►
From YouTube: Kubernetes Resource Management WG20170418
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
A
Costs
are,
is
they're
unable
to
see
my
screen,
yep,
ok,
so
just
logisitics.
Before
we
get
going
with
the
war
interesting
details,
I
was
able
to
find
a
room
for
us.
The
room
can
sit
about
30
people.
It
should
have
all
the
necessary
facilities
that
we
need
or
for
having
a
day-long
meeting.
There's
a
cafe
close
by
so
breakfast
lunch,
I
love.
What
developer
actually
is?
Basically,
lunch
should
be
there
and
location.
This
isn't
funny,
then
I
think
earlier.
A
I
said
like
Google
Moulton
you,
but
it's
going
to
be
inside
there,
but
it's
pretty
close,
probably
about
like
three
to
four
miles
from
the
previous
location.
I
also
mention
a
few
nearby
hotels
that
are
closely
linked
for
folks
who
are
going
to
travel
from
outside
of
the
area
and
like
the
request.
Medicos
are
mentioning
that
that
could
be
some
socializing
events
after
the
day-long
event,
so
we
basically
have
section
where
you
can
add
yourself
if
you're
interested
in
knowing
who
isn't
staying
on
and
meeting
them
or
having
a
drink
with
them.
A
So
feel
that
no
question
wings,
everything
is
fine,
so
I
guess
dishes
have
to
go
or
dodge
internal
and
filigree
have
two
goals
with
advil.
Today,
one
is
like
finalize
the
different
set
of
items
that
we
need
to
discuss.
We
might
have
to
prune
this
list
because
we
don't
have
enough
time
and
also
come
up
with
a
schedule
discuss
these
items.
It
has
to
make
sense
large,
yes,
and
it
also
has
to
need
some
of
the
travel
constraints
with
some
of
us
have
the
wreck.
C
What
what
mission
I
spent
some
time
actually
doing
was
trying
to
coalesce
the
input
topics
that
people
provided
and
kind
of
bucket
them
a
little
bit
more,
and
so
he
and
I
have
a
certain
view
of
what
these
bullet
points
align
to.
So
maybe
we
can
just
talk
to
reach
one,
and
then
people
can
ask
questions
or
clarifying
and
then
ideally,
we'd
want
owners
coming
into
these
face-to-face
that
have
put
together
some
materials
to
help
guide
the
discussion.
So
we're
not
all
getting
there
on
on
step,
one
I
guess
yeah.
A
I
yeah
I
agree
like
that's
a
really
important
point,
because
you're
not
going
to
have
that
much
time
and
like
you
would
lose
concentration
after
like
two
or
three
hours
so
better
to
be
up
to
date
before
the
meeting
and
then
have
a
more
productive
session
during
the
meeting.
Rather
like
try
to
understand
what
the
requirements
are.
C
A
Sure
the
first
items
actual
added
by
Jamie
I'm,
hoping
that
his
own
is
going
to
be
presenting
it
and
like
driving
down
prescription
the
the
main,
the
main
goal
there
is
to
list
the
different
types
of
workloads
that
critters
is
expected
to
manage
and
then
figure
out
at
high
level,
what
sort
of
workloads
with
revellers
be
managing
natively
and
which
one
it
would
not
and
and
that
will
influence
the
different
extensions
that
you
have
to
provide
to
open
source
convertors.
In
order
to
satisfy
the
addition
of
love.
Your
constraints.
A
A
A
A
A
A
Then
we
sort
of
bike
on
some
other
items
are
like
more,
but
if
you
really
want
to
come
up
with
a
conclusive
answer,
I
think
we
should
give
ourselves
enough
time
to
go
or
all
the
different
options
and
make
all
the
different
approaches
that
that
everyone
in
the
room
is
going
to
come
up
with
and
then
choose
the
best
one.
So
I
feel
like
we
should
give
enough
time
for
each
topic,
and
if
you
happen
to
finish
one
earlier,
then
that's
good.
Maybe
you
can
get
into
further
get
the
next
time
area.
