
the
one-way
latency
as
the
condition
for
the
data
scheduling-
and
we
can
see
that
if
the
client
wants
you
to
transfer
transmit
the
data
to
the
server
in
case,
we
can
see
that
the
path
one
have
the
lowest
one-way
latency
compared
to
the
other
two
paths.
So
the
data
will
be
scheduled
in
the
past
one
based
on
the
one-way,
one-way
latency.
R
So
we
periodically
update
the
one-way
latency
of
all
paths
and
schedule
the
data
again.
Next,
we
will
explain
the
scheduling
scheme
for
in
three
step
use
for
indeed
health.
So,
firstly,
first,
let's
see
the
initialization
stage
so
here
during
the
initialization.
So
we
have
no
path
characteristics
for
both
of
the
past
so
and
also
MPDC
setup,
the
connection
of
the
primary
path
first
afterwards
and
then
setup
the
connection
of
the
other
path
path,
one
by
one.
So
at
that
time
we
don't
we.
R
We
can
transmit
data
redundantly
for
a
period
of
time,
for
example,
1
seconds
next
and
after
after
after
we
obtain
the
one-way
latency.
So
we
schedule
the
packet
according
to
the
one-way
latency,
and
we
can
see
that
before.
If
we
only
consider
one-way
latency
of
one
path,
you
may
think
there
is
a
time
synchronization
problem
there,
but
here
because
we
want
to
select
the
past
with
the
lowest
one-way
latency
so
by
comparing
by
comparing
the
one-way
latencies
of
the
two
part.
R
The
time
synchronization
issue
won't
exist
and
more
because
the
difference
of
the
absolute
time
clock
will
not
exist
will
be,
will
be
subtracted.
So
here
we
compare
the
past
one
and
pass
to
you
of
the
one-way
latency.
So
if
pass
one
has
the
lower
one
where
latency.
So
we
select
one
pass
one
to
transmit
the
data
next.
R
As
we
know,
the
network
environment
will
change,
because
in
the
wireless
environment
we
may
have
different,
we
may
have
packet
loss,
random
packet
loss
and
also
we
have
lots
of
flows
coexist
in
the
network.
Some
flows
come
in
and
some
flows
go
out,
so
we
better
gp2u
to
obtain
the
one-way
agencies
and
to
see
whether
they
change
or
not.
So
here
we
use
a
periodical
redundant
transmission
to
get
the
updated,
one-way
lenses.
Here
we
define
every
10
seconds.
We
go
back
to
you,
the
result
transmission
and
to
see
whether
the
one-way
latencies
change
or
not.
R
Here
why
we
use
the
result
transmission
to
do
the
measurement
of
one?
Well,
isn't
it?
The
reason
is
that,
because
one
other
reduction
transmission
transmit
the
data
in
all
the
paths
same
time,
so
you
can
get
a
one
with
a
disease
of
the
older
past
at
the
same
time,
at
the
same
time,
you
do
not
need
to
measure
them
one
by
one
and
the
other
question.
The
other
reason
is
you
the
read
on
transmission
because
they
transmit
the
data,
so
they
won't
introduce
extra
packets
and
the
third
reason
is
that
you,
the
result.
R
Okay,
there
is
a
case.
You
have
to
immediately
activate
the
redundant
transmission
here,
because
if
the
past
one
is
selected
for
the
data
transmission
and
then
you
can
know
whether
the
one
validity
of
that
path
changes
or
not,
if
you
found
that
the
one
relay
disease
of
that
selected
path
is
not
the
lowest
anymore.
R
Here
the
can
occurs
30,
which
means
that
this
is
my
past
history
option
and
we
define
the
subtype
as
the
direct
eight
since
and
since
the
0
X
0
to
0
X
7,
having
defined
okay
next
so
for
the
one-way
latency
calculation,
and
so
first
in
negotiation
is
needed
to
ensure
them
both
the
same.
There
is
a
receiver
suppose,
my
past
one-way
latency,
and
here
we
use
the
sender
to
calculate
one
will
Edison.
