►
From YouTube: DASH Behavioral Model WG Jan 5 2023
Description
DASH API in SAI Master - Reshma approved this morning
Add DASH API by marian-pritsak · Pull Request #1590 · https://github.com/opencomputeproject/SAI/pull/1590
DASH/pull/310 Add ADDR_FAMILY attr generation (for any table that has IP address, IP prefix, or IP prefix list as a key)
https://github.com/Azure/DASH/pull/310
API used/needed for SONiC CRM (Critical Resource Monitoring)
SAI API to query resources left in the ASIC for sai_object_type
github.com/opencomputeprojct/SAI/blob/master/inc/sai.h
API allows you to look more closely into resources that are left
Can raise alarms in the db
Underlay Routing - in DASH we have complex LPM routing and the 2 default routes w/2 ToRs.
Do we need to implement another LPM table?
No: this is possible to implement w/o an extra LPM. Implementation can perform the hashing on the 5-tuple & algorithm? Can we assume ECMP-in and return packet on the same port w/o a hash? (We have a default hash assigned by the SAI definitions - check opencompute project). What we DON'T want is for an implementation to PREFER a port. We can either 5-tuple hash or send out the same port (add to HLD).
Is there a way to have 2 default routes & decide routing w/o ECMP?
We can clarify in HLD the default route we learn via BGP from the peer.
Expectation is not a lookup and route (take care of both cases w/o a lookup in the DPU)
1. default route
2. ECMP
What if there are multiple hops out of a port? (Not what Azure does, however it is a good question).
Using this tech in another enterprise w/o a CLOS (equal cost) network, they may have to design this differently via the DPU implementation.
Next week come to the call with ST/PL questions if you all have any :)
A
A
B
A
You're
welcome
you're,
welcome,
yep
and
Ryan
last
time
we
were
here
and
you
weren't.
It
was
before
Christmas
and
we
were
talking
and
what
were
we
talking
about.
Gerald
was
maybe
going
to
look
into
getting
some
support
internally
for
bmv2
and
we
had
some
people.
We
have
some
tests
that
we
have,
that
are
we're
marketing,
don't
run
in
bmv2
and
that's
what
I
remember
talking
about
and
then
intel
was
telling
us
about
their
P4
dpdk
support.
A
C
A
Problem
no
problem,
I
was
just
displaying
the
notes
from
last
time
before
we
left
for
Christmas,
so
and
yeah
so
I
think
did
we
have
something
yesterday
in
the
community
call
that
we
were
going
to
talk
about
today
in
the
in
the
behavioral
model,
meeting
I'll
have
to
switch
screens
and
stop
sharing
my
screen.
That's
anyone
recall
what
that
was
or
I
can
pull
it
up.
D
C
D
D
D
All
right,
so
it's
this
is
a
small
change
to
the
psychiato
generation,
doesn't
really
change
anything
for
the
from
the
configuration
point
of
view,
it's
adding
a
read-only
attribute,
and
this
is
needed
for
for
Sonic
CRM,
which
is
using
the
size
CRM
apis,
so
that
will
go
to
site.
There
is,
let
me
find
it
sorry
object,
I
guess.
So
there
is
an
API
to
query
how
much
resources
are
left
in
the
Asic
per
object,
type.
A
D
All
right,
I'm
talking
about
this
Psy
object,
type
get
availability,
so
this
API
allows
you
to
see
how
my
our
last
per
object
type
and
in
that
we
have
a
new
object.
New
object
types.
Of
course,
new
entries,
like
outbound
routing
entries
in
bond
marketing
entries
I,
want
CA
to
p8
mappings,
inbound,
PA
validation,
entries,
HCL
rules,
ACL
groups
Etc,
so
those
those
tables
can
be
their
remaining
size
required
with
this
API.
D
So
you
pass
the
dpus
switch
ID
that
was
generated
during
the
API
initialization
and
creating
a
new.
Can
you
pass
object?
Type
of
your
interest
like,
for
example,
album
route
entry
and,
additionally,
you
can
pass
attributes
which
are
resource
selectors,
basically,
which
are
provide
a
more
granular
query
to
a
specific,
let's
say,
flavor
of
this
resource
and
an
example
of
that
would
be
all
of
the
entries
that
are
that
have
IP
addresses
as
part
of
their
key.
