►
From YouTube: IETF100-MPTCP-20171114-1550
Description
MPTCP meeting session at IETF100
2017/11/14 1550
https://datatracker.ietf.org/meeting/100/proceedings/
A
A
Okay,
we'll
wait
another
two
or
three
minute
and
there
starts
up
meeting
and
be
home
and
we
are
looking
for
note
taker
and
JavaScript
Pro.
If
anyone
want
to
brought
your
a
me
know,
anyone
if
you
are
interested
in
taking
notes,
could
you
raise
your
hand.
A
A
So
let's
get
started
so
welcome.
This
is
multi
pass.
This
multi
pass
TCP
working
group
meeting.
My
name
is
Yoshi
Ishida,
one
of
the
co-chair
of
the
multiple
TCP
working
group
and
unfortunately,
Phil
cannot
come
to
this
meeting.
I
think
he
might
be
in
the
remote
ok
and
we
already
got
a
note
one
note
taker
thanks
rod
in
JavaScript.
So
before
we
start
the
meeting
I
would
like
to
keep
one
things
remind
you
of
a
news
pick
up
at
the
mic.
A
A
A
This
is
the
agenda
for
the
today's
meeting
and
at
first
she
will
talk
about
working
group
status
and
so
on
and
after
this
we
have
three
presentation.
The
first
of
all
needs
consideration
for
MPD's
the
operation
in
5g
and
the
presenter
is
October
and
then
after
this
Jing
will
talk
about
a
productive
approach
to
avoid
a
Home
Instead
relation
on
MPT
CP
and
then
after
this
Brad
will
talk
about
Sox
version
6.
So
for
some
reason
we
have
representation
this
time.
So
para.
A
Okay,
maybe
on
so
let's
talk
about
the
working
group
status,
so
we
are
trying
to
complete
the
base
draft
right
now
and
then
in
chairs
understanding.
There
are
several
remaining
items
before
we
finalize
that
removing
address
identifiers
from
mp3,
oh
this
one
has
been
proposed
in
the
last
meeting
and
I
think
we
got
enough
support
for
this
one
and
then
this
one
is
already
incorporated
into
the
best
route.
A
So
I
think
this
is
already
done
and
guideline
about
MP,
TCP
and
TF
for
interactions
and
in
there
for
this
one
we
are
waiting
for
the
proposed
text
from
the
authors.
So
let's
wait
the
text
from
the
authors
and
then
once
we
cut
the
text
from
the
authors,
we
can
review
it,
we
can
discuss
it
and
the
next
one
is
fast
cross
update.
A
So
there
are
several
discussion
on
the
first
groves
approach,
which
is
slightly
different,
the
button
what
is
written
in
the
best
draft
and
right
now
there
are
several
discussion
and
then
this
morning
or
ebf
sent
a
new
proposal
stickies
to
the
mailing
list.
So
please
read
his
name
and
provide
some
comments
and
then
we
can
continue
the
discussion
and
then
I
think
Olivia
is
in
the
mode.
So
if
you
I
don't
recommend
to
the
very
detailed
technical
question
right
now,
because
no,
we
don't
have
us
right
and
they
obey
the
remote.
A
A
Think
we
need
to
see
proposal
before
we
need
to
discuss
whether
this
is
going
to
be
struck
or
not
so
I
think
they
also
should
creepier
proposal
in
order
to
proceed
this
one.
So
these
are
the
items
that
still
think
remaining
items
and
if
there
is
another
Ezra
stuff
that
should
be
discussed
in
the
piece
draft,
please
read
us
know.
B
A
Okay
seems
to
be
good
and
then
the
progress
of
this
draft
is
Bill
no
getting
derailed
these
days.
So
this
is
just
a
brief
idea.
That's
not
feel
and
I
are
discussing
about
having
wintery
meeting,
let's
say
in
January
or
February
timeframe,
and
if
people
think
this
will
be
useful
to
accelerate
the
process
of
the
finalizing
Mistral,
how
you
feel
and
I
will
happy
to
arrange
the
interim
meeting.
So
if
you
have
any
opinions
or
suggestions,
please
let
me
know
anything.
