►
From YouTube: SONiC DASH Workgroup Community Meeting Dec 7 2022
Description
PR289 needs volunteer
PR290 merged
PR293 submitted
Egresss port selection convo
A
A
B
I
think
pull
request.
290
I
I
feel
like
it's
ready
to
be
merged.
Gohan
left
two
comments
about
a
couple
of
Fairly
innocuous
items
more
like
questions
than
than
problems,
and
we
responded
they're,
just
backlog
items
overall,
so
I
don't
think,
there's
any
major
objections
to
what's
in
there.
B
The
content,
not
here
today,
to
finish
up
anything
but
there's
there
was
no
serious
complaints,
no
requests
for
changes,
so
it
seems
to
me
like
it's
ready
to
be
merged,
and
this
is
like
the
fourth
go
around
yeah,
so
you
know
we
propose
we
merged
it
and
then
you
know
something
comes
up,
we'll
we'll
address
it
coming
forward.
Just
like
anything,
I.
A
Think
so
too,
and
you
know
I
think
I
probably
have
ability
to
do
that
as
well.
So
let
me
share
a
different
screen
here:
yeah
the
screen.
C
A
A
challenger
yep,
so
we
just
had
a
couple
of
different
comments
that
like
Chris
said,
are
fairly
innocuous
and
if
it
doesn't
get
looked
at,
maybe
I'll
just
go
ahead
and
approve
and
merge.
B
A
Okay,
great
so,
if
necessary,
I'll
see
if
I
can
do
this
later
on.
Today,
thanks
yeah
yep
yep.
D
A
Oh
wonderful,
okay!
Let
me
just
finish
up
here:
thanks
reshma,
that's
awesome!
Oh
you've
got
it
right
here.
Okay,
do
you
want
me
to
display
or
should
I
go
ahead,
and
let
you
guys
do
that.
E
It
actually
doesn't
matter
because
you
can
display
at
least
a
file
list.
I
will
describe
briefly.
Oh
content
is
here
actually,
so
that's
in
the
description
of
the
pr.
So
now
I
want
to
just
pay
attention
that
it
is
dropped
because
I
have
some
question
but
like
before
question
I
I.
So
the
content
is
that
we
are
adding
New
Year
night
test
cases
and
in
in
combination
with
the
spawn
itself.
E
So
right
now
it
will
be
easier
to
review
because
we
pushed
test
plan
217
with
prepare
217
with
this
plan,
but
like
it's
not
reviewed
without
any
comments
for
a
while,
so
I
decided
to
like
do
not
wait
for
you
and
I'm
starting
pushing
the
code
of
status
cases
and
I'm,
pushing
like
next
to
the
test
case
code
as
as
well
as
plan.
So
it
will
be
easier
to
review
to
see
what
is
the
content
and
how
code
looks
like
in
addition.
E
Also
right
now,
there's
five
files
changed
because
we
moved
helper
functions
to
a
separate
file,
PSI
Dash
utils.
So
this
will
be
some
helper
functions
that
are
going
to
be
used
in
all
other
test
cases.
So
like,
basically,
if
you
see
PSI,
Dash,
ACL
and
vinette
changes
just.
E
F
After
why
don't
we
just
have
the
the
pr
separated
like
any
refacting
can
be
done
as
a
separate
PR
versus
new
a
functionality
as
another
one,
because
otherwise
I
think
it
will
be
hard
for
reviewers.
E
Right:
okay,
if
it
simplifies
I
I,
can
create
a
separate
PR
where
I'm
moving
out
common
functions
and
then
I
will
have
a
PR
with
only
two
files.
Okay,
so
not
not
a
problem.
I
will
do
this
so
tomorrow.
It
will
be
like
that
so
because
I
believe
that
yeah
so
moving
out
common
functions,
it
will
be
extremely
easy
to
approve
because
no
changes
at
all
yeah.
E
Okay,
okay,
good
I,
agree:
I
I
will
lose
it,
so
that's
easy.
Okay,
then
I
will
rebate
that
PR
and
you
will
see
only
two
files
and
next
comment.
E
I
actually
had
is
that
my
idea
was
that
those
test
cases
should
pass
on
CI,
because
once
we
will
merge
this
PR
it
in
those
test
cases
and
those
Scots
will
be
running
on
the
pr
check
and
I
found
that
not
all
these
cases
are
passing
on
the
BMV
too,
and
my
question
is
the
actually
what
so
right
now
failing
get
attributes.
