►
From YouTube: IETF101-NFVRG-20180321-1330
Description
NFVRG meeting session at IETF101
2018/03/21 1330
https://datatracker.ietf.org/meeting/101/proceedings/
A
A
If
you
don't
intended
to
be
here,
is
about
time
for
you
to
leave
the
room
if
you
intended
to
be
here
police,
seeing
the
blue
sheets
that
you
will
receive
enough
and
I
such
a
moment.
Those
of
you
who
have
the
blue
sheets-
please
circulate
them.
I
will
remind
these
couple
of
times
during
the
session.
Well,
I'm,
developing
one
of
the
chairs
of
the
NF
ERG
to
my
right,
federal
and
other
visa
is
acting
co-chair
today.
A
A
A
few
a
few
reminders
of
fleeing
set
cetera
where
the
materials
all
those
lights
that
have
been
prepared
for
today
are
available
and
the
material
sites,
including
pointers
to
the
drafts
that
will
be
discussed.
The
links
for
for
remote
participation,
I
know
you
don't
need
it
for
the
sake
of
those
who
are
connecting
from
outside
the
Middle
East.
If
you
are
not
subscribed
yet
web
and
weekend,
and
if
you're
curious
about
the
note
that
Pedro
right
Linda
was
Linda,
she
told
me
she
was
going
to
be
here.
A
A
Thank
you,
and
just
if
any
of
you
is
willing
to
connect
to
the
jabber
room
and
act
as
a
scribe,
simply
that's
simply
paying
attention
to
what
others
say:
go
to
the
mic
to
transmit
any
question
from
the
people
participating
by
Java.
That
will
be
very,
very
much
appreciated.
It's
something
very
simple!
You
have
just
to
keep
reading
a
screen
that
we're
very
seldom
gets
any
questions
and
bring
the
question
to
the
mic.
You
won't
be
hall
responsible
for
anything
that
is
said
in
Java,
you're,
just
a
messenger.
Anyone.
A
Well,
another
announcement
about
research
related
events
in
the
old
days
we
we
had
at
least
at
least
here.
This
has
changed
because
we
have
too
many
lately
well,
which
is
a
good
thing
and
what
we
use
is
there
is
a
list.
The
any
virjil
is
for
an
ounce
and
is
any
call
for
papers
any
call
for
the
most
contributions,
whatever
please
use
the
list
for
that.
But
please
limit
the
announcement
to
research
relevant
events
not
to
announce
a
new
release.
Is
a
software
not
to
announce
for
sure
seminars
of
from
products
or
whatever.
A
A
A
Second-
is
an
announcement
that
we
will
have
an
interim
meeting
on
the
week,
17
the
one
starting
on
the
23rd
of
April
in
the
very
area.
We
still
have
to
decide
they
concrete
place,
but
it
will
be
duly
announced
on
the
on
the
list.
The
yes
that
we
I
mean
a
few
of
us
have
been
talking
about
motor-driven,
optimization
how
you
can
rely
on
model
definition
to
to
optimize
placements
and
life
cycle
in
general
of
natural
functions,
and
there
are
a
few
people
working
on
this
there.
There
is
an
event.
A
I
mean
that's
really
when
that
is
happening
in
the
Bay
Area
during
those
days.
So
we
will
take
the
opportunity
to
have
it
will
be
a
half
day
depending
on
the
number
of
speakers.
It
will
be
a
half
day
or
one
day,
event
in
which
the
earliest
represents
use.
Cases
results
on
this
idea
of
more
the
realization.
If
you
happen
to
be
in
the
surroundings,
you
most
welcome.
If
you
are
willing
to
participate
remotely,
you
will
be
most
welcome
as
well.
A
So
I
want
to
draw
your
attention
as
well
to
two
other
relevant
documents
and
results
that
we
won't
have
time
to
be
sent
today,
but
is
something
that
I
bought,
like
the
the
group
chickens,
to
consider
to
make
some
comments
remarks
whatever
whatever
you
have
on
them.
When
is
a
new
version
of
this
raft
that
was
presented
in
Singapore
if
I,
remember
well,
unanticipated
allocation,
I
don't
have
put
a
direct
link
or
explicit
link,
because
you
should
be
easy
to
find
it
on
the
data.
A
Tracker
second
is
a
document
that
was
published
on
this
December
in
the
was
presented
in
Indian
I
Triple
E
nav
ASEAN
conference
that
I
found
very
interesting,
I
invited
the
the
speaker.
We
could
not
make
room
for
him
so
I.
This
is
the
the
reference
is
on
down
the
purification
of
service
chains
and
is
I.
Think
I
think
is
very
much
connected,
or
is
it
advances
the?
A
What
we
have
in
a
if
you
remember
document
has
been
around
is
look
for
quite
a
while
on
precisely
on
service
chain,
modification
and
I
think
is
I
suppose
reading.
Finally,
something
that
has
been
presented
at
the
IC
energy
that
we
we
had
to
balance
here.
The
agendas
so
is
is,
is
a
paper
on
how
to
the
experience
of
deploying
ICN
stacks
using
using
n
Fe
techniques,
so
it
has
been
presented
and
demonstrated
the
IC
energy.
So
if
you
are
interested
go
to
the
material
sites,
you
will
refine
it
on
the
under
under
IC
energy.
A
A
There
will
be
above
propose
group
that
is
called
comes
on
slice,
natural
slice,
orchestration
and,
and
there
will
be
the
second
poem.
This
is
the
moment
in
which
it
will
be
decided
whether
it
will
become
a
work,
a
working
group
or
not.
If
you're
interested
interested
in
network
slicing
pay
pay
a
visit,
we
have
been
discussing
this
morning,
they
are
to
an
SF
capability
mobile.
If
you're
interested
have
a
look
at
it.
A
This
morning
there
was
a
IOT,
F,
open,
I,
know
I'm
late,
but
if
you're
interested
how
IRT
AVC
is
evolving
in
general,
you
can
go
go
for
the
look
at
the
materials
and
there
is
a
barb
off
on
edge
computing
today
as
well.
Whoever
is
interested
I,
don't
know.
If
Shang
is
here,
no
no
I
don't
see
him
so
well.
I
can
give
you
some
details
later
on.
If
you
are
interested
in
participating
whatever
else
you
have,
we
will
go
at
the
at
the
end.
A
Hopefully
we
will
have
10
minutes
for
open
mic
mic
any
matters
related,
so
we
can
discuss
our
days.
We
can
make
a
sure
presentation
of
some
of
the
graphs
and
the
documents
that
were
that
they
mention
before
that
was
are
not
a
part
of
the
agenda
currently,
so
we
will
move
on
and
go
for
Carlos
to
to
take
care.
If
you
don't
mind,
Carlos
deal
with
the
three
your
three
drafts
in
a
row,
so
we
thank.
C
No,
okay,
so
first
quick
update
on
the
on
the
gaps
draft.
So
basically
the
status
is
that
the
version
city
of
six
was
sent
to
the
IRC
four
for
comment.
