►
From YouTube: IETF92-AQM-20150324-1730
Description
AQM meeting session at IETF92
2015/03/24 1730
A
B
A
B
B
Goddess,
so
there
is
going
to
be
a
talk
on
the
evaluation
guidelines.
/
characterization
guidelines
is
a
case.
Maybe
I
can
benefits
and
quick
Chicago
status,
so
the
recommendations
graph
draft
is
the
one
thing
that
we've
been
pretty
close
to
finishing
up
now
over
the
iesg
ballots
are
green
and
I.
Believe
Martin
is
working
with
the
general
with
Lee
Davies
right
on
making
sure
his
comments
are
all
10
kills.
B
So
we
have
a
bunch
of
any
working
group
items
now
and
a
lot
of
them
seem
to
be
ready
for
a
last
call
and
things
that
I
think
we
want
to
do
today.
We
will
just
trying
to
get
a
sense
of
how
you
should
do
this
son
of
glass
cause.
We
got
the
five
things
cute
up
there
and
I
think
the
natural
ability
to
do
them
in
would
certainly
not
sighs
in
parallel
by.
B
The
time
it'd
be
like
the
Next
Iron
Chef
meeting
before
the
day
with
last
call's,
so
we
have
say
overlap
in
the
last
calls
between
one
or
two
documents
at
a
time,
but
I
think
the
natural
way
that
they
would
be
the
evaluation
guidelines.
The
ekend
benefits,
probably
fq
implementation
in
the
two
algorithms
and
thicken
to
last
calls
on
each
disease.
B
D
C
B
Assuming
that
two
weeks
is,
you
know
the
time
it
takes
to
collect
up
all
the
comments
and
close
well.
D
B
B
B
A
D
B
A
A
A
B
D
D
B
B
E
Okay,
so
this
is
about
an
update
on
the
Aiken
characterization
guidelines
draft.
Next
we
have
had
some
feedback
at
the
ITF
in
Honolulu
and
also
we've
got
under
version.
0
we've
got
some
feedback
and
full
review
from
Alfred,
then
some
comments
on
the
version
1
from
Wolfram,
and
then
we
submitted
the
version
to
just
the
9th
march.
E
Two
revisions
are
submitted
one
and
two
based
on
this
feedbacks
that
we've
brought
in
the
idea,
so
they're,
mainly
editorial
work
and
minor
fixes.
But
the
main
thing
is
that
that
we
provided
more
clarity
in
the
definition
of
the
end-to-end
metrics
that
are
being
defined
in
the
draft,
so
they
are
right.
Now
they
come
up
with
a
definition.
What
what
is
good
good
put,
what
is
loss,
what
is
Jeter
so
and
they
they
are
citing
other
rfcs,
and
also
that
we
have
more
clarity
on
how
to
report
the
data
related
to
the
end-to-end
metric.
E
E
So
so,
if,
if
you
have
a
hardware
that
that
gives
you
the
you
know,
the
queue
size
and
you
can
get
it
from
the
hardware
and
if
its
impact
on
the
aqm
mechanism
itself
is
negligible,
then
drag
the
draft
encourages
of
collecting
those
data
as
or
number
of
easy
n
marks
or
whatever's
in
the
middle
boxes.
So
whenever,
whenever
it's
possible,
then
it
kind
of
has
the
new
draft
it
kind
of
has
this
mindset
of
real-life
test
testing
things
on
a
real
system?
E
So,
with
regards
to
the
metrics
and
and
their
data
measurements,
we
defined
a
set
of
metrics
that
needs
that
their
data
needs
to
be
collected.
So
previously
it
was.
The
draft
was
saying
that
you
actually
can
choose
what
metrics
you
want
to
report
on
so
right
now.
The
way
it's
written
is
that
it
actually
says
that
you
should
actually
report
the
maximum
number
of
metrics,
that
you
can
record
them
and
they
are
relevant
to
the
context
of
the
evaluation
and
to
the
scenario
that
you
evaluating.
