►
From YouTube: Kubernetes Resource Management WG 20170404
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
B
A
E
Cool
yeah,
so
my
parts
going
to
be
really
short
before
I
end
up
soup,
yet
who's
going
to
give
the
update.
I
just
wanted
to
note
that
we
wrote
a
design
doc
to
go
with
the
branch
that
we
have
and
that's
linked
in
the
in
the
agenda
right
now.
So
you
know
after
the
meeting
feel
free
to
take
a
look
and
add
comments,
concerns
questions
all
that
kind
of
stuff
yeah.
Take
it
away.
F
So
this
is
a
design
document,
what
we've
prepared
with
Connor.
So
the
main
update
update
since
the
last
inning
part
is
that
we
handled
all
of
the
sea
group
resource
control
to
the
couplet.
There
is
no
right
inside
the
isolator
we
just
in
our
POC,
enabled
the
let
me
find
it.
It
will
be
right
here
we
enable
the
CP
0
sets
views
and
cubelets
is
itself
is
pinning
the
odds
to
the
specific
CPUs.
F
Somehow
scheduling
this
isn't
decision,
so
maybe
I
will
go
with
the
flow,
so
the
isolator
is
advertising
the
epic
resources,
which
was
discovered
during
the
discovery
phase
at
the
start
of
the
isolator
and
the
then
the
Copic
integer
resources
are
being
advertised
to
the
note
and
then,
when
the
user
of
the
kubernetes
cluster
wants
to
use
the
resource,
for
example,
CPU
said
CPUs:
did
he
just
use
it
in
the
manifest?
And
then,
when
it
comes
to
our
is
laterz
isolator
it
it
need
to
received
the
pods
back.
F
F
We
are
almost
sure
that
we
can
pin
the
spot,
and
then
we
just
specify
the
specific
course
so,
for
example,
one
and
five
just
with
the
format
that
cpu
sets,
are
accepting
and
then
we
return
these
information
in
the
event.
Reply
event
replies
here
with
a
cigarette
resource
and
this
type
and
then
the
cube,
let
is
doing
the
writing
part.
We
enabled
this
subsystem
in
the
calculus
and
clip
container
is
pinging
pots
to
the
CPU.
F
F
I
can
show
you:
maybe
this
one
would
be
better
this
PR
right
now,
okay,
so
this
is
we
just
say
that,
for
example,
in
the
epic
resource
right,
she
would
say
that
this
isolator
is
advertising
cook
or
affinity.
For
example,
if
you
have
eight
cores,
it
will
say
eight
course
and
then,
if
the
manifest
is
coming-
and
it
says
like,
for
example,
I
want
for
the
exclusive
cpus,
the
isolator
will
a
pass
to
cube.
Let
information.
F
For
example,
you
can
take
from
zero
to
three
cities
and
then,
if
there
is
another
pot
which,
for
example,
once
and
and
four
and
of
course
there
outback
integer
resources
will
be
decremented,
because
this
is
how
we
discover
secures
and
then
we
keep
the
state
of
the
CPU
topology
in
memory
and
we
we
always
know
which
cpu
to
give
to
the
pot.
So
we
always
give
a
unique
cpus
to
the
application
right
now.
So.
E
E
You
know
a
certain
number,
of
course
to
each
pod
that
asks
for
it,
and
then
you
know,
if
you
wanted
those
to
be,
you
know
exclusively
used
by
those
user
pods.
You
would
have
to
do
extra
work
like
you
will
CPUs
and
maybe
tanks
to
keep
other
other
pods
off
of
the
node
that
don't
consume
the
oirs,
but
this
was
meant
to
be
just
an
example
of
how
the
API
could
be
used.
D
Yeah
I
mean
this
is
awesome,
I
think
Megan's,
whatever
whatever
is
showing
today,
we
sort
of
understand
how
it's
possible
have
my
custom
custom
hardware,
technology
alignments
actually
there's
also
another
use
case
of
like
feeding
the
neck
additional
hardware
back
especially
clickers.