R
The
reason
is
that,
because
the
sender
need
to
compare
which
path
has
the
lower,
so
we
just
let
the
woman
in
the
sender
to
do
the
calculation
and
there
are
the
reason
it
not
because
if
you
use
the
sender
to
calculate
the
wobble
NZ
and
which
means
that
the
data
packet
do
not
does
not
need
to,
you
carried
the
timestamp
of
the
sending
time.
Okay.
B
A
This
is
your
serum
floor.
So
if
the
pass
is
completely
disjoint
here,
it
makes
sense
to
send
redundant
models,
make
sense
to
measure
a
long
wait
array.
But
if
the
past
is
jointed
and
then
you
use
redundant
mode,
then
old
data
is
on
the
same
path
and
then
does
in
in
this
case
still
does
this
works
for
one
wait,
break
one
way,
do
a
measurement.
What
it
does
it
effect.
Have
you
ever
considered
this
kind
of
scenario,
I'm.
A
R
Because
the
TCP
time
option
is
already
has
its
own
meaning
and
for
the
data
Brockie
it
carries
that
timestamp
as
a
senator
but
foot
in
ACK,
so
it
carries
to
timestamp
stream.
One
is
for
the
sending
time
of
the
data
and
the
other
one
is
for
the
folder
for
copy
there's
any
time
of
the
data.
Then
what
is
the
the
sending
time
of
attack?
So
we
do
not
want
you
to
change
the
meaning
of
the
timestamp
of
the
option
in
the
TCP
option,
but
it's
okay.
G
G
R
G
G
There
is
a
data
difference,
but
it's
a
small
one
right.
If
you
just
use
this
like
or
delay
I,
could
it
won't
be
if
you,
if,
okay
yeah,
we
had
this
program
but
in
any
case
I,
don't
think
that
measuring
one-way
delay
is
the
specific
function
you
need
for
all
right.
I
see
the
use
case
for
MB
TCP,
but
it's
it's
a
functionality
that
should
not
provide
it,
be
provided
in
an
MP
TCP
sub-options.
So
if
you
want
a
different
time,
stamping
mechanism
in
TCP,
it
should
be
a
TCP
option.
G
S
S
If
you
should
either
each
interface
needs
to
separate
you,
congestion,
control
or
if
you
try
to
send
transmitted
back
a
simultaneous
e
you've
either
got
to
do
some
kind
of
joint
congestion
control
across
both
interfaces
or
have
one
of
them
drive
the
other
one
and
I
just
don't
quite
understand
how
you
propose
it
to
work.
I.
R
S
So
if
you
switch
into
your
redundant
mode
for
say
a
second
but
the
second
path,
you've
just
added
is
going
say
ten
times
slower.
It
will
then
take
ten
seconds
to
transfer
the
data
that
the
first
path
sends
in
one.
Second,
do
you
give
up
at
some
point
and
say:
that's
just
going
to
be
too
slow,
or
what
do
you
do?
Yeah.
R
S
B
T
Right
hi
then
I'll
give
you
a
quick
update
on
stocks
version
six.
So
next
slide,
please.
So,
as
you
may
probably
know,
suck
six
stock
seats
we've
been
busy
trimming
down
the
arty
arty
T
overhead
right
now,
if
you're,
using
your
RTG
authentication
or
no
education
at
all,
the
overhead
is
zero
in
terms
of
protecting
against
malicious
third
parties.
T
We've
decided
to
take
a
step
back
and
rather
than
do
encryption
within
the
protocol,
we
just
run
it
over
TLS
and
now
we
start
version
six
also
mitigates
early
data
replay
attacks,
so
it's
also
safe
to
use
over
this
to
use
TLS
early
data
and
new
in
the
latest
version
of
the
draft.
We
added
whether
the
stock
up
like
mechanism
and
we've
used
it
to
implement
a
couple
of
MP
TCP
use
cases.
So
next
slide,
please
so.
First
opt-in,
this
in
this
latest
version
of
the
draft
we've
also
specified
how
to
perform
plaintext
password
authentication.