We
got
very
good
reviews
from
Michael
wells
and
a
foreign.
Then
we
have
the
rest.
Those
comments
in
the
releases
on
the
update,
seven
and
eight
and
then
on
version.
Eight.
We
also
got
a
review
by
Vincent
raka,
mainly
related
to
what
they're
worth
different
comments,
but
Sandin
were
very
critical
to
address
were
directed
to
the
IE
RTF.
C
C
That
there
were
some
things
that
we
didn't
follow,
so
we
addressed
those
in
version
9,
so
we
leave.
The
document
is
now
ready
since
the
beginning,
I
think
the
the
document
has
been
reviewed
by
more
than
twenty
people,
so
I
think
it
reflects
the
the
research
group
view
so
I
from
the
author
size.
We
believe
that
is
ready
for.
A
C
Okay,
so
this
is
an
update
on
a
document
that
has
been
around
for
for
a
while.
This
is
about
multi
domain
visualization,
multi
domain
orchestration.
Originally
this
working
from
one
project
that
we
are
in
most
recently
in
this
version.
Actually
we
merge,
or
we
join
team
up
forces
with
another
project
that
is
now
basically
taking
all
everything
on
on
the
results
from
from
the
redeploy,
so
you
can
see
that
the
the
number
of
authors
has
increased,
and
now
we
have
people
from
CT,
TC
and
EC
SSA
networks,
volantis
and
interdigital.
C
So
the
rationale
of
the
document
I
will
spend.
This
is
about
the
need
for
multi-domain
orchestration,
so
there
are
scenarios
even
for
the
pure
point
of
view
of
infrastructure,
where
you
wanna
orchestrate
resources
from
or
belonging
to
different
domains
and
also
from
the
point
of
view
of
services
that
you
may
want
to
pull
in
or
join
or
put
together
functions
that
are
managed
by
different
domains.
C
So
this
was
the
base
rationale
for
for
this
kind
of
document.
Then
there
are
things
that
need
to
be
done
that
are
not
yet
done
from
civilians
in
a
V
architecture.
In
order
to
enable
this
multi
domain
or
castrated
orchestration,
and
there
are,
there
is
room
for
ITF
protocols
to
feel
some
some
of
the
gaps
that
we
that
we
found
there.
C
So
with
that
in
mind,
we
started
this
document
with
basically
with
this
fight
the
exploit
that
is
operating
with
multi
domain
orchestration
in
particular,
and
then
now
it's
also
getting
input
from
another
point
that
is
qualify
a
transformer
which
is
also
doing
this
in
a
more
global
scope
for
slicing
for
vertical
industries
and
considering
federation
of
domains.
So
here
just
to
show
what
has
chain
one
single
slide,
you
can
see
that
there
is
a
lot
of
new
common.
Even
the
the
size
of
the
document
has
increased
from
16
pages
to
33
pages.
C
Very
simply,
this
is
architecture
that
we
started
to
look
at
in
5dx
like
okay.
You
have
a
multi
domain
kind
of
Orchestrator
x,
Orchestrator
that
is
then
dealing
with
the
domain
orchestrators
and
also
interacting
with
other
multi
provider
orchestrators,
this
interface
in
the
middle
and
with
whatever
customer.
It
is
in
the
interface
that
you
have
on
the
north
pal.
So
this
was
the
immediate,
the
initial
architecture
and
we
basically
in
the
document
describe
each
of
these
particular
interfaces.
What
are
the
things?
C
The
functionalities
that
has
to
be
provided
and
fy'y
transformer
basically
follows
a
very
similar
approach.
I
will
not
go
into
details
for
the
sake
of
time,
but
is
basically
the
same
having
a
new
functionality
with
this,
this
vertical
slicer,
which
is
basically
the
functionality
building
with
with
the
final
customer
setting
in
the
case
of
this
approaches,
are
the
vertical
industry.
So
you
wanna
have
I
slice
for
each
held.
You
want
a
half
a
slice
for
industry,
for
the
0.
C
You
wanna
have
a
slice
for
being
cooler,
so
that's
basically
the
role
of
the
vertical
slicer
to
get
that
request
and
to
manage
the
slice
in
one
single
domain
on
multiple
domains
for
if
that's
require,
so
that's,
basically
it
so
next
steps
inside
the
XY
evaluating
software
proof
of
concept
of
this
architecture,
so
that
is
running
code.
Most
of
it
is
open
source
you
can
check
in
on
the
website.
This
will
be
taken
by
file
transform
and
as
I
mentioned,
so
there
is
commitment
from
from
people
to
leverage
under
result
an
uninvolved
in
Fayette
transformer.
C
We
will
be
also
to
in
open
source
mainly
to
commit
in
to
main
open
source
project
like
osm
or
others
like
on
up.
So
there
are
something
that
will
be
moved
there.
So
question
is,
or
request
is
for
the
people
here
to
provide
feedback
on
this.
We
believe
that
is
a
lot
of
content
that
may
be
useful
in
a
very
valid
topic
that
is
multi
domain
orchestration.
C
C
And
the
last
presentation
is
about
new
work
that
we
are
looking
at
about
discovery
of
resources
and
basically
the
dynamic
association
between
those
resources,
so
they
built
own
infrastructure
manager
and
then
a
view.
So
basically,
motivation
here
is
that
we
are
looking
order.
Some
scenarios
in
which
resources
are
dynamically
appear
and
disappearing
and
volatile
are
mobile.
So
you
cannot
assume
like
data
center
environment,
which
everything
is
kind
of
fix
unknown
and
you
come
pre-configured
things,
but
it's
more
about.
C
Okay,
I
have
resources
on
my
mobile
device
on
my
car
and
the
excess
infrastructure,
and
then
I
connect
a
name
maybe
discovering
those
resources
that
I
can
potentially
used
and
orchestrate
so
with
that
idea.
Basically,
what
we
want
to
look
is
into
dynamic,
Association
and
discovery
of
built
on
infrastructure
Minaya
from
an
Orchestrator
point
of
view.
Basically,
we
understand
that
the
beam
discriminated,
of
course
has
na
VI
resources
associated.
So
what
we
discover
is
the
theme
from
the
NA
VI
point
of
view
or
the
other
way
around
from
the
appeal.
C
With
from
the
beam,
we
discover
potential
any
views
in
a
mobile,
dynamic
environment.
So
this
we
what
we
will
have
in
mind,
so
some
examples
of
the
resources
of
the
scenario
story
that
we
are
considering
on
the
top
part
of
the
figure
you
see,
that
is,
in
the
net
worth
alyssum
orchestration
capabilities,
of
course,
with
some
resources
available,
but
abs
to
also
discover
resources
that
are
more
vile,
that
care
attached
to
the
network
that
can
potentially
be
used
by
that
orchestration
intelligence
in
the
network.
C
So
there
is
how
to
make
those
resources,
mobile
resources,
advertise
and
discover
by
the
network
on
the
top
on
the
bottom
part
of
a
figure.