For
some
other
reasons.
That's
not
Nora
Cupid
trying
to
inject
that
in
or
in
the
right
possible
means
another
open
feature.
Oh
I,
hope,
topology
and
like
identifying
new
hardware,
would
sort
of
work
with
each
other
is
also
a
very
senior
ski
movie.
D
So
saying
that,
like
we
visit
one
use
case
here,
which
is
that
of
like
having
custom
isolation,
policies
based
on
like
hardware
topology
other
use
cases
other
use
cases
that
like,
if
you
have
additional
hardware
devices
but
I,
not
understood
like
titanfall,
how
do
we,
how
do
we
inject
them
into
pods
and
I
call
it
a
PA,
represent
earth
and
I
think
if
IRS
is
sort
of
a
means
for
representing
them
but
I?
How
do
we?
How
do
we
actually
inject
those
resources
like
that's?
E
E
Ok,
I
know
you
have
you
had
time
to
read
it
with
you
and
we
just
put
it
in
there
two
seconds
ago,
but
yeah.
Definitely,
you
know
adding
additional
information
and
you
reply
that
the
isolator
sims
back
to
the
cube.
'let,
I
think,
is
you
know
it
seems
like
it
queer
way
to
add
support
for
for
other
devices
or
you
could
add.
You
know
things
to
the
device
whitelist.
You
could
inject
environment
variables.
You
could
maybe
add
volume
ounce
tube
to
mountain,
necessary
libraries
and
whatnot
I.
E
Think
there's
a
lot
of
possibility
here
and
the
fact
that
the
cubit
ends
up
being
the
writer
of
the
sea
group
kind
of
you
know
simplifies
some
of
the
questions
about
how
all
the
Isolators
are
supposed
to
coordinate,
because
the
cubic
can
encode
a
policy
of
you
know
of
precedence
where
maybe,
if
no
isolator
returns
information
for
a
specific
vector,
then
maybe
the
default
policy
just
takes
over.
For
example,
there
could
be
other
policies
we
could
think
of
to
you,
but
those
give
us
a
place
to
sue
work.
Maybe
yeah.
D
I
completely
agree,
I
think
I'm
post
my
thoughts
or
comments
in
the
Donald.
Okay,
it's
easier
to
coordinate
I,
just
won't
add
one
more
thing.
Since
you
mentioned
policies
you
this
I,
don't
see,
I,
don't
see
the
update
start
account.
You
saw
the
part
creation
workflow,
but
then
there's
the
part
where,
like
as
new
pods
get
added,
you
might
want
to
change
your
previous
allocations
or
you
might
want
to
rebalance,
and-
or
maybe
you
just
like-
keep
continuously
changing
host
of
the
system
is
behaving
based
on
like
a
fellow's
and
things
like
this.
D
So
if
the
culet
are
to
have
its
own
inbuilt,
topology
alanis,
for
example,
it'll,
probably
be
rebalancing
or
and
so
I
don't
see
how
I
thought
by
the
isolator
in
this
case
would
would
be
able
to
rebalance,
but
that's
something
that's
not
something
to
consider
so.
E
There's
a
it's
listed
as
a
possible
future
extension
right
now.
The
only
events
we
admit
we
omit
to
the
isolator
are
pre-start
pod,
which
is
during
pods
here
per
pad
level.
Container
creation
and
post
stop
pod,
which
is,
after
the
less
container
in
the
pod,
goes
terminal,
but
we
could
add
a
sink
because
you
know
we're
catching
calls
from
ensure
exists
periodically,
and
so
we
could
maybe
delegate
to
the
isolator
at
that
point.
Okay,
but
yeah.
There
need
to
be
some
kind
of
coordination.
