►
From YouTube: IETF99-TSVAREA-20170717-1740
Description
TSVAREA meeting session at IETF99
2017/07/17 1740
https://datatracker.ietf.org/meeting/99/proceedings/
A
Okay,
so
we
have
the
blue
sheets
going
around
welcome
to
TSV
area.
We
have
the
blue
sheets
going
around
and
David
like
is,
are
described
as
headlight
and
do
we
have
at
ever
slide,
or
do
we
need
one,
maybe
maybe
maybe
maybe
not
I
will
I
will
join
I
will
join
the
Medeco
and
watch
the
jeffer
from
up
here.
A
A
A
Because
here
you
guys
will
see
that
many
times
this
week,
but
this
is
another
one
we're
in
Monday
afternoon
in
session
3,
and
we
want
to
do
a
little
bit
on
TSP
overview
and
status
and
want
to
talk
about
a
proposed
research
group.
That's
probably
people
in
TSP
area
will
be
interested
in
at
least
some.
Then
we're
going
to
have
issue
Oh
Honda
talk
about
network
stacks,
integrating
with
envy
marketing
abstractions
and
then
we're
going
to
have
some
open
mic
time.
A
We
always
like
to
put
up
the
slide
for
who's
on
the
TSP
area
review
team
and
how
many
reviews
they've
done.
This
has
actually
been
a
fairly
light.
Ietf
cycle
for
reviews
and
actually
what's
coming
to
the
IHG,
has
been
kind
of
light
also,
but
they
say
we
wanted.
We
appreciate
every
one
of
them.
I
would
have
to
say
also
that
the
reviews
that
we've
been
getting
in
the
last
couple
of
ITF
cycles
have
had
a
lot
of
influence
on
hell.
Mary
and
I
fill
out
ballots.
So
we
really
appreciate
that
yeah.
A
I
ppm
is
doing
post
char,
recharter
discussions
that
would
cover
the
adoption
of
Instituto
am
capacity
for
is
doing
recharter
discussions
after
a
really
productive
couple
of
IETF
cycles.
Thank
you.
Thank
you.
Guys,
quick
is,
has
a
first
implementation
drat
that
the
hackathon
yesterday
ad
comment
as
well.
A
Taps
has
the
last
two
working
group,
graphs
or
past
working
group
last
call
so
moving
deep
into
implementation
and
API
space
I
have
no
say
I.
Catch
was
the
first
working
group
by
charter
and
I
have
never
been
as
happy
about
taps
as
I
am
now.
Thank
you.
All
tram
is
finishing
up
this
deliverable
final
deliverables
and
TSB
working
group.
When
you
got
14
active
working
group,
drafts
plus
17
active
related
drafts,
some
of
which
have
been
out
put
out
for
adoption,
calls
well.
B
So
I
will
only
mention
the
highlights,
and
one
highlight
is
that
ATM
is
basically
closed,
still
listed
here.
I
didn't
push
the
button,
but
all
documents
it
and
I'm
waiting
just
for
one
single
person
to
update
the
last
document,
then
so
TCP
Inc
is
has
just
submitted
their
main
documents
for
publication,
so
they
are
kind
of
done.
They
have
like
some
site
documents.
They
want
to
work
on
then
also
RM
ket
is
doing
really
good
progress.
B
B
A
Thank
you.
Our
documents
published
since
ITF
98
has
been
like
I,
say
a
little
bit
light,
but
that's
not
that.
That's
not
that's
not
unique
to
transport.
It's
just
in
kind
of
a
light.
Ietf
cycle
since
I
think
we
just
had
a
light
IETF
meeting
and
that
kind
of
stuff,
but
we
have
gotten
stuff
through
and
yeah.
A
Anyway,
but
again
think
you
know
they
say,
thank
you
all
for
the
work
you
do.
We
didn't
have
one
request
for
feedback
from
people
and
I
said:
I
would
throw
up
the
slide.
This
is
the
details
on
it,
but
there's
a
thread
already
started
on
the
TSP
area
list.
If
you
have
something
to
add,
please
follow
up
there
and
it's
it's
already.
You
know
the
people.
Look
like
they've
already
received
several
helpful
comments.
So
thank
you
for
that
anything
else
that
we
need
to.
B
A
B
Just
telling
you
so
before
this
meeting
there
was
a
banana
buff.
So
we're
not
reporting
on
that
and
we
don't
announce
it
because
it's
already
gone,
but
it
has
relation
to
you
and
transport.
