►
From YouTube: DASH Behavioral Model WG Dec 8 2022
Description
Vincent:
SAI Build = x86 specificity (SAI Thrift Server?) - if there is a problem we can flag this?
Running SAI Thrift on ARM server?
maybe check opencompute/SAI
For reference from SONiC: https://github.com/sonic-net/SONiC/blob/master/doc/platform_api/new_platform_api.md This is the HLD for platform 2.0 APIs
A
A
A
A
A
Great
yeah,
so
the
last
week
and
reshma
can
help
me
out
here,
but
there
was
a
conversation
around
you
guys.
That
said,
you
know
the
deployment
topology
for
the
Microsoft
would
be
helpful.
When
you
know
packets
were
coming
into
the
dpu
and
transiting
the
ports
and
how
do
we
select
the
egress
support
and
I
needed
to
take
this
to
the
sdn
team?
A
Etc
and
a
dpu
is
different
than
an
appliance
than
a
Smart
Switch,
and
how
are
we
going
to
choose
and
I
did
take
to
the
sdn2
and
luckily
Gohan
and
Prince
joined
the
call
yesterday
and
pointed
out
that
in
the
Sonic
hld
here
in
this
documentation,
if
you
go
down
to
three
I
think
it
was
3.3.5.
Something
like
that.
Let
me
scroll
down,
keep
going
under
layer
routing
here
we
go
so
it
talks
about
whether
it's
an
appliance
or
or
even
a
Smart
Switch.
A
You
know
this
was
how
we
were
thinking
that
the
the
egress
Port
would
be
selected
Etc
and
this
section
here,
and
it
turns
out
that
you
know
we
maybe
needed
a
little
more
clarity
that
was
going
to
come
from
I
believe
Prince
was
gonna.
Add
a
little
bit
more
clarity
into
this
part
of
the
document,
but
we
were
talking
about
okay.
Do
we
because
we
wanted
to
update
the
destination
Mac
and
bmv2
and
like
do
we
need
a
full
prefix
table
full
routing?
A
Are
we
getting
a
full
routing
table
full
LPM
lookup
table?
Are
we
trying
to
be
a
router?
Can
we
optimize?
You
know
blah
BL,
this
long
conversation
right,
and
so
we
figured
out
that
we
maybe
want
to
update
this
section
here
to
talk
a
little
bit
more
about
that.
So
I
wanted
to.
Let
you
guys
know
we
did
kind
of
follow
up.
With
that
conversation
there
did
I
get
that
right.
Reshma,
yeah,
perfect!
Yes,.
E
A
A
A
A
F
C
A
A
A
A
F
H
F
E
A
E
B
Well,
what
what
is
let's
see?
First
of
all,
this
is
just
a
behavioral
model
right
so
and
then
I
guess
of
all
these.
After
all
these,
you
know
artifacts
or
just
to
do
a
simulation,
I
guess
on
a
PC
and
I
guess:
I,
guess
it
there's
an
assumption
that
yeah
x86,
maybe
I,
don't
know
if
we
need
to
be
able
to
simulate
it
when
the
simulations,
in
other
other,
let's
say
server
or
workstation
kind
of
environments.
I,
don't
know.
Is
there
a
demand
for
that
I.
H
F
B
B
You
my
Titanic
tongue,
got
in
the
way
and
that's
an
inside
joke
so
I
I.
Guess
if
there's
a
specific
problem,
you
identified
be
good
to
flag
it.
You
know
and
maybe
figure
out,
there's
a
motivation
to
do
something
about
it,
but
any
anyone
is
welcome
to
enhance
things
to
work
on
different
platforms
if
they
want.
H
B
G
B
Mean
sure-
and
you
know
it
might
actually
Point
all
the
way
back
to
the
PSI
repository
itself
or
the
sci
Thrift
server.
You
know
all
that
code
really
lives
there
and
I'm
just
borrowing
it
so
to
speak.
So
that
might
be
a
good
conversation
to
eventually
held
it
to
the
cadet
community.
And
if
you
come
up
with
some
improvements
that
would
be
awesome.
A
G
A
G
A
Not
on
Hardware
necessarily,
but
you
know
virtualized,
and
so
this
is
our
project
and
our
work
item
task
list
for
that
we
all
work
on
and
down
below,
are
the
items
we've
completed
and
then
the
bugs
we've
solved
and
then
the
items
that
still
need
work,
if
that
makes
sense
so
and
then,
as
we
as
we
go
through,
we
try
to
get
resources
to
figure
out
how
to
get
this
moving
forward.
So
if
you
have
any
questions,
let
us
know
yeah.
G
B
Well,
no,
the
data
plane
model
is
in
P4,
but
the
platform
elements
like
bmv2
itself,
which
is
you
know,
C
plus
plus
code
and
then
there's
the
compiler
tool
chain
to
handle
the
new
language
constructs.
So
that's
p4c,
so
there's
p4c,
compiler,
again
behavioral
model
and
then
our
data
plane
code
and
those
are
all
the
moving
parts
yeah
yeah.
Well
then,
the
apis
are
on
top
of
all
that
right.
All
the
client
software
Etc
yeah.
B
I
just
wanted
to
continue
the
conversation.
I
started
yesterday
with
a
little
more
clarity.
I
I
brought
up
in
yesterday's
meeting
the
desire
to
have
a
few
kpi
metrics
readable
so
that
during
testing
we
can
understand
the
platform
and
I
just
pasted
down
to
the
chat
two
metrics
that
the
mercha
indicated
would
be
very
useful
for
him,
and
that
would
be
the
dpu
data
plane,
compute
utilization.