So
that's
create
delete
works
like
delete
with
one
issue:
I
I
will
I
explained
it
earlier.
E
I
will
repeat,
but
get
set
doesn't
work
and
my
question
actually
do
we
expect
it
work
at
all
like
in
dash
I
mean
that
we
can
get
any
attributes
that
we
are
creating
for
eni.
We
net
everything
like
this
or
what
is
expectation.
Sorry
I
didn't
find
this
in
documentation
if
I
missed
something.
Please
Point
me
out.
A
A
The
question
was,
it
seems
that
a
couple
of
which
commented
test
cases
in
bmv2
continue
to
fail.
Is
that
right
and
do
we
exist
to
work,
is
that
the
right
am
I,
saying
it
right.
E
Yeah
some
specific
test
cases
forget
attributes
only
get
attributes
so
like
create
eni.
Creative
net
does
work,
but
if
I
try
to
get
some
attributes
from
the
eni,
for
example
the
get
comma
the
get
request
fails
is
so
do
we
expect
expect
that
get
for
any
attribute
should
work
for
dash
yeah.
B
So
we're
talking
about
bmv2
right,
yeah
yeah.
Well,
it's
known
backlog
that
bmb2
lib
PSI
needs
git
and
set
accessors
they're
not
implemented,
never
have
been.
So
it's
not
just
it's
not
a
bug,
it's
enhancement
or
a
feature,
completion
issue
and
also
we
need
the
bulk
apis
set
and
remove
bulk,
which
there's
a
PR
for
that.
But
it's
been
I,
don't
know
sitting
around
I
can't
remember
the
status
of
that
that
Marion
did
so
yeah.
It's
a
feature.
We
need
that.
D
Makes
sense
only
it
looks
like
Anton
tried
some
get
and
some
of
the
get
is
working
for
v-net
and
BNI,
maybe
in
the
backlog
the
task
list
that
we
have
for
bmv2
for
the
Get
Set
item,
we
can
add
more
details
there
that
eni
attributes
get
especially
specifically
is
not
working
and
that
way
we'll
be
able
to
focus
to
add
those
functionality.
D
B
G
B
Only
thing
you
can
get
are
a
few
dummy
attributes
that
I
hand
coded
in
libside
over
the
last
few
weeks
to
support
test
initialization,
like
you
know,
default
VLAN,
those
kinds
of
things
it
returns
dummy
objects
just
to
get
the
python
Frameworks
to
function,
but
there's
no
code
I
mean
the
code
says
not
implemented
right
in
the
code.
Okay,
so
it's
attack.
B
C
The
intent
of
all
right,
quick
question:
what's
the
intent
of
the
set
I'll
understand?
Obviously,
if
you
want
to
configure
your
tables
and
all
that
that's
needed
the
intention
of
the
gate,
I
don't
understand
because
whatever
you're
gonna
get
from
your
data,
pane
I,
suppose
it's
something
that
the
configuration
layer
has
and
push
down
so
not
sure
I
understand
the
notion
of
the
get.
D
To
basically
verify
that
the
data
plane
has
actually
implemented
the
you
know,
configurations
and
at
any
time
be
able
to
do
the
verification
Telemetry.
Whatever
is
needed,
yeah,
that's
mainly
the
reason
for
get
Anton.
Is
that
the
same
that
you're
expecting
in
the
test
case?
Like
suppose,
if
somebody
does
a
show
command,
you
know
we
should
be
able
to
get
that
yeah.
C
E
C
G
C
B
Yeah,
that's
really
a
philosophical
issue.
I
mean
I.
Think
it's
worth
it.
You
know,
there's
different
schools
of
thought.
I
happen
to
fall
in
the
camp
that
anything
you
can
set.
You
should
be
able
to
get
is
just
almost
like
a
configuration
loopback,
but
you
know
it's
arguable.
It's
a
pragmatic
issue.
Like
you
mentioned
Vincent,
there
might
be
a
transformation
when
you
do
the
set
and
you
don't
want
to
store
the
original
value
because
you
know
it
should
have
to
cash
it
forever.
B
B
You
know
and
that's
just
a
group
decision-
do
we
support
it
or
not?