A
C
A
C
A
A
Here,
the
main
requirement
is
like
providing
better
SLO
for
access
to
compute
resources,
and
today
we
try
to
provide
some
basic
isolation
based
on
user
requirements,
but
we
do
not
guarantee
anything
latency
for
any
apps.
So
it's
difficult
for
an
application
author
to
provide
like
say,
39,
44
lines,
because
you
don't
even
know
how
much
of
availability
that
the
underlying
system
is
providing
them
for
access
to
basic
computer
resources.
So
this
gets
into
the
shoes
of
like
dealing
with
hardware
topology
and
then
how
do
we
Akalis?
What
kind
of
work
loads
can
be
support?
A
Natively
workloads
can
be
run
together.
What
sort
of
like
requirements
would
be
met
for
these
natively
supported
modules?
Some
of
the
an
extension
to
this
is
like
requiring
very
specific
modes
of
isolation.
For
example,
an
app
is
like
super
latency
sensitive,
so
it
requires
a
dedicated
code
because
it
cannot
like
tolerate
any
sort
of
any
sort
of
like
interference
from
from
other
apps
sharing
same
core.
So
those
are
examples,
or
like
other
example
that
we
came
up
with,
was
like
running
on
CP.
U
0
specifically
for
like
network
application,
so
she
have
more
examples.
C
Of
this
topic,
yeah
anything
I
think
/
chat
yesterday
right
we're
talking
about
like
we
need
to
understand
what
we
think
the
default
posture
is
going
to
be
for
the
cube
list
and
then
that
will
help
us
inform
where
we
would
potentially
look
to
differentiate
from
that
default
right.
So,
like
I
know,
yesterday
noon
hour,
chatting
I
mean
I
have
a
posture.
That's
awesome,
like
I'm.
A
A
A
The
next
topic,
which
is
a
segue
from
segue
from
Lord
Erica
mentioning,
is
that
we
need
extension
both
of
the
third
cluster
level
and
at
the
node
level.
This
is
to
support
like
workloads
that
are
either
workloads
that
are
not
natively,
supported
by
cabelas
or
mrs.
support,
or
its
true,
like
have
custom
behavior
for
for
both
clothes
that
are
supported
by
curators,
for
example,
of
the
native
the
native
scheduling
scheme.
A
When
native
resource
management
teams
aren't
adequate
or
they
are
in
performing
this
for
the
given
hardware
given
like
Jason
to
requirements,
maybe
you
want
to
run
your
own
scheduler
and
an
isolator
at
the
node
level,
so
this
topic
is
meant
to
cover
those
scenarios.
I
mean
the
prime
example.
Here
is
like
providing
no
level
isolation
which
into
spokes
are
been
climbing
quite
a
bit
recently.
A
It
steady
that
leg
derek
was
having
a
classic
example
of
like
a
submarine
just
like
that's
running
apps
that
are
specifically
fine-tuned
to
use
underlying
hardware,
which
is
which
is
a
classic
example
of
like
doing
manual
binding.
How
do
we
deal
with
that
right,
like
this,
is
extreme
ever
going
to
suppose
such
a
scenario
I
think
no
but
like?
If
that's
a
requirement
for
for
a
given
set
of
users,
then
like?
How
do
we
achieve
that
requirement?
Is
there
going
to
be
a
recommended
means,
and
do
we
have
the
necessary
extension
for
that
I?
A
A
Think
of
this
more
as
like
you're
going
to
launch
some
specialized
app,
that's
going
to
require
a
specific,
dedicated
core
that
that's
closer
to
a
set
of
GPUs
and
your
SPS
and
unique,
and
if
it
happens
that
it
cannot
find
a
node
I
call
all
nodes
already
have
apps
that
are
like
running
on.
Different
life
isn't,
and
you
happen
to
not
be
able
to
start
this
job,
then
it's
a
problem
for
you.
So
how
do
you?
How
do
you
ensure
that
job
can
will
be
guaranteed
resources
of
a
future?
A
There
are
some
ideas
and
thoughts
on
how
to
deal
with
this,
but
this
is
a
fairly
open
topic
that
we
haven't
explored
as
a
community.
Yet
the
next
topic
think
dark
out
of
this,
which
is
the
world
like
defaulting
policies.
