►
From YouTube: IETF99-NFVRG-20170717-1740
Description
NFVRG meeting session at IETF99
2017/07/17 1740
https://datatracker.ietf.org/meeting/99/proceedings/
A
B
C
A
A
D
A
Thanks,
okay,
welcome
this.
As
you
can
see
there,
this
is
the
NE
veerji
meeting.
This
is
going
to
be
according
to
a
little
bit
shorter
than
those
of
you
who
are
part
of
the
usual
suspects
will
remember.
This
is
going
to
be
a
just
one
hour
meeting
with
that.
One
presentation
on
technical
matters,
the
discussion
of
one
documents
and
some
open
mic
period
for
on
the
on
what
to
do
in
the
in
the
coming
times
of
the
air,
with
their
working
with
their
research
group.
A
So,
just
two
introductory
these
as
to
remind
you
about
the
IRT
FIP,
our
policy,
this
is
essentially
the
same
as
at
the
ITF.
The
other
one
said
this
morning.
You
have
to
remind
that
everything
you
say
here
is
a
contribution
and
is
under
the
IPR
policies.
Be
aware
of
that,
this
is
the
usual
call
for
collaboration
in
reviewing
documents,
and
these
are
some
data
about
where
we
are.
What
are
we
doing
when
you
come
find
some
the
mailing
list
slides
where
are
available
for
you
on
the
web
and
the
wiki?
A
D
We
have
a
tradition
in
BMW
G
that
anybody
sitting
in
the
back
row
is
an
automatic
note,
taker
or
Jabbar
strive.
So
if
you
sit
in
the
back
row,
I
guess
you're
signing
up
yeah
gentlemen.
Looking
at
his
phone
over
there
in
the
left
hand
corner
no,
no
hands,
no
beer.
Can
we
bribe
you
with
a
beer,
no
takers?
D
A
A
We
take
a
dime,
don't
complain
about
the
agenda,
the
minutes
not
reflecting
exactly
with
you,
okay.
So
let's
move
on
this
is
the
usual
request
that,
whatever
the
research
event,
you
are
aware
of
just
use
the
list,
but
usually
only
for
research
events
limited
not
to
just
to
make
announcements
or
whatever
of
other
matters,
and
if
you
have
any
doubt,
ask
us
if
you
news
about
well,
do
you
know,
and
then
in
the
list
we
were
discussing,
we
discuss
about
this
in
Chicago
and
we
have
been
discussing.
A
The
angle
is
we're
about
to
sing
our
first
document
of
year
to
energy
for
publication,
and
we
are
now
that
we
have
under
these
milestone,
one
of
the
years
that
we
were
discussing
with
the
other
she
chair
was
precisely
to
focus
a
little
bit.
The
the
activity
in
the
group
focusing
in
a
few
matters
that
suitable
for
produced
research
results
in
the
midterm
to
try
to
organize
shorter
meetings,
but
more
frequent,
not
only
not
limited.
A
Only
to
the
to
the
gatherings
here
in
the
ATF
meetings
trying
to
assemble
the
the
group
or
part
of
the
group
wins
in
the
in
the
opportunity
of
a
scientific
conference
etc,
and
we
will
talk
about
this
later
on
the
agenda
and
in
a
few
remarks.
Just
for
you
to
be
aware
of
this,
one
is
that
there
was
before
this
session.
A
There
was
a
ball
phone
network
slicing,
the
in
which
well
many
one
of
the
things
that
I
would
remark
is
that
someone
other
than
Farrell
highlighted
that
the
ITF
is
not
a
research
group
and
that
many
of
the
of
the
problems
that
were
discussed
during
this
afternoon
or
network
slicing,
probably
would
be,
will
have
to
do
with
the
with
our
concerns
here
in
the
NF.
U
G
so
just
consider
this
as
a
potential
input
to
all
the
later
later.
A
This
evening
there
was
a
meet-up
on
distributed
infrastructure,
which
is
about
enabled
say,
I
will
say
it
blockchains,
basically
blockchain
Sun
and
related
distributed
mechanisms.
So
if
you're
interested
I
can
bring
you
to
the
to
the
meeting
place,
there
is
a
there
is
another
before
or
I'm,
not
sure.
A
If
it
is
a
former
Bourne
Identity
enable
networking
on
another
another
that
proposed
research
path,
aware
networking,
where
precisely
and
if
we
can
have
a
role
to
play
during
the
bits
and
bytes
there
will
be
a
demo
of
the
effect
GX
project
that,
if
you
remember
from
the
previous
meeting
in
Berlin,
there
was
this
robots
that
were
Danes
dancing
around
the
the
lines
and
there
were
cloud
base
and
running
a
service
chamber
based
on
menifee
images
and
improve
of
that.
An
improvement
of
that.
A
There
is
a
presentation
on
this
Friday
on
mostar,
which
is
something
that
I
believe
could
be
used
for
for
establishing
trust
in
a
new
environment,
and
there
will
be
a
bob
of
on
edge
computing.
No
I,
don't
know
exactly
when,
where
but
again,
if
you're
interested
just
tell
me-
and
whatever
you
want
to
remark,
it
would
be
I
will
be
glad
to
include
it
here.
Well,
moving
to
the
agenda
first,
this
is
we
have
to.
We
are
talking
about
these
about
we
are
in
the
welcome.
A
Then
we
will
have
a
Carlos
presenting
the
draw
some
challenges
to
us,
IRC
discussion
for
publication,
and
then
we
have
a
talk
from
Buffalo
and
color
Iran
flexibility
as
another
measure
for
in
the
case
of
in
Fe
and
finally,
we'll
be
talking
about
be
a
little
bit
about
the
future
and
how
we
will
achieve
this
refocusing
of
the
NFA
og.
So
before
we
start
with
Carlos.
Just
let
me
remind
you
that
I
gave
the
the
blue
sheets
in
both
wings.