Who
wants
to
put
the
work
in
if
for
a
Sonic
implementation,
you're
going
to
have
to
support,
get
whether
you,
you
know
everything's
expected
to
be
bi-directional,
so
I
don't
think,
there's
going
to
be
a
debate
there,
the
platform
will
be
assumed
because
Sonic
can
do
whatever
it
wants
with
the
PSI
interface
right.
B
So
you
know
just
put
that
out
there
be
prepared
for
counters.
It's
a
behavioral
model
issue
right.
We
need
to
Define
and
Implement
counters.
Do.
A
D
Okay,
so
just
to
complete
the
the
eni
question
from
Anton
I
just
like
want
to
mention
that
eni
has
a
lot
of
attributes
right,
so
it
will
be
good
to
know
how
it
is
implemented,
whether
it
is
implemented
and
and
set,
should
work
for
eni
attributes.
That
would
be
very
helpful.
A
And
since
you're
on
the
call
you
are
still
on
the
call,
okay
good,
we
are
thinking
we're
pretty
much
ready
for
290..
Have
you
taken
a
look
at
290?
A
You
know
no
I
did
okay
I'll
bring
you
through
it
real,
quick,
there's,
there's
just
a
couple
of
very,
very
innocuous
comments
that
basically
this
these
questions,
that
the
answer
is
that
they're
in
our
backlog
and
it's
pretty
much
ready
to
go
now,
and
so
would
it
be
okay
with
you
guys
if
I
go
ahead
and
approve
and
merge
this
one's
been
waiting
quite
some
time
and
it's
kind
of
holding
up
everything
else.
So.
A
F
F
A
A
Merry
Christmas,
okay,
thanks
Prince
for
showing
me
that
so,
okay
and
so
Anton
you're,
saying
I
need
to
add
this
to
bmv2.
Here.
E
Yeah
yeah,
but
I
would
like
to
add
here
a
few
comments
that
see
that
still
checks
for
my
PR,
a
passing
because
I
added
additional
flag
to
identify
that
this
is
a
bmv2
and
set
get
test
cases
I
actually
skipped
on
bmv2
but
like
if
someone
wants
to
run
this
on
harder.
Something
like
that.
So
that's
with
changing
that
flag.
It
will
be
possible
to
like
to
run
everything,
so
you
can
also
take
a
look
and
check.
E
Yeah
I
have
to
check
this
mechanism.
Okay,
so
by
tomorrow,
I
will
update
this
PR,
so
it
will
contain
only
like
two
files,
so
it
will
be
easy
to
review
and
easy
to
actually
see
how
this
flag
works.
B
I
wanted
to
raise
another
topic,
and
maybe
behavioral
model
meeting
can
talk
about
it
too,
but
yeah
the
notion
of
getting
platform
metrics,
let's
say
kpis
or
attributes
about
the
platform
that
the
dash
is
running
on,
for
example,
would
be
nice
to
be
able
to
ask
in
a
fender
agnostic
way
for
a
list
of
you
know.
Let's
say
memory,
memory
pool
size
and
utilization,
CPU
loading,
all
kinds
of
things
like
that,
the
typical
linuxy
type
of
kpis
and
I
don't
know
how
this
is
done.
B
Let's
say
in
the
standard
Sonic
world
or
if
it's
done
it
all,
and
it's
probably
not
done
through
PSI,
but
I
think
we
could
add
we're
talking
about
an
API
that
just
says
you
know,
get
platform
attributes
and
it's
just
a
let's
say,
unspecified
list
of
key
value
Pairs,
and
we
could
settle
on
some
conventions
where
you
know
a
common
set,
that
every
platform
should
Supply
and
then
vendors
can
add
their
own.
That
would
be
very
useful
when
we
start
getting
into
serious
testing
on
real
Hardware,
which.
H
So
yeah,
so
that's
why
I
think
it's
handled
by
Sonic.
So
we
have
this.
You
know
Telemetry
container,
which,
which
you
put
you
know
the
CPU
memory
and
every
process
utilized.
You.
H
Yeah
those
things
I
think
it's
it's
it's
been
unified,
so
once
the
vendor
running
the
Sonic,
you
know
the
the
the
HP
file
yeah,
so
so
I
think
the
we
have
some
tests
in
the
Sonic
management
report
to
you
know
to
to
allow
you
to
how
to
get
the
the
data
from
the
Telemetry
I
think
there
was
a
Chia
nmi
client
allowed
you
to
get
those
data.
