►
From YouTube: IETF104-HOTRFC-20190324-1800
Description
HOTRFC meeting session at IETF104
2019/03/24 1800
https://datatracker.ietf.org/meeting/104/proceedings/
A
All
right:
well,
it's
kind
of
trickling
now
so
I
might
as
well
get
started,
so
my
name
is
Aaron
funk,
and
this
is
the
request
for
a
conversation
event.
Hot
RFC.
This
is
I,
think
the
fourth
time
that
we've
done
this
event
and
people
seem
enthusiastic
about
it,
so
I'll
be
interested
in
getting
feedback
at
the
end
of
the
night.
A
If
you
think
this
is
a
good,
the
the
point
of
this
event
is
to
have
a
series
of
lightning
talks
that
gives
folks
who
have
an
idea
or
a
proposal,
or
something
going
on
here
and
they're,
looking
to
find
more
collaborators
or
get
more
input
a
chance
to
sort
of
get
the
the
gist
of
their
idea
out
there
and
then
provide
some
coordinates
so
that
more
conversation
can
happen.
So
the
idea
here
is
not
to
have
to
do
any
kind
of
a
thorough
presentation
or
or
have
really
have
any
dialogue
here.
This
is
a.
A
This
is
really
a
broadcast
session
right.
You're
gonna
hear
about
a
bunch
of
different
things,
the
their
abstracts
for
all
the
talks
that
are
posted
in
the
agenda
and
we're
gonna
go
through
the
order
as
they're
written
in
the
agenda,
all
the
slides
are
posted
on
the
material
site
as
well.
So
if
you
see
something,
you
can
always
go
back
and
refer
to
it,
there
and
I've
asked
the
presenters
to
include
in
the
abstract
and
also
in
their
presentation
some
indication
of
what
forum
they'd
like
to
continue
the
conversation
in
so
some
people.
A
A
So
the
format
is
is
simple:
yeah
and
you
know
crisp,
you
get
four
minutes
and
that's
it
at
the
end
of
the
four
minutes
when
we're
doing
the
large
Eggert
time
management
rules.
So
I
want
to
thank
my
eye.
Our
TF
chair
successor,
Lars
Eckert,
for
how
to
do
this,
which
is
when
you
hear
the
timer
go
off.
A
I
would
like
everybody
to
start
clapping,
that's
the
end
of
the
presentation.
Okay,
that's
it!
So
you
get
four
minutes
and
then
it's
over
so
try
to
say
what
you
got
to
say
in
four
minutes.
If
you
don't
get
to,
if
there's
something
that
you
haven't
finished,
you
can
continue
the
conversation
somewhere
else
yeah.
B
Hi,
my
name
hi,
my
name
is
brett
jordan.
Today,
we're
gonna
be
talking
a
little
bit
about
course
of
actions
and
automated
play
books,
and
what
does
that
mean?
We
do
have
a
mailing
list
that
you
can
monitor.
You
can
subscribe
to
you
to
contribute.
We've
done
an
introductory
draft.
You
can
read
to
get
some
more
information
about
this
and
then
we're
also
having
up
off
on
Friday
at
9:30.
B
B
So
defense
is
slow,
it's
typically
very
manual
and
very
siloed,
and
so
when
we
think
about
the
way
a
threat
actor
is
going
to
permeate
a
network
they're
going
to
come
in
and
in
this
example,
we're
going
to
hit
a
mobile
set
of
devices
and
then
the
threat
actor
will
move,
laterally
and
compromise
throughout
the
network.
And
what
typically
happens
is
all
these
different
groups
need
to
be
involved
at
various
stages.
B
To
do
the
mitigation
remediation
and
then
future
prevention,
so
there's
a
lot
of
work
that
goes
on
and
a
lot
of
people
need
to
be
involved.
So
what
is
Kakao?
It's
collaborative
automated
courses
of
action
or
play
books.
It
is
a
proposed
standard
to
deal
with
the
creation
and
distribution
and
then
the
monitoring
of
the
action
so
that
you
know
if
they
were
successful
and
if
not,
what
would
be
the
back
out
rules
or
the
conditional
logic
around
them.
So
it's
fundamentally
around
prevention
mitigation
and
remediation
of
threats
in
the
cyber
domain.
B
So
when
our
playbook,
so
if
you're
not
familiar
in
sock
or
at
the
security
operational
side
of
it,
typically,
you
have
a
set
of
documents
that
describe
how
you're
going
to
respond
to
a
threat,
and
these
will
be
documented,
they're,
typically
in
a
binder
on
a
shelf
or
they
might
be
in
a
knowledgebase
article.
And
so
you
have
to
work
with
a
lot
of
different
groups.
B
If
you
have
different
business
units
or
different
enclaves,
you
know
that
information
needs
to
be
shared
and
it
does
not
allow
organizations
to
respond
quickly
or
efficiently,
and
so
what
we'd
like
to
do
is
design
a
data
model
and
a
methodology
for
allowing
users
to
collaborate
and
to
work
on
them
and
to
share
them.
So
when
an
industry
example
in
response
to
something
like
fuzzy
panda,
you
could
have
various
entities
or
various
groups
to
find
solutions
for
certain
products
in
a
portfolio.
B
They
can
all
digitally
sign
them
and
then
they
could
turn
around
and
disseminate
them.
So
an
end
organization
would
have
some
level
of
trust.
So,
instead
of
going
out
to
some
blog
post
and
trying
to
figure
out
what
you
need
to
do
for
Windows
10
service
pack
3,
you
could
potentially
link
this
into
threat
intelligence
and
push
this
out
through
a
nice.
Our
Isacc.
B
B
A
C
D
My
name
is
Linda
Dunbar,
so
here
I'm
just
going
to
talk
a
little
bit
about
some
of
the
activities
for
IC
UN,
because
it
being
scattered
in
different
places,
different
areas.