So
maybe
you
should
look
at
it
if
you're
not
aware
of
it
and
then
the
other
thing
that
might
be
interesting
for
Transport
is
there's
proposed.
Research
group
called
passive
where
networking
and
Brian
is
sharing
that
and
has
yeah
Pichet.
C
D
C
C
The
thing
that
led
us
to
put
this
together
as
a
research
group
proposal,
was
we've
seen
a
whole
lot
of
work
recently
so
buffs
as
well
as
working
groups
both
in
the
internet
and
transport
areas,
where
there's
an
element
of
path,
awareness
where
the
endpoints
of
a
communication
actually
are
participating
more
in
sort
of
past
selection
or
path,
control
or
getting
more
information
about
the
past.
So
an
obvious
one
of
these
is
multipath
TCP.
C
There
are
other
there's
other
work
in
the
internet,
the
internet
area,
so
there's
spring.
So
if
we're
doing
source
routing
with
ipv6
segment
routing,
there
are
ahead
of
there's
a
giant
long
list
of
these
there's
a
lot
of
work
sort
of
in
in
Sdn,
which
gives
you
enabling
technologies
for
these
things.
So
we're
gonna
have
an
initial
proposed
work.
C
If
we
have
the
scope
of
this
right,
this
is
something
that
we're
looking
to
start
off
as
a
way
to
coordinate
work
on
a
path
awareness
at
these
two
layers:
research
on
path,
awareness,
these
two
layers
to
bring
that
into
the
IETF,
but
we've
taken
a
very
much
a
transport
focus
on
this,
because
I'm,
a
trans
person
and
an
internet
layer
focus
on
this
and
we're
looking
to
see
if
they're
sort
of
like
comments
on
how
that
should
be
scoped.
So
please
come
by
on
Wednesday
afternoon.
Thank
you
very
much.
A
B
E
E
E
F
E
E
J
All
right
so
good
afternoon,
everybody-
and
this
is
a
little
bit
about
the
off
topic
from
transport
layer,
because
this
is
not
particle
on
some
protocol.
But
given
that
these
days,
I
found
transport,
so
networking
API
for
transport
layer
is
quite
hot
topic
in
transport
area.
So
the
I
decided
make
presentation.
K
L
J
J
So
basically
the
main
memory
becomes
persistent,
a
lot
of
work
done,
for
example,
in
Terry
the
announcing
3d
crosspoint,
which
is
like
memory,
but
data
stays
there
all
the
time,
even
after
reboots
and
so
the
daily
lot
of
such
kind
of
normal
memory
technologies.
But
common
characteristics
is
they're
all
byte
addressable.
You
know
that
this
is
not
byte
addressable,
so
you
have
to
access
physical
disks
disks
by
the
brokenness
granularity,
but
the
persistent
memory
about
that
vessel,
and
also
they
are
very
low
latency
compared
to
disks.
J
So
they
only
access
latency
for
disks
is
basically
the
millisecond
or
hundreds
of
microseconds,
even
for
SSDs,
but
non
grata
in
main
memory.
Access
latency
become
tens
to
thousands
of
nanosecond.
That
is
a
very
very
long.
So
the
in
the
community
of
storage,
the
wrote
of
work
is
going
so
in
addition
to
research
paper
introduced
here,
they
were
lot
of
work
from
industry.
J
J
But
mean
when
you
have,
and
on
top
of
that,
you
have
operating
system
stack
like
memory
management,
Network
stack,
including
tcp/ip,
and
on
top
of
that
you
have
socket
API
which
resembles
pirate
yes
for
networking,
and
you
also
have
fire
systems
and
application
running
or
on
top
of
it,
and
when
we
plug
non-volatile
main
memory
into
the
system,
then
just
currently
the
system
looks
like
following.
This
is
how
Linux
and
windows
works
today.
J
So
once
you
plug
in
via
memory
device
into
main
memory,
then
the
you
can
create
fire
on
top
of
it,
because
you
can
create
fire
system
on
top
of
it,
but
once
this
fired,
em
mapped
to
user
space,
application,
user
space
directory
access
to
data
by
load
and
store
instructions.
So
if
the
data
is
disk,
then
you
have
to
copy
data
on
to
buffer
cache
or
memory.
One
stand:
you
can
only
access
there,
the
difference
and.
J
J
So,
though,
when
you
move
data
from
the
NIC
to
the
final
destination
or
some
file,
then
how
to
date,
what
is
data
comes
to
the
NIC
and
the
goes
up
to
the
network
stack
and
application,
read
that
data
and
application
memory
copy
that
data
from
DRAM
to
non-volatile
main
memory.