B
Yeah
so
that
says
kind
of
by
suspected,
because
I
know
there's
a
lot
of
platform
info
at
the
Sonic
level.
Is
it
worth
talking
about
standardizing
that
at
the
PSI
interface
or
is
there
no.
H
B
If
we
were
doing,
let's
say
rigorous
testing
of
dpus
using
site
Thrift
right,
which
is
the
Microsoft's
preference
to
have
you
know
a
single
common
way
of
testing
all
dpu's
data
plane,
we've
we
as
we're
testing
and
we're
interacting
with
vendors.
You
know
one-on-one,
we
do
queries
using
platform
commands
to
get
information
about
memory
and
other
things
which
is
useful
in
our
dialogues
with
them.
B
And
if
we
had
to
wait
for
Sonic,
we
wouldn't
you
know
we
wouldn't
be
able
to
do
that
in
a
standard
way,
so
we're
having
to
do
it
in
a
vendor
dependent
way.
It
seems
like
it's
an
opportunity
to
think
rethink
that,
and
maybe
it
doesn't
need
to
be
a
distraction
but
I
think
it's
something
we're
thinking
about.
So
there's
a
let's
say
a
NOS
independent
way
of
getting
some
of
this
stuff.
B
But
it
you
know,
that's
just
an
idea.
G
H
No,
no,
that's
a
CRM,
no,
that's
a
CRM
right,
so
we
have
this
critical
resource
monitor
right.
So
you
know
if
you're
talking
about
the
async
table,
usage,
I,
think
yeah,
you
know.
If
that
hasn't
been
defined
in
the
size,
then
we
should
Define
the
size
right.
So
you
know
for
those
data
plan.
You
know,
statistics
of
cricket.
Let's
say
you
know
how
many
uni
tables
are
available.
What
what
is
the
percentage
of
the
United
tapers
available?
B
H
C
Maybe
quite
a
few
there
I
killed
a
few
side
objects.
You
know
Sweetwater
at
least
that
is
there
any
way
to
get
more
in
the
data
on
that
cool,
yeah.
C
H
I
That
is
yeah.
There
is
some
sort
of
a
mind
set
change
that
we
need.
Basically,
what
happens
is
that
in
Sonic
generally,
we
have
this
CPU,
which
is
an
external
CPU
memory,
is
an
excellent
CPU
to
the
from
the
Asic
point
of
view
right.
So
when
we
come
to
the
dpu
case,
here
we
have
the
dpu
Asic,
but
essentially
the
the
CPU
inside
this
dpu
is.
It
is
embedded
CPU
and
also
the
memory
is
also
embedded
so
but,
however,
they
will
be
treated
or
they
should
be
treated.
I
However,
the
the
embedded
CPU
and
all
the
the
other
memory
that
essentially
is
is
required
for
the
Sonic
and
Sonic
execution.
That
is
again,
it
should
be
monitored
through
the
telemetry,
which
essentially
has
a
gnmi
interface,
that
that
we
can
actually
use
right.
I
Yeah
I
agree:
I,
think
that
you
know
it
does
actually
require
this
kind
of
a
the
API
or.
However,
however,
it
may
be
right
much
like
how
you
have
any
platform
apis.
It
should
be
the
it
should
basically
be
coming
through
the
platform
apis
today.
What
gnmi
is
providing
through
the
Telemetry
is
eventually
is,
is
mapping
to
some
platform
API,
to
give
you
the
the
CPU,
Telemetry
or,
or
you
know,
fan
power
supply
whatever
be
the
case
right
same
thing
is
going.
B
To
happen
here
right,
but
I
think
when
it
comes
to
accelerator
I,
think
it's
opening
up
the
question
a
bit
because
you've
got
stuff.
That's
a
lot
more
intensive
and
detailed
than
fan
and
power
supply
and
temperature
right.
It's
almost
like
data
plane
platform
stuff
and
it
could
be,
it
could
be
represented
by
an
API
and
just
remove
the
ambiguity
because
I
think
saying:
well,
we
have
a
Northbound
interface
gnmi,
give
us
all
this
stuff
and
not
helping
specify.
