►
From YouTube: Antrea Community Meeting 08/16/2021
Description
Antrea Community Meeting, August 16th 2021
A
Yeah
perfect
so
good
morning,
good
afternoon,
good
evening,
thanks
for
joining
this
week's
andrea
community
meeting-
and
we
have
quite
a
good
agenda
for
today,
we
have
at
least
three
topics
and
we
will
start
with
a
discussion
of
the
multicast,
auto
discovery,
design
and,
and
that
will
be
done
by
waning,
and
then
we
have
a
hong
yang
that
will
talk
to
our
talk
with
us
about
benchmark
results
for
antria
proxy
with
the
tc
flower,
and
you
know,
then,
a
discussion
of
approaches
of
proxy
based
on
ip
tables
versus
dc
flower.
A
So
therefore,
without
much
further
ado,
let's
get
started
with
today's
agenda.
It
seems
that
there
are
no
other
proposed
topics,
but
you
know
feel
free
to
to
propose
anything
else.
If
there
is
anything
else
that
you'd
like
to
discuss
and
we
can
get
started
so
winning
whenever
you're
ready,
please
go
ahead.
B
I
want
to
share
my
tester's
design
with
some
for
last
questions.
The
first
is
for
the
multicast
group
discovery.
We
originally
want
to
discover
muscular
skills
using
in
australia,
but
it
is
not
following
the
generic
solutions
for
multicast,
so
we
have
changed.
The
discovery
solution
to
using
the
packed
in
and
picked
out
for.
The
multicast
group
is
auto
discovery
to
discover
the
medical
group.
B
First
and
three
agents
should
send
a
portable
entry
on
the
right
bridge
to
require
the
hmt
message
to
send
that
sent
to
the
controller
I
mean
the
unsure
agent
and
then
enter
agents
will
pass
the
medical
school
from
the
message
and
second
and
three
agents
of
periodically
to
send
a
generic
identity.
Query
message
to
os:
it
is
using
the
packet
out
mechanisms
and
the
message
is
forwarding
normal
in
the
west
bridge.
B
Then
the
local
pulse
will
receive
the
agent
query
and
then
they
will
record
its
much
a
package
group
in
the
ignp
report
message
having
the
open
flow
entries
we
set
up
in
the
step,
one.
The
rjmp
report
report
message
will
be
sent
to
the
interior
agent
and
then
enter
agent
could
pass
the
group
from
the
from
the
packet.
We
want
entry
agent
to
support
the
agent.
He
will
promote
the
with
three
verses.
B
I
mean
version
one
two
and
three,
but
we
don't
expect
to
implement
the
agmp
mysteries
fully
in
the
first
release.
It
is
because
some
soft
specific
group
records
in
the
message-
and
I
think
it
will
be
much
cost
in
the
coding
so
to
support
the
fully
the
agmp
victory.
We
will
put
it
in
the
next
release,
but
for
the
first
release
version,
this
could
support
to
pass.
The
group
message.
Grouping
address
from
either
v1
v2
or
v3.
B
B
Which
is
used
in
hmpb3
and
we
match
the
destination
address
with
the
0.2,
which
means
all
routers.
Both
two
of
the
launchers
will
install
on
the
left.
To
let
entry
agent
reveal
the
hmp
report
message
and
for
the
actions
we
require
will
use
controller
as
a
normal
action,
for
the
first
controller
is
to
take
in
the
message
and
for
normal
is
to
let
it
always
to
learn
the
hrmp
message
and
always
could
use
multicast
looping
feature
for
untradition
after
it
has
received
an
agency
report
message.
It
has
processed
the
message
in
the
following
logic.
B
First,
if
it
is
a
new
group
found
from
the
message
until
agent
will
add
it
into
its
local
catch
and
trade
agent
maintains
the
local
cash
with
the
multicast
group
and
the
last
prop
tab
for
the
first
first
time
founded.
The
group
group
address.
It
is
excited
into
the
local
catch
and
enter
agents
will
configure
the
transport
interface
to
join
the
multicast
group,
then
for
the
physical
switch
or
router
it
could
forward
the
multicast
traffic
to
the
node.