What
we
get
is
a
mobile
orchestration
mobile
operator
that
is
connected
to
a
network
and
basically,
that
orchestration
capability
or
intelligence
wants
to
discover
resources
that
can
be
orchestrated
to
be
used
for
a
particular
functionality.
So
the
mobile
nabo,
connecting
to
a
network
and
trying
to
discover
potential
resources
that
could
be
could
be
used
in
the
infrastructure
or
even
by
other
mobile
nodes
that
are
availa.
C
So
that's
in
a
1-minute
summary
the
idea
behind
the
draft.
So
this
is
not
a
standardized
in
at
Siena.
We,
as
far
as
we
understand
we
read
that
is
room
for
these
two
to
enable
protocols
to
convey
this
discovery,
to
enable
this
association
between
the
beam
and
FBI
and
any
view,
sorry
in
dynamic
environments.
So
again,
maybe
is
that
for
you,
if
you
are
interested
in
working
with
that,
let
us
know
we
will
now
start
looking
into
into
potential
solutions.
There
are
some
projects
where
we'll
be
doing
two
of
concepts
of
this
solution.
C
It's
imagine
that
you
have.
You
have
a
modern
architecture,
a
straight
row.
You
have
resources,
but
you
want
to
find
more
resources
that
you
can
use.
So
there
may
be
some
there
may
be
some
other
devices
that
are
mobile,
I
have
available
resources
and
you
have
our
ties
them
in
the
network
for
others
to
potentially
use
associate
those
resources
and
use
for
a
given
time.
So
that's
you
basically
advertise
for
potential
users
of
those
resources
that
you
have
so.
C
A
Well,
I
have
another
question:
is
physically
I
understand
one
of
the
directions
understand
some
infrastructure
announcing
availability,
but
I
can
hardly
imagine
they
need
to
discover
an
Evo
I
mean
any
of
you,
something
that
is
requires
a
strong
trust
foundations.
Reviews
is
going
to
control
everything
that
you
deploy
a
fan.
It's
a
little
bit
unusual
or
a
strange.
F
C
B
B
See
no
okay,
I'll
tell
you
the
next
one,
so
this
was
presented
in
Berlin
last
time
was
already
full
I
I
got
some
volunteers
to
read
through
it.
So
thank
you
very
much
for
the
ring.
I
then
integrated.
All
the
comments
I
got
from
the
people.
Most
important
thing
is
that
we
have
a
real
life
implementation,
which
is
completely
finished
and
available,
and
we
get
up
there
mean
if
you
want
drawback,
is
that
it
needs
an
open
daylight
to
run.
B
On
top
highlights,
it
generates
output
for
the
legacy
open,
mono
and
osm,
and
provides
a
very
clean
template
for
adapting
for
any
other
platform
or
format
you
might
need
next.
Please
so
I
have
to
recognize.
I
have
a
love-hate
relationship
with
our
development
and
one
of
the
things
this
project
really
needs
is
to
have
an
update
for
more
recent
frameworks
in
No
Deal
or
even
to
make
it
or
deal
independent.
B
That's
something
we
we
are
looking
for,
how
to
to
see
how
to
progress,
and
then
the
other
challenge
for
all
plans
for
us
is
to
apply
it
in
real
life.
So
we
have
a
European
project,
which
is
the
mommy
project,
doing
measurement
and
stuff,
and
we
are
moving
that
into
an
NLP
environment
and
in
that
happy
environment.
What
you
have
is
change
is
a
very
long
change,
because
you
have
producers,
you
have
middleboxes,
you
have
many
things
and
that's
in
normal,
OS
m
or
open
Manolo
would
be
a
very
complex
yang.
B
So
we
are
going
to
be
using
that
scenario
to
see
how
easy
it
is
to
use
this
new
language
to
do
things,
because
again,
the
ml
is
quite
tricky.
It's
extremely
time
consuming
and
very
error-prone
next
and
last
I.
Think
yes,
so
I
continue
to
be
open
to
further
requests
or
suggestions.
So
if
someone
reads
the
document,
please
come
forward
and
if
reasonable
time
there's
none
I
would
like
to
see
this
profit
by
planning.
B
A
H
Overall
goal
of
the
fiber
gene
fire
is
basically
to
build
an
extensible
of
5g
NFB
Beza
ecosystem
focus
mainly
and
veg,
going
to
search
for
this
case.
This
platform
is
based
on
two
main
principles.
First,
one
is
the
architectural
standards,
basing
the
ETS
I
own,
a
TV
reference
architecture.
Ii
second,
is
basing
the
open
source
technology
to
validate
and
also
to
initial
the
platform
we
are
focusing
to
occur,
induces
the
5g
automotive
industries
and
ours.
A
smart
city
induces
the
vertical
actually.
H
Currently,
we
already
have
five
eg
in
Fido
persist
in
our
deployed
and
operational
and
I
will
talk
later.
We
have
all
read
the
open
calls
where
we
are
looking
for
extended
the
functionality
as
we
are,
the
new
infrastructure
ease
to
this
five
g5e
plataform
and
those
making
available
for
new
experimenters.
H
Here,
as
you
can
see
in
this
village,
I
presented
a
the
overall
5g
firing
appearance
mode
of
our
factory.
That
is
based,
as
I
said
before,
in
that
existing
open,
so
sparsh
like
OpenStack,
open
daylight
as
as
a
Sdn
controller
and
dasa
for
the
orchestrator.
We
are
working
open
source
in
mono
basin,
the
ETS
dia
various
architectural.
These
architectures
of
5g
fire
is
composed
of
four
main
building
blocks,
the
first
one.
H
We
introduced
a
new
concept
of
Vienna's
that
we
call
we
call
an
experiment,
experimental
instances
of
verge
cows
where
we
define
the
V,
except
in
this
case
we
are
not
looking
just
for
network
functions.
We
are
looking
to
go
beyond
from
the
application
layer,
all
the
functionalities,
from
the
application
layer
and
to
the
net
onto
to
the
network.
Later,
then,
we
have
the
automated
the
management's
and
orchestrations
that
is
ETS
a
I
complain.
H
And
finally,
we
have
the
NL
s,
the
third,
the
building
to
have
the
packaging,
five
design
and
architecture
from
work
by
itself
where
we
can
define
a
new
api's
and
extend
the
employee
new
facilities
as
well
as
new
functionalities.
The
last
one
four
building
blocks.
We
have
the
network
function,
virtualization
infrastructure
that
included
a
compute
node,
the
network
node
only
the
storage
node
by
itself
this
this
reference
model
I,
have
here
the
core
architectural
components.
H
These
figures
show
the
corner
of
tetro
components
of
the
five
gene
file
and
mapping
to
the
ETS
I
reference,
a
Couture
article
in
every
what
you
can
see
here,
that
we
have
in
fact
a
file,
a
web
portal
where
you
can
experiment
and
apologia
network
serves
descriptive
as
well,
vnf,
descriptive
to
upload
the
experiment
and
also
through
this
portal
we
can
do
the
5g
for
administrator
can
manage
the
50
platform
by
itself.