E
E
It's
good
to
gather
the
data
related
to
the
Q
level
metrics,
but
the
main
the
layout
of
the
evaluations
should
be
based
on
the
gathering
data
from
the
end
to
end
point
of
view,
and
also
based
on
the
RFC's,
how
to
measure
the
packet
loss
in
the
system.
So
we've
got
two
metals
on
how
to
measure
the
packet
loss,
whether
you
assume
that
the
packet
would
arrive
within
a
certain
delay
bound
based
on
the
RFC
2680,
and
if
it
doesn't,
then
you
assume
that
it's
lost
or
based
on
the
RFC
2544.
E
You
actually
can't
all
the
packets
sent.
Then
you
can't
all
the
non
duplicated
packets
received,
and
then
you
measure
what
was
lost.
Then
the
definitions
of
good
put
one
day,
one
way,
delay
and
jitter.
They
already
embedded
into
the
text,
and
we
came
up
with
recommendations
on
how
so
that
does
sort
of
time
intervals.
If
you're
going
to
report
good
food
over
time,
how
are
you
going
to
do
it?
Different
people
may
measure
good
food
/
different
time
intervals,
so
here
it
provides
certain
recommendations,
for
example
every
multiple
RTT.
E
A
E
E
Where
you
go
the
asymmetric
link
scenarios,
it
should
investigate
the
bi-directional
traffic
first,
it
looks
at
a
scenario
when
you
have
the
viaduct
by
bi-directional
traffic,
when
you
have
an
acre
on
only
one
direction.
You
may
also
want
to
look
at
a
more
complex
scenario
where
two
instances
of
a
QM
on
both
directions,
and
then
you
want
to
also
look
at
the
impact,
as
the
TCP
Act
drops
by
a
QM.
E
A
very
more
critical
part
of
that
is
to
how
to
define
different
condition,
levels
and
rule.
In
the
last
scientific
provided
a
Lincoln
and
a
paper.
We
try
to
read
it.
It
was
highly
complex
mathematically
and
it
actually
was
very
difficult
to
translate
it
into
something
that
is
quantifiable
under
every
link.
So
we
came
up
with
so
previously
it
was
uses.
It
was
using
a
certain
packet
loss
ratio
assuming
that,
if
you
have
had
a
drop
tail
queue,
so
it
would
have
translated
to
certain
level
of
congestion.
E
What
we
want
to
test
is
that
is,
it
is
able
to
bring
down
the
condition
on
the
link.
So
previously
was
really
actually
very
high
percentage
ratios
for
defining
what
is
mild
with
what
is
medium
and
what
is
what
is
heavy
congestion
sure
we
would
reduce
it
10
times
less
so
to
make
it
more
realistic.
In
the
version
number
two
and
next
week,
and
at
last.
F
Aggregates
safe
that
paper,
I
referred
to,
you
was
a
bit
too
heavy.
The
packet
loss
levels
can
I
suggest
using
us.
You've
reduced
them
to
more
realistic
levels,
but
maybe
a
slightly
wider
spread
of
them
would
be
with
wire.
Like
you've
got
point
1.51
to
see,
maybe
going
to
you
notice
a
point
Oh
5.1
1.5%,
or
something
like
that.
The
kiss.
E
If
you
have,
the
link
that
is
like
a
100
millisecond
base
are
two
key
and
10
mb.
You
know
that's
very
classic
scenario
and
a
PDP
of
offering
then
you're
gonna
have
like
around
18
19
flows
for
a
heavy
congestion
which,
which
makes
sense
actually
and
for
a
mild
congestion.
You
would
have
around
for
flows
and
that
is
actually
compatible
to
what
we
have
in
mind,
for
that
is
specific
link
scenario,
but
with
the
previous
selling
that
was
like
higher
have
like
19
flows
were
actually
a
mild
condition
which
would
never.
C
E
E
B
So
I
one
comment,
yeah
I
didn't
think
about
it.
I
saw
you
last
night
that
we
should
be
careful
that
this
is
the
guidelines
and
recommendations
on
evaluating
the
algorithms
or
not
on
designing
the
algorithms
right.