A
Okay.
So
so
far
there
is
no
preference
to
having
the
interim
meeting,
at
least
in
this
room,
okay
moving
on,
and
then
there
are
several
discussions
about
mp40
activities
these
days,
and
currently
we
have
two
draft
and
the
one
is
us
socks
version
six
approach.
This
is
purely
application
protocol
and
then
next
proposal
is
MP.
Tcp
converter.
A
This
trust
is
more
focusing
on
the
MP
TCP
to
align
with
our
chata
text
and
para.
The
combat
a
text
has
some
kind
of
generic
mechanism
and
then
because
of
this
generic
mechanism,
in
my
my
apply
this
almost
same
architecture
to
other
TCP
extensions
and,
for
example,
you
know
we
might
be
able
to
apply
this
technology
into
TCP
ink,
which
might
be
interesting
use
cases.
A
So,
after
the
know,
think
about
this
kind
of
possibility,
working
area,
direct
and
or
MPT
CPG
and
ppm
chair
has
several
discussions
and
then
decide
to
have
a
you
know
small
discussion
on
in
at
each
p.m.
working
group
on
Saturday.
So
if
things
goes
well,
you
know
we
are
thinking
about
running
adoption
core
on
this
draft
on
the
TCPA
marking
group,
meaning
risk,
but
still
you
know,
we
need
to
discuss
it.
So
the
main
focus
on
the
Saturday
discussion
in
ppm
is
to
discuss
the
future
direction
of
this
draft.
A
So
if
you
are
interested
in
some
opinions
about
future
direction
of,
it
is
left
whether
this
is
no
only
focus
on
PTT.
What
it
should
have
MPT
c4
should
add
more
like
a
generic
mechanism
for
any
kind
of
transport
extensions.
Please
provide
your
feedback.
Okay,.
D
Miku
Levine
and
say
this:
the
mechanism
described
in
the
document
is
actually
generic
to
any
kind
of
TCP
option
might
be
most
useful
for
MB
TCP.
However,
given
this
is
a
generic
mechanism,
we
want
to
see
if
there's
interest
in
a
TCP
and
working
group,
we
go
straight
to
for
an
adaption
call.
Unfortunately,
others
cannot
be
here
to
give
a
presentation.
That
would
be
much
easier.
However,
we
run
an
adoption
call
if
the
adoption
call
fails,
which
means
there's
like
no
interest
in
that
group.
A
So,
basically,
we
will
discuss
a
future
direction
of
the
grass
in
TCP
M.
So
we
don't
say
you
know
we
are
no
accommodate
this
or
of
the
in
TCP
right
now.
We
just
try
to
discuss
figures,
people
think
that's
a
motivation
and
then
so
other
media
said
no.
They
also
is
not
entertaining
so
the
Thursday
meeting,
it's
not
a
chance.
A
A
Just
I
would
like
to
make
sure
okay,
then
implementation
news.
So,
as
you
might
notice,
no,
we
got
new
MP
disappearing
acts,
Connor
implementation,
the
bottom
number
is
0.93
and
it
has
been
released.
No
big
buzz
second,
and
this
release
mainly
focus
on
standardization
Rock,
a
stabilization
work,
I
mean
bug,
fixes
and
then
catching
up
with
upstream.
So
as
a
result,
the
current
MP
disappearing
acts.
Implementation
is
based
on
Linux
kernel
bottom.
A
Forty
four
point:
nine
and
then
this
because
of
this
it
allows
the
user
to
utilize
MP
TCP,
which
TCP
PPR,
and
this
might
be
related
to
the
jingles
application
today
and
and
then
also-
and
they
are
also
focusing
on
the
you
know
catching
up
upstream,
so
they're
prone
needs.
Next
release
will
be
based
on
bass
on
4.14,
so
do
some
real
refactoring
in
order
to
catch
up
the
absence,
that's
seems
to
be
their
plan,
so
any
comments
about
implementation,
news
or,
if
you
know
other
implementation
news,
please
let
us
know.