D
So
when
you
pass
this
attribute
and
say,
I
want
to
know
any
itv4
entry
and
still
create
of
that
type
or
how
many
IPv6
entries
I
can
create
of
that
type.
This
using
this
attribute
the
API
for
availability.
Query
passing
it
over
here.
It
will
give
you
a
different
result
depending
on
the
implementation.
It
could
be.
D
The
same
could
be
different
could
be
like
a
common
memory
pool
and
then
those
two
results
will
be
dependent
on
one
and
another,
and
it
means
that
it
only
gives
you
the
amount
of
entries
that
provided
that
no
other
types
of
entries
will
be
created
or
they
can
be
separated.
And
you
know
that
this
is
your
result,
number
of
entries
so
all
depending
on
the
implementation.
D
But
this
API
allows
you
to
look
more
closely
into
how
much
you
have
left
and
what
Sonic
does
is
that
it
periodically
queries
all
those
object,
types
provided
and
now
it
will
be
able
to
query
them
where
I
T
address
family
and
in
case
the
some
of
the
reasonability
gets
lower
than
a
threshold.
By
let's
say
there
is
less
than
10
percent
of
table
left.
It
will
raise
an
alarm
in
the
database
so
that
controller
infra
can
read
that
and
take
a
court
a
take.
D
D
Oh,
it's
it's
Sonic,
it's
Sonic
implementation
and
yes,
it
is
favorable,
so
you
can
look
into
the
Sonic
critical
resource
monitoring
implementation.
It
has
to
basically
it's
obtainable
things.
First,
one
is
falling
frequency
for
those
resource.
Availabilities
I
think
Focus,
something
like
five
minutes
and
then
threshold
is
yeah.
It
is
also
configurable,
it's
whatever
you
think,
is
safe
enough
in
your
network,
so
you
won't
run
out
of
resources.
D
Yeah,
it
will
automatically
to
add
that,
and
another
thing
is
like
vendors-
are
not
necessarily
required
to
implement
that
right
away.
Sonic
infrastructure
if
some
availability
is
not
implemented,
yet
you
can
just
return,
not
implement
it
and
it
will
just
it's
not
implemented
and
so
on.
So
it's
not
like
anything
will
be
broken.
If
you
don't
have
it
right
away.
D
G
I
H
E
Okay
yeah,
so
this
one
is,
you
know
the
get
API
to
just
to
look
at
the
yes,
so
and
then
all
the
other
logic
is
all
about
the
NOS
level
to
see
that
okay,
you
know
what's
the
use
and
then
what
to
do,
how
often
to
Monitor
and
what
to
do
and
so
forth.
So
this
is
just
what
it
is,
but
thank
you
so
much
I
think
this
is
pretty
good.
So
so
today
you
know
what
I
see
that
Marion
that
you
have
actually
created
a
code
to
Auto
generate
the
API
right
now.
E
D
Yeah
yeah,
it
can
be
done.
I
think.
The
question
here
is
how
original
it
is
for
bmv2,
because
there
are
different
kinds
of
tests
that
we
can
do
like
just
tests
isolated
from
Sonic,
where
you
run
traffic
on
the
mv2,
which
is
directly
configured
through
PSY,
and
another
thing
that
we
want
to
do
is
to
write
in
V2
actual
as
a
as
a
new
Target
in
Sonic
and
in
that
case,
having
this
API
implementation.
D
Actually
that
would
be
more
preferable
because
it
would
it
would
be
used
by
Sonic
and
the
in
the
bmb2
Target.
So
I
am
not
doing
this
implementation
yet,
but
when
we
will
finish
the
integration
with
Sonic,
which
is
I,
think
a
plan
not
only
from
me,
but
also
from
Microsoft
side
to
have
as
like
another
virtual
switch.
E
D
A
site
yeah
yeah,
it
will
be
auto-generated
I
just
didn't
put
that
there
yet
because
so
there
are
two
things
there
is
a
size
of
the
table
which
is
available
from
Q4
runtime.
Also
the
size
I
think
we
can
also
the
number
of
entries
that
are
already
created
before
runtime.
So
implementing
this
can
be
auto-generated
as
well
sounds
good.
D
A
F
A
D
C
C
H
Sorry
I
was
distracted,
but
if
I,
if
I
hear
you
correctly,
you
are
asking
about
the
follow-up
from
yesterday's
meeting
right.
Yesterday's
discussion
on
on
the
underlay,
yes
yeah
I,
think
I
will
clarify
that
in
in
the
actually
with
with
a
few
more
details.