So
the
aqm
should
support
ecn.
I,
don't
know.
If
that's
what
you
want
to
say
that
we
want
to
say
that
you
know
the
tests
and
evaluations
of
a
qm
should
enable
ACN
and
provide
the
results.
B
H
Bilson
sank
like
I,
just
apologize,
I'm
not
entirely
after
hadn't
reviewed
the
whole
thing
I
was
just
looking
through
it
to
find
out
like
how
many
flows
you'd
be
looking
at
I
know
it
like
home
gateway
was
mentioned,
then
17
flows
was
considered
a
large
number,
but
you
know
what
about
evaluating
this
like
a
core
router
or
somewhere,
where
we're
seeing
100,000
flows
or
is
that
is
their
scope
of
this?
Is
it
is
targeted
to
a
specific
domain?
That
is
this:
whatever
flows
should.
E
H
E
That
wasn't
about
algorithms
but
more
about
different
condition.
Levels,
and
you
know
what
is
the
condition
level
is
kind
of
fuzzy
definition,
so
it
could
translate
it
into
good
things.
We
have
to
kind
of
come
up
with
a
way
or
how,
to
you
know,
produce
rather
like
a
fuzzy
kind
of
mild,
medium
and
heavy
condition
levels
so
I
mean
I,
don't
know
how
bad
this
can
go
away,
but
I
mean
if
you
have
a
reference
or
anything
that
help.
We
are
very
happy
to
incorporate
that
so.
I
Really
go
well
in
my
head,
so
why
don't
you
just
remove
it?
Why
don't
you
just
say
like
we
have
heavy
load,
we
use
a
large
number
of
flows
and
you
just
make
up
a
number
because
it
actually
it's
not
that
important,
which
number
we
have
it's
just
a
large
number
and
then
we
have
a
medium
number.
We
have
a
small
number
yeah.
I
J
Michael
Abramson
so
I've
seen
some
specs.
C
B
A
K
K
E
J
So
hi
I'm,
Alan,
Morton
and
I've
got
my
benchmarking
methodology
co-chairs
had
on
at
the
moment.
I
tried
there
to
stimulate
some
review
from
the
other
benchmarking
with
folks,
by
sharing
my
review
with
them,
and
so
far
it
hasn't
been
successful.
But
there's
a
I
would
say:
there's
as
much
overlap
between
this
work
and
I
ppm
is.
There
is
with
benchmarking
methodology.
So
maybe
we
can
try
one
more
time
to
get
some
folks
from
BMW
G
to
this
as
well
up
there
to
comment.
J
I
would
make
it,
which
is
on
the
previous
slide,
if
you
wouldn't
mind,
just
put
them
back
their
West
Briton.
Thank
you.
Yes,
yes,
traffic
mix,
so
at
least
for
the
for
the
UDP
stateless
stuff.
We
have
an
RF
seeing
the
number
escapes
me,
but
it's
the
I
mix
genome
which
can
help
here
this.
This
provides
repeatability.
If
you
use
different
mixes,
we
think
it's
at
least
it's
our
best
shot
of
a
beautiful.
J
I
Me
again
and
I
just
want
to
explain
further
what
I
meant,
sir.
First
of
all
what
I
like
about
this
document,
it
that's
very
clear
and
very
focused
and
I-
think
it's
really
useful
for
people
who
want
to
do
something
so
I
also
regarding
to
the
IP
p.m.
stuff.
It
should
be
in
line
with
IVA
p.m.
but
it
shouldn't
make
this
couple
this
document
more
complicated.
I
It
should
stay
as
it
is.
So
that's
that's
a
nice
period,
and
so
that
was
also
why
I
made
this
proposal,
because
the
congestion
level
is
the
term
itself
is
very
confusing
to
me
because
you
calculated
based
on
drop
ship
and
then
we
applied
the
menu.
You
get
a
number
of
flows.
You
apply
this
number
flows
using
a
different
ATM
and
get
a
different
loss
right
out
of
it,
which
other
people
will
again
see
as
a
congestion
lever,
and
so
that's
just
it
just
makes
I
understand
what
it
is.