A
A
E
He
said
no
and
he
would
be
interested
in
knowing
what
would
what
would
what
use
case
that
might
occur
in
so
that
that's
basically
what
we
tried
to
document
in
this
draft,
so
I
think
that
you
know
everyone
here
is,
of
course
the
experts
on
MPTV.
So
basically
the
the
basic
idea,
though,
obviously
is,
if
you
have
multiple
interfaces,
meaning
practically
speaking
Wi-Fi
and
cellular.
Most
of
the
time.
E
So
even
if
you
have
mobility
going
around
at
the
radio
access
you,
the
anchor
only
sees
or
only
assigns
one
IP
address,
/
/
ue4,
the
what
they
call
the
PDU
session
and
the
mobility
is
taken
care
of
more
or
less
at
the
radio
access
network,
and
your
IP
address
doesn't
really
change
because
of
that.
But
what's
happening
for
5g
in
3gpp
is
they're
moving
to
a
more
distributed
and
closer
to
the
edge
mobility
model
and
basically.
E
What
they're
going
to
do
is
to
support
the
make
before
break.
Essentially
the
make
before
break
mobility
is
they're
going
to
allow
multiple
anchors
to
exist
simultaneously
for
a
given
UE,
and
we
gave
the
reference
of
the
3gpp
document
that
describes
this.
So
you
can,
you
know,
confirm
for
yourself
if
that's
your
same
interpretation
but
I
think
the
text
is
pretty
clear
there
and
basically,
what
will
happen
is
for
the
equivalent
make
before
break
in
5g
you'll
now,
in
certain
cases,
have
multiple
IP
addresses
being
assigned
to
the
given
UE.
E
So
from
the
you
know,
interface,
if
we
call
the
whole
cellular
network,
the
interface
you
might
have
multiple
IP
addresses
coexisting
for
a
given
UE
at
a
given
point
of
time.
Basically
to
you
know
to
keep
it
simple
to
have
make
before
break
with
them
with
the
mobility
anchors
pushed
towards
the
edge
so
anyways,
the
3gpp
alice
has
very
specific
terminology.
So
here
you
know
they
have
this
branching
point
and
PDU
session
and
all
this
which
is
basically
taken
directly
from
their
Tia
I,
believe
it's
a
TS,
the
3gpp
document.
E
That's
referred
there
in
the
draft
and
really
we
just
wanted
to
bring
it
up
to
the
group.
If
this
really
does
cause
any,
you
know
potential
issue,
I
guess
I
guess:
MP
TCP
will
still
work
because
at
the
worst
case,
you'll
have
multiple
IP
addresses
on
a
given
interface,
and
you
know
you
might
think
you
have
a
backup,
but
you
don't,
you
might
think
you're
doing
bandwidth
aggregation,
but
you
don't
but
I
guess
the
algorithms
will
will
will
still
take
care
of
it,
but
you'll
get
the
lowest.
E
You
know
possible,
throughput,
I,
guess
to
that
you
have
an
interface
and
you
might
have
a
problem.
If
you
try
to
fall
back
and
then
actual
fall
back
IP
address
is
also
disappearing.
So
really
all
we
wanted
to
do
is
follow
up
on
the
question
we
had
asked
because
we're
following
what
happens
in
in
3gpp
and
basically
want
to
bring
it
up
to
the
group
whether
they
also
see
an
issue
and
whether
there's
anything
to
be
done.
C
Shaylee
cisco,
so
I'm
just
curious,
like
from
the
endpoint
today,
when
you
open
in
mp
TCP
sessions
right,
would
you
pick
the
address
based
on
an
interface
or
would
you
would
you
pick
multiply
piece?
When
would
you
why
would
you
take
multiple
IPS
from
the
same
interface?
When
you
are
doing
link
aggregation,
you
will
do
Wi-Fi
and
LTE
link,
aggregation
or
LTE
and
DSL
link,
aggregation
or
whatever
right,
whatever
radius
interfaces
that
you
have
in
each
interface
even
today
and
wireline
I
can
get
an
ipv4.
E
Sorry
to
interrupt
she,
what
I
understood
you
know
from
the
it
was
the
Apple
developer
last
time
who
spoke
and
when
we
had
the
question
and
answer,
he
was
saying
that
they
did
not
assume
it
that
that's
any
was
that's
what
I
understood.