If
you
can
pass
it
to
the
center
to
behavior.
A
E
So
this
is
a
brief
recap
of
the
history
of
the
draft
from
individual
submission
back
in
in
Prak,
actually
in
2008,
EF,
93rd
and
then
was
adopted
in
Yokohama
well
after
Yokohama
and
has
been
reviewed
several
times
and
gone
through
several
revisions.
Here
we
are
in
version
zero.
Six
briefly,
couples
on
the
main
goals
of
the
documents.
E
Owning
goals
are
to
identify
some
research
challenges
for
or
in
the
area
of
network
visualization,
based
on
a
gap
analysis
that
we
perform
and
then
that
document
may
be
used
as
a
basis
for
new
activities
that
are
still
not
cover
in
ATF
I
RTF
covering
those
those
challenges
and
in
this
version,
as
we
will
briefly
show
later,
we
also
try
to
match
those
research
challenges.
Those
gaps
with
the
focus
areas
that
we
are
discussing
in
in
the
group.
E
So
this
is
the
structure
of
the
draft,
as
of
today
in
in
red,
I
am
trying
to
identify
the
ended,
the
changes,
the
new
sections
that
has
been
introduced.
This
has
been
based
on
the
reviews
that
we
got.
We
got
nine
reviews
since
the
previous
version,
so
thanks
for
all
the
reviewers
I
am
listing
them
in
a
in
a
synchronous
light.
So
we
introduced
some
additional
background
on
itu-t
SDM
function,
architectural
fashion
architecture.
We
need
some
updates
on
the
ABC
due
to
the
the
changes,
then
that
have
been
happy
taking
place
on
on
Emma
C.
E
So
again,
thanks
a
lot
for
all
the
comments
and
I've.
Even
I'm
not
listing
all
the
changes,
because
there
are
quite
a
lot,
we
try
to
address
all
the
comments.
Basically,
in
addition
to
this
changes
in
the
in
the
structure,
we
also
tried
to
put
more
Academy
references.
This
was
something
that
was
requested
by
more
than
one
reviewer,
so
we
try
to
do
that.
E
Well,
basically,
this
is
also
an
update
on
the
on
the
gaps
and
the
potential
ATF
efforts
after
reviewing
the
new
content
with
the
new
section.
So
this
is
the
table
there.
There
are
more
more
details,
of
course,
in
the
draft
and
the
update
that
we
did
regarding
the
the
focus
areas,
how
to
match
the
open
research
topics
that
we
identify,
based
on
the
gap
analysis
into
these
four
areas
or
four
focus
points
that
that
was
proposed
in
the
on
the
mailing
list,
and
this
is
basically
it.
We
got
very
positive
reviews
and
very
comprehensive
review.
E
A
C
E
E
C
A
Well,
we
are
doing
well
in
terms
of
time
and
what
our
next
speaker
is
a
ball
and
he's
from
the
Technical
University
of
Munich
and
well
has
been
working
for
a
while
and
something
that
I
think
is
when
we
have
been
discussing
about
the
different
ways
in
which
we
should
work.
One
is
our
measurements.
The
pond
is
that
measurement
is
something
that
is
being
well
addressed
by
working
groups
in
the
in
the
80s,
but
precisely
this
is
in
a
thank.
You,
sir,
is
an
interesting
opportunity
of
measuring
something
difference,
but
is
equally
important.
Probably
please.
A
F
Thanks
Diego
for
inviting
me
so
what
I'm
presenting
here
that
I'm
going
to
entertain
you
a
little
bit
this
late
session
here,
is
about
measuring
something
different
that
because,
in
a
larger
project
that
I'm
started
and
have
not
finished,
this
is
a
five-year
project
and
just
in
the
second
year
now
what
we
are
doing
here
we
are
thinking
about.
Measuring
flexibility,
which
comes
at
at
most
importance.
Now,
in
all
respects,
are
in
network
function,
virtualization,
Sdn,
virtualization,
and
all
these
things,
let
me
take
you,
try
to
take
you
on
a
journey.
F
What
we're
doing
I
have
you
have
more
examples,
then?
Basically,
I
have
then
detailed
equations,
or
something
like
that.
So,
let's
start
the
journey
at
where
we
are
and
what
we
face,
and
that's
also
why
we're
here
for
network
through
virtualization
we
face,
and
we
see
the
network
coming
with
new
requirements.
Vertical
industries
are
dynamically:
changing
user,
behavior,
global
digitalization
and
so
on
and
so
forth.
F
What
is
less
addressed
nowadays
is
the
flexibility
to
react
to
these
dynamically
changing
requirements
and,
in
the
end,
less
the
adaptation
of
the
network
is
of
a
concern
today,
but
actually,
what
we
know
is
that
those
species
are
they
prevail
in
the
environment
that
are
most
flexible
to
adapt
to
changes.
We
know
this
from
us,
sir,
so
we
basically
can
handle
anything.
If
this
is
good
or
not,
we
can
discuss
at
another
point,
but
we
applied
flexible
to
do
something
new.
F
This
is
also
what
we
do
in
research,
and
this
is
something
I
would
like
to
bring
you
now
how
this
can
be
reflected
in
networking.
I
will
start
presenting
after
some
more
motivation.
Our
idea
of
how
flexibility
of
a
system
can
be
measured
that
you
is
really
that
you
have
a
design
choice,
a
you,
have
a
design,
choice,
B
and
now
you're,
not
asking
for
performance
only
but
you're
asking
for
which
design
choice
is
more
flexible
than
the
other
one,
because
I
might
in
future
face
requirements.
F
New
requests,
I,
don't
know
yet
now,
and
how
can
my
system
support
them?
I
have
a
bunch
of
use
cases
with
me.
Pending
on
how
much
time
allows-
and
this
is
researcher
I'm-
probably
raising
more
questions
than
I-
can
give
answers
at
this
stage
of
the
project.