I
But
it's
really
it's
a
weird
term
to
me
and
I
also
don't
think
that
is
needed.
So,
of
course,
I'm
using
over
different
vendors
today
products
using
the
same
number
of
flows,
relieved
to
what
you
call
different
projections,
efforts,
because
all
the
thing
it
doesn't
matter
to
have
always
the
same
tradition
ever
it's
okay,
if
you
just
test
a
small
number
of
laws
and
be
remember,
flows
and
a
maximum
Bo
feels.
E
I
Records
so
you
have
100
milliseconds
and
to
end
today
10m
bit
link
and
then
you
calculate
your
numbers
of
floats
and
that's
in
this
case
10.
Now
you
take
the
same
and
you
have
different
scenario.
You
have
a
similar
bandwidth,
so
the
number
you
broke
I
could
a
disguising,
but
like
half
half
the
bandwidth,
so
it's
five
instead
of
10.
Yes,.
A
I
Is
important
is
that
you
compare
the
algorithms
in
these
scenarios,
like
you
have
10
megabits
per
second
half
your
10
mega,
a
second
hundred
meetings
and
10
flows,
and
then
you
can
play
all
the
apms
and
then
the
next
scenario.
It
doesn't
actually
matter
if
you
have
five
or
ten
flows,
but
you
have
to
have
the
same
number
of
flows
and
that's
why
I
just
say
pick
a
number
which
is
okay
for
all
sinners.
Just.
E
E
I
Actually
want
and
the
end
is
to
cover
in
the
use
cases.
So
if
you
even
so,
you
have
a
large
link
and
you
only
have
two
flows
for
like
or
you
have
that
Smalling
and
you
have
hundred
flows
and
that's
very
heavy
congestion.
Let's
review
q's
case,
if
you
really
can
have
a
link
where
you
have
found
it
flows,
then
you
should
test
it.
I,
don't
care
about
the
congestion,
never
because
the
congestion
has
influenced
by
the
ATM
I
care
about
the
scenario
yet
and
I
want.
I
F
You
can't
actually
make
it
at
some
kind
of
comparison
based
on
what
happened
with
this
condition
label
my
metric,
because
it's
a
relative
measurement
within
the
within
the
new
apprentice
of
a
scenario,
but
the
number
of
flows
you
require
for
those
two
scenarios
is
radically
different
right.
Yeah
for
the
team
taylor
scenario
I
mean
you
know,
I
know
so.
F
I
C
When
I
was
whether
at
a
serious
problem,
is
you
go
to
the
large
dimensionality
space,
you
are
simplified
and
I
will
suggest
that
being
one
of
the
primers
you
have
a
pic
is
the
na-
for
shear
rate.
That
is
the
aggregate
bandwidth
divided
by
the
number
of
rows.
I
guess
with
these
ways
said
it,
but,
but
but
that
is
the
commander
is
from
something
to
be
held
constant
because
that's
the
parameter
that
the
users
care
about
as
you
scale
the
system
all
right.
E
G
So
go
first
sup,
my
neighborhood
and
okay,
so
I
time
to
read
this
quickly
through
again
and
some
of
the
terminology
here
is
a
little
bit
different
for
it.
In
the
ATM
guidelines
recommendation
recommendations,
guy
who's
vcp
will
talk,
we
just
top-rated
beginning.
Is
it
worth
trying
to
align
this
and
not
talk
about
switches
but
talk
about
network
devices
and
to
use
the
same
definition
for
a
qm,
because
you
have
a
kind
of
more
fuzzy
definition,
and
there
are
things
in
here
that
we
could
make
changes
to?
How
can
we
make
comments?
B
G
G
Right
and
I
need
to
be
professor
at
the
very
beginning
and
Michael
gave
me
the
job
of
doing
the
last
editing,
uploading
and
I
managed
to
mess
up
the
upload
procedure
by
uploading
me.