I
was
told
it's
good
to
come
here,
get
more
people
to
be
aware
where
they
are
and
how
to
contribute.
So
if
anybody
are
in
industry,
probably
many
people
most
of
you
have
heard
sd1
night
sounds
very
simple:
is
policy
pays
falling
based
on
the
application
and
policies
and
reaching
the
cloud?
D
So
that's
like
a
30,000
foot
view
of
sd-1
from
the
surface
doesn't
seem
to
have
much
working
as
a
ETF,
but
if
we
peel
the
onion,
if
you
look
in
detail-
and
actually
there
are
lots
of
new
things
being
introduced,
so
first
we
break
the
sd1
into
two
categories.
One
is
basically
over
the
internet,
like
you
have
a
small
branch
office
or
you
have
a
virtual
router
in
the
cloud
data
center
to
reach
all
the
branch
offices,
so
everything
has
to
be
encrypted.
D
Another
case
could
be
service
providers,
opening
a
new
port
for
the
internet,
offload
for
less
important
work
and
also
like
they
have
connections
to
the
private
network
which
doesn't
need
encryption.
So
the
key
things
here
is
really
the
one
port
management,
because
the
sd1
notes
can
be
started
in
a
call
center
in
a
shopping
mall
and
so
require
zero,
attach
provisioning.
D
So
for
the
control
plan,
part
is
basically
I
just
want
to
give
you
update
on
the
current
work.
When
is
in
the
routing
area,
we
have
a
problem
statement
and
gap
analysis
in
RTG,
a
working
group
and
in
the
in
the
idea
we
have
a
draft
to
propose
a
new
address
family
for
the
web
port,
facing
the
ISPs
and
for
the
bass
working
group.
There's
a
secure,
VPN,
secure,
Elstree
VPN
talked
about
how
do
we
stretch
the
PE
from
the
private
links
to
the
public
Internet
and
in
the
security
area.
There
are
two
activities
there.
D
One
is
the
IPSec
Sdn
control
a
kassak
management.
Another
one
is
a
hard
way
to
the
pairwise
key
management
wiki.
And
how
do
we
utilize
the
SDN
controller
to
do
that
in
the
ops
area?
There's
a
service
model?
How
do
we
start
I,
steal
and
services?
So
here's
some
of
the
counter
work
and
if
anybody
interested
please
come
to
those
working
groups
or
you
can
talk
to
me
in
the
hallway.
Thank
you.
Thank
you.
E
E
I'm
going
to
speak
about
something
that
may
be
interested
to
hear
web
security
systems
I've
been
working
with
web
security
protocols
since
HS
back.
You
can
see
that
I
guess
but
I
start
with
XML
as
everyone
else,
but
did
the
switch
to
Jason,
and
it
was
relatively
painless,
but
there
were
things
that
we
didn't
got
from
the
XML
world.
One
thing
we
didn't
got
was
a
complexity,
and
that
was
good,
but
there
were
other
things
that
we
may
need
like
this
one.
Everyone
said
everybody
sends
messages
with
JSON
over
HTTP.
E
Some
want
this
message,
this
message
to
be
signed,
but
there
is
no
standard
for
doing
that.
So
those
who
are
interested
they
essentially
build
their
own
stuff.
Today,
there's
absolutely
nothing
wrong
with
that.
These
solutions
work.
The
question
is:
couldn't
there
actually
be
room
for
a
standard
for
this,
and
that
is
what
is
talked
about.
E
This
is
then
my
contribution
to
this
whole
stretches
for
Assigned
HTTP
requests.
It
is
not
enough
to
just
sign
the
body
of
an
HTTP
message.
If
there
is
a
body,
you
must
also
sign
other
things
that
make
up
the
complete
message.
Like
the
URI.
The
method-
and
if
you
have
information
in
headers
that
are
important,
you
should
assign
them
as
well
and,
of
course,
a
timestamp
is
always
good
to
have.
This
is
a
JSON
based
system.
That
means
that
the
body
must
be
in
a
zone.
E
One
of
the
things
that
are
not
specific
with
this
is
that
the
whole
sign
request
becomes
a
serial
serializable,
a
self-contained
object.
That
means
it
can
be
embedded
in
other
objects.
Okay,
my
time
is
gone.
Okay,
okay
and
for
the
cryptography
I
have
not
invented
anything.
You
just
reused,
Yossi
stack.
E
E
Can
organization
is
needed
because
there
are
different
dialects
or,
let's
say,
variations
that
are
completely
valid,
JSON
wise,
but
would
create
problems
when
you
sign
so
the
end
points
the
one
that
signs
and
the
one
that
verifies
they
must
have
a
normalized
or
chemicalized
version
of
the
message
on
the
wire.
You
can
still
send
your
own
as
soon
as
today.
Here
is
a
real-world
example,
although
little
bit
ridiculous
with
a
payload
of
two
properties,
and
then
there
is
a
security
element
holding
the
important
things
of
the
transaction.
E
Everything
here
is
signed,
except
the
lost
element
that
contains
JSON
verbal
signature
element
a
detached,
so
it's
not
very
complex
compared
to
the
XML
star
that
the
via
stack
or
so-
and
my
mission
here
is
to
see.
If
is
there
an
interest
here
in
IDF,
to
make
this
or
something
similar
a
standard,
and
it
is
not
just
a
theory
it.
You
also
have
live
code
to
look
on
and
you
can
even
test
this
program
immediately
on
an
online
service.
If
you
hit
the
long
URL,
you
are
actually
doing
a
web
secure
request
using
this
system.
C
F
C
G
Okay,
my
name
is
Michael
I
am
a
scientist
and
adding
a
new
feature
to
HTTP
that
changes
an
old
design
assumption.
So
the
web
was
designed
for
pages
that
don't
change
very
often,
and
so
we
HTTP
a
request
response
protocol,
where
the
client
asks
the
server
for
a
page.