Let's
say
the
southbound
interface
to
the
data.
I
B
Which
would
help
you
with
things
like
you
know,
complex
table
allocation
and
memory,
pool
items
and
and
CPU
performance
inside
of
a
dpu,
I
shouldn't,
say
CPU,
dpu
or
accelerator
items.
There
might
even
be
architecture,
specific
items
that
can't
be
exactly
standardized
but
they're
still
of
Interest
right.
E
B
A
C
I
haven't
heard
anything
I
know,
I.
Think
Anton
made
some
comment
there
that
he
fixed
a
couple
things
but
I
haven't
heard
anything
regarding
that.
C
A
Okay
thanks
thanks
so
much
you
know,
Gohan,
we
had
an
interesting
conversation
yesterday.
A
Was
it
reshma
honey
for
all
of
us,
John
Kearney
about
when,
when
we're
not
looking
at
the
DP
specifically,
but
when
we're
looking,
which
has
you
know
the
two
ports
but
we're
looking
at
a
Smart
Switch
implementation
where,
where
you
know,
there's
multiple
up
to
30
something-
and
you
know
the
packets
come
in,
and
the
in
transit
Port
one
and
we're
they
need
to
pick.
You
know
how
they're
going
to
egress.
A
We
were
thinking
about
how
the
selection
happens
there
as
we
move
forward
and
we're
wondering
if
we
had
thought
about
that.
Much
in
the
future
in
the
future,
I
was
wondering.
Is
that
something
we
talk
about
here
in
this
meeting.
H
H
A
Know
describe
yeah,
it's
basically
the
decision
for
the
exiting
part
and
it's
not
not
just
the
dpu,
but
when
it
comes
to
like
a
Smart,
Switch
form
factor
and
and
I
kind
of
talked
to
Michael
about
how
it
works.
Today
you
know
with
the
ports
and
the
load
balancing
and
the
cross
wiring,
but
freshman.
Maybe
can
you
ask
the
question
or
make
the
statement.
D
Sure,
very
quickly
we
were
discussing
in
the
behavioral
model
meeting
that
for
the
packet
that
egresses
after
the
Capa
mapping,
we
would
need
to
update
the
MAC
address
and
we'll
need
to
implement
the
basic
routing
in
the
BMV
too.
So
there
were
questions
whether
we
can
just
to
default
routing
and
so
on
and
so
forth
or
yeah.
There
were
questions
about
how
the
packet
aggresses,
whether
it
should
egress
coming
in
from
one
port
egress
to
another
port,
or
should
there
be
ecmp
should
there
be?
D
F
I
think
we
have
one
section
in
the
the
sonic
hld
that
covers
this
underlay
routing.
Did
you
get
the
chance
to
look
at
that
I
mean
if
there
are
updates
required.
We
can
modify
that
section,
but
let.
B
D
Also
saying
something:
I,
oh.
H
No,
no
I
was
just
saying
that
you
know
Prince
has
that
I
think
the
underly
routing
requirement
to
find
right
so
yeah.
G
A
Got
it
got
it
okay,
so
this
is
the
appliance
and
with
that
map
also
to,
for
example,
a
switch
form
factor.
H
D
So
basically,
we
need
to
update
the
destination
Mac
in
the
packet
when
we
are
doing
the
routing
there.
That
part
is
not
implemented
in
bmv2,
so
there
were
some
questions
is
Mukesh
on
the
line
or
Vincent
might
also.
C
F
So
I
think,
if
there
is
no
mapping
on
the
one
that
we
have
captured
here,
is
like
use
the
incoming
packet
source
and
destination
and
just
swap
it
and
send
it
back
on
the
same
port.
F
I
thought
you
were
asking
to
how
to
find
how
to
derive
the
destination
Mac
right,
correct,
yeah,.
D
Yeah
some
more
details
on
that.
If
we
can
add
it
will
be
good
how
it
is
connected,
we
heard
from
Marianne.
Maybe
it
will
go
to
a
gateway
gateway
to
Port,
something
like
that,
and
you
know
whether
to
use
default.
Gateway
things
like
that,
so
some
more
detail
on
that
would
be
good.
We
don't
have
the
MAC
address
part
resolved
yet
so
that
needs
to
be
implemented
in
BMV.
So.