A
new
version
of
the
individual
submission
round
in
the
working
group
1
and
nobody
told
me
any
different
until
after
the
deadline
had
passed
and
the
ID
tracker
will
now
be
changed
to
not
let
you
do
such
a
stupid
thing
because
of
the
people
don't
as
well.
G
This
hey
ed
should
spin
the
archive,
and
then
we
have
to
conquer.
Do
it
releases
to
to
make
sure
it
is
replaced
by
the
right
one
and
all
sorts
of
silly
stuff
and
finally
admitted
the
right
version
and
I
already
took
this
version,
which
is
version
wrong
to
respond
to
comments
provided
by
the
working
group
and
a
powerful
one.
So
when
you
talk
about
this
now
purpose,
though,
the
draft
is
to
present
a
vision
for
why
people
should
be
using
ECM.
G
If
you
think
this
is
the
wrong
popular
t,
you
think
which
will
do
something
different.
Then
you
need
to
speak
up
because
the
perps
this
document
is
to
preserve
our
vision.
Please
use
easier
if
I
as
the
key
benefits
and
it
could
potentially
identify
the
pitfalls.
That
was
the
second
comment
we
got
from
West
and
it
closest
to
think
again
because
talking
about
the
profiles
in
a
document
which
is
effectively
a
manifesto
for
saying
turn
on
ECM
seems
to
be
able
to
put
contradictory.
G
This
document
still
approaching
the
RFC
series
trying
to
say
gzzz
em,
neither
based
approach
book
describing
current
pitfalls,
appeared
to
be
maybe
not
the
right
language
version
2
that
tries
to
avoid
the
word
pitfalls
without
removing
those
sessions
have
talked
about
the
bad
things
that
can
go
on.
It
just
changes
the
language
now
challenges.
D
A
G
D
G
G
Potentially
confuse
people
is
I,
move
the
bleach,
your
little
box
requirements
earlier
in
the
document,
then,
because
I
put
it
with
deployment
issues,
because
I
thought
this
was
a
deployment
thing
that
you
should
tell
people
not
to
bleat
cheese
here.
So
I
do
put
it
at
the
end
where
we
talked
Brett
mechanisms.
Okay,
everything
else,
the
conclusions.
K
G
The
challenges
to
be
motivating
the
use
of
ECM.
This
is
a
statement
to
people
that
we
should
be
using
it.
Yet
it's
not
a
certain
problems
of
going
away.
The
ITF
11
working
on
it.
In
fact,
I
intend
to
write
drafts
about
TCM,
to
make
it
better
or
to
make
it
work
effectively,
which
I
think
is
a
little
bit
of
that
challenge,
but
Sir.
K
A
K
Because
after
I
tweeted
something
about
the
talk,
I
gave
an
ICC
orgy
I
got
called
out
in
Twitter
about
breaking
the
internet
for
telling
people
to
turn
it
on,
but
it
a
good
for
you,
TCB
anycast,
when
you're
bleaching
or
when
you're
copying
the
boss
fight.
So
many
things
are
wrong
with
not
that
I
can't
even
start.
D
K
So
the
point
is
the
point
that
I'm
making
is
that
if
we're
talking
about
okay
well,
we're
thinking
about
all
the
problems
and
repeating
people
at
this
document,
there
are
people
in
the
community
who
will
say:
ECM
is
broken
because
queasy
feeling
that
I
want
to
do
what
I
saw
once
or
I
have
some
broken
equipment
or
I
had
a
bad
idea
and
I'm
continuing
to
do
the
bad
idea.
I,
don't
know
that
we
want
to
put
that
in
there,
but
if
this
is
supposed
to
be
the
one,
stop
shop
where's
the
line,
I.
G
F
Hendra
McGregor,
as
far
as
we
know,
we
haven't,
got,
had
a.
F
G
F
B
Sorry
I
got
your
perspective
on
that
and
let
you
know
that
this
isn't
supposed
to
be
spec
for
you.
Seeing
that
modifying
you
know
anything
about.
You
see
end
it's
about
explaining
the
people
arrive,
they
might
want
to
use
it,
so
it
would
be
the
wrong
place.