E
D
D
Natively,
then
the
policy
needs
take
into
account
that
until
a
TI
has
to
reflect
a
lesson
like
if
the
Isolators
is
specifying
a
cpu
expect,
for
example,
and
the
cuba
has
already
made
the
decision
that
shouldn't
be
like
conflicting
situations
on
the
note,
so
there
should
be
a
means
to
like
express
what
capabilities
are
being
granted
to
isolator
and
and
that's
what
I
Slater
should
be
working
towards.
What.
C
A
I
think
the
human,
even
in
the
long
term,
the
keyboard
itself
as
it
becomes
more
I,
guess
apology
aware
needs
to
support.
Grace
release
chatting
off
its
defaults.
Write
you!
So
it's
not
a
matter
of
just
granting
to
the
isolator.
What
you
can
do.
We
all
see
me
to
be
able
to
tell
the
qubit
not
to
do
what
it
might
desire
to
do.
Yes,.
D
I
I
completely
agree
like
how
that
API
would
work
or
something
we
can
think
of.
It
might
just
be
through
configuration
which
is
external
to
the
cubelets
undo,
the
isolator
or
it
might.
It
could
be
some
negotiation
that
the
Cuban
devices
in
order
does
so,
but
that
other
are
the
fashion
that
we
have
to
capture
a
little
bit
at
least,
and
we
can
I
trade
on
that
over
time.
D
But
I
just
want
to
say
it
out
loud
I'm,
hoping
that,
like
as
far
as
the
letting
go
part
of
our
agenda
like
once,
we
decide
on
like
whether
cubed
is
going
to
take
on
some
form
of
like
Harvard
4g
awareness
in
the
near
future
and
if
it
does
and
I'm
hoping
that
like,
we
would
focus
our
energy
on
the
default
implementation.
First
before
we
come
up
with
like
multiple
other
alternatives
to
that,
because
it
will
I
completely
fragment
echo
system
running
and
it's
you'd
be
sending
confusing
signals
to
our
users.
I
think.
A
What
we
need
to
do
we
decide
that
they
keep
what's
going
to
do.
The
default
is
identify
clearly
which
workload
the
qubits
optimizing.
Its
default
for
I
think
it's
very
likely
that
you
know
in
passing
an
extremely
presented
like
10,
different
verticals.
It
felt
like
I'd
like
to
understand
which
of
those
verticals.
Do
you
think
the
key
looks
default
is
best
fit
because
I
think
when
you
guys
discussed
this,
it's
like
the
board
use
case
and
I
think
there's
a
lot
of
verticals
where
that
is
not
the
best
fit
I
I.
A
So
I
think
what
we
need
to
do
before
we
say
like
where
engineers
should
prioritize
their
time
and
effort
like
I'd
like
to
understand
it's
like
it
q.what
default
ideal
for
animation
studios
is
the
khelat
default
ideal
for
network
operators.
Right
is
a
ideal
for
teaching
these
workloads,
so
I
think
realistically,
vendors
are
going
to
have
to
go
do
stuff
in
the
spaces
for
the
personals.
If
so,
when
yeah.
D
A
E
D
B
D
Is
there
doing
able
to
see
the
doc
yep
so
TL
DR
is
that
we
decided
on
a
schedule
that's
going
to
be
from
first
of
may
to
the
third
of
mate.
So
it's
for
monday
through
wednesday,
the
location
was
not
a
new
I'll
go
over,
why
we
decided
the
schedule
and
the
location
so
based
on
initial
survey,
we
got
a
bunch
of
topics
and
given
the
extent
of
topics
we
had
be
because
that,
like
we
need
at
least
three
days-
and
I'm
really
not
sure
we
cover
all
of
tennessee-
is
where
I
keep.
D
This
have
actionable
fork
at
the
end
of
the
three
days
which
we
can
take
and
in
like
we
can
can
continue
working
right
like
if
you're,
not
you're,
not
having
the
issues
that
we
currently
having
with
virtual
meetings
are.