He
said
he
said
they
did
not
assume
on
cellular.
You
could
have
more
than
one
IP
address.
You.
F
The
reason
is
they
don't
want
to
use
it
for
bandwidth.
The
sharing
or
anything
like
that.
This
is
specific
use
case
where
mobility
happens
like
for
existing
3G,
4G
use
cases
where
Wi-Fi
and
cellular
IP
addresses
are
there
in
two
separate
interfaces
that
is
not
changing
in
Phi
G.
What
is
changing
in
Phi
G?
Is
you
be
of
mobility?
What
we
described
we're
temporarily
sub
flow
has
to
be
created
and
the
worlds
of
flow
has
to
be
removed.
There
is
no
bandwidth
sharing
or
aggregation
at
this
point.
G
You,
steel,
chair
sure,
from
Apple
I'm,
not
the
Apple
engineer
you
were
talking
about,
but
he's
one
of
my
colleagues
you,
you
asked
the
question
here
about
how
MP
TCP
would
have
to
change
or
whether
it
would
work
and
I
would
put
it
completely.
The
other
way
around
the
issue
that
our
current
code
only
uses
a
single
address,
I
think
it's
a
minor
engineering
detail,
it's
nothing
fundamental.
G
What
is
fundamental
is
if
a
mobile
device
changes
addresses
as
it
moves
around
then
multipath
TCP
allows
it
to
maintain
a
connection
by
bringing
up
new
sub
flows
on
the
new
addresses
and
dropping
the
old
ones.
So
it's
traditional
TCP,
that's
going
to
fall
on
its
face
and
break
here,
which
makes
this
a
very
compelling
argument
for
more
use
of
multipath,
TCP
I.
Think.
E
E
No
we're
not
proposing
a
solution,
we're
basically
just
trying
to
identify
the
use
case
and
if,
if
there
is
anything
or
if
maybe
a
steward
is
saying,
there's
nothing.
It
just
makes
everything
better
that
that's
cool.
That's
fine,
too!
So
we're
just
trying
to
follow
up
on
trying
to
understand
how
the
system
will
work
in
5g.
Ok,.
F
There
was
a
discussion
like
few
years
back
for
the
UPF
complete
exactly
the
case.
You
documented
here
wait
20250
at
5:01
document
at
the
branch
point
and
UPF
they
were
discussion
that
they
wanted
to
do
the
mobility
to
the
UE.
That
is,
like
you
know,
not
angry
mobility
that
the
existing
3gpp
they
want
to
do
to
them.
P
TCP,
all
the
weight
for
the
mobility
management,
but
that
was
not
done
so
MP
TCP
only
used
for
3gpp
3gpp
are
non
gtp
access
and
the
regular
anchor
point
probability
is
still
there
in
the
3gpp.
F
H
Mark
Townsley
from
Cisco
I'm
gonna
channel
some
of
the
people
I'm
working
with
some
of
his
colleagues
and
some
over
at
Google
working
on
the
Android
platform
and
I
think
it's
RFC,
79
34
the
details,
all
the
reasons
why
you
might
have
more
than
one
ipv6
address
on
a
given
interface,
it's
not
just
in
the
5g
case.
It's
a
general
thing.
C
B
C
E
Basically,
the
the
reason
we
were
interested
in
this
is
because
we
have
colleagues
from
our
company
that
go
to
3gpp
and
I.
Don't
think
it
was
formally
discussed
there,
but
there
was
some
discussion
about
since
the
model
for
5g
is
changing,
apparently
from
what
3gpp
does
for
4G
and
3G
was
there
going
to
be
any
impacts
on
the
IP
applications,
TCP,
etc?
So
that's
how
we
were
originally
interested
in
the
problem
and
that's
why
we
had
asked
the
question
in
the
last
IETF
meeting
MP
TCP
meeting,
and
so
basically
that
was
our
motivation.
E
I
As
far
as
n
systems
are
concerned,
where,
for
example,
Stuart
Stuart's
case
the
what's
going
to
go
to
Apple,
probably
isn't
going
to
get
broken
out
early.