How
do
we
deal
with
like
defaulting
for
resources
like
you
have
the
correct
me
if
I'm
wrong?
The
way
I
understand
is,
except
my.
C
Take
on
this
last
bowl
is
I.
Think
the
first
part
of
the
the
face-to-face
we're
going
to
be
discussing
a
bunch
of
primitives
and
things
actually
end
users,
don't
really
want
to
think
too
deeply
about
and
then
I
think.
We
should
ideally
close
up
the
face
to
face
with
how
do
we
make
this
easier
for
users
to
consume,
and
so.
C
Yeah,
that's
basically
like
if
we
end
up
believing
this
we're
saying
we're
going
to
have
a
more
complicated
resource
requirements,
section
of
a
container
or
a
more
complicated
resource,
attributes
part
and
we're
not
doing
anything,
to
make
it
easier
for
users
to
reason
about
these
things
or
write
their
prospects.
Then
I
think
we
haven't
like
solve
the
whole
problem.
A
Yeah,
all
those
things
actually
I'm,
just
writing
it
down
just
so
that
and
make
sure
that
we
are
on
the
same
page,
yep
yeah,
like
the
record,
saying
I,
think
we
have
some
problems
and
some
of
this
might
be
covered
in
the
representation
aspect
of
them.
That
I
like
you,
need
to
have
some
some
same
default
across
different
application
classes
which
might
make
like
application
deployment
much
simpler.
A
I
guess
maybe
you're
just
writing
all
of
them
as
blood
sugar,
but
you
need
like
some
some
some
defaulting,
that's
not
just
at
a
namespace
level,
for
example,
and
then
you
also
need,
like
some
portable
resource
names.
This
is
similar
to
like
storage
classes,
so
where
you
have
certain
classes,
so
then
those
classes
back
to
like
certain
sort
of
devices
example
here
could
be,
like
you
say,
tesslar
GPUs,
that
that
falls
in
the
same
class
alone.
A
That
could
not
do
one
of
five
or
different
kinds
of
GPUs
because
they
all
happen
to
be
compatible
with
each
other
and
and
uncool
I
guarantee
son.
So
that's
that's
one
requirement
there
or
you
have
a
nikka
like
you
have
classes
of
nixon
and
you
have
high
performance
akin
and
you
make
sure
that
you
get
the
big
the
last.
C
The
original
issue
that
the
owner
had
opened
up
was
basically
I,
have
big
data,
jobs
and
I
have
traditional
web
jobs
and
I
want
to
have
them
both
share
the
same
cluster,
so
one
runs
at
night
and
one
runs
during
the
day,
and
then
how
do
we
like
support
fair
resource
sharing
across
new
class
of
workloads?
Now.
E
They
give
you
david,
has
written
a
document
and
adding
priority
in
and
registering
in
general.
He
sure
did
already
I
believe,
but
couple
days
or
maybe
several
days
ago,
with
the
team
I
believe
I
have
to
link
to
the
document
I.
Can
you
send
it
out
but
anyways
that
that
document
is
sort
of
like
a
high
level
idea?
It
doesn't
go
into
like
more
details.
E
I
am
I'm
going
to
write
another
document
which
actually
a
little
bit
more
into
the
details
of
how
we
are
going
to
add
priority
to
the
API
and
how
the
preemption
is
going
to
work
based
on
the
priorities
yeah.
So
we
have
one
document
ready
right
now,
which
we
can
share
them
and
that's
the
one
that
paid
it
is
written.
E
C
A
C
Traditional
resource
quarter,
quarter
right,
yeah,
yeah
I
mean
more
like
if
I
have
to
I
guess
I
want
to
know.
If
we
imagine
not
be
any
overlap
between
this
and
like
what
we're
going
to
do
with
the
future
Claude
so
like
if
they
see
a
two
pods
that
land
on
the
same
know
that
asked
for
competing
priority.
Access
to
resources
like
depending
I
would
be
dedicated
gps
or
something.