You
know
like
discussing
a
few
things
and
we
go
back
to
our
own
book
and
then
we
come
back
I
like
I
made
progress
so
hopefully
that
that
wouldn't
happen
again
the
same
person.
So
we
decided
that
we
need
at
least
three
days
and
then
based
on
the
based
on
the
feedback.
D
The
alternation
survey
people
are
spread
apart
across
various
reasons,
and
west
coast
seems
to
be
the
most
dominant
one
along
certainties,
who
actually
responded
this
away
and
I
know
for
a
fact
that
erection
who
guess
who
did
not
respond
to
the
survey
because
of
cube
corner
culet
events
that
are
going
on.
So
with
that
in
mind,
we
decided
that,
like
the
area
might
be
the
best
location.
So
for
now
I've
listed
google,
aster
I,
don't
think
there's
a
requirement
is
not
happen
in
Google.
D
D
Sig
was
meeting
in
doing
the
week
of
lemon,
and
there
was
talking
on
and
chippy
to
are
going
to
talk
on
and
then
last
week
of
april
did
not
work
for
a
few
red
hackers
and
meet
fox
valley,
acid,
and
it
was
oscar
and
there's
like
whole
bunch
of
God.
Every
week
basically
has
some
content
or
other.
So
it's
basically
a
question
of
like
which
we
do
we
choose
and
like
what's
the
best
compromise.
D
So
we
decided
that,
like
you
choose
a
first
week
of
me
since
it
worked
for
most
of
our
Nikon
sign-in
book,
very
own
ancestor
I
really
apologize.
If
that's
not
the
best
time
for
some
of
you
but
I
mean
you
really
don't
have
much
options
here.
So
please,
RSVP
I,
think
you
should
have
gotten
RSVP,
so
I
CB
survey
already
RSVP
that,
because
we
have
to
like
make
sure
that
tourists
hosting
has
all
the
names,
and
so
they
can
take
our
security.
We're
going
to
use
this
talk
to
coordinate
the
agenda.
D
If
you
cannot
make
it
two
or
three
days,
we
will
try
to
accommodate
your
topic
of
interest
on
the
data
and
you
would
be
attending.
But
given
the
given
amount
of
interest,
we
have
to
limit
the
number
of
topics
so
I
again,
I
apologize
earlier
on
it
you're
in
your
favorite
topic,
cuz
I'm,
being
business.
We
will
also
try
to
have
a
VC
seven,
but
our
see
personally
speaking,
VCS
haven't
been
great
for
large
meetings
in
the
past.
D
So
if
you
are
really
really
interested
in,
if
you
want
to
actively
participate,
then
I
would
recommend
showing
at
that
person,
but
if
you're
just
going
to
hearing
that
NBC
may
just
work
for
you,
so
we're
curating
the
goals
here,
like
you're,
trying
to
figure
out
what
the
main
topics
are
listed.
A
few
Jeremy
assisted
a
few.
D
D
D
D
A
Okay,
so
I
think
that
covers
the
fifth
is.
Hopefully
a
number
of
people
are
able
to
join
you're
interested
in
this
topic
and
I
really
think
some
the
stuff
that
the
person
impeller
doing
are
helping
to
experience
on
the
topics
would
be
talking
about.
So
thank
you
guys
was
that
I
think
the
next
item
on
our
agenda
was
was
the
pod
precepts
discussion.
Oh.
I
Okay,
so
this
proposal.
I
I
J
I
So
I
kind
of
any
1.6
is
a
really
a
column
objection
policy
object
which
serves
a
very
similar
purpose,
but
it's
not
exactly
I'm
gonna
work
for
us
because
it
just
discriminate
based
on
label.
We
want
to
do
the
same,
but
for
for
when
I
pod
actually
uses
a
resource.
So
in
this
specific
case
a
bunch
of
example,
here
I
want
to
run
tensorflow
under
one
have
to
send
out
no
mo
file.