So
you
might
think
of
that.
B
As
you
know,
deep
you
percentage
loading
and
dpu
data,
plane,
memory,
utilization
and
and
the
reason
those
are
of
Interest
would
be
as
we're
creating
really
really
large
scale
tests.
It
can
be
very
useful
to
see
how
close
we
are
to
100
percent
and
we
provide
that
feedback.
Even
to
you
know
the
vendors
engineering
teams
and
say
you
know
we're
running
this
test
and
we're
getting
up
to
90
or
95
or
80.
B
It's
it's
valuable,
and
if
we
could
consider
these
two
items-
and
maybe
there'll
be
other
ones
that
come
forth
as
readable
through
the
side
layer
and
not
have
to
depend
on
Sonic
for
all
statistics,
that
would
be
valuable.
In
fact,
Sonic
wouldn't
necessarily
present
these
anyway,
because
you're
not
really
part
of
Sonic
platform.
Yeah.
B
B
A
B
B
Yeah,
like
DQ
platform,
attributes
or
something
better,
probably
read
only
but
maybe
there'd
be
a
case
for
some
read,
write
attributes
at
some
point.
If
we
want
to
control
the
platform
a
little
bit
so
I
think
this
would
fit
into
that
effort
of
the
the
general
effort
of
adding
new,
fixed
Dash
related
apis.
C
B
G
C
B
B
Yeah
CRM
could
even
pull
these
right
later
in
Sonic,
but
for
testing
it
would
be
valuable
to
happen
so
yeah
I,
don't
I'd,
be
happy
to
help
in
this
effort.
I'm
just
looking
for
is
there
General
consensus,
consensus
that
seems
useful.
These
two
simple
items
is
there
any
new
feels
like
they
can't
be
provided,
or
they
don't
make
sense
in
their
platform,
for
example,.
F
I
think
it's
the
memory,
one
I,
don't
think
is
as
much
of
an
issue,
but
the
load
one
is
is,
might
be
very
difficult
to
interpret
what
it
actually
means,
because
right
like
for
example,
say
you
know
the
dpu
is
running
at
like
50
of
its
capacity.
You
know,
processing
capacity
like
that's
under,
like
the
current
like,
like
offered,
you
know,
load
but
like
every
kind
of
like
load
like
you
know,
like
V6,
might
take
more
processing
capacity
than
V4.
F
B
F
F
Think
that
that's
like
doable,
I
I
think
that's
doable.
I.
Think
like
that.
You
know
like
if
it's
like
a
number,
that's
like
a
percentage
or
something
like
it's
just
really
kind
of
giving
you
sort
of
like
a
an
idea
of
where
things
are
rather
than
like.
Knowing
like
with
certainty.
If
it
says
50
that
you
could
like
double
the
offered
load
of
exactly
the
same
kind
of
traffic
and
they
that
it'll
fit
it'll
work
and
it'll
fit
sure.
B
Understood
in
fact,
there's
going
to
be
a
lot
of
heuristic
results
and
I.
Imagine
one
if
we
had
these
two
numbers
available
on
every
platform.
As
we
tested
different
platforms,
we
probably
have
graphs
showing
here's
this
test.
We
ran
and
by
the
way
here
was
these
two
numbers
indicated
and
it's
just
information
right.
It.
B
E
B
B
Yeah,
but
there
may
be
even
implementations
that
don't
even
distinguish
like
maybe
there's
one
core
out
of
128,
that
happens
to
be
the
control
CPU,
but
the
you
know
the
overall
stats.
How
do
you
lump
them
all
together
and
average
them
or
produce
some
number
I?
Don't
know,
but
you
know
Lisa
got
a
conversation
going
about
it
and
we
there's
no
urgency
here.
I.
Think
as
as
we
talk
about
this
over
time,
I
think
it'd
be
useful
to
come
up
with
something
at
least
see
if
it
sticks.
C
For
example
right
you
are
talking
about
like
power
supply
and
all
of
those
things.
So
if
you,
if
you
want
to
have
this
thing,
then
essentially
you
know
we
need
to
have
that
thing
be
approached
through
PSI
itself.
That
would
make
sense,
because
of
the
fact
that
you
know
it
will
vary
from
one
vendor
to
the
other.
It
will
not
be
a
different
thing.
So
I
think
this
is
something
that
you
know.
We
can
probably
have
an
item
to
discuss
and
then
come
up
with
some
as
a
community
right
and
then
say:
okay.
B
A
Let
me
think
here:
I
think
that
is
about
all
I
had
for
the
day
until
we
can
get
Marion
on
the
call
to
update
his
items.
I
can
try
and
get
with
him
separately.
I've
had
yusuf's
kind
of
been
helping
going
through
closing
the
items.
When
you
have
an
issue,
you
actually
have
to
go
ahead
and
close
the
issue,
and
then
you
have
to
actually
change
the
status
as
well.
So
if
you
see
anything
weird
coming
through
your
email
from
GitHub,
that's
just
us
trying
to
comb
through
and
you
know
close
the
backlog.
D
D
D
D
A
A
A
D
D
A
A
A
Yeah
I'll
just
make
a
comment
saying
discuss
with
the
community
good
to
close
something
like
that.
Yeah,
you
guys
understand
my
shorthand
blah
blah
blah.
Okay,
okay,
thanks
thanks
Yusuf
yep,
no
problem.
Thank
you!
Yeah
yeah!
Do
you
guys
have
anything
else
for
today
or
Andy?
Did
you
have
anything
from
PNA
or
P4?
You
wanted
to
bring
up.