This
can
be
mapped
mapping
to
the
ETSI
reference
architectures
regarding
to
the
operation
of
support
system
or
business
support
system.
H
Here
we
have
there
all
the
orchestrations
of
the
vnfs
of
the
exurbs,
where
you
have
the
same
manner
where
you
have
the
capability
of
the
selfs
orchestrate
under
his
source
orchestrations
and
the
SCA.
The
configurations
of
the
vnf
another
map
by
itself
is
their
energy
I,
the
infrastructure
by
itself,
where
we
have
the
store,
node,
the
computing
level
node
and
the
network
node.
H
Basically,
to
validate
these
are
infrastructure
disable
system.
Finally,
existing
we
have
started
with
tool
to
a
Beach,
condo
service
and
the
same
time
are
for
a
use
cases.
The
first
one
is
the
automotive
industry,
where
you
deploy
account
will
take
use
cases
to
validate
the
data
for
this
car.
We
take,
we
use
the
vehicle,
our
networking
that
was
deploying
it.
H
The
first
one
we
use
the
pinky
Ponk
is
a
well-known
and
networks
FC
the
bug
it
has
a
infrastructure
mainly
that
when
we
have
the
multi
site
or
multi
domain
orchestrations,
then
we
have
our
CR.
Unified
gateway
is
another
use
case
and
a
case
where
we
deploy
becomes
a
5
g,
5
Gina
box,
where
we
deploy
in
Bristol
the
LTE
architecture,
where
we
have
window
to
be
okay
and
they
EPC
the
core
of
the
network.
Where
we
have
done
here.
H
H
Basically,
here
I'm
gonna
go
to
show
quickly
here
how
work
is
they
fight
which
Oh
ecosystem
in
terms
of
experiment?
And
then
we
have
some
animation
here.
If
he's
managed
that,
if
you
imagine
it
that
this
blue
blue
ball
is
a
experimenter
and
he
has
experiment
that
he
want
to
deploy
right
and
then
we
are
expect
that
he
has
this
experiment.
Descriptive
could
be
armed
or
Tosca.
H
Then
he
will
deploy
these
to
the
5g5
for
Tao,
in
fact,
in
Port
au.
He
has
this
need
to
be
registered
before
to
deploy
something
once
the
experiment
is
registered,
he'll
be
available
for
him
all
the
villages
debatably,
except
of
via
net
for
the
beach
cow
industries,
where
he
can
instantiate
modified
and
updated
using
his
experiment.
H
Once
they,
the
user,
has
submitted
his
network
self-descriptive
vnf
descriptor.
He
can
do
all
the
configuration
where
he
can
choose
in
which
facility
he
won't
run
the
experiment,
since
we
would
work
with
much
demand.
In
our
case,
we
have
deploy
a
facility,
a
testbed,
anitya
peru.
I
improved
to
go
another
one
d5
five
torniquet,
madrid
and
another
one
team
will
be
in
briefs
universe
of
Bristow.
Then
he
can
choice
here
where
he
want
to
run
his
experiment.
At
the
same
time,
he
can
schedule
it
his
experiment
what
he
says
to
be
meet
his
experiment.
H
They
are
the
monitoring
orchestration.
Your
dark
is
OS
hermano.
We
will
receive
this
via
API
from
the
post.
The
portal
connect
the
same
man,
the
network
serves
descriptive
and
then
you
proceed
of
the
right
allocation
of
resources
and
which
facility
we're
going
to
go
to
run
a
scheduled
experiment
of
the
users.
H
Of
course,
as
we
are
well
using
the
manifest
money's
basic,
we
received
a
the
service
of
the
state.
We
will
receive
the
request
for
the
port
au
5g
website
port
out
and
then
we'll
communicate
with
the
resource
orchestrations
and
the
vnf
configurations
and
abstract,
if
necessary,
to
deploy
the
experiment
of
the
users
and
they
specific
acidity
or
mooch
facilities.
H
Once
this
help
of
figuration
is
done,
as
I
said,
we
can
deploy
this
d5
tonic
and
Madri,
or
you
can
deploy
it
available
to
God
or
in
Bristo,
or
you
can
deploy
this
and
mooch
side
chauka
station
at
the
same
time
since
they
use
at
the
experiment
and
can't
have
our
to
VN
apps.
That
can
be
a
interconnect
between
them,
but
each
vnf
can
be
different
sites.
H
H
Once
you
have
allocated
the
resources
of
the
experiment
and
they
specifically
facility
too
much
facility,
then
they,
the
experiment,
is-
is
almost
ready
to
run
his
experiment
because
we
have
everything
connected
now
you
have
provided
a
connection
for
the
right
vnfs
of
he
accepts.
Then
he
can.
He
can
run
the
experiments.
H
Basically,
we
are
looking
to
extend
our
the
functionalities
and
say
welcome
to
new
facilities
that
can
enjoy
the
5g
our
platform.
For
this
end,
we
are,
we
have
managed
our
what
we
call
open
calls
and
you
have
to
open
calls.
Basically,
the
first
one
is
a
year
one
tube.
This
is
approach
of
two
years.
They
this
first
open
Co,
is
a
experimental
and
the
new
infrastructure
and
they
are
their
house
for
the
provider.
That
was
a
Jew
shinato
do
vxf
of
each
car,
which
going
to
see
the
second
openko
is
the
year
2/3.
H
We
are
expected
to
receive
a
new
experiment
that
can
explore
it,
make
better
use
of
the
5g5
plataform
they
total
founded.
The
amount
is
here
budget
available
for
the
third
part
that
we
are
doing,
support
them
to
join
us
in
this
pipe
dream.
Failure
persistent
and
helps
to
extend
the
overall
the
platform.
H
Basically,
the
first
open,
coil
phase,
one
was
is
already
closed,
was
opening
December
of
2017
and
closing
are
in
February
2018?
Then
you
won
they
last
to
deface
it
to
you.
Belong
should
end
the
spring
2018
they
open
concept,
the
second
open
call
and
phase
one
will
be
autumn
2018.
Basically,
the
openko
include
two
categories.
The
first
category
is
the
goal:
is
invite
experimenter
to
use
and
explore
the
5g
facility
that
we
have
a
red
line
et
aveiro
five,
don
Madre
and
Bristow
the
second
category.
H
We
are
looking
to
invite
the
open
source
development
of
vxf
developers
to
join,
to
extend
the
capability,
the
functionality
of
hydrogen
fight,
and
also
we
are
looking
for
invited
designers
to
join
us
as
I
extended
the
new
facility.
Currently,
actually,
we
have
three
facility
operational
one.
The
fourth
one
is
in
Brazil,
but
we
want
to
add
more
facility.
Then
then,
whatever
to
include
new
facilities
coming
as
I
said
before,
the
5ug
file
over
rocks
platform
or
ecosystems
is
completely
open.
Source
is
available
bed.