A
server
gives
it.
G
But
then
what
happens
if
the
page
changes
that
connections
already
gone,
and
so
the
user
just
has
to
click
reload
and,
as
you
know,
they
request-
and
this
made
a
lot
of
sense
when
most
pages
were
written
by
hand,
they
don't
change
very
often,
but
then
you
know
just
five
years
later.
Pages
are
changing
like
every
week
or
two
and
they
start
to
get
generated
from
a
database
and
now
to
hook
up
the
database
to
HTTP
people.
G
Okay,
so
there
we
go
so
now
to
connect
the
database.
You
have
a
lot
of
you
controller
framework.
This
is
the
first
thing
people
had
to
learn
outside
of
HTTP
itself
to
start
building
dynamic
sites.
Then
the
site
starts
changing
without
even
clicking
reload,
and
for
that
we
have
this
javascript
thing
you
have
to
figure
out
and
then
you
have
to
connect
it
to
the
server
with
an
expo
HP
request,
and
now
today's
sites
are
changing
all
the
time.
G
So
we
tied
together
with
more
stuff-
and
this
is
why
web
programming
really
sucks
today-
and
most
of
this
is
just
junk
outside
of
HTTP,
so
we're
gonna
make
HTTP
handle
change
so
that
you
don't
need
all
this
junk
and
the
way
you
handle
change
is
you
add
synchronization,
there's
a
bunch
of
new
technologies
out
there
help
someone's
interested
in
them,
and
this
is
the
result,
so
all
that
junk
inside
it
all
those
little
gray
boxes
turn
into
black
boxes.
Those
are
all
stuff.
That's
on
the
protocol.
Now
it's
automatically
done
for
you.
G
You
don't
need
to
learn
all
this
crap.
You
don't
need
to
write
the
code.
You
don't
need
to
wire
it
all
together.
In
my
experience,
it
cuts
down
the
code
you
write
by
about
70
percent,
seven
zero
percent
and
since
all
this
stuff
is
now
on
the
protocol,
that
means
that
every
little
piece
of
state
in
here
has
a
URL
and
it's
all
on
the
open.
G
So
you're
not
just
linking
pages
together
in
HTTP,
the
pages
have
URLs
and
you
can
link
a
page,
not
a
page,
but
now
the
state
inside
of
your
site
can
be
synchronized
with
state
inside
another
site,
and
so
we
can
share
our
state
together.
This
makes
the
whole
web
much
more
open.
You
can
and
here's
how
it
here's
how
it
works.
So
this
is
why
we
call
it
upgrade.
So
this
is
a
so
every
little
patch.
You
know
it's
communicating
and
ifs
or
patches.
Each
patch
is
a
little
color
here.
G
The
patches
connect
together
in
a
braid
and
when
you
have
to
when
you
have
two
computers
that
modify
the
same
state,
the
braid
Forks,
and
then
it
merges
back
together
down
and
flattens
into
a
single
dimension,
and
this
is
basically
this.
This
braid
is
representing
space-time
and
space-time.
Forks
becomes
relative,
and
then
it
collapses
back
down
and
by
collapsing
space-time
you're
able
to
program
as
if
multiple
state
pieces
of
state
on
multiple
servers
are
in
the
same
place.
G
It
makes
it
that
easy
to
write
your
code
and
it
solves
all
these
problems
and
in
the
web,
caches
don't
get
stale
anymore.
The
network
is
much
faster.
Every
text
area
becomes
a
collaborative
editor
like
Google
Docs.
Multiple
people
can
edit
it
at
the
same
time
you
get
an
offline
mode
for
free,
and
so
this
is
my
first
time
here,
I'm
interested
in
finding
anyone
interested
in
any
of
this
stuff
and
we're
meeting
on
Tuesday
evening
at
dinner
and
information
is
on
this
website.
H
The
issue
is
first
congestion
response
in
data
center,
so
the
problem
is
that
we
want
to
address
is
that
there
are
high
links
piercing
the
data
centers
are
making
network
transfer
complete
faster
in
fewer
oddities
short
data
post
requires
low
latency,
while
when
their
data
transfer
requires
a
throughput,
the
current
congestion
control
is
using
ECM
and
is
reactive
on
the
congestion.
The
switch
makes
is
marks,
the
ECM
beats
and
receive
and
notify
the
sender.
H
The
center
is
reduces
the
transmit
rate
and
increased
it
back
based
on
predefined
policy
and
that's
basically
based
on
some,
not
complete
information,
but
just
information.
Partial
information,
DC
TCP,
the
facility
CP,
extends
the
ECM
processing
just
estimate
the
fraction
of
bytes
that
encounter
congestion
and
scales
the
TCP
congestion
window.
Based
on
this
estimate,
quick,
an
RTP
report
back
the
number
of
ec
n
mark
Becca
and
there's
also
working
on
for
do
it
also
for
TCP
data
centers
in
telecom
public.
H
A
communication
protocol
is
in
most
cases,
Rocky
V
2,
which
is
InfiniBand
over
UDP
now,
currently,
the
the
link
layer
flow
control
based
on
I
Triple
E
8o
2.1,
is
used
to
provide
a
lossless
Network
in
the
data
center,
but
the
problem
with,
but
there's
problems
with
head-of-line
blocking
there's
an
example.
Here,
for
example,
X
can
be
the
try
to
sense.
H
H
H
A
So,
let's
pause
here
for
a
second
I,
see
folks,
would
know
you
can
come
on
up.
Folks
are
standing
three
deep
in
the
back
and
there
are
lots
empty
seats
up
in
the
front
for
you
through
a
few
rows.
So
if
you
want
to
have
a
seat,
sit
on
the
floor,
something
like
that
tired
of
standing
we'll
take
a
break
now,
you
can
come
and
have
a
seat.