So
this
is
how
the
application
data,
reception
and
data
store
looks
like
today,
but
this
is
a
free
program
and
I'm
not
sure
how
this
is
problematic.
J
So
the
rest
think
about
some
simple
transaction
between
client
and
the
server
this
guy
is
the
client
send
some
data.
Hey
I,
want
to
keep
this
data
even
after
you
crash
I
want
you
to
recover
the
data.
So
in
this
case,
what
we
have
to
do
is
I
think
about
grant
client
wants
to
say
1
kilobyte
of
data,
the
how
the
thing
looks
like
is
granting
some
data
and
this
data
arrives
at
Annique.
Then
the
data
is
DMA.
K
J
To
DRAM
and
processed
by
TCP
and
IP
stack,
then
the
application
read
this
data
and
write
to
some
file.
After
that,
this
is
the
important
step
it
has
to
issue
some
everything
or
the
fighting,
the
already
a
mapped
it
has
to
issue
the
imaging
to
make
sure
data
is
actually
on
the
physical
disk
device
disk
or
a
safety
whatever
yep.
J
J
So
when
we
have
a
safety
or
disk,
this
process
took
milliseconds
so
out
of
that
so
networking
apart
or
in
this
picture
with
that
step
number
four,
it
took
only
only
40
nanoseconds,
a
40
micro
second,
so
there
was
nothing
to
do
with
the
networking,
but
once
we
press
that
the
disk
non-volatile
memory,
this
process
takes
only
40
to
microsecond
in
your
measurement.
So
the
now
the
storage
was
sharing.
Data
is
very
cheap,
so
networking
is
a
hypotonic,
so
we
have
to
add
it.
J
So
this
is
a
bit
more
synthetic
measurement
and
we
did
some
tests
like
having
one
client
on
one
server
and
client
sent
continually
the
HTTP
POST
to
save
one
kilobyte
of
data
and
server
takes
the
steps
to
store
data
persistently
and
we
changed
the
number
of
TCP
connection.
So
the
compared
to
direct
precinct.
J
Only
the
the
white
box
or
bars
and
the
dark
grey
box
and
path
so
in
creating
select
means
the
throughput
and
the
right
means
latency,
average
latency
so
clearly
compared
to
the
networking
only
case.
The
savings
storing
data
transition
entry.
The
code
is
a
lot
of
pattern,
a
lot
of
programs,
so
reducing
radiation
are
reducing
throughput
a
lot
and
increasing
latency
a
lot.
J
So
the
usually
the
copying
data
to
so
some
people
say
copy
the
expensive,
some
people
say
copy
even
copy
is
not
expensive,
so
it
depends
on
the
situation
so,
for
example
like
this
picture
so
reading
from
network
buffer
and
some
kind
of
buffer
to
application
buffer,
but
the
over
and
over
and
over,
but
if
the
application
buffer
stage
in
the
same
memory
address
or
same
location,
it's
actually
pretty
cheap,
because
application
bar
father
stays
in
the
cache.
But.
J
J
Imagine
that
is
the
destination
file
is
already
a
mapped
to
the
application
address
space.
So
here
the
result
of
the
where
the
cache
misses
happening
in
the
system.
So
the
when
we
don't
pass
this
the
data
or
we
don't
perform
data
copy
from
application
buffer
to
this
the
logo
file,
then
the
cache
misses
cache.
Miss
ratio
is
only
0.02,
0%
and
most
of
cache
misses
are
happening
in
the
network
stack
but
they're
only
a
little
bit,
but
once
we
start
saving
the
positioning
data,
then
the
lot
of
cache
misses
happens.
J
So
the
one
thing
we
so
the
our
integrated
network
stack
and
persistent
memory.
Abstraction
is
basically
pretty
simple,
so
the
one
important
thing
is:
we
have
to
allocate
packet
buffers
where
Nick
TM
HD
into
the
non-volatile
memory
memory
region.
It
is
just
a
file
created
and
after
that
we
have
to
employ
the
zero
copy
API,
because
we
have
seen
that
the
copy
the
expensive
so
the
overstep
you
track.
Data
arrival,
satanic
and
application
read
that
data
after
tcp/ip
processing
without
zero
copy,
the
end
application.
J
So
the
danger,
since
the
memory
is
already
the
persistent,
so
the
application
just
somehow
write
metadata
of
particular
buffer,
for
example
the
packet
buffer
number
one
length
and
offset
so
I'm
going
to
show
you
in
more
detail
next,
so
the
this
is
how
our
system
looks
like.