H
H
J
Think,
what's
ambiguous
in
here?
Is
it's
not
clear
like?
Does
this
mean
that
there's
either
just
a
single
default
route
or
no
default
route?
Or
does
this
mean
that
there's,
like
you
know
a
full
prefix
table
and
that
may
or
may
not
have
a
default
route
like
it's
not
clear
from
like,
like
I,
think
this
is
trying
to
describe
something
simple
but,
like
you
could
interpret
this
as
meaning
like
you
just
need
to
do
full
routing.
F
In
the
underlay
we
can
have
a
default
route,
that's
pointing
to
a
Gateway,
but
it,
but
even
if
it
is
not
present,
then
it's
like
like
it's
mentioned
it
can.
It
can
derive
the
the
mac
and
and
Port
from
the
receiver
package.
So
I
think
maybe
we
can
have
a
conversation
to
see
all
the
all
the
scenarios
and
and
I
can
update
the
document
here
to
cover.
C
H
Yeah
correct,
correct,
I
think
maybe
the
clarification
is
that
you
know
they
don't
know
whether
the
plans
will
get
full
routing
table.
I
mean
you
know.
Maybe
it's
one
route,
it's
doable.
You
know
if
it's
a
full
routing
to
a
whole
LPM
blue
card
table.
Maybe
that's
not
necessary.
I!
Guess
that's
the
question
you
know,
then
we
can
clarify
that
you
know
you
know
for
the
underlay.
You
know
we
only
put
the
default
around
there,
so
they
just
do
the
default
route
and
to
the
ecmp
harsh
and
send
it
out
right.
So
so.
D
D
That's
exactly
what
we
need
Gohan,
whether
we
need
a
default
route,
what
kind
of
ecmp
we
are
expecting
and
how
to
update
the
destination
Mac
address
and
that
detail.
If
we
can
add
that
will
I'll
be
very
clear
to
everyone.
H
Something
I
think
maybe
Prince
I
think
it
seems
clear
when
you
seem
to
clarify
that
exactly
one
right,
so
you
know
get
rid
of
all
this
full
rounding
tables.
You
know
that
let's
clarify
that
right,
so
there
should
be.
We
do
not
expect
the
underlying
have
a
full
routing
table.
There.
H
And
then
you
can
explain,
you
know
the
you
have
the
default
route
and
then
pointing
to
the
next
hop
right.
So
it
will
be
two
next
top
and
then
from
next
up
you
get
the
neighbor
and
then
get
the
MAC
address.
Therefore
you
know
you
get
the
MAC
address
and
and
saying
and
then
you
will
have
the
you
know
the
gpmp
right
so
in
terms
of
ecmp
I
guess.
J
H
H
I
mean
even
even
in
the
Smart
Switch
pipeline
I
I,
think
that
that
the
scenario
is
also
possible
because
you
know
I
understand
you
know
that's
an
internal
link,
but
we
do
observe
you
know
CRC
arrows
and
those
internal
links,
and
in
that
case
we
do
have
a
mitigation
to
shut
down
those
internal
links
right
so
I
think
that's
possible.
C
J
Look
I
I
mean
from
from
our
perspective.
We
can
do
anything
okay,
so
it's
not
like
it's.
The
the
pushback
is
just
to
try
to
make
things
simple.
It's
not
like
that!
It
can't
that
that
it
can't
be
done.
It's
just
that
if
the
actual
requirement
has
like
constraints
that
allows
it
to
be
simple,
we'd
like
to
exploit
that.
If
the
actual
requirement
says
it
has
to
be
completely
flexible
and
do
everything
that
that
a
router
can
do,
then
that's
just
more
complicated
I.
H
Think
it's
totally
agree.
Okay,
how
about
this
right?
So
I
think
you
know
we
do
have
a
prince.
We
do
have
some
idea
right.
So
let's
bring
the
proposal
on
the
table
right
so
because
I
think,
if
you
don't
have
a
proposal,
then
people
don't
have
a
way
way
to
start
discussion.
So
let's
bring
a
proposal,
a
specific
proposal
on
the
table
and
we
can
discuss
around
it
right
so
and
then
see
you
know.
Maybe
you
know
there
will
be
some
pushback
on
some
points
and
let's
discuss
and
close
okay.
H
No,
no,
let's
not
not
go
there
right.