But
from
what
I
understand
we
don't
need
a
an
LPM
table
right
because
it's
just
the
default
route
that
we
are
going
to
get
from
the
pair.
So.
H
G
H
End
of
the
day
we
don't,
we
are
like.
The
expectation
is
not
to
do
a
proper
lookup
and
you
know
match
on
a
longest
prefix
and
those
things
right.
It's
basically
how
do
you
want
to
forward
in
the
underlay
right?
So
there
are
only
two
cases
that
we
need
to
to
consider.
One
is
Regular
routing
like
a
default
route
that
is
pointing
to
one
next
stop.
The
other
is
default
route
pointing
to
two
next
stops
right,
the
ecmp
case.
H
C
H
J
H
So
so
that
also
in
in
the
traditional
switch
switch
world,
we
we
do
like
LPM
lookups
right,
like
even
even
the
ip2me
routes
are,
are
looked
up
in
the
LPM
table
and
then,
if
based
on
the
next
stop,
which
is
for
IP
to
me,
it's
a
it's
a
CPU
Port
it
gets
trapped.
But
that
also
is
not
lookup
that
we
would
need
to
do
in
TPU,
yeah
I
mean.
Does
this
make
sense
or
or
does
anyone
has
any
concern
like.
C
H
F
F
F
K
H
J
Mean
having
a
hash
anxiety
is
not
a
problem.
There
are
plenty
of
them
out
there.
I,
don't
think
Psy
defines
like
do
you
do
three
double.
This
is
five
whatever.
Usually
that's
configurable
on
the
router
in
this
case
here
I'm,
not
sure
sorry
defines
API.
Maybe
it
does,
but
the
bottom
line
is:
why
do
we
even
need
to
compute
a
hash?
I
mean
Computing.
Hash
is
not
it's
not
free
right.
J
F
J
H
E
H
E
You
have
to
basically
go
ahead
and
then,
but
clearly
you
know
you
are
leaving
it
out
onto
the
to
routing
to
to
decide
in
terms
of
you
know
how
it
wants
to
do
the
hashing
or
whatever
the
algorithm.
That
may
be,
underneath
to
to
say
that
you
know
how
am
I
going
to
choose
one
out
of
the
two
in
order
based
on
the
packet
that
I'm,
you
know,
or
a
flow
that
I'm
I'm,
going
to
pin
on
and
I'm
going
to,
pin
that
that
flow
onto
a
single
route.
H
H
Now
for
the
non-default
or
the
the
route,
then
you
will
have
a
a
an
LPM
table
required
right,
so
so
even
the
LPM
triple
what
it
it
reduces
your
final
next
jobs
right.
So
that
is
one
thing
that
we
can.
The
implementation
can
taken
like
you,
don't
need
to
do
a
lookup
in
the
LPM
table
to
find
the
underlay
routing.
So
instead,
based
on
the
information
provided
to
the
the
PSI
and
SDK,
it
can
determine
how
the
packet
forwarding
has
to
like
to
the
packet
to
be
sent
out.
H
E
So
so
Prince,
what
you're
talking
about
is
that
the?
How
someone
is
going
to
implement
the
lookup
part
of
it,
where
the
lookup
is
going
to
happen,
whether
it's
in
the
LPM
table
or
any
other
table,
whether
it
can
be
either?
You
know
the
full
match
or
you
complete
match
or
whatever,
with
the
cash,
whether
it
can
be
hash
table
or
it
can
be
a
LPM
table.
E
B
B
F
F
Yeah,
you
don't
want
that
to
happen,
because
then
you
don't
have
the
bandwidth,
but
you
know
they
could
either
hash
I
think
they
could
either
hash
or
they
could
do
what
somebody
suggested
and
just
send
it
at
the
same
port
because
it's
already
been
hashed
already
by
the
switch
and
sent
to
your
particular
report.
So
it's
already,
you
know
pre-done
that
hash
has
been
done
already,
so
you
could
do
it
that
way
too.
I,
don't
think,
there's
anything
wrong
with
that.
F
B
A
C
F
A
G
K
K
F
Always
the
same,
unless
you
have
a
transient,
you
know
condition,
but
they're
generally
the
same
I
guess
you
know
what
if
somebody
else
did
not
have
that
kind
of
a
network
set
up
where
everything
is
equal
cost,
like
maybe
they're,
using
a
slingshot,
whatever
some
other
network
configuration
and
then
what
would
happen
then?