B
Packing
message
for
the
response
time
and
agent
will
loop
in
the
local
cache
and
find
the
sale
medical
group
entry.
Then
it
will
remove
it
from
the
local
cache
and
require
the
transport
interface
to
leave
the
group.
If
the
eventual
agent
has
received
an
ignp
group
specific.
B
B
If
no
membership
report
has
received
for
the
target
group,
ensure
agents
will
process
it
as
the
last
group
member
and
remove
the
group
from
the
cash
and
the
conveyor
support
interface
to
leave.
The
group
about
oil
is,
for
the
other
discovery
solution
for
multi-test
projects
and
then
a
multicast
group
discovery
and
for
the
multicast
traffic
forwarding.
B
It
is
almost
the
same
as
what
we
have
designed
in
the
last
meeting,
we're
still
using
the
used
os,
multicast
new
snooping
feature,
and
they
will
provide
open
flow
entries
and
to
forward
the
multicast
traffic
to
local
powers
and
they
forward
the
multicast
of
traffic
if
it
is
to
ensure
it
gets
away
in
space
if
it
is
sent
out
from
localhost.
B
Then
agent
also
configure
the
multicast
routing
entry
on
the
host
for
the
the
group
for
for
the
multicast
group,
which
is
found
from
local
pulse.
It
will
add
a
routing
cut,
multicast
residential,
like
in
under
holes
like
this.
It
will
require
the
source
address
with
the
powder
drives
and
the
the
target
group
invoice,
which
is
what
it
has
found
and
it
required
packages
forwarded
from
unturned
general
to
the
transport
interface
and
as
for
the
external
multicast
traffic
entry
agents
will
leverage
the
imp
message.
B
So
the
packet
could
be
forwarded
to
the
local
pulse.
It
is
for
the
routing
entry
configuration
and
for
the
first
release
for
automatic
feature.
We
will
support
the
knowing
cap
mode.
Therefore,
for
in-cap
mode
it
will
be
implement
implemented
in
the
next
version
because
of
the
the
multicast
invoice
is
not
able
to
share
between
different
nodes
in
cap
mode.
B
If
we
don't
use
other
central
central
mechanisms
like
entry
controller,
it
is
for
the
design
for
multicast
traffic.
C
A
D
D
I
mean
I
don't
know
much
about
multicast,
but
do
you
know,
can
you
explain
why
igmp
v3
is
not
going
to
be
supported
at
first
and
is
it
a
big
issue
or
do
you
think
igmp
v2
is,
is
the
most
common
version
of
the
protocol
in
use.
B
B
Also,
we
don't
want
to
differentiate
the
sources
of
the
multicast
traffic
universal
release,
so
we
don't
want
to
support
this
feature.
That
is
why
we
don't
plan
to.
We
don't
mean
full
support
for
agmp
history,
but
this
requirement
is
this
feature
is
only
introduced
in
ionic
stream,
not
included
in
re2,
so
we
don't
need.
We
don't
say
that
argentina
base
2
is
not
responsive.
D
D
So
do
sorry
do
do
we
do
we
support
do.
Do
we
not
support
a
gmp
v3
at
all,
or
we
only
support
igmpv3
if
sources
are
not
provided.
B
Where
is
the
pause
part
of
the
ignp3?
I
mean.
D
B
No,
no,
nothing
to
your
eyes
at
it.
You
can
go
ahead.
C
B
Yeah
yeah,
I
think
we
will
finally
support
it,
but
it
is,
it
should
be
a
long-term.
I
think
it
might
have
some
something
similar
like
the
multicast
surveys,
network
policy
or
all
some
security
requirements.
A
B
A
E
Sorry,
can
you
see
the
screen?
Oh,
yes,
okay,
thanks
hello,
guys!
I
will
first
introduce
you,
know
new
design
to
android
proxy
for
not
with
linux
tc,
but
ap
tables,
and
I
make
a
new
request.
E
First,
the
new
design
is
much
simpler.
This
is
much
simpler
for
for
not
hot
or
not
powder.
We
just
need
the
episode
to
to
to
editable
entries
and
routing
entry.