F
F
Adaptation
is
slow
to
new
requirements,
any
particular
reaction
to
dynamic
changes,
as
we
can
see
here.
The
point
is
not
so
well
is
very
difficult
to
do,
but
we
are
faced
with
more
and
more
changes,
more
and
more
reasons
for
adaptation
and
now
not
arguing
for
a
future
Internet
I'm
just
arguing
and
showing
you
well.
We
have
tools
now
available.
Now
these
tools
are
network
function,
virtualization,
they
are
software.
F
Fine
networking
and
those
tools
actually
addresses
are
not
presenting
now
new
concepts
for
flexibility,
but
we
have
these
concepts
they
promise
to
create
and
adapt
networks
and
functions
in
software.
We
call
this
also
network.
Software
is
a
shoe.
You
all
know
this
also.
You
have
in
this
example
here
you
have
some
traffic
and
we
have
you,
this
green
flows
and
this
yellow
flow
is
coming
and
then
we
shift,
while
some
control,
or
by
or
some
Orchestrator
network,
a
slice
or
the
network
slow
to
other
places
and
also
and
shift
network
functions
there.
F
So
we
can
react
to
these
changes,
so
everything
fine,
not
what
I'm
doing
here.
Well
now
the
questioning
starts
are
all
problem
solved.
Are
we
already
fully
flexible
with
these
concepts?
The
question
should
rather
be,
and
this
is
what
we
should
ask
in
particular
as
academics.
How
far
can
we
go
with
these
flexibility?
What
is
the
right
network
design
that
gives
us
more
flexibility
than
another
network
design,
and
this
is
what
we
are
targeting
in
our
project.
F
We
are
striving
for
a
more
fundamental
understanding
of
how
to
provide
flexibility,
and
we
are
even
now
I
will
show
you
a
first
glance
on
a
quantitative
measure,
for
flexibility
for
and
against
certain
design
choices.
So
what
is
flexibility
in
networks?
Well,
it's
like
in
everyday
life
for
networks.
We
say
it's
the
ability
to
support
new
requests,
so
requests
that
we
are
not
given
in
there
a
set
of
requirements
when
the
system
design
has
been
built
so
to
support
new
requests
to
change
design
requirements
in
a
timely
manner.
F
Now
design
requirements
might
be
a
Janus
might
be
traffic
pattern
that
are
coming
you've
seen
this
in
the
previous
slide,
that
might
be
change
of
latency
requirements
are
low,
delay,
one
millisecond
and
so
on
and
in
a
timely
matter,
because
time
Matassa,
if
I,
allow
your
infinite
time,
you
might
all
everybody
might
be
able
to
do
something
different,
but
we
would
like
to
have
it
quickly
done
and
how
we
can
do.
This
is
because
the
system
design
allows
us
to
adapt
network
resources,
topology
capacity
and
so
on
if
needed.
F
Now,
our
very
first
definition
of
our
flexibility
measure
could
look
like
the
following.
Taking
into
account
of
what
I
said
now,
we
measure
the
flexibility,
Phi
T
T
stands
now
for
the
time
for
the
time
limiter
offer
system
s
as
the
supported
new
requests
within
a
given
time
over
the
total
number
of
given
requests
given
new
requests
to
the
system.
So
we
challenge
the
system
with
new
requests
and
then
we
simply
count
how
many
requests
the
system
can
fulfill
in
a
given
time.
F
So
that
allows
us
to
compare
a
system
a
and
a
system
B,
of
course,
under
the
same
given
requests.
Otherwise
we
are
not
comparable.
So
the
fraction
of
the
number
of
new
requests
in
a
certain
time
interval
time
matters,
but,
of
course,
what
we
know
now
when
we
move
the
time
to
infinite,
then
we
also
get
some
glance
on
the
system.
Then
it's
basically
how
many
requests
can
be
support
over
all
even
requests.
So
this
is
basically
then
approaching
to
it
when
we
do
time
to
infinite.
Ok,
so
think
about
this
one.
F
This
is
now
a
proposal
for
how
to
compare
different
designs
for
flexibility.
Let
me
show
you
now
is
a
very
simple,
for
example,
of
course,
with
some
Network
functions
of
how
this
could
work.
Let's
look
at
a
very,
very
simple
network.
We
have
four
nodes:
the
node
capacities,
NCR
1,
1,
2
3
and
we
have
a
controller-
could
be
an
SDN
controller
there.
You
see
this
is
residing
at
the
left
side
and
this
as
the
end
controller,
is
now
requiring
a
larger
capacity.
F
So
the
new
controller
capacity,
because
it
has
to
do
more,
is
now
suddenly
requiring
an
old
capacity
of
2.
Now
the
question
is:
is
the
system
flexible
enough
to
support
this
new
request?
For
example,
we
could
migrate
the
controller
to
a
node
with
a
higher
capacity,
and
now
there
is
some
time
constraint.
This
should
be
quick
and
we
only
allow
now
for
simplicity,
I,
don't
have
seconds
yeah
I
have
now
time
in
hops.
I
say
we
only
can
allow
to
migrate.
One
hop
over
one
link.
Won't
this
be
doable.
F
Well,
is
this
migration
time
it
is
doable?
We
have
one
node
one
hop
away
which
has
an
old
capacity
of
2
now,
so
we
are
able
to
say
yes,
this
request
can
be
fulfilled
and
that
we
can
counter
our
Phi
is
one
new
request,
supported
over
all
even
new
requests.
Of
course,
this
is
a
bit
boring
if
I
only
have
one
request
and
this
one
is
fulfilled,
so
we
can
extend
this
and
say
we
have
more
requirements,
more
new
requests.
F
Now,
at
the
left
side,
we
have
this
one
we
have
just
seen
in
the
middle
we
say
now:
a
new
controller
capacity
required
is
3.
There
is
a
node
that
has
a
controller
capacity
of
3.