So
we
said
we'll
we'll
go,
go
with
a
special
proposal
and
we'll
discuss
that
I
can
I
can
imagine
there.
There
are
some
concerns
that
you
see
in
piano
like
okay,
do
you
need
to
looking
inner
header
all
these
kind
of
things
right?
So,
but
let's,
let's
bring
a
specific
proposal
that
everybody
knows
what
exactly
it
is
and
then
we
can
discuss
right.
H
H
No
I
mean
that's,
that's
a
point.
I
want
to
discuss
right,
so
you
know
so.
J
The
the
thing
about
Dash
is
that,
like
it's
symmetrical
right,
the
amount
of
Ingress
traffic
is
like
equal
to
the
amount
of
egress
traffic
right.
With
the
exception
of
some.
You
know,
policy
drops
or
whatever,
and
if
you're
doing
ecmp
into
the
dash
data
plane
then
like.
Why?
Wouldn't
that
be
the
same
basis
for
doing
ecmp
out
of
the
dash
data
plane
like
it
just
seems
like
if
it
was
your
intention
to
do
ecmp
into
it,
you
could
use
like
whatever
Port
the
packet
came
in
on
as
the
basis
for
the
port.
J
H
J
J
H
Yeah
because
because
you
know
a
lot
of
things,
trade
off
right,
so
there's
different,
you
know
things
to
consider
yeah,
so
we'll
we'll
bring.
You
know
the
proposal
and
justifications
and
and
we'll
discuss.
H
I
think
not.
A
A
So
it's
9
46.
We
have
14
minutes
left
on
any
anyone
else
on
the
call
have
anything
that
they.
D
Want
to
cover-
maybe
we
need
some
more
time
so
if
we
can
discuss
in
the
behavioral
model
meeting
or
any
other
meeting
next
time.
The
scale
aspect
that
we
discussed
yesterday
yeah.
A
Number
two
remember:
two:
yesterday
we
were
looking
at.
You
know
that
bigger
table
was
that
was
the
hero
of
test,
which
that's
just
a
test
to
stress.
You
know
that.
B
A
Numbers
used
to
stress,
and
then
the
scale
and
requirement,
if
you
want
to
call
it
that
is,
is
set
in
the
other
documents.
The
Sonic
hld
hld
and
the
the
packet
transforms
doc.
So,
but
yet
Niraj
was
going
to
help
me
with
that.
He
said
and
make
sure
we
go
through
it
all
together.
So
sure.
A
A
Yeah
and
then,
if
we,
if
I,
go
back
and
go
into,
this
is
document
here.
This
is
the
other.
You
know
scale
per
dpu,
slash
card
and
since
March
is
on
the
call
I
will
pop
into
the
document
where
we
have
it
all
laid
out.
This
was
here.
G
D
Maybe
it
was
enjoy,
maybe
to
someone
else,
sorry
about
that.
The
interpretation
was
that
the
200
gig
dpu
when
we
mentioned
these
numbers,
they
were
per
h
a
pair,
but
if
it
is
per
card,
that's
fine
too.
Just
want
some
and.
G
My
my
understanding
here
is
that
64
numbers
and
these
numbers
will
be
when
one
of
them
goes
down.
So
basically
under
normal
usage,
Condition,
it's
32
with
half,
but
when
one
of
them
goes
down
in
the
you
know
keeps
in
and
one
is
to
take
over
all
the
work
it
needs
to
support
up
to
64..
So.
D
During
failover
got
it
yeah
that
makes
sense
yeah.
We
can
confirm
that
then
it's
fine
John
did
you
have
any
other
question
related
to
this
yeah.
J
I
mean
I
think,
like
my
my
concern
was
more
with
the
connections
per
second
and
how
to
interpret
connections
per
second
within
the
context
of
right,
and
does
it
mean
that
you
know
one
dpu
is
getting
3.75
million,
the
other
dpu
is
getting
3.75
million
and
then
they're
each
sinking,
they're
3.75
million
to
the
others
dpu
or
is
each
getting
half
of
that
and
then
syncing
that
to
the
other
one
and
I
think
at
least
pin
Sando
yesterday
said
that
in
the
context
of
H
A,
the
3.75
was
for
the
combined
pair
of
of
data
planes,
and
so
you
know,
that's
I,
think
I
think
we
said
yesterday
that,
like
that,
should
just
be
clarified
in
this
in
this
table,
because
it's
not
clear,
you
know
what
the
3.75
million
means
within
the
context
of.