Well,
I
decide
for
is
used
to
install
node
powder
ipod
protocol
information,
and
the
episode
is
used
as
the
destination
for
this.
For
this
tour.
For
these
two
I've
tabled
into
its
chain
entry
output
is
for
package
from
localhost
and
the
chain
entry
promoting
is
for
package
from
remote.
E
D
E
D
E
We
can
see
we
can
see
the
benchmark
to.
We
can
see
the
benchmark
to
explain
the
question.
Okay
and
this.
There
are
three
three
clusters.
First,
the
first
two
are
anti-proxy
four,
and
we
see
we
can
support
service
from
out
from
outside
of
the
cluster.
For
this
this,
this
data
there
are
some
17.
E
This
is
the
list,
is
the
test
data
and
with
testing
and
testing
this
this
and
this
this
group,
these
two
group
data
anti-proxy,
oh
sorry,
cookie
proxy,
is
remote
and
then
for
array
for
our
cases
I
have
tested
10
times
and
remove
the
highest
and
the
lowest
one.
E
E
And
for
tcpr
a
little
performance
degree
next,
for
there
are
many
services,
for
example,
thousand
two
thousand
and
one
hundred
services
we
can
see
that
the
tcp
stream
is
not
you
know,
is
not
changed
changed
a
lot
also
about
a
five
percent
performance
degree.
E
The
entry
proxy
full
are
much
better
because
this
test
will
create
a
new
tcp
connection
for
every
contraction,
and
there
are
two
thousand:
there
are
about
two
thousand
services
until
proxy
default
use
uses
cookie
proxy
to
do
the
net
another
and
the
time
complexity
for
certain
ip
tables
is
easily
renewable.
Well,
actually,
proxy
4
is
housing.
F
I
think
we
need
to
clarify
one
thing:
that
is
not
an
entry
process
implemented
by
actables.
Otc
is
just
that
the
traffic
reduction
is
implemented
by
ib
tables
or
tc,
the
the
service
load,
balancing
so
of
the
scenario
in
all
of
these
scenarios,
the
service
loan
balancing
is
still
handled
by
open
switch.
E
F
I
think
to
further
answer,
antonio's
question:
that
the
reason
why
we
introduced
tc
to
handle
the
traffic
reduction
is
that
we
thought
it
would.
It
would
be
simpler
and.
F
We
have
to
install
many
tcp
filters
in
every
interface,
for
example,
if
you
house
three
in
external
interfaces,
we
will
do
the
same
thing
three
times
and
if
you
have
three
ips
on
the
interface,
then
the
tc
flower
rules
will
will
will
be
three
times
than
one
ap
case,
and
that's
one
complexity
and
another
is
how
that
reply.
F
Traffic
will
be
handed
because
when
they
open
with
switch
output,
the
response
to
and
take
it
away
in
the
face
we
will
need
to
figure
out
which
interface,
which
is
turning
the
interface
we
need
to
redirect
to
so.
It
introduced
some
long
flow
to
handle
this
case,
and
it's
a
little
complicated
too
and
by
switching
to
app
tables.
F
It
turns
out
that
the
performance
is
not
affected
a
lot,
as
you
can
see
that
in
some
cases
it's
even
better
than
tc
redirection,
or
at
least
another
much
worse
than
it
and
with
we
can
still
benefit
from
the
advantage
of
opengl
switch
handling
the
tcp
crr,
and
you
can
see
that
when
there
are
more
than
2000
services,
regardless
of
the
redirection
mechanism,
we
can
still
get
much
better
metrics
than
cube
processing
the
tcps
for
the
tcps
metrics.
So
that
that
way
we
are
considering
switch
back
to
a
table.
Spaced
redirection
mechanism.
D
Thanks
for
the
clarification
out
of
curiosity,
do
you
know
why
in
the
last
benchmark,
tcp
rr,
it
seems
that
tc
is
actually
degrading
faster
than
the
iptables
implementation?
Do
you
know
why
you
know
I
can
see
zero
zero
point
nine,
so
ten
percent
for
tc
and
about
five
percent
for
ip
tables.