However,
this
node
is
2
hops
away,
so
time
constraint,
not
fulfilled,
not
flexible,
given
our
limited
time
and
on
the
right,
we
have
a
controller
capacity
demand
of
4
new
requests
can
also
not
be
fulfilled
because
there's
any
way
no
node
in
this
system.
F
If
you
now
relax
the
time-
and
we
say
we
don't
care
about
these
time
constraints-
we
put
the
time
constraint
to
infinite.
Then
you
see
it
in
the
bottom
row
and
if
you
are
not
constrained
in
the
number
of
ups,
then
you
can
see
that
also
the
middle
one
can
be
supported
and
in
the
end
we
have
two
requests
out
of
three
given
new
requests.
So
we
have
2/3
or
66%
of
flexibility
of
the
system.
If
time
doesn't
matter.
F
Now
you've
seen
some
glands
on
this
flexibility
as
a
measure.
Now
you
might
ask
yourself:
how
would
this
work
now?
Can
I
compare
flexibility
with
flexibility?
Isn't
this
apples
and
oranges?
Yes,
you
have
to
be
careful
now.
Flexibility
as
a
measure
as
kind
of
a
quality
of
flexibility,
is
similar
with
quality
of
service.
If
you
cannot
say,
I
have
more
data
rate,
you
have
lower
delay.
That's
why
your
quality
of
service
is
higher
than
my
quality
of
service
and
similarly
for
flexibility,
we
also
have
to
decide
case-by-case.
F
We
propose
flexibility
aspects
and
we
say
in
order
to
be
able
to
quantitatively
compare
different
systems
designs.
We
have
to
say
we
are
comparing
these
systems
with
respect
to
their
adaptation.
Flexibility
on
flow
steering,
like
with
software
affine
networking
all
with
their
flexibility
on
funk
replacement.
How
easy
is
the
system?
F
F
So
these
are
gateway
functions,
they
have
data
plane,
they
have
control
plane.
Now
what
we
can
do
is
to
be
more
flexible.
We
can
take
those
functions
nowadays,
they
are
sold
in
racks
are
in
Hardware
special
hardware.
Is
these
functions
on
and
using
the
concept
of
network
function?
Virtualization
yeah?
We
have
a
data
center,
you
just
move
the
whole
gateway
function
into
the
cloud.
Now
we
need
some
network
element
there.
We
have
a
virtualized
gateway.
Then,
of
course
we
have
to
take
care
of
core.
F
Could
be
good
solution?
There
is
another
solution
if
you
are
concerned
with
the
network
traffic
and
the
delay
that
we
can
try,
this
separation
of
concerns,
separation
of
data
plane,
a
new
plane,
more
the
Sdn,
like
concept,
move
the
data
plane
back
on
the
network
element
now
and
then
we
have
a
decomposed
gateway,
he
might
have.
We
have
less
data
plane
latency,
because
we
don't
move
the
data
plane
in
the
cloud.
We
are
just
concerned
with
control
plane.
The
question
is:
what
is
the
delay
here,
for
example,
or
in
the
effort
we
don't
know
yet.
F
This
is
something
we
have
to
find
out
if
the
thing
grows
more
complex,
because
we
have
more
of
these
gateways
that
we
have
to
decide
to
be
all
moved
them
into
the
cloud
to
be
all
move
the
data
plane
back.
Do
we
have
a
combination
of
them
this
way
or
the
other
way?
And
so
on
and
so
forth?
I
don't
want
to
go
into
the
detail.
It's
a
complex
function,
decomposition
and
function,
chaining
and
function.
Placement
problem
we've
run
a
lot
of
studies.
This
was
a
paper
published
two
years
ago
at
a
CMC
calm.
F
In
a
workshop,
we
coined
it.
We
coined
there
the
function,
placement
problem
where
we
ran
a
huge
simulation
on
several
topologies
to
find
out
which
of
these
concepts
is
better
with
respect
to
performance
and
with
respect
to
latency
and
now
I.
Take
this
example
in
order
to
run
another
measurement
and
try
to
find
out
which
of
these
is
more
flexible.
F
So
again,
remember
we
have
this
three
possibilities.
We
can
have
this
Sdn
design,
keep
the
data
plane
plane
down,
only
move
the
control
plane
into
the
cloud
we
have
to
end
of
redesign.
We
basically
move
all
the
gateway
into
the
cloud,
and
then
we
have
to
put
all
the
data
plane
traffic
into
the
cloud
or
we
can
have
a
mixed
design.
Think
about
it.
What
do
you
think
would
be
your
favorite?
What
concept
would
be
more
flexible
than
the
other
one?
F
In
order
to
find
out,
we
have
now
challenged
the
system
with
new
requests
now,
whereas
the
systems
might
have
been
built
for
a
certain
latency
in
mind,
we
are
now
challenging
the
system
with
different
latency
requests.
The
system
has
not
been
planned
for
and
we
varied
the
data
plane
latency
and
the
control
plane
latency
each
with
ten
different
delay
limit
limits
10
by
10
is
100,
so
we
have
now
a
baseline
of
100
new
requests.
F
Now
you
can
see
easily
I'm
hinting
at
some
percentage
when
counting
the
number
of
requests
that
can
be
fulfilled
in
the
end
to
compare
the
different
designs
in
calculating
5
of
the
placement
of
these
functions
over
these
different
designs,
as
basically
some
of
the
feasible
solutions
over
the
sum
of
all
solutions
in
total
100
and
now
this
is
the
outcome.
I
don't
go
into.
Details
is
all
can
be
read
in
the
paper.
We
had
this
at
some
at
infocomm,
some
info
Kahn
workshop
published
last
year.
What
do
you
see
here?
F
We
have
now
different
boxes
here.
Different
plots,
different
rates
on
whether
we're
more
interest
in
the
data
plane,
latency
or
reservoir
more
interested
in
the
control
plane.
Latency
two
things
we
can
see.