Would
they
both?
J
A
F
Now,
I
think
that
this
is
not
an
Azure
issue.
We
always
have
equal
costs
right,
it's
a
cost
Network,
but
maybe
other
networks
don't
have
equal
costs,
so
I'm
just
I
think
they're
just
bringing
up
the
point.
That
is
there.
Does
this
just
work
in
Microsoft
and
maybe
not
somewhere
else,
because
they're
not
equal
costs
to
the
destinations
and
I,
don't
know
the
answer
to
that.
But
it's
a
good
thing
to
think
about
from
a
Microsoft
point
of
view,
it's
not
an
issue.
H
H
H
F
H
B
So
in
the
I'm
not
too
familiar
with
the
cost
Network
or
these,
but
if
the
in
the
case
of
if
it
is
not
an
equal
cost,
Network
then
with
the
the
Nic
or
a
dpu
ipu
need
to
implement
the
actually
proper
LPM,
because
there
will
be
different,
I
guess,
sort
of
you
know
the
hierarchy
or
LPM
may
need
to
be
implemented.
Is
that?
What's?
Is
that
correct?.
L
B
L
A
L
A
A
F
Yeah
I
would
just
like
encourage
people
that
the
discussion
yesterday
was
really
good
because
it
introduced
the
two
new
Services
private
Lincoln
service
tunnel
before
next
week.
If
people
have
questions
from
that,
because
there
was
some
questions
but
not
a
lot
of
questions,
if
people
have
questions,
maybe
they
can
send
them
in
before
next
week,
so
that
prints
will
update.
You
know
based
on
yesterday's
questions,
but
if
you
want
to
have
more
questions,
it'd
be
really
good
that
we
could.
F
We
could
make
sure
that
everybody
has
the
same
understanding
of
those
two
services
and
the
other
thing.
I
have
a
question
for
Prince.
If
he's
still
on
yeah,
so
I
was
told
and
I
just
want
to
validate
that
venet
to
v-net
the
apis
that
were
created
for
that
actually
are
sufficient
for
v-net
peering
and
Global
v-net
peering.
H
F
Because
of
that,
do
we
have
any
description
in
hld
just
at
least
describing
you
know,
okay,
so
if
they're
sufficient
what
what?
What
were
those
like?
What
what
is
it
doing
so
that
the
community
actually
knows?
You
know
what
that
means,
because
you
know
especially
Global
v-net
appearing
I
doubt
if
anybody
on
this
call
even
knows
how
that
works.
G
H
F
F
H
We
have
in
as
a
understanding
there,
but
in
the
hld
it's
generic,
like
in
the
in
that
the
schema.
If
you
look
like
you,
look
up
a
particular
route
right
in
in
for
an
eni
and
your
destination
is
either
the
same
unit
or
the
another
winner.
It
it
it's
it's
generic.
So
if
it
is
another
v-net,
it
means
it's
a
peering
scenario,
but
that
part
I
didn't
capture
explicitly
like
if
it
is,
if
it
is
a
different
v-net.
That
means
it's
going
to
a
different
like
it's
part
of
a
peering
configuration.
H
F
H
G
H
E
No,
that
is
great
thanks
a
lot
Christina
for
doing
this,
and
thank
you
for
bringing
this
question
because
so
far
you
know.
I
I
was
thinking
that
okay,
we
were
doing
we
needed
to
win
it
and
I'm
just
trying
to
really
figure
out
what
was
the
difference
between
enet
to
v-net
and
v-net,
bearing
I'm
sure
other
people
know
it,
but
I
I
wasn't
really
sure
about
the
exact.
E
You
know
differences
between
the
two,
and
now
we
have
a
third
one
which
is
global
peering,
so
we
need
Global
peering,
so
it
will
be
great
to
just
you
know,
have
something
very
small,
just
to
say:
okay,
what
are
these
and
what
are
the
differences
among
all
three,
that
way
we
will
know
okay
yeah.
Well,
we
we
already
have
done
all
of
those
things,
but
it's.
F
Just
that
yeah
now
we
know
the
differences
I
have
the
question.
Also
a
lot
of
people
have
a
question
because
they
want
to
make
sure
that
they're
testing
properly.
You
know
if
you're
going
to
do
Global
Vena
appearing.
They
want
to
know
that.