All
the
code
is
available
in
github.
H
They
use
cases
that
we
have
deployed,
find
cava
takes
Maxis
say
that
you
use
ffmpeg
bag.
Video,
transcoder,
Vienna
and
OpenCV
Vienna
is
available
in
public
in
the
each
hub
as
well.
We
have
the
week
and
the
website
for
new
xperia
x
parameters
the
release
that
we
have
been
used
so
far,
in
fact
in
file
ecosystem.
The
OSM
magnitude
is
true,
but
we
are
right
now
trying
to
be
great
to
release
a
tree.
H
H
I
Have
one
repair
on
the
roof,
yeah
Yakov,
Stein
I,
not
quite
sure,
I
understand
exactly
what
kind
of
experiments?
No
it's
where
the
vnfs
can
be.
Is
this
just
in
the
core,
or
could
it
be
in
Iran
if
I
want
to
it?
For
instance,
the
the
use
case
that
came
into
my
mind
is
changing.
Splits
could
I
do
some
of
the
functionality
of
the
radio
access
network
as
a
vnfs,
as
can
be
done
in
a
virtualized
G
node
B
in
this
project,
yeah.
H
You
can
the
idea
5
D,
Phi
D
is
provided
like
repository
of
the
access
for
Victoria
cavero
industry.
In
this
case.
In
this
case,
they
will
have
the
experiment
that
only
can
use
a,
for
example,
the
existing
the
accept
that
we
have
their
gap,
and
you
have
the
vxf
develop
that
can
provide
a
new
functionality
capability.
They
can
develop
a
a
new
VN
f
of
XF
in
deploying
this
is
positive
once
with
the
point
is
depository
we
need
to
choose
and
which
fastest
you
can
deploy.
This
now.
I
I
understand
that
you
could
put
theoretically
the
function
there.
The
question
is:
where
do
you
have
the
points
where
the
function
can
run?
Do
you
actually
have
deep
into
the
network
towards
the
G
note
bees
someplace
to
put
the
functionality?
Are
you
all
always
in
the
back?
You
have
your
servers
in
some
centralized
site,
yeah.
H
A
You
know
me
one
of
the
things
that
a
Louis
you
mention
is
with
this
unified
gateway.
The
idea
is
that
we
do
we'll
be
able
to
play
with
virtualized
functions
as
well
under
for
the
radio
environment.
Okay
is
I
mean
not
in
the
real
natural
we're
not
talking
about
the
production
network.
I
understand,
but.
I
It's
just,
but
no,
no,
it
is
something
which
would
be
really
useful
to
do,
because
there's
no
experience
right
now
with
the
differences
between
the
different
splits
and
how
they
can
affect
different
applications.
When
you
have,
let's
say,
delay,
sensitivity
or
something
and
there's
no
way
to
do
it
really
accept.
It
is.
A
H
H
H
B
J
So
that
picture
it's
this
one,
that
picture
shows
basically
a
multi-site
orchestration
scenario
and
it's
pretty
centralized
right.
So
you
have
one
Orchestrator
deployed
and
five
tonic,
and
then
you
have
distributed
Vince
now
exactly
I
assume
that
after
instantiation
of
the
different
in
apps,
there
is
quite
some
traffic
between
the
different
sites,
so
not
only
for
weather
reports
between
vnfs
and
we
enough
manager
and
the
manager
is
supposed
to
be
on
the
ocean
side
here.
J
H
J
Because
that's
something
that's
wrong
with
me.
Always
that's
a
picture.
You
typically
see
in
terms
of
boxes
and
lines
in
between,
but
I
was
always
wondering
if
people
consider
offloading
some
of
the
enough
amount
of
features
to
the
different
sites
just
to
reduce
the
load
on
the
different
sets
in
terms
of
well,
which
kind
of
traffic
goes
on
between
yeah.
H
K
A
L
H
Effective,
actually,
this
is
something
that
we
are
doing
now
manually.
We
go
to
the
OpenStack
dashboard
and
then
we
create
they
are
the
network
tennis
there
and
we
are
located
the
network
tent
there,
but
in
the
future
we
are
trying
to
automate
our
auto
mile
up
to
provide
autumn
automation
in
these
and
these
tasks.
Ok,.
F
A
F
H
M
Brings
me
my
mind,
close
to
mobile
its
computing,
where
the
different
and
vnfs
are
managed
by
different
actors,
and
this
is
related
to
the
question
about
slicing.
So
I
would
imagine
that
a
different
entity
actor
operates
in
Vanity's,
an
application
level
component
as
opposed
to
a
network
function,
so
I
have
to
question
based
on
this
observation.
The
first
one
is
whether
you
have
any
plans
of
formatting
extending
this
environment
to
also
cater
for
mobile.
It's
computing
aspects
like
location,
API
is
and
stuff
like
that,
which
is.
H
The
first
one:
yes,
we
intend
to
extend
the
capability
to
the
meek.
Note
ok,
for
example,
the
use
case
that
I
told
about
this
maths.
It
is
saved
where
we
were
working
with
206
digging
cameras
where
the
cameras
move
around
of
the
city
and
we
use
the
Vienna
visit
transcode.
We
are
trying
to
bring
the
computation
taker
of
the
computations,
the
Vienna
video
transcoder,
and
they
face
detection
of
origin
from
the
cloud
and
bring
too
close
to
the
use
in
this
case
to
the
edge.
But
we
are
I.
H
M
N
I'm
sorry
I
brought
from
much
innovation
center
University
of
Surrey.
Today
I'm
going
to
explain
to
you
our
nearly
finished
European
Union
project,
H
2020
project,
which
has
a
lot
of
correlations,
which
I
see
with
five
trillion
file
project.
This
is
about
a
federated
and
orchestrating
orchestrate
virtualization
testbed
for
network
functions,
virtualization
and
software-defined
networks,
especially
in
the
context
of
file
systems.
N
So
the
project
basically
has
put
a
lot
of
effort
in
orchestrating
a
multi-site
deployment
across
Europe
in
multiple
countries.
This
involves
Germany
sites
in
primary
focus
and
Technical
University
of
Berlin
University
of
Suribachi
I,
see
Center
and
Ericsson,
particularly,
and
also
the
show
telecom
was
mauled
and
later
on.
We
had
other
partners
like
assembly
data
systems
from
Italy.
N
This
slide
essentially
shows
you
the
overall
architecture
of
the
middleware,
which
is
actually
the
innovation
that
this
project
brought
about.
Essentially,
on
top
of
the
EM
orchestration,
which
binds
the
different
testbed
sites,
we
each
running
and
OpenStack
women
we
had
so
far
middleware
running,
which
essentially
is
responsible
for
interpreting
an
experimentation
package
and
and
then
making
sure
that
the
experimenters
intended
test
bits
are
involved
and
their
intended
operations
are
involved,
such
as.
N
If
the
experimenter
is
particularly
interested
in
programmability
or
using
so
far,
Defined
Networking,
then
the
experiment
manager
would
actually
involve
its
specialized
manager,
which
is
an
SDR
manager.