I
Can
show
you
the
proposed
changes
to
these?
What
you
see
is
the
purpose
of
diabetes
being
existing
object.
It
lets
you
select
path
by
label
and
then,
if
there's
a
match,
but
my
latest
story,
and
then
he
lets
you
inject
environment,
variables,
volumes
and
volumes.
So
the
idea
is
that
you,
as
an
administrator
as
a
user,
create
this
once
per
class
or
per
namespace
and
then
all
the,
for
example.
All
the
parts
they
have
a
specific
label
will
be
modified
at
creation
time.
A
H
A
I
had
to
propose
that
not
only
did
I
want
pod
presets
to
handle
environment
and
volume
out
defaulting,
but
also
resource
requirements,
because
in
that
conflict,
with
what
you'd
be
looking
to
target
cuz
like
if
it's,
if
it's
also
the
one
that's
going
to
be
tossing
resources,
then
it
can't
be
selecting
on
those
resources.
I
guess.
I
A
I
I
I
I
Yes,
I'm
sure
it's
one
of
the
open
issues
that
I
mentioned
and
then
this
enterprise
to
the
pub
presets
right
now
their
pockets
is
applied
in
order
like
oldest
to
newest.
I
believe
so
and
one
example
argue
that
if
you
manipulate
a
path
variable
you're
gonna
be
in
trouble,
because
you
know
you,
one
policy
sets
a
specific
value
for
the
path
variable
and
then
the
next
one
will
just
override
it
and
the
you
here.
I
I
I
D
I
D
Not
sure
if
it's
necessary
to
like
have
the
API
object
represents
all
Sean
cape,
like
collections
I
mean
even
though
it
is
not
like
I'm,
not
sure
that
it
is
already
implemented.
But
there
was
discussion
on
like
having
a
deposit
right
begins.
It
could
just
be
a
custom
defaulter
that
we're
unnecessarily
have
to
represent
okay.
F
D
I
You
know
like
that
would
be
something
like
you
know,
being
able
to
publish
a
nap
on
on
the
Harry
thaw
and
even
have
to
worry
about
the
you
know
the
specifics
of
where
some
directories
are
on
them.
The
notes
really
yeah.
Could
this
live
in
an
admission
controller
controller?
So
the
countries
are
already
isn't
an
extra
mission
controller,
so
the
two
antennas
are
we
either
extend
existing
object
and
controller,
or
we
write
a
parallel
one,
which
does
almost
the
same
things
so
yeah
it's
a
trade-off
there.
I
D
I
D
I
A
A
user
to
you
know,
depending
on
the
case
now,
I
want
to
be
careful
there,
so
the
it's
very
much
an
end
user
exposed
and
today
I
mean
so
when
you
a
lot
it
for
the
background
for
pod
preset
was.
It
was
a
part
of
the
Service
Catalog
effort
where
folks,
who
are
not
where
the
federal
reserve's
catalog,
let
you
go
and
look
at
a
catalog
Osiris's.
A
You
can
provision
one
of
those
images
and
then
you
can
find
one
of
those
services
into
your
pod
and
the
bunny
information
is
the
secret
or
it's
also
default
vine.
There's
a
default
value,
man,
Justin
and
so
the
pod
preset
came
in
here,
because
people
don't
want.
You
know
how
to
mutate
their
pods
guys
because
on
the
services
they're
binding
from
so
at
least
from
anyone
who's
in
the
Service
Catalog,
delicious
ones
like
when
you
create
a
bind
to
a
sort
of
a
pod
resettle
position
for
you
in
the
namespace.
A
I
I
D
I
D
A
H
A
D
D
I
A
I
D
I
could
think
of
like
three
different
recording
systems
that
we
need.
I'm
like
what
preset
seems
to
be
one
of
those
like
a
saying
like
there's
the
translation,
from
what
vp8
of
the
user
speak
to
the
APA.