So
it's
not
obvious
to
me.
You
would
actually
have
an
issue
there.
A
lot
of
the
early
breakout
scenarios
are
concerned
with
non
mobility
and
Internet
of
Things,
so
yeah.
E
E
I,
don't
know
if
I
call
survive
for
the
within
the
network,
I'm
not
sure.
What's
going,
they
have
the
the
branching
points.
If
you
follow
the
3gpp,
the
latest
TSS
I,
guess
they
have
this
common
that
well
they
call
this
branching
point.
So
my
understanding
is
basically
that
that's
the
thing
in
the
network
that's
going
to
collect
from
the
multiple
anchors
and
they're
gonna,
put
it
all
together,
either
on
the
downlink
or
the
uplink
and
give
it
one
face
to
the
mobile.
So
that's
my
understanding
how
3gpp
is
approaching
it,
and
the
other
thing
is.
E
E
F
Actually
you
proposing
that
you're
showing
the
branching
point
solution
in
five
zero
one
because
of
the
fact
that
if
we
showed
that
that
means
that
they
are
not
considering
for
UPA
publicly,
they
are
not
considering
MP
TCP.
That's
one
thing:
you
have
to
bear
it
in
mind.
If
it
is
wrong,
tell
me
that
is:
can
you
just
expand
just
please
because
of
the
fact
that
they
are
using
branching
point
and
2
P
2?
U
PF
s!
That
means
they
have
a
network
solution
for
the
UPF
mobility.
F
They
are
not
going
to
come
to
MP
TCP
for
that
solution,
so
MPT
PMP
TCP.
Today
they
are
using
for
Wi-Fi
to
cellular
and
that
is
being
used
and
there
was
discussions
for
using
UPF
mobility,
but
because
of
the
fact
that
you
are
showing
the
branching
point
and
to
PU
UPF,
they
have
a
network
solution.
Okay,.
A
Point
one
comment
I've
been
previously
proposed:
maybe
you
know
some
useful.
The
proposal
may
be
related
to
your
approach.
So
what
have
you
proposed
is
vanu
mptp
and
the
point
announced
the
IP
address.
I
basically
proposed
to
put
some
kind
of
communication.
I
let's
say
I
have
a
for
IP
addresses
and
then
a
and
the
B
has
the
same
interface.
A
I
can
see
the
same
communication
I
did
to
the
address
so
that
they
can
use
autobot
ID
others,
you
can
know
a
and
B
are
same,
belong
to
the
same
interface,
a
C
and
D
under
defending,
so
that
you
can
have
some
strategy
to
find
how
you
set
up
the
subfloor.
That's
kind
of
approach.
She's
mentioning
is
it's
related
to
your
proposal.
Maybe
you
can
utilize
it
yeah.
E
A
E
J
J
J
So
we
we
already
have
existing
work,
the
which
is
design
that
Chris
called
opportunistic
retransmission
and
the
panel
penalisation
solution
and
this
solution,
which
is
try
to
react,
I'm,
acknowledging
data
from
this
low
path
and
to
the
fast
path,
and
afterwards
it
have
the
congestion
window
of
the
slow
path.
However,
even
with
this
method,
we
found
that
a
performance
degradation
still
exists,
because
we
found
that
this
may.
This
solution
is
only
triggered
when
the
performance
high
stick
rated,
so
this
is
a
kind
of
reactive
solution
and
we
found
that
this
agree.
J
So,
hence
we
think
we
need
a
new
solution
and
we
require
this
solution.
It
should
take
proaction
a
proactive
actions
based
on
the
path
attribute
and
and
also
we
are
required
to
solution,
can
adaptively
employ
both
the
path
or
only
the
best
path.
According
to
the
past
attributes,
however,
this
proactive
action
sometimes
can
be
counterproductive,
so
because,
because
we
need
to
get
the
Super's
of
every
single
path
and
also
the
aggregated
path,
this
Drupal's
over
an
estimated
by
the
same
window,
a
sim
window
by
the
rtt.