E
I
Hi
everyone,
so
this
lightning
talk
presents
the
pay
to
your
eyes
scheme.
So
what
is
a
pay
to?
You
are
just
how
you
can.
Click
on
email
addresses
to
typically
open
your
email,
client.
The
goal
of
pay
to
addresses
is
to
be
able
to
click
on
this
URI
and
instead
of
saying
an
email
opening
a
banking
application
to
send
a
payment.
Sir
fax,
for
example.
Here
in
this
URL,
you
would
send
a
donation
of
25
euros
to
this
European
Sapir
accounts
de
1,
2,
3,
4
5.
I
So
formally,
these
URLs
represent
a
prefilled
pro
forma
invoice
and
when
you
reference
the
reference
them
there's
a
dispatcher
that
will
invoke
a
handlers
specific
to
the
payment
target
type.
That,
then,
will
ask
you
to
complete
the
payment.
The
syntax
of
this
is
pretty
simple.
It's
Pedro
colon,
slash
slash,
then
the
payments
target
accounts
type
which
currently,
we
have
defined
this
for
SEPA,
for
Bitcoin,
for
American
ACH
and
for
the
Indian
ubi
payment
system.
I
I
For
example,
if
I
have
accounted
various
European
banks
say,
I
want
some
central
dispatcher
my
operating
system
to
handle
this
situation
correctly
and
to
ask
me
which
banking
application
is
open
and
also
all
these
your
eyes
consider
common
concerns
for
syntax
naming
and
so
on,
and
so
doing
this
with
all
kinds
of
your
eyes,
schemes
for
different
payment
systems
would
be
very
cumbersome.
So
now
you
may
ask
there's
this
web
payment
standard.
What
about
this?
This
is
completely
orthogonal
to
pay
too.
I
So
web
payments
are
a
standard
for
things
in
the
browser
that
have
a
JavaScript
API
and
it
has
different
goals,
so
it's
basically
an
API
to
get
customers
payments
information
to
manage
it
and
to
payment
flows
inside
the
browser.
So
if
this
sounds
interesting
to
you,
please
read
the
draft
draft
all
pay
too
and
provide
feedback.
I
Also,
if
you
have
experience
with
payment
systems,
we
would
be
interested
to
hear
whether
you
can
think
of
any
more
target
types
or
maybe
generic
options,
and
you
can
implement
this
in
your
own
application
and
discuss
this
with
us
via
email
at,
for
example,
on
the
your
I
review,
mailing
list
and
yeah.
Here's
the
link
to
the
draft
and
feel
free
to
talk
to
me.
Thank
you.
J
Hi,
my
name
is
Marcus
Armand
and
it
soukous
guess
from
the
dominating
color
on
the
front
page
I'm
from
Deutsche,
Telekom
and
I.
Today,
I
want
to
talk
about
multipath
transmission
for
a
UDP
or
IP
traffic,
which
is
also
subject.
Often
three
drafts
mentioned
on
the
front
page
front
page
and
yet
just
submitted
before
this
ITF
and
yeah.
What
is
our
motivation?
What
is
our
driver?
That
is
a
basically
multi
multi
connectivity,
network,
architectures
and
exemplary
of
two
network
architecture?
J
K
So
how
we
got
here
is
basically
in
the
late
1990s
and
early
2000s.
We
were
finding
a
lot
of
link
character.
The
links
that
did
not
work
very
well
for
TCP
Aaron
had
the
tcp
over
sat
satellite
working
group
and
working
with
a
key
problem
with
very
long
RTT
interaction
with
low
loss
recovery.
We
I
I
actually
met
Aaron,
bringing
a
proposal
for
TCP
over
cellular
and
the
IDs
suggested
a
structured
approach
instead
of
TCP
over
food
links.
So
we
did
performance
implications
of
link,
characteristics,
working
group
and
we
links
that
were
slow
links.
K
K
So
what
I'd
like
to
do,
since
the
code
alone
is
for
brainstorming
I'd
like
to
brainstorm
a
bit
I
want
to
have
a
side,
meaning
in
the
code
lounge
after
the
last
session
on
Thursday,
to
talk
about
whether
it's
time
for
a
discussion
about
paths
and
performance
and
if
so,
what
problems
are
in
scope.
What's
research
was
engineering
which
parts
of
the
problem
space
didn't
already
chartered
working
groups
and
research
groups?
K
I
know
there
are
some
and
what
are
the
next
steps
between
IETF
104
and
ITF
105
I'm,
stepping
down
from
enough
positions
this
week
to
where
I
am
losing
a
dot
a
day
until
Wednesday?
So
please,
please
feel
free
to
grab
me
anytime
after
30.
You
know
Thursday
morning
or
later,
but
the
site
that
I
said
the
sign
meeting
would
be
Thursday
after
the
last
session.
Thank
you
all.
Thank.
D
This
each
Olli
I'm
going
to
talk
about
this
localized
optimization
on
path
segments
which
we
call
the
loops
okay,
the
loops.
Actually,
the
basic
function
is
try
to
provide
a
local
best
effort,
reliability.
So
here
local,
actually
is
not
a
local
area
network
rather
than
end-to-end
is
means
between
two
overlay
nodes.
You
can
look
at
the
picture
to
show
the
fundamental
use
case
of
that
to
talk
about
some
backgrounds
and
the
opportunities
in
current
practice.
D
The
paths
performance
in
has
in
practice
it's
becoming
to
an
end
with
an
increasing
deployment
of
the
encryption
and,
at
the
same
time,
in
our
notes,
including
the
virtual
nodes
which
are
created
from
the
cloud
workload,
are
becoming
more
and
more
powerful.
So
it's
a
viable
trader.
The
trade
processing
power
against
and
the
paths
are
segments
quality,
so
loops
is
trying
to
capture
those
opportunities
to
provide
them
optimization
within
the
segments
of
the
end-to-end
path.