Oh
that's,
a
IP,
V6
overlay,
then
they're
going
to
put
some.
You
know
more
emphasis
on
on
that,
but
you
know
it's
because
we
always
say
these
seven
Services,
which
I
don't
think
are
really
seven,
but
we
always
used
to
say
that
we
have
ticked
about
five
of
them
up
right
now.
F
That
are
mostly
defined.
The
last
two
prints
just
introduced
this
week,
but
you
know
over
the
next
couple
weeks
to
be
able
to
finish
that
and
then
it's
just
a
matter
of
for
the
community
to
know
so
how
many
else
are
how
much
more
is
coming,
at
least
in
the
initial
run
and
I
know
SLB
what
some
people
call
I
think
they
call
it
a.
F
F
F
A
I
think,
what's
a
challenge
for
me
too
guys
is
you
know
we
can?
We
have
the
descriptions.
You
know
like
these.
These
documents
that
I'm
pulling
up
here
it
tells
you
what
it
is,
but
this
does
not
tell
us
as
technical
contributors.
What
is
the
difference
in
the
packet
transformation?
That's
going
to
happen,
you
know
that's
what
I
don't
have
like.
This
is
great
telling
us
what
it
is,
but
at
our
level
when
we're
trying
to
say
okay,
we
need
to
pop
this
header.
A
F
A
H
F
A
A
F
A
F
F
H
F
F
F
A
Yeah
I
have
I,
have
questions
into
Niraj
and
pranjal
to
ask
about
that,
but
yeah
and
I
can
easily
add
that
comment
that
you
just
mentioned.
I
could
pop
it
in
here
no
problem,
but
you
know
this
is
this
is
what
I
have
and
I'm
just
saying
you
know
as
an
overall
description.
This
is
this
is
cool.
We
can
talk
about
what
it
is,
but
when
we're
talking
about
the
Transformations,
that's
what
our
people
need
to
know
so.
F
C
A
F
So
I
think
we
can
do
that.
I
was
thinking
more
the
typical
side.
Are
we
missing?
Does
the
implementers
need
to
know
anything
else
than
what
we
just
said,
because
those
end
caps
seem
to
be
drawn
and
and
we're
going
to
put
them
a
little
bit
more
verbal
stuff
in
there
for
the
first
three
services
and
does
anybody
need
anything?
In
addition
to
that
is
the
question.
E
I,
just
wanna,
you
know
I
mean
so
far
as
you
were
doing
this
v-net
and
we
had
to
be
net
and
we're
not
bearing
it
was
more
of
a
synchronous
right.
You
know
you
would
see
that
okay,
what
sites
are
essentially
doing
the
same
thing
now
as
we
are
going
towards
services
such
as
you
know,
shared
services
we
talked
about,
for
which
we
have
private
links
and
service
tunnels.
Is
this
also
symmetric?
In
other
words,
the?
Are
we
expecting
the
services
to
be
running
on
the
dpus
as
well?
E
A
C
E
E
F
Yeah,
the
one
thing
keep
in
mind
that
even
even
with
peanut
beanet,
maybe
one
VM
has
is
enabled
with
Sirius,
doesn't
mean
the
other
sites
enabled
both
serious
you.
Don't
ever
have
that
kind
of
symmetry.
You
can't
be
guaranteed
that
there's
a
dpu
anywhere
else,
you
might
be
the
only
dpu
enabled
VM
in
the
entire
network
for
all
it
cares.
So
it's
the
same
thing
with
with
storage
I
mean
maybe
it
has
a
dpu
someday.
F
H
Now
I
think
the
key
point
is
the
the
VM
that
is
the
source
like
where
the
dpu
is
servicing
right.
The
VM
sends
the
packet
to
the
dpu
and
dpu
sends
it
to
this
shared
service
like
storage.
The
return
path
also
comes
through
this
dpu
for
this
VM,
so
the
VM
is
the
one
that
is
serviced
by
the
TPU
and
not
like
it's
not
guaranteed
that
other
site
also
has
a
dpu.
F
F
But
if
you
have
questions
like
this,
like
so
I
was
saying,
if
you
guys
have
questions
or
you're,
not
sure
about
how
those
two
Services
work,
write
them
down
and
get
them
in
before
next
week,
and
then
maybe
Prince
can
be
prepared
to
answer
some
of
them.
Because
I'm
sure
people
have
questions
for
sure.
But.