J
J
J
Hence,
so
that's
why
we
consider
to
combine
our
solution
with
the
PBR,
but
there
is
still
a
challenge
that
how
to
get
the
throughput
of
each
pass
and
mud
pumps
efficiently.
Currently
we
have
existing
solutions.
One
is
a
modeling
solution,
but
this
modeling
solution
requires
the
past
attributes
such
as
the
RTG,
the
PIR
and
the
bandwidth,
but
this
modeling
solution
is
relevant
to
the
specific
congestion,
control
and
flow
controls.
So
another
solution
is
measurement.
J
Ok,
we
call
our
solution
pad,
which
is
the
aggravation
of
the
proactive
approach
and
to
avoid
the
performance
degradation,
and
this
solution
consists
of
two
parts:
throughput
measurement
and
the
mode
selection
for
the
throughputs
measurement
we
employ
two
modes,
one
is
the
one,
is
the
redundant
mode
and
the
other
is
the
empty
mode
for
the
redundant
mode.
We
use
the
redundant
schedule
and
to
simultaneously
to
major
of
the
Super's
of
each
part
and
the
best
part
for
the
NP
mode.
We
use
the
mean
RTD
schedule
and
you
measure
the
aggregate
into
throughput.
J
Of
course
you
can
use
all
those
schedules,
as
you
like
after
we
measure
the
throughput,
we
compare
the
support
and
then
we
select
the
mode
the
mode
with
the
higher
some
food
will
be
selected.
So
if
the
redundant
mode
is
selected,
so
we
need
to
select
the
path
with
the
higher
throughput
to
send
the
live
data
and
if
the
empty
mode
is
selected,
so
we
found
all
the
paths
together
to
send
the
data.
Okay,
we
also
need
to
consider
how
to
design
for
the
two
stages
of
the
path
solution
either
slow
start
stage.
J
We
only
use
the
redundant
mode.
The
reason
is
because,
at
this
stage
we
know
nothing
about
the
past
attributes
so
and
also
for
each
after
each
RTD
and
the
sim
window
increases
twice,
and
thus
we
know
of
each
pasta
increases
as
a
lady,
so
du,
which
is
due
to
the
different
RTD
and
different,
stop
flow
connections
at
a
time.
That's
why
we
use
the
redundant
mode,
which
is
try
to
make
sure
there
is
no
performance
degradation
on
this
stage
and
after
we
found
that
the
pass
is
fully
utilized.
J
So
we've
been
to
the
congestion
avoidance
stage
at
this
stage,
because
we
have
to
remember
that
the
network
is
always
time
variant.
So
at
this
stage
we
take
turns
to
one
the
two
most
and
to
measure
the
throughput
of
fasten
a
monocle
posh.
So
after
we
measure
the
seafood
and
then
we
compare
and
select
the
mode
afterwards
we
select
the
best,
but
did
the
past
okay.
Now,
let's
see
the
results
we
in
the
left
figure.
The
barber
is
largely
enough.
We
can
see
that
me.
J
Our
duty
schedule
performs
good
and
our
pad
solution
performs
almost
the
same
as
in
the
majority,
and
naturally
it
achieves
about
more
than
94%
of
the
mean
our
duty
schedule
and
in
rap
figure
the
buffer
is
limited.
So
you
can
see
that
the
Ming
RTT
is
not
good.
It
only
performs
at
twenty
twenty
six
percent
of
the
best
parts
for
our
path
solution
can
achieve
almost
the
same
as
the
best
path.
J
J
J
J
Because
the
PBR
is
a
rate
basis,
so
we
do
not
need
to
which,
which
is
not
window
based.
So
we
do
not
need
to
cover
with
the
two
paths
and
also
this
measurement
is
dependent
on
any
other
congestion
control
schemes
or
any
other
flow
control
schemes.
So
it's
not
work.
Okay,.
D
I
mean
the
the
goal
of
the
coupling
is
that
you
basically
always
get
at
least
as
much
performance
as
you
would
get
on
them
on
the
better
or
on
the
best
flow
right
and
you're,
not
using
more
resources
than
you
would
otherwise
get.