D
So,
in
the
main
feature
where
we
are
expecting
from
loops
is
to
do
the
local
intent
worker
recovery
by
either
the
retransmission
order
FEC
for
whatever
correction.
So
there
are
basically
three
elements
of
loop
solution:
the
first
one
to
do
the
local
record
recovery,
as
indicated
in
the
picture
in
our
particular
usage
scenario,
is
between
to
overlay
nodes
in
their
two
building
blocks
in
order
to
perform
the
local
recovery.
D
The
first
one
is
to
do
the
local
measurement
Merida's
document,
the
delay
and
the
delay
variation,
and
this
kind
of
environment
can
help
to
decide
whether
a
packet
was
lost
due
to
a
local
congestion
or
some
other,
rather
than
I
mean
some
other
like
a
random
events
in
the
second
building
block
for
local
recovery
is
a
contrast
or
interaction.
That
means
how
to
interact
with
other
layer
into
n
congestion.
D
So
we
we
are
thinking
to
use
the
easy
for
congestion,
constant
loss,
so
we're
expecting
the
loops
were
a
benefit
most
to
the
non
congestion
loss
in
terms
of
the
support
for
the
longer
flows
and
for
in
terms
of
the
flow
control
flow
completion
time
in
further
for
the
shorter
flows
so
rescheduling
aside
meeting
on
Wednesday
afternoon
1:45
in
the
in
this
room,
a
salona
in
the
purpose
is
to
discuss
the
technical
problems
and
trying
to
discuss.
Are
we
ready
for
above
in
this
area?
D
M
I'm
Paul
Congdon
here
to
talk
about
what
we
can
do
to
make
high
performance
computing,
an
already
a
network
scale
on
the
order
of
cloud
networks.
Today,
famous
quote
from
a
very
recent
activity
in
the
industry.
The
future
data
centers
of
all
kind
will
be
built
like
high-performance
computers.
As
you
might
know,
Nvidia
just
made
a
big
investment
in
Mellanox.
M
So
traditionally
high-performance
networks
were
you
know,
kind
of
custom
small.
You
know
I'm
finna,
ban
type
networks
that
have
their
own
link
layer
flow
controls
to
be
lossless.
Those
protocols,
weren't
really
tolerant,
of
loss.
More
recently,
you
know:
we've
done
things
to
run
those
protocols
over
IP
protocol,
so
we
have
I
warp.
We
have
a
whole
suite
of
things
here
in
the
IETF
and
then
in
the
InfiniBand
group.
We
have
rocky
version
2,
which
runs
over
UDP
reason
why
this
is
more
important.
M
Is
that
the
info,
the
results
that
we're
getting
out
of
these
networks
are
no
longer
just
solving
weather
problems
or
cracking
DNA?
They
were
actually
doing
everyday
jobs
for
us,
so
this
is
really
important
to
get
large-scale
high-performance
computing
into
the
mainstream.
A
famous
slide
from
the
sneer
organization
that
talks
about
all
the
different
stacks
and
shows
the
evolution
of
our
DMA.
You
know
we
started
with
InfiniBand
pure
InfiniBand.
On
the
left
hand
side,
there
was
obviously
that's
a
separate
network.
When
you
want
to
run
other
applications,
we
needed
a
different
NIC.
M
Then
they
started
to
run
InfiniBand
protocols
over
Ethernet,
but
it
wasn't
a
routable
protocol.
So
that
became
a
problem.
We
had
a
very
complex
layer
to
congestion
control
that
was
in
the
end
that
was
built
as
well,
that
wasn't
very
widely
deployed.
So
now
we're
in
this
rocky
v2
phase,
where
we've
run
that
over
UDP
and
we
get
IP
routing
and
all
that,
so
we
have
had
some
scale,
but
there's
still
some
incomplete
specifications
on
congestion
and
and
we're
still
relying
on
layer,
2
flow
control,
which
has
lots
of
problems
in
the
ITF
course.
M
We
have
I
work
but
to
implement
there's
a
lot
of
tricks.
You
need
to
do
to
TCP
like
DC,
TCP
and
others.
There's
some
concerns
about
slowstar.
So
what
so?
What?
If
there
was
a
rocky
v3
or
or
an
I
war,
plus
plots
something
that
we
could
improve
upon.
So
that's
what
we're
talking
about.
So
by
hyperscale
we
mean
again,
clouds
fill
thousands
and
thousands
of
things
being
able
to.
You
know
rapidly,
deploy
things
automatic
provisioning.
M
You
know
having
a
mix
of
this
traffic
along
with
other
types
of
traffic,
but
still
maintaining
the
high
throughput
and
low
latency
that
you
need.
There's
been
a
bunch
of
papers
and
work.
I've
cited
several
of
them
here
that
talk
about.
You
know
how
we
can
scale
our
DMA
using
commodity,
Ethernet,
running
I
warp
over
over
connectionless
protocols.
That
was
talking
about
several
years
ago
how
to
tweak
ecn
for
data
centers
and
and
and
then
again,
just
sort
of
looking
at
what
changes
we
might
make
to
get
already
made
to
run
over
network.
M
So
there
is
a
bunch
of
work.
That's
happened
in
the
research
community.
There's
a
draft
okay!
There's
a
draft
here
that
describes
a
problem
statement
and
talks
about
some
things
that
we
could
do.
That's
that's
available
here
at
the
idea.
So
what?
If
you
know
we
could
do
this?
You
know
what,
if
we
didn't
have
to
be
completely
lossless,
how
could
we
get
low
loss?
You
know
what
if
we
ran
I
warp
over
UDP
like
a
quick
like
thing
instead
of
PCP,
would
that
solve
some
problems?
M
So
you
know:
can
we
do
this
hyper
scale,
so
there's
some
thoughts.