We
can
first
see
that
a
mixed
in
the
left
side,
part
of
each
of
these
graphs
that
are
mixed
as
the
N
and
every
design
is
more
flexible.
F
Not
because
we
have
here
a
higher
flexibility
up
to
60
70
%
of
the
cases.
This
is
also
what
probably
you
thought
is
right,
because
mixed
always
sounds
like
more
flexible.
It
is
what
you
get.
However,
there
is
something
written
I'm,
not
sure
if
you
can
read
it
underneath
logically
centralized
here,
two
data
centers
or
they
didn't
tell
you
before,
is
we
also
tried.
F
Now
is
different
system
parameters
and
on
the
left
side
we
only
allowed
two
data
centers
on
which
the
fan
actions
can
be
placed
and
there,
of
course,
the
mix
deployment
can
take
with
make
most
for
the
flexibility.
However,
we
also
found
there
is
a
cap
if
you
throw
eight
data
centers
now
at
the
system
and
allow
eight
potential
places
in
our
us
Evelynn
network
that
we
used,
then
there
is
not
much
of
a
difference
with
respect
to
flexibility
with
these
three
designs
and
you
can
use
the
more
Sdn.
The
more
is
an
every
design.
F
F
Now
you
might
ask:
where
is
the
time?
He
tells
us
now
something
about
flexibility
in
these
days,
even
a
certain
time
threshold?
This
has
to
be
done
well.
This
was
another
first
example,
and
on
this
years
Infocom
we
had
another
publication
where
we
were
striving
now
for
a
dynamic
placement
of
functions.
F
What
is
this
difficult,
more
complex
examples
from
Jiji
PP,
but
we
just
took
as
before
and
Sdn
controller
as
a
network
function
in
this
second
example.
So
you
all
know
the
controller
placement
problem.
You
are
we
place
one
or
n
as
the
end
controllers.
We
now
say
this
is
our
dynamic
controller
placement
problem,
replace
as
the
end
controllers
for
time,
varying
traffic
input
and,
of
course,
in
order
to
optimize
the
delays
from
the
switches
to
the
controller.
F
F
Randomly
traffic
profiles
requests
for
flows
into
the
system
that
required
controllers
in
order
to
be
most
efficient
to
migrate.
So
we
had
be
compared
now,
a
system
with
one
controller
to
three
or
four
controllers
in
order
to
find
out
which
one
is
more
flexible
and
you
can
already
think
about
which
one
is
your
favorite
yeah?
Do
you
think
the
one
with
one
controller
or
one?
F
These
four
controllers
is
more
flexible,
flips,
flexible
with
respect
to
what
well,
we
have
an
algorithm
running
after
each
flow
that
finds
the
optimal
controller
placement
and
flow
to
controller
assignment
minimizing
the
average
flow
set-up.
Time
so
meant
this
is
our
performance
indicator.
You
would
like
to
minimize
the
overall
delay
in
the
flow
set
up
in
the
system,
and
now
we
say
we
check
how
many
controllers
can
be
migrated
in
a
given
time.
T.
This
is
our
limitation,
and
this
is
our
flexibility.
As
you
don't
know,
what
will
be
a
good
time
T?
F
What
will
be
the
limit
for
allowing
a
controller
or
several
controllers
to
migrate
in
our
analysis
shown
on
the
next
slide?
We
simply
vary
through
the
time,
and
then
you
can
see
with
respect
to
which
time
threshold
which
of
these
designs
is
one
controller
to
controller
three
controllers.
Four
controllers
is
more
flexible
so
which
setting
of
how
many
controllers
can
accommodate
can
support
more
of
the
new
requests
of
these
flows
flow
requests
that
are
coming
into
the
system,
and
this
is
what
I
show
here
yeah
in
this
draft
on
the
upper
one.
F
You
see
the
flexibility.
This
is
the
migration
success
ratio
and
you
can
observe
two
things
here.
Blue
is
the
system
with
one
controller,
as
we
have
a
very
very
short
time
constraint,
120
milliseconds
here,
even
with
155
milliseconds,
he
can
observe
that
one
controller
is
surprisingly
more
flexible
in
these
cases.
F
What
you
probably
thought
is
the
four
controller
cases
if
you
relax
the
time,
but
still
it's
a
tough
time
constraint
a
little
bit
more
then
for
considerable
t,
more
controllers
are
more
flexible.
What
does
it
mean
if
you
have
very
very
strict
time
constraint,
now,
keep
the
controller
in
this
given
setting
at
the
place,
and
you
can
then
serve
still
a
decent
number
of
designs.
F
However,
what
is
the
trade-off
of
this
flexibility,
as
you
can
imagine,
without
migrating,
the
controller
too
much
performance
is
not
very
good.
This
is
what
we
show
in
the
lower
graph
here,
yeah,
so
performance
now
lower
is
better.
This
is
the
average
flow
set-up
time
you
see
here.
The
blue
is
always
staying
here
at
this
level.
Now
it's
not
really
performant.
This
is
also
what
you
can
expect,
but
though
it
depends
what
you
would
like
to
have
now.
You
would
like
to
adapt
to
a
certain
situation
now.
F
I
would
like
to
migrate,
always
have
the
best,
or
rather
now
be
careful
with
your
time
and
keep
the
tight
time
constraints.
What
we
also
find
out
is
that
there
is
a
cap
in
the
game
you're
not
drawing
this
area.
A
cap.
You
see
also
a
little
bit
in
performance.
This
cannot
improve,
there's
also
kept
in
flexibility
and
what
we
don't
show
here
that
when
we
go
back,
migration
is
also
at
a
cost
at
at
some
point.
We
don't
gain
anymore
in
performance
or
in
flexibility.
F
The
only
thing
that
is
rising
is
the
cost,
because
we
more
or
less
migrate,
migrate
migrate.
Now
this
is
also
something
that
have
has
to
be
taken
into
account
when
considering
flexibility
and
so
far
we
have
not
seen
this
in
most
of
the
publication
publications
too
taken
care
of
yeah,
and
this
slide
brings
me
basically
all
ready
to
sum
up
now
what
I
have
said.