So
basically,
we
have
the
nickering.
Nickering
is
just
to
the
nick
hardware
and
we
have
some
the
file
containing
a
packet
buffers
and
we
have
a
tcp/ip
implementation,
so
the
sink
apart,
the
Nick
is
Nick
with
receiving
the
for
some
requests
to
be
stored.
J
In
that
case,
first
packet
is
sorry:
Nick
just
receive
the
full
packet.
Each
of
them
includes
data
or
client
request
and
think
so
then
the
first
packet
is
processed
by
tcp/ip
input
and
after
that
and
application
read
this
data
so
application.
Fine.
All
this
data
is
the
important
or
we
have
to
store
this
data.
Then
the
application
opens
another
five
to
store
metadata
or
some
descriptor
for
that
packet
and
say:
I
want
to
say
packet,
buffer,
zero,
offset
105
and
rings
1000
some
batch
okay,
but
after
that,
the
day
the
one
was
dead.
J
So
you
know
that
today
the
nicks-
this
is
a
good
thing,
so
today's
Nicks,
when
today's
Nick's
dmas
packet
into
memory,
it
only
logically
DMS
into
main
memory,
because
the
data
actually
states
in
last
level
cache
so
the
way
Nick
receives
packet
data
is
not
actually
on
memory
yet
so
it
it
is
on
the
Nick's.
It
is
on
CPU
cache,
so
the
application
must
explicitly
fresh
the
data
and
this
metadata
to
actually
is
persistent,
restore
data.
Yes,
so
the
next
packet
is
also
the
considered
of
data
to
be
persisted.
So
the.
J
So
after
that,
if
we
think
about
the
third
packet
was
something
some
data
which
is
not
to
be
stored.
In
that
case,
when
application
find
all
this
data
is
not
important
or
some
key
value
store
get
in
that
case,
it
simply
skips
this
data
to
be
written
back
to
memory,
and
the
final
data
was
also
considered
as
data
to
be
persisted.
So
it's
simply
the
fresh
data
and
metadata.
J
J
So
here's
some
result,
so
we
compared
this
using
some
some
extended
net
map
API
and
not
extended
net.
My
baby
yeah
but
extended
tcp/ip
API
and
using
the
method
data
persisting
method,
I
explained
before
so
the
by
avoiding
memory
copy
and
the
weekend.
Sorry,
the
white
bar
is
not
storing
data
at
all,
so
theoretical
the
limit
of
performance
and
the
white
right
way.
Part
and
boxes
are
the
storing
data
by
copying
data
and
the
gray.
One
are
the
our
system.
A
J
Sana
bottlenecks,
so
we
must
be
designing
networking
api.
So
actually
you
know
the
socket
API
today
resembles
disk
api,
so
we
can
read
and
write
and
we
can
send
a
file,
but
there
is
no
official
Emma
for
network
stack
efficiently.
So
we
need
some
more
memory
friendly
api's
for
tcp/ip
as
well,
because
it
becomes
a
bottleneck
and
we
have
a
lot
of
ongoing
work.
J
So
we
are
brushing
up
a
ps4
beta
integration
with
net
map
and
we
are
also
currently
the
packet
buffer
must
be
a
single
fire,
but
we
found
it
is
actually
not
very
convenient
because,
for
example,
like
a
database
application,
it
has
multiple
files
to
keep
two
breaks
or
chunks
file.
So
we
have
to
support
the
packet
buffer
formula
file
system
and
also
we
are
building
local
use,
gates,
applications
like
key
value
stores
and
also
some
softer
switches
as
well.
So
but
again
so
the
proposing
the
network
stack
and
some
persistent
memory.
J
We
are
building
this
picture
like
the
architecture
so
application.
The
database
API
application
using
net
map
because
it
is
networking
a
passion
of
a
map
to
a
file,
but
underneath
we
have
post
network
stack
and
file
system
and
we
have
non
brought
a
main
memory
and
mix
so
I
know
that
the
transport
working
group
basically
designing
a
new
transport
API
because
to
better
utilize.
The
rich
transport
protocol
features,
but
persistent
memory
actually
coming
not
only
to
data
center
service,
but
also
mobile
devices.
J
N
J
J
I
found
that
the
taps
working
group
are
defining
and
how
to
better
you
try
transport
feature
and
but
I
think
if
we
think
about
API
change,
it
quite
big
thing.
So,
the
if
we
design
we
redesign
API
I
think
we
have
to
not
only
take
into
account
some
transport
protocol
features,
but
also
some
upcoming
devices.