A
A
One
is
going
to
be
Q
blitz
and
maybe
like
some
other
metrics
sources,
and
then
the
other
one
is
going
to
be
rescheduling,
which
is
going
to
really
like
try
to
rebalance
notes
based
on
based
on
the
observed
performance,
because
if
the
node
cannot
do
anything
beyond
what
it's
trying
to
do,
it
either
has
to
have
it
or
it
has
to
wait
for
three
scalar
to
do
that.
I
usually
think
like
three
scheduler
is
the
one
that
should
be
dealing
at
that
problem.
A
E
Generally
true,
so
maybe
I
didn't
I
didn't
understand
the
first
question
correctly,
but
if
you
have
like
priorities
already
and
assuming
at
it
task
that
with
task
which
is
like
to
the
lower
level,
priority
is
running
on
a
note
and
then
suddenly,
a
new
task
arrives
with
a
higher
priority
or
somehow
deep
resource
requirement
of
having
high
priority
task
changes.
I'm,
not
sure
if
google,
it
is
currently
supports
this
change
of
research
department
but
anyways.
C
E
C
Like
the
only
thing
we
really
have
is
cause,
and
then
most
of
my
understanding,
I'm
paranoid,
gumption
and
I'm
perfectly
fine
with
that.
So,
but
I
think
it
relates
to
why
I
am
ok
with
this
being
at
the
end
of
the
agenda.
Is
that
it's
more
like
a
scheduling
fit
question,
and
then
how
do
I
give
it
active
pods
from
that
node
in
response
let
fit
versus?
How
do
I
keep
both
toads,
both
pods
resident
on
the
node,
and
maybe
only
give
temporal
priority
to
one
versus
the
other
like
we're?
Not
we're?
C
Not
I'm
at
a
prior
to
scheme,
that's
going
to
go
like
tell
us
ever
going
to
reuse,
TFS
shares
differently
or
something
right
or
yeah,
we're
going
to
get
the
real-time
scheduler
or
some
other
thing.
That
impacts
the
note
itself.
It's
more
just
cluster
wide
scheduling,
use
which
to
me
seems
less
controversial
than
like
those
things
that
blend
how
we
manifest
it
on
the
node
and
then
how
it
scheduled
to
load
so
yeah.
E
A
E
Is
this
is
not
finalized
yet
and
actually
I,
just
david
hasn't
shared
a
document
with
everyone
and
I
will
let
him
do
that.
He
told
me
that
he
would,
but
maybe
he
has
forgotten,
but
anyhow
I
send
a
link
to
the
doctor
and
when
I'm
not
sure,
if
you
guys
will
have
access
to
it.
Yes,
but
anyhow,
and
at
the
end
of
the
document
he
goes
through
a
few
possible
options
about
preemption
pre-empting
printing
various
parts.
One
is
like
letting
cube.
E
Let
do
the
preemption,
which
is
a
little
bit
more
controversial
as
you
go,
sir,
and
the
other
one
is
to
collect
the
scheduler
decide
about
preemptions,
and
the
priority
of
our
priority
is
something
that
is
defined
by
the
fight
user
by
letting
the
scheduler
decide
which
cause
to
effect
based
on
given
priorities
and
preempt.
The
other
ones
which
are
not
necessarily
needed
for
I
mean
they
are
not
ideal
if
they
are
holding
other
high
priority
jobs
from
getting
enough
resources.
E
E
Exactly
though,
basically
the
performance
is
sort
of
imply
in
the
priority.
If
you
really
need
to
guarantee
performance
for
your
pods,
you
need
to
give
them
higher
priority
and
you
need
to
ask
for
enough
resources.
That's
how
performance
is
so
again
so
anyways,
so
the
other
work
uppal
of
options,
but
most
probably
everyone
is
sort
of
like
inclined
towards
going
but
scheduler
deciding
about
the
preemptions.
A
F
A
F
A
I
would
love
to
like
discus
and
contributed
know
if
you
can
share
more
details.
Yes,
certainly
yeah,
I
blister
sentatives
lets
you
have
time,
then
we
can
get
to
it
and
we
can
discuss
more
yeah,
okay,
so
so
in
terms
of
so
I'm
guessing
like
the
list
that
we
have
is
already
sorted
in
terms
of
priority,
because
there's
some
like
logical,
logical
hierarchy
between
these
different
requirements.