Underneath
you
can
see
five
managers
here,
Sdn
manager,
monitoring
manager,
any
fee
manager,
security
manager
and
physical
device
manager.
The
management
structure
is
essentially
required
modular,
so
you
can
actually
bring
about
your
special
managers
structure
if
you
want
to,
and
the
project
also
why
it's
open
source
software
provides
how
to
extend
this
modular
middleware
by
defining
and
building
your
own
managing
software.
N
Underneath
you
would
see
open
button,
which
is
an
Etsy
manual,
compliant
Orchestrator,
and
it
also
has
seen
a
lot
of
developments
in
the
past
years
and
it's
proven
to
be
very
powerful
or
Orchestrator
now,
and
it's
fully
integrated
with
the
different
test
base
in
this
project.
For
more
information,
you
can
see
our
whitepaper
publication
on
our
websites,
which
is
on
softwares,
approached
experimentation.
Why
and
how.
N
The
basic
view
of
an
experimenter
of
the
users
of
the
software
middleware
is
a
dashboard
and
on
the
dashboard,
so
far,
user
or
experimenter
would
see
a
number
of
resources
that
are
readily
available
for
them
to
use.
The
experiment
essentially
prepares
a
cloud
archive
file
which
contains
their
experiment
definition
yamo
format,
along
with
an
another
archive
file
which
includes
their
network
service
descriptors.
That
always
includes
there
be
an
original
function
descriptors
as
well
as
any
scripts
that
require
orchestration
lifecycle
events,
and
apart
from
that,
there
are
also
some
metadata
information
about
the
experiments
overall.
N
The
experiment,
access
and
support
essentially
was
a
big
big
part
of
the
project,
because
each
experimenter
comes
with
their
own
requirements,
which
are
not
always
feasible
with
the
platform
you
would
want
to
you
make
available.
Sometimes
you
need
to
make
compromises.
Sometimes
you
need
to
make
some
special
cases
for
them
and
provides
some
flexibility,
but
other
times
you
would
have
to
say
no,
because
that
would
explain
the
idea
of
shared
resources
and
a
shared
platform.
N
Just
like
the
fashion
empire
project
presented.
We
had
open
calls,
and
during
this
open
course
there
were
multiple
experiment
is
running
on
the
platform
to
provide
them,
even
here
with
sufficient
support
and
timely
support.
We
had
specialized
as
dedicated
slack
channels
for
each
experimenter,
as
well
as
shared
slack
channels.
Among
all
open
call
participants.
N
N
The
past
experimentation
waves
on
so
far
hit
a
lot
of
topics.
This
required
a
lot
of
back
back-end
support
for
various
topics,
so
you
would
need
to
really
work
with
your
experimental
experimenters
24/7
almost
because
they
would
need
to
reach
our
platform
whenever
they
want
to
in
the
with
net
remount
process.
So
you
need
to
reply
to
them
in
time.
Otherwise
they
can
be
off
the
track,
but
the
good
thing
is
supporting
this.
N
N
One
of
the
partner
use
cases.
Actually
this
was
performed.
In
fact,
I
see
facilities
was
network
slicing
in
the
factory
for
basically
fasciitis
center.
They
will
not
be
a
number
of
software
pieces
following
3gpp
standards,
essentially
for
a
control,
plane,
use
of
separation
and
dedicated
user
plain
slices,
two
different
experimenters,
so
that
they
would
actually
be
assigned
with
a
certain
user
plane
size,
effectively
providing
isolation
of
a
network
or
services
to
them
and
context
awareness
as
well.
N
So
the
project
would
like
to
share
its
experience
with
other
prior
projects
and
other
funky
themed
projects,
especially
PCM
community.
Please
check
out
our
open
source
middlebury
software,
which
we
are
now
revising
and
we
will
be
upgrading
soon
and
we
would
like
to
really
contribute
to
standardization
bodies,
especially
IETF,
which
is
why
I'm
here
today
we
have
a
lot
of
white
papers
and
which
some
of
them
are
upcoming
and
also
the
technical
deliverables.
A.
N
B
O
N
Is
really
very
important
topic
actually,
especially
in
terms
of
the
mobile
core
network.
We
performed
a
couple
of
benchmarking
measurements
about
the
effect
of
virtualization
on
the
performance
user
playing
performance
between
an
entrant.
Essentially,
it's
been
an
actual
mobile
device,
communicating
good
through
the
core
network
and
to
intranet
and
Internet
endpoints.
N
N
It's
a
performance
measurement
of
actual
orchestrated
deployment
of
mobile
core
network
deployments,
in
which
case
the
deployment
was
performed
remotely
from
Germany,
and
so
the
experimenter
is
anywhere
in
the
world
uses
the
dashboard
to
essentially
use
deploy
user
plane,
functionalities
of
a
mobile
call
with
the
orchestrated
look
at
North
Germany
in
our
UK
facility,
and
we
have
demonstrated
the
and
to
end
functionality
between
different
Network
slices
over
the
federated
test
beds
during
many
demonstration
events.
So
the
benchmarking
results
were
essentially
for
mobile
core
networks
and
so.
F
P
N
A
P
D
K
N
I
I
A
I
We're
starting
here
a
quick
reminder
of
something
that
I
showed
you
in
the
first
NF
ERG
meeting
long
time
ago,
I
differentiate
between
a
regular
communication
service
like
we've
known
in
the
past,
which
simply
says
take
this
number
of
bits
move
it
from
here
to
here,
maybe
there's
a
delay
constraint.
Maybe
you
need
at
least
so-and-so
bandwidth
that
was
pretty
much
it
and
we,
like
all
our
rich
communication
service,
which
is
a
combination
of
computation
and
communications.
I
I
I
The
thing
that
people
really
are
interested
in
from
a
communication
service
and
a
rich
communication
service
is
the
quality
of
experience.
Quality
of
experience
is
a
subjective
metric
and
is
only
an
end-to-end
metric
and
it's
recognized
by
the
user.
Now
what
we
measure
in
general
is
what's
called
quality
of
service,
QoS,
kpi's,
QoS
parameters,
QoS
parameter
are
simply
things
which
are
easy
to
measure.
They
don't
directly
mean
anything
to
the
end-user.
I
However,
the
reason
they
are
relevant,
the
reason
people
measure
packet
loss,
delay,
packet,
delay,
variation,
etc,
is
because
there
is
this
belief
that,
given
the
application,
there
is
a
relationship
between
a
bunch
of
QoS
parameters
and
the
quality
of
experience.
As
the
user
sees
this
now,
we
have
actually,
over
the
last
few
years,
discovered
actual
equations.
We
have
the
mathematical
relations
for
a
bunch
of
different
applications,
for
instance
for
conversational
voice
voice
over
IP.
You
tell
me
what
the
codec
pipe
is.
I
You
tell
me
what
the
packet
loss
is
and
the
fat
loss
distribution
tell
me
the
delay
and
I
can
give
you
something
which
correlates
very
well
with
the
mean
opinion
score.