That
system
understand
so
they're
like
can
charge
for
portable
and
then
the
next
step
is
like
having
some
understand
defaulting
like
to
like
four
pieces
for
nigh
on
impossible,
Thank
You,
the
creative
driving
up,
say,
wallah
semen
on
all
these
boys
are
things
like
enhancing
or
I
think
might
require,
like.
B
D
A
So
that
mine
31
assets,
then
sugar,
visiontek
polyps
on
the
policy.
Okay,.
H
I
K
Just
to
continue
a
bit
on
this
and
the
volume
problem
and
the
binary
problem,
and
basically
they're
clever
and
probably
something
that
we
would
solve
in
the
next
version
well
in
their
solution,
which
can
be
a
container
and
so
I,
don't
think
it
would
be
good
with
our
solution.
It
will
live
in
an
admission
control
it
with
the
anode
by
a
runtime.
Oh.
K
So
you
have
more
control
because
then
you
can
specify
if
you
want,
for
example,
compute
or
graphic
libraries
and
yeah,
you
don't
need
to
specify
the
volume,
because
we're
in
directing
ourselves
under
at
the
difference
when
x,
will
add
the
runtime
level.
So
you
don't
need
to
specify
it
to
darker
or
you
don't
need
to
specify
it
to
rocket
or
if
you
don't
need
to
specify
single
driver
tea
and
we're
just
under
that
y
al.
H
An
angel
enjoyed
our
approach
required
a
wrapper
on
top
of
docker
and
with
the
new
approach.
If
you
can
ask
current
time
and
I
think
your
CI
being
extended,
everyone
is
version,
and
so
we
can
are.
They
need
to
describe
what
we're
in
mind
quickly,
another
the
documents
even
out
you
want
to
stop,
or
maybe
going
to
come
back
to
the
right
time
later.
Well,.
K
D
K
One
of
the
things
we've
well
one
of
them,
the
the
knobs
we've.
Actually,
we
were
talking
with
the
different
users,
deep
learning
teams
and
in
vegas,
and
what
we
think
and
believe
that
the
users
need
is,
of
course,
the
number
of
GPUs.
That's
something
that
needs
to
be
that
that's
already
been
anything,
but
the
different
other
knobs
that
users
usually
want
to
team
and
I
have
a
impacts.
Performance
would
be
capability
and
topology
and
then
usually
different
users
want
to
have
arrow
click,
correction
and
the
minimum
memory
a
GPU
has
and
the
last
thing
being.
K
H
K
I
think
the
minimum
member
GPU
memory
is
something
that
really
needs
to
go
into
remedies.
If
you
see
is
something
that's
that's
going
to
that
shouldn't
be
very
tricky
to
handle.
The
driver
version
is
also
something
that
shouldn't
be
very
tricky
to
handle
and
I
think
the
world.
What
we
believe
that
is
the
most
important
part,
and
where
was
the
discussion
yesterday
is
sick
schedule
have
beat
apologia.
Where
should
the
schedule
at
the
kapalua
and.
A
K
The
number
of
the
number
GPUs
are
global
to
homo.
The
complete
capabilities
are
dependent
on
each
GPU
and
the
GPU
memory
is
dependent
on
each
GPU.
The
eris
error
correction,
ECT,
is
dependent
on
each
GPU
and
the
topology
is
dependent
home
node,
a
la
casa
y
espero
more
than
that
and
the
driver
version
is
dependent
on
the
node.
So,
okay.
H
So
yeah
yeah
to
answer
required
the
previous
question.
We
have
these
cases
where
people
might
not
know
80
GPUs
on
Oscar
nine
gb
is
on
a
node,
oh
I
when
well,
one
is
a
graphic
one,
a
small
one:
that's
not
a
crazy
using
shoot
because
for
tradition
that
and
let
for
computer
to
use
that
our
me
is
not
memory
and
so
on.