Yes,.
N
N
O
Tommy
Polly
Apple,
so
yeah,
thank
you
for
showing
this
is
really
cool.
Coming
from
working
on
some
of
the
stuff
and
taps
at
Apple,
you
know
we're
doing
a
lot
of
work
on
trying
to
find
the
API
layers.
Just
recently,
we
deployed
on
our
mobile
devices
the
fact
that,
underneath
the
higher
level
API
is,
we
are
using
kind
of
the
shared
memory
model
here
with
the
user
space
stack
mm-hmm
very
so
much
for
this.
O
O
J
O
O
B
Miracle,
even
so,
we
had
a
site
discussion
which
was
related
to
the
fact
that
we're
thinking
about
using
a
message
based
API
in
intercepts
instead
of
a
streaming
API
and
if
I
get
it
right,
I
think
that
would
actually
be
favorable
for
your
approach
right
or
is:
do
you
think,
there's
any
big
difference?
Yeah.
J
So
actually,
the
the
major
drawback
of
using
net
map,
API
or
kind
of
memory
based
API
for
networking
is
some
very
extremely
easy
support
of
streaming.
The
stream
says
is
be
difficult
because
application
has
to
the
chanc
data
by
the
application
has
to
segment
data
by
themselves,
but
I
think
I
mean
the
community
or
world
is
moving
into
accepting
this
kind
of
thing,
for
example
in
acts.
How
the
KCM
today
kind
of
multiplexer
today,
which
enables
message
based
application,
can
send
traffic
over
solid
TCP
application.
K
J
M
J
Haven't
actually
consider
using
the
PDK,
because
the
BTK
completely
runs
in
user
space,
even
device
driver,
you
know
so,
but
I
I,
actually
don't
like
the
perfect
by
complete
bypass
approach
or
I
want
to
take.
The
good
part
of
the
Canon,
like
Anya,
has
a
good
memory
management
system
and
very
good
the
protection
and.
J
M
J
J
N
P
I
have
a
little
evil
question
so
as
we
sing
that
a
message
based
the
api's
seems
to
be
much
better
for
most
kind
of
networking.
Terrific-
and
you
say,
memory
based
avi
is
also
much
better.
Do
you
see
there's
still
room
for
some
kind
of
screen
based
at
the
eye
in
the
future
in
the
future,
or
is
it
just
to
go
away
sometime
so.
J
K
J
So
you
know
that
we
do
spring,
maybe
I
own,
say
today's
TCP,
the
we
of
course
stream
data,
or
we
can
write,
say
10
megabyte
of
data
and,
like
the
changing
data
internally
I
think
we
can
also
write
10
megabyte
of
data
into
a
UDP,
socket
I.
Think
then,
the
in
that
case
UDP
layer,
the
segment
this
data
into
you
know,
IP
fragments
so
and
anyway,
we
have
at
some
point.
We
have
to
segment
some
data
I.
Think
many.
J
B
B
A
It
was
what
you
know.
That
was
one
of
the
big
one
of
the
considerations
for
me
being
willing
to
serve
a
third
term.
As
your
area,
director,
we've
gone
from
the
first
blow-up
that
I
had
on
after
joining
the
I
on
the
is
G,
was
the
MPLS
over
UDP
IETF
last
call.
Some
of
you
may
remember
that,
and
we've
moved
from
that
kind
of
relationship
with
the
rest
of
the
IETF
to
we
are
now
getting
requests
for
early
transport
review
from
a
working
group.
That's
concerning
adopting
a
draft.
You
know
it's
like.
A
B
A
Q
A
It's
also
worth
just
me
mentioning
so
my
co
area
director
will
be
up
for
her
position
will
be
up
for
Nam
Kham
review
this
year.
Please
be
considering
your
feedback,
because
Tom
come
really
appreciates
feedback
about
about
the
incumbents
and
about
and
about
the
needs
of
the
area.
Don't
let
don't
forget
that
part
they're
supposed
to
be
balancing
not
like
do
people
think
this
person
is
doing
a
good
job?
A
A
And
if
you
don't
put
your
name
in
this
year,
I
will
be
up
for
review
the
following
year
and
the
only
feedback
that
I
got
from
Nam
Kham,
which
they
met
with
the
people
that
the
the
willy
nominees
and
said
this
is
the
kind
of
feedback
we
got
about.
You
was
that
serving
a
third
term
was
a
really
big
hurdle
for
them
to
get
over.
He
said
there
will
not
be
a
fourth
it's
if
you
take,
you
know
if
it
takes
you.