F
So
what
you
should
have
understood
now
is,
of
course,
this
is
what
you
know
is
we
were
faced
with
new
requirements
in
networking
research
from
all
these
emerging
industries
and
so
on
and
so
forth,
and
this
is
what
you
know.
Flexibility
is
a
key
point.
You
actually
serve
eight
now
and
in
the
past
c-calm
papers,
one-third
of
the
papers,
mentioning
flexibility,
somehow
as
some
of
the
design
goals,
but
only
very
few
of
them
actually
say
how
this
flexibility
is
reached.
F
F
So
we
argue
there
is
a
measure
needed
in
order
to
be
able
to
compare
different
designs,
and
this
was
is
a
first
proposal,
as
I
said,
this
is
part
of
a
longer
project
and
now,
in
the
second
year
of
this
five
years,
project
and
I
would
like
to
put
this
out
to
the
community.
Also
give
me
your
feedback,
or
do
you
think
about
it?
F
I
think
it
is
very
important
to
get
on
to
the
flexibility
and
to
be
able
to
compare
different
designs,
or
they
have
not
shown
today,
because
of
limited
time
is
also
be
aware
that
this
flexibility
measure
is
only
valid
now
with
respect
to
performance
and
in
particular
cost.
So
we
don't
know
yet
in
the
green
line,
how
flexibility
is
evolving
over
different
designs
and
whether
the
cost
now
is
there
a
key
point
where
we
say:
ok,
there
is
a
break-even
or
even
the
cost
is
then
rising
to
infinite.
F
H
G
Very
interesting
subject
and
actually
I
agree
with
you.
That's
it's
it's
an
important
topic,
but
I
I
think
it's
a
complex
challenge,
and
initially,
when
I
saw
flexibility
on
the
title,
I
I
thought
about
something
else,
I'll
get
to
that,
but
I
I
think
one
of
the
challenges
that
I
see
with
with
this
and
considering
as
a
measure
is
the
fact
that
it
really
depends
on
the
function.
What
function,
what
flexibility
means,
so
you
can
and
and
and
and
depends
on
the
function
and
depends
on
the
implementation,
so
I'm
really
curious.
G
Who
do
you
think
the
audience
of
this
measure
is?
Is
that
the
vendors
that
we
put
RFC
rfp's?
Is
it
the
internal
design
organizations
within
the
network,
operators
that
are
deploying
these
things?
Is
it?
Is
it
just
the
research
community
and
I'm
really
trying
to
understand?
Okay
I
got
a
measure
assuming
that
we
got
the
answer.
Okay,
so
what
do
I
do
with
it
and
what
that
measure
means?
Is
it
hundred
kilometers
an
hour?
You
know
it
versus
fifty
kilometers
or
now?
What
does
it
mean
for
a
car?
G
I
know
what
it
means,
but
for
for
you
know
for
your
example,
with
a
with
a
gateway,
I
think
it
makes
sense,
but
if
I
take
the
same
measure,
apply
it
to
a
totally
different
function,
that
measurable
stem
and
and
and
so
the
what
I
thought
was
a
flexibility,
and
maybe
this
is
something
you
can
look
into
it
as
we're
migrating
to
cloud.
First
solutions:
you
know
the
flexibility,
it
kind
of
comes
down
to
the
granularity,
and
how
are
you
composing
compose
things,
and
that
gives
me
more
flexibility.
G
How
I
designed
them
I
can
spread
them
out.
I
can
keep
them
together
if
I
can
have
a
solution
that
allows
me
to
do
all
these
things,
but
it's
my
choice
to
choose
that
flexibility.
That's
up
to
me
and
then
I
can
that
comes
with
some
measures
are
associated,
that's
even
better,
but
I,
just
kind
of
found
that
something
was
missing
and
I
I
know
it's
a
it's
a
long
journey
myself.
Good
luck!
Thank
you.
Thank.
F
You
yeah
it's
a
long
journey
and
a
lot
of
questions
that
you
put.
Let
me
try
to
take,
maybe
one
or
two
just
to
clarify
so
you're
totally
right
that
we
cannot
take
this
flexibility
as
a
measure
as
saying
you
were
flexible.
Eighty
percent,
you
have
flexibly
seventy
percent,
that's
it
so
I
tried
to
hint
at
that,
and
there
is
a
more
complex
thing
behind
it.
F
I
just
didn't
have
enough
time,
I'll
be
over
here
home
Siri
behind
it
and
also
a
mathematical
foundation
for
that,
and
we
have
to
use
these
aspects
and
exactly
as
you
say,
there
is
a
flexibility
and
decomposition
there's
a
flexibility
in
flow
steering.
There
is
a
flexibility
for
function,
placement
and
we
can
only
view
them
in
their
area.
We
can
also
not
say-
and
this
is
what
I
try
to
explain-
your
quality
of
service
is
better
than
his
I.
Always
you
would
probably
ask
me:
what
do
you
mean
eeeh
you
mean
my
latency.
F
Do
you
mean
my
data
rate
or
what
and
similarly
I
think
for
flexibility?
We
have
to
go
with
that
and
we
might
have
in
the
end
some
challenges.
Is
it
what
we
are
try
a
striving
at
for
which
case
which
kind
of
request
set
and
and
basically
request,
distribution
can
be
used,
and
then
the
idea
is
really
giving
a
tool
to
the
vendors
to
the
operators.
F
So
they
are
aware
of
this
flexibility
because
in
the
end
it
is
something
that
we
always
wanted,
but
nobody
really
could
put
it
into
numbers
and
in
the
end
it
was
always
the
cost
and
the
cheapest
system
was
then
bought.
I
had
worked
for
an
operator
myself
for
quite
some
time
and
I
had
struggled
a
lot
in
the
research
department
with
this
flexibility
versus
cost,
and
that's
why
giving
some
tool
where
you
can
argue
to
your
boss,
but
we
are
now
more
flexible
hope.