It's
quite
uncommon,
but
this
is
a
use.
A
directive.
H
Well
same
default
here
will
be
to
always
try
to
allocate
0
n1
and
actually
most
of
the
time
you
if
you
cannot,
you
don't
allocate
that
let's
say
only
0
and
to
have
it
available.
Someone
asks
for
two,
and
maybe
you
Saint,
default
to
just
waste
or
motive
used
to
be
available
instead
of
having
a
subdivision
but
Russell
Simmons
cases
where
the
GPUs
are
used
independently
of
each
others,
and
there
are
no
cheap
information
in
these
games.
Alex
I,
a
SNP
and
I
gap,
and
back.
K
Another
use
case
would
be
the
loser
specifying
that
he
wants
one
and
be
linked
or
two
ND
ones,
and
in
that
case
you
wouldn't
suffer.
Example.
He
had
two
nodes
and
the
user
user
would
be
able
to
prepare
petunias
on
the
image,
and
it
would
be
one
of
the
things
our
users
would
be
able
to
specify
is
whether
they
want
their
job
too,
and
at
least
on
a
switch
on
a
switch
base.
K
So,
for
example,
let's
say
you
want
a
job
that
takes
two
cpu
without
on
this,
I'm
not
sure
if
you
can
see
my
cursor,
but
it
would
be
on
load
up,
left,
GPU
and
then
all
the
GPU
accident
we'd
want
to
have
the
drop
in
on
cpu
0,
and
that
is
something
that
we
bet.
That
is
something
that
could
be
decided
that
schedule
available.
H
J
May
add
something
you
know,
you
know
your
your
proposal
makes
sense
and-
and
of
course
it's
a
valid
point
that
really
may
want
to
have
fine
parenteral
all
these
gpus.
But
my
question
is
that
how
realistic
is
this
to
think
that
this
is
going
to
be
possible
on
thousands
of
machines
they're
such
finding
a
fine,
fine,
grained
control
over
like
GQ's
and
exactly
which
court
is
going
to
be
allocated
unless
thousands
of
them
thousands
of
machine
running,
probably
more
than
just
one
workload
equipment
right?
J
K
That
we're
already
doing
somewhat
with
eps,
for
example,
you're
already
having
details
on
how
much
memory
you
want
your
job
to
be
available
and
isn't
that,
where
we're
doing
what
we're
saying
that
we
want
to
be
able
to
keep
you
pinning
we're
already
giving
the
user
fine-grained
details
and
we'll
just
saying
that
it
could
be
great
to
have
the
same
details.
Those
fine
grained
details.
A
A
D
Think
I
interpret
body
sticker
slightly
differently
and
I'm
just
going
to
try
to
paraphrase
that.
The
way
I
answer
that
cool
is
currently
supports
are
like
it's
supposed
to
attributes
at
the
node
level.
It
does
not
support
a
tribute
to
the
resource
level
and,
like
GPUs,
are
bringing
up
that
promise
like
we
have
the
resource
attributes
and
there
is
no
means
for
for
the
system
to
represent
that
are
for
the
users
to
request
or
specific
attributes,
and
the
second
part
is
that
one
of
the
questions
was
like.
D
Does
it
make
sense
to
handle
all
this
for
the
scheduler
level,
and
I
think
I
quali
made
a
great
bonding
that
like
trying
to
eat
with
trying
to
be
with
all
the
existing
predicates
at
the
under
priority
function
with
schedule,
level
is
already
pretty
complex
and
that,
like
you,
could
all
of
them
together.
It's
very
difficult
to
I,
don't
like
how
the
schedule
is
going
to
behave
under
this
current
scenario.
If
you're
going
to
add
hardware
topology
to
that,
then
it
couldn't
look
like
extremely
complex
I'm.
A
D
C
K
A
But
that
goes
back
to
life
is
just
sang
like
we
don't
like.