H
I
agree,
I
certainly
agree.
So
did
we
get
the
clarification
on
that.
D
H
A
We'll
need
to
talk
to
to
Michael
and
Joey
again.
This
is
the.
This
is
just
the
hero
test.
It's
just
a
test.
It's
but
we'll
we'll
get
clarification,
because
when
this
was
written,
I
don't
know
that
we
had
considered
h
a
in
the
in
the
scenario.
Had
we
mercha.
G
For
the
scale
numbers
for
sure
h,
a
was
considered
because
the
discussion
was
that
not
under
normal
operation,
it's
32
Enis,
but
when
one
of
the
currents
goes
down
and
the
other
needs
to
pick
up
all
the
load,
we
need
to
be
able
to
load
in
memory
all
the
64
emis,
since
it's
32
from
1
car
32
from
the
other.
So
the
test
is
testing
kind
of
the
worst
case
scenario,
but
under
normal
circum,
normal
operation
will
be
half
of
these
values.
You
know
day
to
day
use,
but
again
the
test
test.
G
J
Yeah
I'm
going
on
what
pensando
said
yesterday
so
I,
but
but
if
that's
not
correct,
then
that
we
need
to
know
that
okay.
A
The
other
comment
yesterday:
yeah
yeah,
okay,
well,
sorry,
guys,
I'm,
still
sick,
good,
talk,
good
talk
unless
we
have
anything
else
for
today.
Anyone
else
Chris
did
you.
She
had
something
on
Deck
before
at
the
end
of
December.
Were
you
thinking
you
wanted
to
do
something
in
the
next
couple
of
weeks.
A
B
K
A
question
here
like
I,
said
for
the
dash:
do
we
have
any
like
they
say
Milestone
or
targeting
honestly,
when
we
have
the
API
will
be
released
or
those
kind
of
like
I,
say
Target
date
or
anything
like
we
can
leveraging?
This
is
the
edit
from
the
Alibaba
side.
We
have
a
couple
we're
also
interesting
to
using
the
dash
apis
in
our
project.
Oh.
H
Oh,
we
have
some
API
already
defined,
I
mean
you
know.
Maybe
we
can
talk
offline.
You
know
those
things
yeah.
K
Yeah
yeah
yeah,
okay,
yeah,
basically
I
heard
like
say
a
Microsoft
that
has
some
the
Smart
Switch
will
be
released
in
this
year,
so
it
will
be
like
the
early
zero
and
then
later
this
year
this
year,
oh
sorry,
next
year,
the
top
2023
so
I
just
wondering
like
from
timeline
point
of
view,
will
we
be
able
to
leverage
anything
from
here
or
because
right
now,
I
always
see
the
dash
is
more
focusing
on
the
the
vi
use
cases.
While
Alibaba
was
more
looking
for
the
necking
and
the
the
no
balancing
use
cases.
H
I
think
maybe
you
know
one
thing:
I'm,
sorry,
I'm
I'm,
if
you,
if
you
would
you
know
if
you,
if
it's
a
if
it's
possible
that
I
mean
it
would
be
great?
If
you
can,
you
know,
describe
your
scenarios,
you
know
in
this
forum,
so
we
understand
that
the
weather,
you
know
whether
your
scenario
is
one
of
one
of
them
has
already
been
discussed
right,
so
they
they
yeah.
That
would
be
great
yeah.
H
K
Plan
to
I
had
a
chat
with
the
Christina
before,
like
we
probably
want
to
have
one
slot
like
to
describe
that
about
the
use
cases
all
right
yeah.
So
we
can
continue,
probably
like
say
maybe
the
new
year
after
the
January
sure
like
we
pick
in
the
first
or
second
like
say
week
of
the
January,
we
would
present
not
about
my
use
case.
Okay,.
H
And
maybe
you
can
work
with
Christina
to
close
that
date
and
we
can
put
down
the
agenda.
K
K
A
K
A
I'm
gonna
go
ahead
and
stop
the
recording
then
and
don't
get
mad,
but
I
forgot
to
record
the
meeting
yesterday.
So
that
was
my.