Actually,
if
we
had
guidance
about
that
to
stuff
it,
but.
F
A
G
I
C
G
F
D
F
A
I
I
The
other
ways
that
you
used
to
complete
a
different
contraction
control
for
the
ecn
enabled
through,
and
that
needs
also
some
kind
of
agreement
sanitization
between
what
between
the
ATM
and
the
system,
because
both
need
to
know
that
they
have
to
do
something
different
and
I.
Don't
think
that's
something
we
have
right
now,
so
I'm,
not
sure
if
it
should
go
into
this
document,
I
mean
I
would
happy
people
think
about
it.
But
I'm
not
sure
if
go
on
this
w.
No.
J
So
it's
been
said
that
this
is
not
supposed
to
be
a
specification.
The
specification
exists
and
recommends
for
now
to
mark
rail
would
otherwise
drop
and
I
think
we
should
not
changed
out
our
proposed
to
change
that
in
despair
and
remove
any
text
that
says
anything
about
possibly
marketing
area.
That's
level
email
grant
will.
G
G
I
Maybe
okay
I
agree.
I
I
would
still
like
to
see
like
a
side
remark.
If
you
changed
this
spec
saying
not,
you
have
not
have
to
do
the
same
like
rub
anymore.
If
you
change
it
and
you
change
also
the
congestion
control,
you
can
have
additional
benefits
just
making
work
somewhere
that
you
know
there's
more
coming
up,
which
might
be
interested
in
something
that
that
would
be
I.
Think
that
would
be
nice,
because
if
people
read
this,
they
say:
okay,
that's
basically
all
stopping
you,
but
what's
the
real
benefit.
G
Okay
and
when
we
just
try
one
second
longer
poker,
and
is
it
a
new
benefit
or
is
it
a
greater
benefit
only
because
the
benefits
we
already
outlined
in
some
top-level
describing
the
different
ways
you
can
benefit
I
believe
what
you're
saying
actually
is
a
good
thing
and
I've
got
research
also
in
this
space,
but
it
actually
just
increases
the
benefit.
In
those
particular
cases.
I
The
the
of
the
whole
system,
if
you
have
a
different
aqm
and
the
different
congestion
control,
then
there's
a
new
benefit
that
might
be
doable,
because
today's
congenital
needs
a
certain
amount
of
buffering
trachea,
which
I've
utilization,
and
if
you
want
to
have
higher
education
and
no
delay,
you
might
need
to
change
those.
But
that's
like
this
I'm,
not
sure
that
might
not
belong
it.
I
G
G
G
A
F
No
show
slides
and
uses
micro,
sis
closer,
the
car
will
draft
we
have
at
this
point.
We
have
a
number
of
known
editorial
changes
that
need
to
be
made
a
whole
bunch
of
reference,
that's
and
a
few
other
such
things,
but
in
its
current
form,
its
kind
of
reading
for
review,
understanding
that
there
are
some
known
editorial
issues,
stick
particularly
references
and
when
those
are
resolved,
we
believe
it'll
we're
ready
for
a
score.
So
if
you
go
to
insert
a
link,
if
you
poignant
now,
that's
that's
perfectly
fine
and
it
won't
be
too
long.
F
A
C
J
Martha
lesson
I.
I
have.
I
have
some
reservations
about
the
behave
that
happens
when
it
enters
this
drop
mode.
The
part
of
the
square
root
I've,
seen
strange
things
and
I
can
share
them
off
run
and
we
can
have
a
discussion
about
it,
but
I
think
about
black
tab.
More
justification
about
this
part,
not
the
problem,
the
first
choice,
but
to
repeated
dropping.
F
The
idea,
the
idea
with
this
document
is
to
document
coral,
as
as
it
was
originally
published
in
his
here
out
there
in
the
wild,
rather
than
do
further
development
on
it.
Now
all
right
you
further
algorithm
development
is
a
good
thing,
but
and
we'd
certainly
like
to
know
about
that
and
probably
publish.