A
C
A
A
D
A
D
A
So
studio
is
definitely
necessary
and,
like
maybe
by
tomorrow,
I
can
I
can
come
up
with
an
initial
proposal
which
we
can
learn
their
trade.
Using
this
talk,
I
feel
like
more
higher
priority.
Need
right
now
is
finding
owners
for
each
of
these
items,
because,
like
you
mentioned,
someone
needs
through
the
creative
work
of
like
is
identifying
existing
issues
or
proposal
or
like
TRS,
are
actually
like.
Coming
up
a
few
proposals
to
deal
with
these
problems,
I
mean
this
is
this
thing
could
be
just
have
a
compilation
of
different
requirements
for
each
of
these.
A
So
for
first
one
like
Jeremy
you're,
there
are
like
you're
hoping
that
you
would
take
care
of
everything
there
and
for
the
second
one
I'm
interested
in
looking
at
it
is
if
someone
else
wants
to
collaborate,
I'm
happy
I'm,
happy
to
work
with
that
person,
but
so
feel
free
to
speak
up
or
on
yourself.
If,
if
you
want
to
collaborate
on
this
topic,
I
can.
E
A
B
Maybe
if
you
could
gets
information
about
what
the
the
google
policy
is
like,
at
least
in
broad
strokes,
just
as
a
just
that
we
could
use,
as
maybe
a
discussion
point,
simply
adds,
in
their
words,
30
classes
of
workloads
that
were
supported.
It
was
just
an
idea
of
kind
of
the
considerations
that
went
into
that.
Oh
yeah.
A
A
C
A
Culet
can
discover
devices
and
there
and
the
distance
or
like
some
notion
of
like
connectivity
between
these
devices
cost
of
connectivity.
Then
it
can
be
a
generic,
a
model
right
leg,
because
performance
isolation
is
at
the
point,
just
become
binary,
except
if
it's
like
eglee
link
or
something
that's
a
little
bit
more
complicated.
A
C
Guess
it's
counter
if
you
and
I
want
to
pair
up
on
that.
That's
fine,
because
I
guess
what
I'm,
what
I'm
anticipating
this
is.
There
will
be
a
dividing
line
where
you
say
no
and
that'll
be
what
I
want
to
be
able
to
ensure
I
can
plug
into
this
isolation,
but
and
that
dividing
line
could
be
dedicated,
cpu
sore
that
dividing
line
to
be
Numa
reservation,
so
I'm
show
kind
of
tease
that
out.
A
C
C
C
I
guess
for
my
own
desires,
I'm
not
like
interested
ensuring
that,
like
you,
don't
overcommit
your
quota
relative
to
your
cluster
or
something
so
like
to
me.
That's
a
priority
preemption
discussion.
I
am
interested
in
knowing
that,
like
if
I've
made
a
scheduling
decision
that
I
can
make
that
scheduling
decision
it
like
a
granular
enough
level
that
I
know
that
I'm
not
like
dynamically
being
moved
around
on
the
local
node.
A
C
Like
I
would
to
me
guaranteeing
future
availability
to
me
means
well,
maybe
the
phrasing
here
could
be
off
what
I
want
to
have.
As
a
way
of
saying
that
I
could
write,
I
could
potentially
have
a
topology
aware,
scheduler
and,
as
a
consequence,
I
can
make
apology
aware
binding
decisions
that
prevent
other
pods
from
from
using
that
same
resource.
So
to
me
this
is
like
how
do
I,
how
do
I
find
things
that
we
might
think
the
qubit
is
going
to
do
by
default?
So,
for
example,
oh
and
craig.
C
Think,
in
my
example,
I
have
in
my
head
right
now
is
basically,
if
I
want
the
cluster
scheduler
to
do
the
new
anome
assignment
versus
the
culet
to
do
the
Numa
node
assignment.
How
can
I
assert
that
binding
to
the
Numa
node
in
a
way
that
either
a
custom,
isolator
or
the
scheduler
can
can
manifest?
This
is
to
me.
C
E
C
A
C
Think
this
goes
hand
in
hand
with
maybe
one
of
the
workload
types
that
Jeremy
will
present,
but
like
not
every
workload
in
the
world
is
going
to
drive
density.