T
This
is
actually
viable
now
that
we,
that
also
runs
over
TLS,
and
we
expect
it
to
be
the
de
facto
standard.
So
you
simply
take
the
the
initial
message
of
the
exchange
from
RFC
in
1929,
and
you
place
it
in
our
socks
and
a
socks
request
as
as
as
an
authentication
data
option.
There's
a
small
caveat,
though,
if
the
username
and
password
I
are
exceptionally
long,
they
won't
fit,
and
you
still
have
to
do
the
regular
authentication
that
which
takes
one
RTT
next
slide
please.
So
these
are
the
socket
options
that
I've
been
talking
about.
T
Clients
can
use
them
to
request
certain
behavior
from
the
proxy
on
either
of
the
legs,
either
the
proxy
leg
or
the
proxy
server
leg,
and
these
options
were
heavily
inspired
by
sets
or
copped
and
get
so
copped.
However,
they
are
not
on
a
remote
procedure,
call
for
sensor
copy
and
get
sock,
opt
and
their
individual
semantics
and
have
to
be
individual
options
and
their
semantics
have
to
be
standardized
separately.
T
So
you've
got
socket
levels
for
four
layer
layer
for
their
to
four
layer.
Three
and
therefore
protocols.
We've
also
got
the
socket
layer
for
generic
socket
layer
and
the
option
the
option
the
option
and
then
you've
got
the
option
code
and
the
data
is
data,
depends
on
each
individual
option.
So
next
I,
please
the
simplest
use
case
we
found.
T
Fourth,
for
these
options
is
TFO,
so
there
previously
was
the
field
in
the
request
which
the
client
could
use
to
request
the
proxy
to
use
TF
Owen
in
this
connection
attempts
now
that
field
has
been
removed
and
it's
been
replaced
by
by
an
option.
The
absence
of
this
option
means
that
the
proxy
must
not
use
TFO
next
slide.
Please.
T
We've
implemented
out
an
option
which
can
be
used
for
proxy
bypass
you're,
probably
familiar
with
this
use
case.
So
let's
say
a
mobile
phone
has
both
a
Wi-Fi
and
and
the
4G
interface,
and
it
wants
to
take
advantage
of
both.
However,
servers,
typically
typically,
don't
deploy
MP
TCP,
so
the
client
is
so.
The
mobile
phone
has
to
go
through
a
proxy
first
now,
if
the,
if
the
server
happens,
to
support
the
MP
TCP,
the
proxy
candidate
had
no
by
including
an
MP
TCP
option
in
its
operation.
Replies
next
slide,
please.
So
in
further
connection
attempts.
T
The
client
can
just
contact
the
server
directly
and
forego
the
use
of
the
proxy
and
next
slide.
Please
we've
also
implemented
an
option
which
allows
the
client
to
specify
preferred
MP,
TCP
scheduler,
and
we
currently
support
all
of
the
schedulers
that
are
part
of
the
Linux
MP
TCP
implementation,
so
a
client
can
it
can
place
one
such
option
in
its
request
to
the
proxy
to
use
a
certain,
a
certain
scheduler
or
not
one
leg
word.
Both
legs
and
and
proxy
can
include
one
such
option
in
the
operation.
T
Reply
to
tell
the
client
to
inform
the
client
as
to
what
what
scheduler
it's
using
on
one
or
both
of
the
legs.
We've
found
a
use
case
for
this.
So
if
you're,
a
particular
application
requires
extremely
low
latency
it
can
you
ask
the
proxy
to
use
the
redundant
scheduler
and
have
the
data
be
duplicated
across
all
available
paths,
and
that's
about
it.
For
me,
any
questions.
C
A
T
A
T
T
B
Thank
you,
everybody
for
the
good
presentations
and
discussion
today,
and
you
will
see
some
work
last
call
soon,
which
please
go
in
make
comments.
You
know,
comment,
I've,
I've,
read
it
and
I
don't
have
any
comments,
is
a
good
comment
and
that's
useful
on
all
of
it
or
on
part
of
the
document.
Everything
is
valuable
or
feedback.
Thank
you
very
much
see
you
next
time.