If
the
same
for
streaming,
video
the
same
for
web
browsing,
there's
a
whole
bunch
of
things
which
have
been
done.
These
equations
are
known,
so
they
reinforce
our
feeling
that
QoS
is
relevant.
We
measure
QoS
because
it's
an
easy
thing
to
measure,
but
it
also
is
relevant
to
the
real
thing
that
we're
interested
in
which
is
quality
of
experience.
I
Unfortunately,
all
of
this
information
that
we
have
all
of
the
research
up
to
now
has
been
on
what
I
called
at
the
beginning,
regular
communication
services.
This
has
not
been
looked
into
again
for
rich
communication
services
and
unfortunately,
when
we
started
looking
at
it,
we
found
out
something
very
interesting.
I
We
have
found
out
that
the
quality
of
experience
can
be
proved
not
to
be
in
general,
of
course,
a
simple
function
of
the
application
and
quality
of
service
parameters,
and
that's
what
I
want
to
show
you
here.
I
want
to
show
you
why
QoS
can
be
completely
meaningless
when
I
say
to
us
I
mean
the
typical
parameters
that
one
measures
for
communication
services
like
packet
loss
and
continuity
and
packet
delay
and
things
of
that
sort.
All
that
is
completely
useless
to
predict
quality
of
experience
for
rich
communication
services.
I
That's
what
I
want
to
prove
to
you.
Okay,
this
is
going
to
do
I'll,
do
a
quick
overview
of
some
of
the
proofs
that
we've
done
now.
How
do
you
prove
these?
The
way
we're
going
to
prove
that
it's
irrelevant
is
by
doing
what
we
call
in
physics,
thought
experiments
or
get
dunked
in
experiments.
Einstein
love
that
term.
I
In
any
case,
what
we're
gonna
do
is
we're
gonna,
take
a
KPI
like
packet,
loss
or
delay,
and
then
we're
going
to
show
that
there
is
some
kind
of
network
function
which
is
in
the
path
of
a
rich
communication
service
for
which
that
KPI
is
your
meaningless
or
even
counterintuitive.
I'm
gonna
show
you
the
last
example
that
link
down
can
improve
quality
of
experience,
whether
it's
complete
lack
of
continuity
is
actually
good
for
you.
I
I
For
three,
which
is
15
minutes
from
now
number
one:
let's
talk
about
an
intrusion
protection
system.
What's
an
intrusion
protection
system,
it's
a
fancy
firewall!
Basically,
what
it
does
is
it
takes
packets
and
discourage
them,
and
this
can
be
someplace
deep
in
the
network.
It
could
be
at
the
edge.
It
can
be
different
places
now,
when
it
discards
a
packet.
If
I
measure
packet
loss,
I'm
going
to
see
a
positive
packet
loss,
we
are
used
to
packet
loss
being
bad,
in
other
words
the
higher
the
packet
loss.
I
We
think
the
lower
the
quality
of
experience
is
going
to
be,
but
if
you
think
about
it,
if
I
get
rid
of
a
bad
packet
which
could
insect
your
computer
and
cause
you
to
have
loss
of
experience,
you're
gonna
be
really
really
sad
about
getting
that
packet,
then
throwing
out
that
packet
increased
your
quality
of
experience.
So
what
you're
happy
about
there
being
packet
loss
as
long
as
it
those
packets
I
got
rid
of
okay,
so
here
higher
pocket
loss
means
better
quality
of
experience,
but
that's
a
really
tricky
case.
I
Let's
talk
about
something
which
is
more
convincing
than
now.
Let's
talk
about
the
following
function:
a
TCP
proxy
by
CCP
proxy
I
mean
something
we
put
in
some
place
along
the
indications
path,
which
resegmented
terminates
the
tcp
and
does
it
in
the
other
side
again,
and
the
reason
this
is
done
is
to
Excel
to
accelerate
the
communications.
It
makes
the
delay
of
the
tcp
shorter,
which
means
you
can
get
more
bandwidth.
However
it
resegmented
that
mean.
I
It
means
that
you
can
have
a
TCP
proxy
3
IP
packets
come
in
and
only
to
go
out
because
there's
segmentation
is
done
again
or
three
can
come
in
four
can
go
out
depending
on
the
segmentation
that
TCP
does
now.
If
I
have
three
packets
going
in
and
two
coming
out,
my
PAC
account
how
many
packets
I
received
IC
packet
loss,
but
all
the
bytes
get
there.
There's
no
change
to
core
the
experience
actually
through
better
quality
of
experience,
because
I
really
reduce
the
delay
so
I've
high
packet
loss.
I
Now
look
at
this
three
can
come
in
and
four
go
out:
I
have
negative
packet
loss
and
the
quality
of
experience
is
better
negative
packet
loss.
This
is
ridiculous:
okay,
okay,
okay,
you're
saying:
don't
count
packets,
count
the
bytes
you're
doing
a
trick
on
us
you're
trying
to
pull
one
over
count:
the
bytes
okay.
So
let's
look
at
the
volume
traffic
volume,
the
number
of
bytes,
what
happened.
I
It's
it's
projector
loss
is
not
just
packet
loss.
Okay,
let's
talk
about
the
following
function,
ran
optimization,
meaning
compression
where
I
say
compression
I
mean
he's
a
lossless
data
compression
or
data
D
dupe
or
video
or
audio
compression
or
transcoding,
or
something
of
that
sort.
All
of
these
mechanisms
take
a
certain
number
of
bytes
and
output,
a
lower
number
of
bytes,
but
the
same
information
is
either
heard
or
seen.
So
the
quality
experience
is
the
same
actually
once
again
after
compressing
before
the
compression
I
might
have
had
too
much.
I
Data
and
I
would
have
a
packet
loss
because
of
it
I
compressed
it
down.
Everything
goes
flies
through
I
get
this
really
good
experience
and
there
are
fewer
bytes
received
than
word
cent.
So
traffic
volume
is
once
again,
counterintuitive
I
lose
the
number
of
bytes
I
sent
and
everything
is
better
than
it
was
before.
So
all
the
people
here
who
have
taken
basic
information
theory
courses
are
saying:
ok,
bytes
is
not
right.
We
have
to
look
at
the
Shannon
information
in
those
bytes,
not
count
the
bytes
right.
I
Well,
okay,
let's
look
at
Shannon
information,
but
look
at
the
following
function.
When
optimization
once
again,
but
a
caching
server,
I
see
the
end
right
now,
if
you
think
about
it,
if
I'm
being
said
from
a
CDN
cache,
someplace
and
I
measure
packet
loss,
a
very
high
packet
loss,
even
lack
of
continuity
on
the
path
from
the
original
back-end
server
to
this
to
the
cache
I
could
have
packet
loss
of
90%
and
no
one
would
know
it.
Okay
to
the
continuity
and
the
video
comes
through
fine
and
everyone's
happy,
so
packet
loss
doesn't
mean
anything.