Some
more
clubs
will
will
go
and
and
be
okay,
if
I'm
only
using.
If
there's
eighty
percent
select
resources
on
the
node,
it
was
more
important
to
ensure
that
no
one
else
types
of
Numa
node,
because
I
got
it
well.
E
C
Only
thing
I
want
to
call
it
is
like
to
me
this
is
an
extension
feature,
and
so
that's
why
I
want
to
group
it
under
the
extensions
like
it
and
not
necessarily
the
clustered
default
behavior
actually
and
then
yeah.
If
we
want
about
priority
and
prevention,
I
mean
obviously
there's
other
issues
when
you
do
this
like
how
do
I
exclude
myself
from
critical
pot,
auto
scaling
and
potentially,
how
do
I
exclude
myself
from
priority
and
prevention
like
that
stuff
to
think
tout,
but
like
there
are
people
who
don't
care
about
fragmentation
as
much.
A
C
Not
this
is
why
I
really
feel
like
this
was
going
to
group
into
when
we
decide
what
is
the
default
workload.
Optimizations
are
going
to
make
as
a
key
projects
but
like
even
today,
theorem
and
I.
We're
talking
about
various
examples
of
I,
have
a
two
node
system
and
I
only
ever
scheduled
to
node
1
right,
because
it's
just
more
important
to
me
that
I
always
get
node
1
and
no
one
ever
goes
on
those
two
and
of
that
there
are
environments
where
that's
important
and
I.
C
Don't
think
that
that
should
be
the
cluster
by
default.
But
I
would
like
a
way
of
expressing
more
granular
scheduling
decisions
that
the
cable
it
honors
either
be
a
custom
isolator
someone
does
so
much
stuff,
the
Connors
looking
at
or
what
but
I
to
me.
This
looks
like
an
extension
feature,
so
that's
why
I
think?
Maybe
we
should
describe
them
that
bullet
we
should.
We
should.
E
C
C
If
I
have
a
pod,
oh
yeah
dinner,
and
it
has
two
containers-
a
and
B-
and
they
need
to
share
access
to
huge
pages
and
container
a
needs
to
run
on
Numa
node.
One
in
container
baby
needs
around
Numa
node.
To
like
that's
a
real
world
use
case,
and
to
do
that,
like
none
of
our
acoustic
primitives
that
are
at
the
pods,
go
support
that,
but
I
want
to
be
able
to
support
writing
a
scheduler.
I
can
define
that
nice.
A
Not
just
additional,
and
it's
like
even
if
Scott
isn't
supported,
like
whatever
existing
class
of
workloads,
they're
still
going
to
be
the
problem
of
like
fragmentation
right
like
do
we
deal
with
that.
Maybe
the
answer
is
priority,
but
it's
still
not
clear.
I
mean
I.
Have
some
ideas
on
my
mind,
but
I
want
to
make
sure
that
lays
it
all.
On
the
same
page,
there
all.
C
C
C
B
A
A
E
A
E
A
E
A
A
A
C
Let's
say
so:
I
had
nothing
else.
In
addition
to
add
I,
don't
know
if
folks
are
going
to
have
time
and
for
next
week's
meeting
to
maybe
give
a
preview
of
some
of
the
things
that
we're
talking
about
we're
going
to
discuss,
but
otherwise,
right
now,
I
had
nothing
on
the
agenda
for
next
week.
I
assume
we're
going
to
be
getting
on
the
street
on
this.
Does
anybody
have
anything
that
they
wanted
to
get
more
urgently
discussed
in
next
week's
meeting?
Well,.
F
Only
no
don't
come
to
outside
some
of
the
topics
of
soap,
Sally
brought
god
I,
don't
know.
If
you
can,
we
make
sense
to
kind
of
add
one
level
of
criminality.
In
terms
of
the
note,
it
would
be
more
precise
around
your
goal
like
if
they're
a
decision
you
want
to
make
I
want
to
side
if
you're
going
to
want
to
go
this
route
to
start
this
route
or
goals
enumerate
options
within
for
this
particular
thing,
it
might
be.