You're
gonna
have
a
side
meeting
tomorrow
morning.
10:00
a.m.
couldn't
find
the
perfect
time,
but
tomorrow,
at
10:00
a.m.
in
that
Troy,
yoga,
room
and
I
would
love
to
chat,
love
to
hear
from
operator
for
people
that
run
high-performance
computing.
That
works
as
well
to
understand
why
they
can't
scale
today
and
what
we
could
do
to
talk
about
fixing
that
okay.
Thank
you.
Thank
you.
N
Good
I'm
Jef
Raskin
I
work
on
Chrome
and
I'm
talking
about
web
packaging
this
week.
So
in
a
bunch
of
we
looking
at
a
bunch
of
use
cases
for
for
this
new
system
in
emerging
markets,
lots
of
people
aren't
awful,
aren't
online
the
whole
time,
and
so
we
want
them
to
be
able
to
share
web
applications.
The
way
they
currently
share,
Android
apps
peer-to-peer,
rather
than
download
redownload
them
over
the
Internet
there's
a
lot
of
dislike
for
Google's
amp
technology,
I,
don't
know!
N
If
raise
your
hand,
if
you
don't
like
amp
great
so
we're
trying
to
fix
that,
we
already
have
something
working
where
it
gives
you
the
right,
URLs
we're
working
on
a
way
to
to
avoid
the
need
for
the
amp
format
itself.
In
order
to
do
the
things
that
Google
search
does
with
amp
and
that's
enabled
by
web
packaging.
We
want,
when
a
whole
bunch
of
people
click
a
link
to
one
target
we
want
to
avoid
knocking
off
the
internet.
N
We
want
to
enable
neat
things
like
binary
transparency,
so
that
you
know
that
you
aren't
being
given
a
particular
compromised
version
of
an
application
specific
to
you.
We
want
to
optimize
downloads
of
things
like
J's
modules,
which
are
often
very
small
and
can
be
compressed
better
if
they're
bundled
up
together
and
so
we've
we've
got
these.
These
two
two
proposals
that
add
up
to
web
packaging.
One
is
signed
exchanges
where
we
sign
the
request,
content,
negotiation,
information
and
the
response
headers
and
body.
N
We
use
a
TLS
like
certificate,
that's
not
exactly
TLS
to
avoid
secure
some
security
risks
or
make
all
the
security
risks
opnion.
We
we
expire
it
quickly
and
we
have
a
way
to
update
the
signatures
because
of
that
short
expiration,
we
also
define
bundles,
which
are
a
random
access
and
streamable
way
of
grouping.
Http
exchanges.
N
Chrome
is
shipping
a
version
of
signed
exchanges
in
chrome
73.
This
does
not
mean
that
that
it's
fixed
we've
designed
in
ways
to
upgrade
that
and
drop
old
versions
without
breaking
backwards-compatible
without
breaking
the
user
experience
when
we
do
drop
those.
So
we
you
can
experiment
with
this
today,
but
we
want
to.
We
want
to
get
it
in
a
in
a
fully
consensus.
N
Consensus
form
get
everyone
to
agree
on
on
what
we're
doing
and
then
and
then
upgrade
chrome
support
to
whatever
everyone
agrees
on
so
I'm
talking
about
this
three
times
this
week
and
there's
a
bunch
of
ways
to
give
us
feedback
outside
of
the
meeting
we'll
be
talking
to
dispatch
to
see
where
exactly
in
the
IETF.
This
fits
we
kind
of
expect
AB
off
at
the
next
ITF
meeting.
N
That
dispatch
will
be
will
tell
us
for
sure
we're
having
a
side
meeting
on
Wednesday
to
go
over
everything
in
detail
and
possibly
work
on
a
charter
for
that
off
and
then
we'll
be
talking
presenting,
signed
exchanges
to
the
HTTP
working
group
on
Thursday
HTTP
working
group
will
probably
just
be
a
quick
overview.
It's
not
a
very
long
slot.
N
L
So
I'm
here
to
tell
you
about
some
technology
that
is
currently
available
in
Windows
Mac,
Android
iPhone,
that
you
may
not
know
about,
and
also
the
developments
that
we've
made
on
it
recently
if
you're
familiar
with
DNS.
This
is
gonna.
Look
very
strange
to
you.
Normal
DNS
allows
you
to
look
things
up
by
hostname.
You
can
find
services
like
mail.
You
can
type
in
the
name
of
your
printer
and
it'll
work.
L
L
L
So
if
you
want
to
advertise
services,
there
are
a
couple
of
ways
that
you
can
currently
do
that
you
can
use
multicast
dns
multicast
dns
is
very
automatic.
You
just
plug
your
computer
into
the
network
and
if
there's
a
server
on
the
network,
you
multicast
and
you
find
it
and
then
with
regular
dns.
Currently,
you
would
have
to
type
in
a
bunch
of
information
in
the
zone
file
and
then
your
client
can
automatically
discover
that
information
the
zone
file
using
information
that
the
network
provides
it.
L
So
we
wanted
to
automate
the
DNS
side
of
DNS
SD
so
and
multicast
DNS
is
already
automatic,
but
DNS
is
not
populating.
The
DNS
is
not
so
we
took
two
approaches
to
that.
One
of
them
is
to
provide
the
ability
to
have
an
authoritative,
DNS
server
that
is
answering
queries
using
multicast
DNS
on
the
backend,
so
essentially
multicast
DNS
is
the
database
and
the
cache
is
the
database.
The
other
option
is
to
have
an
authoritative,
DNS
server
that
is
updated
automatically
by
servers,
so
the
server
just
says
hi
I'm
here
with
the
DNS
update.
L
L
So,
basically,
you
have
a
really
dumb
thing
on
your
router
that
a
discovery
proxy
can
talk
to,
so
that
it
can
do
em
DNS
on
a
link
that
it's
not
actually
connected
to
and
then
dannis
update
is
kind
of
a
nice
way
to
solve
the
problem
of
I.