Even
if
you
do
okay
Canada
resources,
you
need
to
ensure
that
the
actual
GPU
end
up
assigning
you
to
satisfies
the
full
set
of
resource
characteristics
requested.
So
I
think
what
I'm,
when
I'm,
not
clear
with
you
on
bishes
like
do
you
think
of
it
as
a?
I
thought,
we're
under
the
impression
that
we
were
going
to
need
to
have
a
way
of
defining
resource
characteristic
more
broadly
than
what
we
have
today
in
the
project.
L
D
D
So
that's
that's
like
saying
if
they
want,
if
I
want
a
GPU
that
has
like
I,
don't
know,
say
hi
for
GV
of
GP
of
memory
just
going
by
the
example
here
and
if
it
I
have
it
have
like
any
link
of
topology
or
lay
in
Wheeling
functionality,
then
I
need
to
find
the
right
node
by
the
GPU
campfire
right
like
if
the
scale
is
going
to
just
try
nodes
arbitrarily.
That's
just
not
going
to
work.
So
there
are
some
decisions
that
the
scheduler
has
to
make
for.
D
Like
has
to
find
the
right
note
that
can
run
this
job,
but
when
it
comes
to
performance,
which
is
very
like
topology
comes
into
play,
mostly
I
mean
I'm,
saying
that
it's
something
that's
sort
of
like
tight
quads
and
given
the
physician
can
only
change
over
time.
It's
probably
preferable
to
make
it
in
the
node
and
I
also
want
a
tour
also
on
less
function.
One
more
point
which
is
set
up
like
thinking
from
the
user
perspective.
Users
are
designing
their
applications
and
they
want
to
know
what
develop
availability.
D
They
can't
relate
to
their
applications
and
they
want
to
know
like
how
many
nine
to
the
library
can.
I
connect.
Can
I
actually
guarantee
some
application,
and
that
starts
on
the
hardware
right
like
so
what
his
spawn
school
is
going
to
provide.
Is
it
won't
say
that
fire
for
a
given
job
in
history
class
I'm
going
to
provide
it
like
Bank,
9.99
percent
availability
for
access
to
cpu
memory,
GPU
and
so
forth?
And
if
you
think
in
in
that
respect,
then
it's
just
question
of
like
how
are
we
going
to
meet
that
fellow
or
SLA?
D
And
so,
if
you
think
on
that
respect,
it
becomes
a
North
problem
to
meet
the
performance
goes
and
becomes.
Scheduler
problem
find
the
right
note
to
make
to
increase
the
probability
of
meeting
the
performance
requirements
and
then
from
an
apiary
standpoint.
One
is
necessary
that,
like
this
is,
can
a
good
question
represent
various
types
of
resources.
H
Okay,
it's
why,
for
the
scalar
level,
I
think
it's
important,
because
if
you
dressed
as
a
skater,
knowing
that
you
are
the
chopped'
wants
to
TP
you-
and
you
know-
oh
that's
what
reviews
on
this
note.
If
you
don't
check
the
topology
and
you
locate
something
that's
cross
socket,
it's
better,
not
to
argue
that
all
I
mean
it's
going
to
be
very
bad
comments
and
as
I
users
will
actually
prefer
not
to
get
this
allegation.
I'm
getting
this
allocation
of
the
movies
I,
don't.
D
J
K
J
K
D
Think
I
think
we
need
something
like
assisting
politics
right.
Like
the
storage,
we
have
api,
which
is
first
formed
which
can
represent
a
lot
of
attributes
and
poise
of
a
flexible
interface
for
choosing
the
right
kinda,
so
it.
Similarly,
we
need
something
like
that
for
GPUs
and
from
a
travelling
like
it's
just
not
a
key
value
pair.
I
just
got
a
integer
resource,
but
it's
more
like
you
can
like
have
more
attributes
associated
with
it.
This
may
be
here.
User
knobs
here
are
something
that
we
have
to
represent
in
the
resource.