F
I
Well,
I'm
Georgia
telecom.
My
question
is
related
to
the
previous.
In
that
way,
as
an
operator
we
have
or
my
question
loss
originally,
is
this
measure
so
flexible
that
I
really
can
compare
two
vendors
or
different
designs
in
it
to
have
here
a
clear
view?
Morally
I
defined
the
parameters,
input
parameters,
but
our
only
aspect
we,
as
a
company
there's
a
new
department,
called
design
to
cost
things.
I
If
we
have
here
also
some
parameters
who
in
a
way
allows
us
to
in
a
way
as
says
and
compare
vendors
implementations
designs,
we
had
to
put
things.
These
are
exactly
questions
that
we
currently
also
investigating
and
try
to
find
solution,
and
my
question
is:
is
have
user
aim
also
to
a
plan
and
this
and
deliver
some
feature
or
a
standardized
functionality
that
I
really
can
compare
flexibility.
Possibly
there
are
other
parameters
which
makes
here
it
more
objective
and
real
comparison,
so
that
I
can
all
conclude.
I
F
The
idea
is
to
get
from
what
nowadays
is
quite
subjective
yeah,
because
we
need
some
basis
in
order
to
some
baseline,
a
basis
where
we
can
make
this
comparison.
Yeah
we're
trying
to
get
as
objective
as
we
can
I'm,
not
sure
if
we
ever
will
find
a
flexibility
where
we
can
just
compare
two
totally
different
designs
because
they
subjected
the
objective
of
this
design.
What
was
this
design
made
for?
This
is
an
inherent
feature
of
the
flexibility
measure,
but
what
we
are
doing
at
the
moment.
F
This
is
what
I
can
say
what
you
can
also
see
in
some
of
our
publications
be
trying
to
get
as
many
cases
the
other
together
cases
try
our
measure
and
please
do
as
well
yeah.
In
order
to
see
what
would
be
the
fundamental
parameters
that
are
most
important
as
a
design
criteria
for
flexibility.
Maybe
there
are
others
they
are
less
important
and
then
we
can
come
up
with
a
baseline
of
let's
say,
parameters
in
features
that
need
to
be
taking
into
account,
but
it's
quite
complex
and.
I
F
One
of
those
projects
that
where
they,
you
is
currently
funding
you
as
an
individual
researcher,
is
quite
some
money
directly
for
your
research
group,
you're
kind
of
free.
What
to
do.
This
is
an
ELC
grant
if
you,
if
you
know
what
is
Islam
basically
funded
for
this
for
five
years,
in
order
to
try
to
find
some
solution,
but
I'm
happy
not
to
get
input
and
discussion.
That's
why
I'm
also
here,
okay.
C
It
so,
first
of
all,
thank
you
very
much.
It's
very
appreciated
to
see
some
lights
in
the
life
of
the
architect,
so
thank
you
now.
I
would
have
two
comments.
One
I
think
is
about
very
often
as
an
architect.
You
see
proposal
design
and
you
don't
know
its
limits.
Where
is
it
that
I
would
have
to
destroy
my
design
to
go
for
something
that
gives
me
more
flexibility?
So
that's
one
question.
The
second
question
is
here
is
about
something
that
took
me.
C
25
years
to
find
in
Russia
in
the
academy
of
science
actually
is
about
the
problem
that
goes
with.
It
is
the
aggregation
of
the
choice
right,
so
you
are
going
to
end
up
with
criteria
like
flexibility,
security,
performance,
cost
stability,
integrate
ability,
blah
blah
blah.
Each
of
them
are
going
to
act
like
voters
and
grab
a
number
of
potential
design
and,
and
so
the
problem
very
often
people
think
Oh
voting
is
simple.
You
just
vote.
The
people
lose
track
that
mathematically
you
are
going
to
have.
C
The
result
of
the
vote
will
be
the
result
not
just
of
the
data
but
as
well
on
which
aggregation
aggregate
or
you
choose
I
could
prove
you
by
ways
that
you,
you
would
fix
the
data
of
each
of
those
voting
guys,
but
depending
the
way
you
actually
vote,
you
will
change
the
results
or
even
worse,
you
will
have
no
result.
So
it's
interesting
to
see
how,
if
you
don't
pay
attention,
we
could
lead
to
the
wrong
conclusion
by
choosing
the
wrong
voting
system.
C
So
now,
to
the
point
of
how
useful
is
this:
when
actually
I
saw
from
a
security
perspective,
some
labs
I
will
not
mention
them
all
so
much
right
off.
That
use
this
type
of
approach
to
calculate.
What's
the
best
security
answer
in
provisioning,
what
so,
what's
the
best
design
so
Imogene
when
you
are
going
to
have
like
in
my
country
Swedes
around?
Let's
say
you
know,
Pareto
it's
5
million
virtual
machines
and
they
got
a
massive
attack.
Guess
what
you
will
not
be
able
to
do
it
by
hand?
C
F
J
K
He'll
John
some
given
University
in
Korea,
so
you
mentioned
that
a
part
of
flexibility,
with
example
of
the
migration,
so
but
to
me
that
it
comes
to
me
the
like,
depending
on
the
use
cases,
for
example,
latency-sensitive
application
may
be
that
PMF
placement
and
a
selection
repeater
proper
measure
of
the
flexibility,
but
just
the
for
you
know,
simple
load,
balancing
or
simple
additional.
You
know,
for
example,
PMF
creation
cases,
for
example,
that
HTTP
PMF,
server
or
Mme.
When
you
know
more
user
is
coming
to
their
mobile
data
at
the
time.
K
What
we
needed
to
just
you
know
that
additional
creation
of
the
PMF
in
that
case
is
flat.
Flexibility
will
be
just
you
know
about
that.