Don't
have
good
multicast
service
or
I.
Don't
want
to
be
sitting
on
the
wire
talking
all
the
time.
L
So
so
we
have
this
extension
of
DNS
update
that
basically
provides
a
very
carefully
constructed
self,
consistent
set
of
updates
that
can
be
validated
so
as
to
so
that
it's
safe
to
apply
the
update.
Even
though
you
don't
really
have
an
authentication,
you
don't
have
a
you,
have
an
established
trust
with
the
client
and
it
provides
the
ability
to
do
first-come
first-serve.
So
you
can
publish
a
key
and
further
updates
to
that
name.
That
will
be
validated
using
the
key.
L
We
have
a
lease
so
that
we
can
garbage
collect,
stale
registrations
and
the
the
name
sticks
around.
So
if
your
service
goes
offline,
the
name
stays
for
a
while
and
does
get
reclaimed
by
some
other
server.
One
of
the
nice
features
of
multicast
DNS
is
that
it's
timely.
So
if
you
look
in
your
UI
and
come
see
us
tomorrow,
that
demo
happier.
L
Okay,
so
some
of
you
may
have
seen
the
dog
food
wiki,
the
dog
food,
both
proposal
and
the
wiki
for
this
ITF,
who
was
a
little
half-baked
I,
didn't
have
time
to
really
do
it.
So
I'm
gonna
talk
about
that.
A
little
bit.
I,
don't
want
to
stress
you
out,
but
I
have
a
real
concern
about
what
we've
been
doing
in
the
ITF
for
the
last
ten
years
and
seem
to
be
continuing
to
do
for
the
next
10
years.
L
L
If
you
look
at
what
happened,
something
like
British
Rail,
for
example,
there
was
a
constituency
they
wanted
it,
and
then
the
constituency
stopped
arguing
for
it
and
it
slowly
decayed
and
turned
into
this
thing
that
got
privatized
and
really
isn't
what
it
used
to
be
back
in
the
good
old
days.
So
I'm
concerned
that
something
similar
might
happen
with
the
IETF.
If
we
go
away,
oh
I
should
be
changing,
slides.
L
Yeah,
so
if
we
go
away,
we're
gonna
have
trouble
right.
This
could
wind
up
getting
sold
off
to
the
highest
bidder
and
so
who's
weak.
The
people
in
this
room
are
ageing.
The
number
of
the
age
of
the
average
age
of
an
IETF
participant
is
growing
over
time.
The
group
is
getting
smaller
and
we're
not
attracting
new
young
participants
who
would
be
interested
into
it.
L
L
So
what
to
do?
Thank
you,
I!
Didn't
there
isn't
a
train
across
the
Atlantic,
sad
to
say:
I
would
love
it
if
there
were
or
a
balloon.
So
ITF
leadership
can't
solve
this
problem
right.
They
tried.
We
had
the
many
couches
mailing
list,
pretty
much
everybody
who
showed
up
was
in
the
ITF
leadership.
They
did
a
nice
job,
but
it
was
kind
of
a
you
know.
We
all
of
these
people
in
this
room.
We
need
to
decide
that
we
want
to
do
this.
We
need
to
be
pushing
the
ITF
leadership.
L
We
need
to
be
telling
them
that
it's
what
we
want.
Otherwise,
it's
not
going
to
happen.
What
I
would
really
like
us
to
do
is
get
to
the
point
where
we
actually
have
an
IETF
in
the
relatively
near
future
that
we
are
only
going
to
do
online.
It's
not
them.
There
isn't
going
to
be
a
venue,
it's
just
going
to
be
online
IETF
112,
maybe
I,
don't
know
that
soon,
but
I
think
we
could
do
that.
L
O
I'm
Wes
Parker
from
USCIS
I
and
I
love.
Ietf
meetings
I've
been
coming
here
for
a
long
time.
They
Rock
I
love
shaking
hands,
I
love,
giving
hugs
to
my
friends,
but
the
reality
is:
is
it's
hard?
It's
hard
for
the
world's
climate
really
hard
for
the
world's
climate.
We're
learning
that
quickly.
It's
hard
for
your
body's
health
there's
been
lots
of
proof
on
that.
It's
hard
for
various
budgets.
I
have
to
justify
my
budget
every
single
time,
I
come
and
it's
hard
for
family
life.
O
O
O
So
current
solution
models
have
have
tried
to
emulate
what
we
have
today
right.
You
have
this
week-long
meetings
you
set
hours
around
particular
meetings
and
and
the
reality
is
you
have
to
wake
up
at
4
o'clock
of
the
meeting
to
have
a
meeting
with
the
group
that
you
like
that
doesn't
work
it's
time
to
think
way
outside
the
box.
So
we
need
a
new
solution.
We
need
entirely
new
directions,
we
need,
you,
know
new
people
as
Ted
was
talking
about
and
and
new
groups.
O
We
need
new
requirements
for
what
it
actually
means
to
meet,
because
the
reality
is
is
when
you
don't
meet
in
a
physical
place.
The
the
whole
logistics
around
a
meeting
are
probably
going
to
be
different.
You
need
new
frame
of
mind
and
what
this
really
boils
down
to
is
you
need
a
revolution
and
how
we
do
business,
not
just
a
simple
evolution
and
it's
gonna
be
commitment
right.
Is
this
not
just
head
and
I
doing
this?
This
is.
O
We've
got
to
get
everybody
on
board
to
actually
make
this
happen,
whether
it's
at
IETF
112
or
some
other
time
so
I'm
gonna
give
you
one
example
of
ideas
that
I've
been
thinking
of
rather
than
the
timetable
you
saw
before
I'd
like
to
hold
every
working
group
last
for
a
week
right
every
working
group
last
week
we
still
dedicate
an
entire
block
of
time.