But
this
is
like
also
differing
discussion,
and
my
second
point
is:
did
you
see
the
presentations
on
PBR
in
the
ICC
Archie
in
the
internet?
Congestion
control,
research?
Group?
Yes,
okay,
so
you
are
aware
of
any
problems
there.
K
J
K
G
So
so,
in
that
case
now
the
performance
we
care
about
is
not
the
most
throughput.
If
that's
a
high
delay
link
we'd
rather
have
the
lowest
delay
link.
So
it's
just
a
comment.
Maybe
use
the
word
throughput
rather
than
performance,
because
performance
ought
to
mean
many
different
things,
depending
on
the
application.
Yeah.
J
L
Since
Hal
from
Hawaii
so
to
your
question
so
agree,
we
would
probably
use
the
throughput
is
better
for
destruct
and
for
the
costing
of
me
RTD.
Actually,
we
use
kill,
use
me
Rd
T
as
the
schedule
for
the
multipath
segregation.
So
in
the
special
case,
morality
is
not
only
for
to
decrease
the
latency
of
the
application
flow,
but
also
in
such
cases
of
the
discrepancy
or
two
links.
It
also
can
help
to
increased
throughput
when
the
two
paths-
really
you
know
different
from
children
two
months.
Yes,.
G
G
One
of
the
reasons
is
because
throughput
is
easy
to
measure
you
run
iperf
and
see
what
number
it
says,
and
sometimes
because
it's
easy
to
measure
and
and
the
home
users
go
to
speedtest.net
to
see
how
fast
their
internet
connection
is,
and
only
the
really
smart
ones
know
that
the
round-trip
delay
matters
as
well.
So
speaking
for
Apple,
we
don't
care
one
bit
about
increasing
throughput.
G
G
Yes,
that
information
is
not
secret.
You
know,
Apple
is
very
secretive
and
other
things
we
don't
talk
about,
but
that
that
was
disclosed
publicly
at
the
Apple
Developer
Conference.
The
videos
are
online.
Our
engineer,
Christoph
Parrish,
who
gave
a
description
about
how
developers
can
use
that
and
we
we
want
to
make
multipath
tcp
available
for
our
developers
to
experiment
with,
because
they
may
think
of
things
that
we
haven't
thought
of
and
that's
why
we
want
all
the
capabilities
available
for
our
own
users
right
now.
G
We
have
not
really
thought
of
a
case
where
we
care
that
much
about
more
more
throughput,
so
I'm
just
saying
for
our
own
use.
We
make
that
available
and,
if
developers,
think
of
a
way
to
use
that
and
I
can
imagine
the
Netflix
client
if
you
saturate
your
Wi-Fi
link
and
saturate
your
LTE
link.
At
the
same
time,
then
you
get
a
slightly
higher
quality,
a
4k
video
on
your
iPhone.
By
using
you
know
both
interfaces
full
blast.
That
might
be
something
that
other
app
developers
think
of
creative
uses
for
I'm.
K
M
M
M
Yeah
so
I'm
gonna
start
by
giving
you
a
quick
refresher,
the
main
motivation
as
to
why
we're
trying
to
bump
the
version
number
is
that
Sox
5
my
makes
use
of
several
round
trips,
and
that's
that's
not
really
necessary,
so
we're
trying
to
reduce
the
di/dt
overhead
down
to
zero
in
the
in
the
optimal
case.
So
next
slide,
please,
yes,
so
the
basic
idea
behind
Sox
six
is
that
the
client
is
optimistic
and
sends
as
much
information
as
possible
upfront.
So
it
tries
to
authenticate
itself.
M
M
Yes,
so
this
is
basically
how
how
Sox
6
looks
when
come
when
compared
to
Sox
5,
so
in
Sox,
5,
the
clients
first
advertises
its
known
authentication
methods
and
the
server
proxy
chooses
one
of
those
methods.
Then
they
goes
to
the
authentication
process.
Then
the
client
makes
a
request
for
the
proxy
to
open
up
a
socket
and
only
then
can
data
flow
in
Sox
6.
The
request
contains
the
client
support,
l,
authentication
methods.