How
the
network
can
adaptively
create
dynamically
its
resources,
so
I
think
that
flexibility,
the
pipe
on
is
that
definition
of
the
flexibility
will
be
vary
depending
on
the
disease
cases.
So
is
that
correct?
Or
it.
F
Is
definitely
use
case
dependent?
Of
course,
this
is
what
I
say
that
you
have
to
clearly
say
what
kind
of
flexibility
are
you
looking
at
if
it's,
the
flexibility
to
I
could
accommodate
new
flows
from
a
different
distribution
yeah
then,
of
course,
you
have
to
fix
how
the
system
is
looking
like
how,
if
it's
migration
or
if
it's
instantiation
of
functions
or
putting
functions,
you
are
on
distributed
cloud,
putting
the
function
beforehand
to
the
edge
cloud,
removing
them
putting
them
centralized.
This
is,
then,
something
that
is
on
the
cost,
balance
I.
K
F
A
Just
let
me
let
me
one
thing:
I
know
that
we
have
a
discussion
on
our
future,
but
one
of
the
things
that
they
have
had
very
often
has
been
that
the
fact
that
we
don't
have
long
enough
discussions
here
in
the
during
our
meeting.
So
if
you
don't
mind
they
will
I
look
I
will
let
the
discussion
continue
and
they
both
suggest.
A
L
Hassan
horribly
and
from
Huawei
Canada
I
think
they
made
of
the
work
you
you
presented
is
basically
it
quantified
the
flexibility
as
I
like
it.
Is
it
separate
performance
criteria
right
so
it
that's,
and
then
for
network
resource
allocation
or
function
placement
or
any
anything
you
want
to
call
it.
The
the
idea
is
to
basically
come
up
with
your
objective
function
and
then
solve
it
right,
so
that
flexibility
is
just
giving
another
dimension
to
the
function
right.
Exactly
fairness,
fairness
is
something
that
was
measured
well
and
then
people
define
image
finest
criteria
right.
F
So
we
have
discussed
is
that
the
question
is
yeah
yeah
it
make
is
about
aggregation,
yeah
I
think
we
had
this
before
the
flexibility.
Definition
is
already
quite
complex,
as
we
see
from
the
discussion
here.
We
have
it
in
mind,
yeah
you're,
thinking
about
it.
There's
always
these
discussions.
How
to
combine
it
with
reliability
was
not
mentioned
before,
for
example,
yeah,
security
and
so
on,
of
course,
would
be
happy,
but
that
try
first
to
go
step
by
step
in
the
project.
F
A
G
Everyone
more
common
general
to
AT&T
to
throw
another
wrench
in
your
complexity.
I
think
part
of
the
factors
to
consider
for
flexibility
is
legacy.
Support
versus
you
know,
greenfield
I!
Think
it's
a
totally
different
or
you
know
you
can
call
it
hybrid
design
versus
you
know,
new
design
that
has
a
significant
factor
and
in
the
end
your
problem,
you're
trying
to
solve
is
there's
a
joke
in
the
project
management
space.
Where
I
can
extrapolate
it
with
like
you
can
have
you
know
you
can
have
it
fast,
flexible
and
cheap,
but
you
can
only
choose
right.
G
M
Proximity
of
our
cloud
placement
is
quite
interesting
and
relatively
new
tournament,
but
it's
going
all
over
the
place.
So
in
ITF,
in
routing
space
in
eight
year
we
use
function
called
PC.
We
choose
this
number
of
objective
functions
to
place
a
workload
somewhere
or
associate
part
of
the
workload.
Have
you
looked
into
implementing
of
your
work
on
top
of
PC?
That
would
use
objective
function
to
compute
the
best
place,
to
initiate
workload.
M
F
A
pass
computation
element:
we
have
thought
about
it,
we
have
not
implemented
it
on
that
one
I
have
to
say
so
so
far.
This
was
more
that
say
on
the
ILP
solving
optimization
problem.
We
have
not
implemented
it,
yet
we
thought
about
it,
because
some
of
these
examples
have
been
taken
from
work.
We
did
with
vendors.
You
took
these
examples
and
re-evaluated
flexibility
right,
so
there's
other
publications
if
you're
more
interested
in
the
placement
problem
with
different
perspectives,
also
multi,
objective,
optimization
and
so
on
then
check
the
website
there.
F
B
Sahana,
I'm
also
from
the
same
university
Technical
University
of
Munich
Oh.
So
if
we
consider
a
node
be,
for
example,
in
a
4G
network
from
two
different
vendors-
and
you
say
you
have
the
same
set
of
constraints
and
then
you
can
say
you
know,
B
of
this
vendor
is
better
than
the
other.
But
when
you
take
the
same,
you
note
B
in
Phi
G,
probably
the
constraints
are
different,
so
are
we
also
comparing
for
the
same
element
across
generations
or
across
domains?
This
is
one
particular
example,
but
this
could
happen
for
other
cases.
F
H
F
Them
the
same,
then
you
can
compare
all
the
designs
of
different
generations.
You
can
find
out
that
maybe
for
G
4.5,
4.8
G,
whatever
is
similarly
flexible
as
v
2,
but
you
have
to
define
the
question
what
you
would
like
to
ask:
what
flexibility
with
respect
to
what
aspect?
This
is
very
important.
Otherwise
it
doesn't
work.
A
Okay,
before
anyone
else
steps
up,
thank
you
so
much
for
length.
Thank
you
so
much
to
you
all.
If
any
of
you
have
not
signed
the
blue
sheet,
please
do
it
so
near
when
you
go
out.
I
will
see
you
and
finally,
since
I
announced
this
idea
of
a
shorter,
more
frequent
meetings,
we
are
planning
to
have
interim
meeting
sometime
in
at
the
return,
the
term
of
the
summer
in
September,
October,
probably
September
in
the
US,
precisely
and
focus
if
we
find
a
proper
speakers
from
all
new
design
patterns
for
NATO
functions.