A
A
I
have
a,
I
have
a
question
regarding
what
chandra
said
about
you
know
the
show
why
initially
tc
was
chosen
over
ip
tables
for
redirection.
You
see
that
in
this
way
we
are
bypassing
the
linux
scale.
Linux
networking
stack,
but
does
this
also
mean
that
with
the
ip
tables
we
will
be
doing
more
kernel
contact
switches.
F
Okay,
this
one.
The
thing
is
that
if
we
enable
four
entry
proxy,
then
in
theory
there
will
there
will
not
be
many
activist
laws
on
the
in
the
host
in
the
loot
network
in
the
space.
So
the
overhead
should
be
small,
unlike
when
kuposi
handles
this,
it
will
have
many
active
slots,
so
it
will
degrade
a
lot.
A
E
I
forgot
I've
got
these
this
time.
I
forgot
to
record
the
cpu
users
and
the
memory
usage.
No
problem
was
just
the
curiosity,
iceland,
as
I
remove
the
quick
proxy,
the
memory
and
the
memory.
A
All
right,
so,
in
light
of
these
results,
are,
I
would
recall,
are
we
recommending
using
ip
tables
for
redirecting
traffic,
because
it
seems
that
it
will
lead
to
an
easier
implementation?
A
more
manageable
implementation?
Is
that
correct.
A
Cool
perfect
is
there
any
other
question
curiosity
from
from
the
community
regarding
these
results
or
the
overall
observations
regarding
you
know:
tc
flower
versus
happy
tables
for
traffic
redirection.
C
Yeah,
I'm
just
thinking
even
channel
I
discussed
and
the
chain
feel
even
we
switched
to
evgf,
probably
snow,
much
performance
gains
compared
to
tcm,
but
I'm
still
thinking.
Maybe
we
should
consider
bppf.
C
Sometime
later,
at
least
in
serial
we
can
see
we
can
use
ebp
to
accelerate
the
traffic
direction
and
bypass
the
host
tlp
stack
and
it
can
be
one
story
for
upl
class
centria.
E
A
Any
other
question
on
this
topic.
I
wait
15
seconds.
F
Right
shuttle
ford
changes
for
the
suggestion.
I
I
think
it
makes
sense,
and
maybe
we
we
could
have
have
the
basic
implementation
based
on
app
tables
which
could
handle
all
traffic
from
all
interface
or
interfaces,
and
if
tc,
eppf
or
xtp
can
really
accelerate
the
redirection.
Maybe
we
could
enable
it
on
a
major
external
interface
that
that
that
we
are
assuming
that
oh
most
of
the
incoming
external
traffic
should
come
from
right.
That
really
makes
sense.
C
Yeah,
that's
also
one
strategy.
I
thought
about
how
it
makes
sense
for
to
cover
the
most
important
cases,
but
I'm
so
thinking
if
we,
if
we
want
to
recommend
ebpf
for
an
epf,
is
supported
by
the
host
os,
I'm
not
sure
we
want
to
cover
more
cases,
but
anyway
I'll,
probably,
let's
think
about
later.
I
guess.
A
A
Cool,
and
so
it
seems
that
we
are
at
the
end
of
the
plan
topics
for
today,
this
meeting
is
there
anything
else
that
we
like
to
discuss
for
today.
C
C
E
A
Yes,
so
so,
just
just
so
everyone
we're
all
on
the
same
page
is
the
pr
already
updated.
Are
you
going
to
update
the
pr
soon.
A
And
I
think
that
with
this
we
have
exhausted
this
topic.
We
still
have
about
20
minutes.
So
if
there
is
a
any
other
topic
that
you
will,
you
would
like
to
bring
up
for
discussion.
Please
go
ahead.
A
Perfect,
it
appears,
therefore,
that
this
is
all
for
today.
So
I
would
like
to
thank
everyone
for
joining
and,
as
usual,
it's
been
a
very
good
conversation
thanks,
a
lot
waning
and
hong
liang
for
your
presentations
and
we'll
meet
again
in
two
weeks
time.
So,
thanks
again
for
joining-
and
I
wish
everyone
a
great
day
good
morning
good
afternoon
and
a
good
evening.