I
Ok
once
again,
you're
saying
that's
a
that's
a
tricky
one,
because
many
of
you
who
are
doing
om
and
things
of
that
sort,
you're
looking
at
tricky
cases,
that's
not
what
you
should
do.
The
way
you
measure
packet
loss
is
by
sending
synthetic
packets
and
those
synthetic
packets
are
not
literally
trampled
on
by
compression
and
no
one's
gonna,
gobble
them
up
and
they're,
not
dead
malware.
So
no
one's
gonna
stop
them
and
that's
the
way
you
try.
You
should
measure
right.
Well,
that's
true!
I'm
not
say
you
can't
measure
packet
loss.
I
I'm
just
saying
synthetic
packets
are
completely
useless
for
measuring
quality
of
experience
because
they
are
not
say
cheering
with
the
real
packets.
The
only
thing
that
affects
quality
of
experience
is
the
real
packets
or
something
that
sake
shares
with
them
and
gets
the
right
parameter.
So
you
can
measure
these
things
sure
you
can,
but
they
don't
mean
anything.
So
here's
a
really
good
exit.
Oh
wait!
A
second
I'm
not
getting
that
I
have
a
few
before
I
get
to
the
really
really
good
one.
That's
gonna
floor!
You!
Let's
talk
about
delay.
I
I
So
depending
what's
in
the
data
packet,
I
can
receive
something
really
fast,
but
it
might
take
a
long
time
until
it
comes
up
on
my
screen
and
the
quality
of
experience
is
based
on
how
long
until
I
see
the
screen
or
until
it's
stabilized
app
bits
for
those
who
you
know
of
it's
a
it
should
be
about
six
to
eight
seconds
a
lot
more
than
eight
seconds
until
my
screen.
Stabilizes
bad
I
can
get
the
packet
if
there
a
second
but
it
didn't
stabilize,
because
the
JavaScript
is
really
working
on
my
CPU.
I
So
therefore,
the
network
delay
is
now
it's
important
in
this
case.
I
have
to
know
what
the
content
of
the
packet
is
and
the
web
browser
is
once
again
it
kind
of
network
function.
It's
just
one
at
the
edge
okay.
That
might
not
convince
you.
Let
me
give
you
the
real
clincher
think
of
the
following
diagram.
I
I
The
end-user
has
a
better
quality
of
experience,
in
other
words,
because
server
one,
which
was
a
long
original
path,
was
either
not
upgraded
or
was
undergoing
a
really
bad
time,
because
a
lot
of
people
were
hitting
it
and
I
went
to
server
2.
Instead,
my
quality
of
experience
increased
improved.
What
caused
it
linked
down
now?
We're
used
to
think
thou,
meaning
my
service
went
away,
I'm
really
unhappy
and
here
think
down
improved
the
quality
of
experience.
I
So
all
of
these
cases
were
all
to
show
you
that
all
these
things
that
we've
taken
for
granted
all
these
years,
that
you
can
look
at
packet
loss,
you
can
look
at
delay
and
you
can
look
at
the
even
link
down
link
up
and
based
on
that
predict,
quality
and
experience.
That's
not
true!
Now.
Actually,
all
of
this
is
even
worse.
You're.
Up
to
now,
I
was
just
talking
with
the
function
that
could
have
been
in
the
appliance.
I
What
happens
for
with
any
of
e
now
or
specifically,
DNF
e
distributed
message
as
long
as
I
had
fixed
appliances,
physical
around
the
place,
you
could
say:
ok,
I
know
that
function
is
there
and
I
can
take
it
into
account
where
I'm
doing
my
calculation,
but
once
I'm
doing
any
of
e.
These
functions
are
moving
around.
They
can
be
there
today
and
not
be
there
tomorrow,
I
don't
know
what's
there,
because
some
Orchestrator
decides
the
best
place
to
put
that
function.
Right
now
is
right
here
and
moment
it
all
with
someplace
else.
I
I,
don't
know
what's
happening
so
I
if
I
measure
these
cues
parameters,
there's
no
way
I
can
figure
out.
What's
going
on,
it's
meaningless.
I
cannot
predict
qoe
in
this
new
world.
So
what
do
you
have
to
do
in
order
to
be
able
to
measure
qoe?
Well,
one
way
is
put
your
back
up
against
the
wall
and
say
Q
OES
end
to
end
its
measured
by
the
end
point.
Let
the
endpoint
measure
it,
in
other
words,
capture.
Only
estimation
like
is
done.
I
Let's
aim
for
IP,
where
someone
is
measuring
the
voice
coming
in
or
you
can
indirectly
estimate
qoe
based
on
the
user's
behavior.
If
a
user
is
using
a
thing-
and
he
says-
and
you
see
all
of
a
sudden,
you
he's
resetting
his
TCP
connection.
That
was
he
saying
this
is
I'm
going
away.
Do
statistics
on
that?
You
can
say
something
you
can
measure
QE
of
synthetic
pockets,
but
they
have
to
be
state
cheering
after
crack
them
in
a
certain
way.
So
they
don't
mean
something
you
can
perhaps
correlate.
Qos
KPIs
with
other
kinds
of
KPIs.
I
For
instance,
look
at
those
and
if
VI
elements
along
the
way,
see
how
the
CPU
is
going
see
how
much
memory
you're
using.
Maybe
in
that
case
you
could
do
something
like
if
you
knew
what
the
function
is.
If
you
don't
know
what
the
function
is,
maybe
machine
learning
can
figure
out
if
I
give
you
a
lot
of
information
about
what
kind
of
CPOE
I
have
how
much
CPU
is
going
on,
etc?
Maybe
you
could
figure
it
out,
I'm,
not
sure,
but
something
that
could
be
researched.
I
I
F
F
I
F
Q
Right,
so
you
know
we're
going
to
packet
loss,
mat,
metrics,
well,
I,
think
Brutus,
I,
think
part
of
the
problem.
There
they're
just
missing
a
bit
of
information
in
the
equation
when
you
say
that
we're
having
problems
of
packets
being
added
to
the
stream
or
remove
from
the
stream.
If
you
were
to
just
take
those
into
account
in
the
quality
metric,
one
of
the
problem
go
away
once.
I
Q
O
O
I
I
think
that
the
fundamental
theme
is
is
measuring,
separate
things
separately
and
that
even
comes
down
to
your
endpoint
argument.
You
measure
the
network
side,
you
measure
the
player
buffer
side
and
you
measure
the
rendering
side
right
and
once
you
know
all
of
those,
then
you
know
sort
of
the
end
to
end
thing.
Well,.
I
That's
what
I
say
here,
however:
no
one
knows
an
explicit
equation
right
now
to
do
that.
I
also
feel
that
if
you
put
more
information
in
here,
then
maybe
you
can
get
to
something,
but
all
the
work
that
people
have
been
doing,
including
myself
over
the
years
of
how
to
make
these
explicit
equations
none
of
that
works.
So
it's
opening
up
a
new
field,
I
think.