It
also
tells
the
proxy
what
server
it
should
connect
to,
and
it
also
sends
a
chunk
of
data.
M
M
M
So
in
our
latest
work
we've
been
focusing
on
security
security,
socks5
allowed
allowed
for
in
protocol
encryption,
we've
decided
to
replicate
that
and
just
run
socks
over
TLS
now
TLS
1.3
has
support
for
early
data.
That
means
that,
on
the
clients
second
connection
attempt
and
from
then
on,
it
can
send
data
as
as
part
of
its
first
as
as
part
of
its
first
packet.
This
encourages
ero
ITT
overhead,
but
it's
it's
prone
to
replay
attacks
now,
since
the
Sox
request
also
has
has
zero
ITT
overhead
and
can
likely
fit
in
that
early
data.
M
M
M
This
whole
negotiation
takes
part.
It's
part
of
require
of
Sox
requests
and
operation
replies,
so
the
authentication
have
nothing
to
do
with
it.
So
next
I,
please
so
the
client
starts
by
requesting
some
tokens
here.
We've
got
the
three
stocks
control
message:
procedures
like
the
request,
the
authentication
reply
and
the
operation
reply.
The
next
I
please.
M
No,
it's
okay,
I've
been
having
feed
issues,
so
the
client
can
include
a
token
request
as
part
of
the
Sox
request
in
which
basically
says.
Please
give
me
that
many
tokens
after
it
finishes
authenticating
the
proxy
replies
with
on
window
advertisement,
which
basically
tells
it
what
tokens
it
can
you
use
so
next
slide,
please.
M
So
the
token
request
just
includes
requested
the
requested
window.
Size
basically
says.
Please
give
me
this
many
tokens.
The
client
can
just
use
it
as
part
of
any
request
if
it
doesn't
have
a
socket
or
if
it
doesn't
want
to
open
a
socket
by
the
way
you
can
just
use
a
no
op
request,
which
basically
does
nothing.
M
This
is
secure
as
long
as
it's
not
done
over
TLS
early
data,
so
client
implementations
should
avoid
using
TLS
early
data
when
making
token
requests
next
slide,
please
so
the
server.
So
the
proxy
applies
with
token
window
advertisement.
Basically,
it
contains
two
integers
the
window
base
and
the
window
size.
The
window
base
is
the
first
talk
usable
token.
The
window
size
is
basically
the
number
of
tokens
that
the
client
can
use.
So,
for
example,
the
base
is
10
and
the
size
is
3.
M
So
once
a
client
has
some
tokens
to
spend,
it
can
start
spending
them
on
requests,
so
the
client
should
include
the
token
expenditure
and,
as
part
of
its
request
and
the
software
basically
sends
back
an
expenditure
reply
as
part
of
the
operation
reply
and
an
optional
window
advertisement
in
case
the
window
has
changed
so
next
slide,
please
so
the
token
expenditure
is
straightforward.
It
just
contains
the
token
that
the
client
is
trying
to
spend
clients
should
obtain
to
spend
the
tokens
in
order.
M
M
M
M
So
we're
gonna
we're
gonna
focus
more
on
npt
CP
for
our
next
draft,
we're
thinking
about
including
options
that,
via
which
the
client
can
request
that
the
proxy
use
a
certain
path
manage
or
packet
scheduler
for
MP
TCP
and
we're
also
thinking
about
adding
better
reverse
proxy
support.
Socks5
only
lets
you
accept
one
connection
from
the
outside
and
then,
if
you
want,
if
you
want
to
accept
another
remote
incoming
connection
from
a
remote
location,
you
have
to
make
another
request.
M
M
So
there's
another
draft
by
oliviere
that
deals
with
a
with
a
similar
issue,
our
two
proposal.
What
our
two
proposals
have
in
common
is
that
there
is
no
control
data
aside
from
the
initial
exchange,
so
afterwards
data
flows
verbatim
via
the
proxy.
However,
we
have
chosen
different
starting
points,
so
Sox
four
six
is
period
layer,
five
protocol.
M
A
A
D
M
A
Okay,
there
is
no
more
questions.
Thank
you
so
much
for
Brad
thank.