We,
you
know,
there's
a
whole
bunch
of
other
parameters
to
go
along
with
it.
Maybe
how
do
we
reproduce
face-to-face
meetings
where
all
of
that
good
stuff
still
happens?
O
O
The
number
of
posts
per
day
so
that
you
know
entire
long
week,
isn't
taken
up
by
a
couple
of
people,
maybe
you're
only
limited
by
two
minutes
per
comment
period.
Maybe
there's
a
webpage
that
allows
you
to.
You
know,
follow
these
streams.
I,
don't
I,
don't
have
all
the
answers.
This
is
yet
one
example
too
one
problem,
I,
don't
know
that
a
week
is
the
right
time
frame
anymore
right.
How
do
we
think
outside
the
box?
So
how
do
we
go
forward
with
this?
We
need
out-of-the-box
thinkers
again
revolution,
not
evolution.
O
We
need
lots
of
people.
That
means
everybody
in
this
room.
Every
single
one
of
you
have
got
to
help
with
this.
We
need
willing
working
groups
that
are
willing
to
be
the
guinea
pigs.
If
you're
a
working
group
chair-
and
you
want
to
be
a
guinea
pig
for
an
interim-
come
see
us,
we
need
a
place
to
hold
these
discussions
about
how
to
even
do
this.
Many
couches
was
one
thing
that
Ted
mentioned
and
again
this
is
actually
has
a
Tuesday.
8:30
p.m.
O
doesn't
mean
bar,
let's
meet
there,
and
let's
talk
about
this
because
this
is
not
going
to
happen
really
quickly
feel
free
to
mail
me
her
to
great
eyesight
edu.
If
you
want
to
make
sure
you
get
on
a
mailing
list
at
some
point,
I
wish
I
had
a
subscribe
link,
so
I
didn't
have
to
handle
all
these
by
hand,
but
I
don't
so
please.
You
know
think
about
this,
because
this
is.
O
This
is
not
something
you're
gonna
solve
overnight
and
if
I
keep
talking
long
enough,
I'll
get
the
really
cool
buzz,
which
is
really
what
I'm
waiting
for
at
this
point,
because
it's
a
cool
car
sound
but
but
think
about
this,
because
this
is
not
something
you're
gonna
come
up
with
an
answer
tonight
right.
This
problem
of
meeting
through
an
entire
week
took
me
a
month
to
even
think
about.
It
was
like
that's.
What's
broken
right,
the
the
physical
world
does
not
match
the
virtual
world.
How
are
we
going
to
get
there?
O
A
P
So
I'm
sitting
here,
relax
4
minutes
between
you
and
dinner.
I'm
gonna
take
the
whole
time
so
I'm
here
to
invite
you
to
a
new
type
of
sessions.
We're
gonna
have
this
week
of
the
ATF
with
input
into
the
slides,
but
this
going
to
be
Wednesday
at
3
o'clock
in
the
Grand
Ballroom.
That's
where
the
hackathon
happened
over
the
weekend.
So
what
we're
doing
is
we
have
this
new
type
of
meeting
just
at
the
calling,
a
deep
dive
knowledge,
a
deep
dive,
the
topic
for
this
time?
P
The
agenda
is
very
simple:
it's
basically,
we
got
three
people
from
three
different
router
vendors
to
come
to
talk
about
router
architecture,
how
our
routers
build,
and
especially
what
are
the
implications
of
that
architecture
in
IETF
protocols
for
in
proto's
that
we
can
develop.
We
have
had
in
the
ASG
several
sort
of
late
surprises
where
we
see
things
that
are
being
done.
That
can't
really
be
happening
on
routers
fragmentation,
for
example,
or
accession
headers
for
ipv6.
Your
different
things
like
that.
We
all
have
in
our
minds
an
idea
of
how
a
router
works.
P
I'm
sure
they're,
all
of
you
do
I
know
I
do
but
not
all
the
routers
that
we
think
we
know
work
the
same
way
or
not
all
of
them
work
the
same
way
in
reality
or
for
the
different
applications
that
we
want
to
use
them.
So
that's
one
of
the
reasons
we
want
to
have
this.
This
is
what
we
want
to
have
with.
This
is
a
deep
dive
into
how
the
routers
are
built
correctly.
P
Before
this,
we
also
had
already
a
talk
by
John
Scudder
from
Juniper
at
the
AAPG,
where
he
talked
about
some
of
these
topics.
This
was
very
well
received.
The
apg
is
a
group
of
operators,
so
we
think
that
it
is
time
for
everyone
in
the
IETF
to
also
be
involved
in
this.
We
are
going
to
be
hosting
this,
as
I
said
on
Wednesday
at
3
p.m.
in
the
Grand
Ballroom.
This
is
the
first
time
we
do
one
of
these
deep
dives.
P
A
Great,
so
that's
the
end
of
our
agenda.
I
would
like
to
invite
you
all
to
come
to
the
Pecha
Kucha
on
Tuesday
night.
This
is
a
non
ITF,
informal
event
which
will
have
I
think
six
or,
depending
on
how
cynical
Brian
Trammell
is
by
Tuesday,
possibly
seven
presentations
of
a
generally
satirical
nature,
and
if
you
have
ideas
on
how
to
make
this
event
better
feel
free
to
catch
me
in
the
hallway
or
send
me
some
email.
A
And
if
you
have
ideas
and
you'd
like
to
get
on
the
agenda
for
next
time,
you
can
just
send
me
a
note
or
send
a
note
to
hot
RFC
at
ITF
tour.
So
sorry,
the
Pecha
Kucha
starts
at
9:30.
I,
don't
have
a
room,
yet
there
will
be
free
beer,
but
it
won't
be
unlimited.
So
there
you
go.
So
thank
you
for
coming,
go
up
and
have
dinner
and
we'll
see
you.