►
From YouTube: Kubernetes Resource Management WG 20170509
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
A
So
yeah
welcome
to
the
mainland
meeting
of
the
resource
management
workgroup
I
know.
Many
of
us
are
fatigued
from
discussing
resource
management
topics
last
week,
but
there
two
items
on
the
agenda
today,
one
as
a
recap
on
what
we
discussed
and
need
some
discussion,
all
next
steps.
And
then
we
have
a
second
item
on
the
agenda
to
do
a
more
detailed
advise
on
the
container
device
interface,
which
I
think
will
dovetail
a
little
bit
into
some
of
the
topics
we
discussed
last
week
as
well.
B
Sure
so
these
are.
These
are
notes
that
the
derek
took
basically
condensed
from
the
raw
notes.
It
looks
like
that
david
asheville
graciously
wrote
for
us
during
last
week,
so
yeah.
So
the
summary
we
had
judge
people
from
a
diverse
group
of
companies
and
there
were
a
few
work
streams
that
were
they
were
decided
out
of
the
out
of
the
three
days.
I
think
we
came
down
to
looks
like
five
are
broken
down
here,
so
the
first
one
was
support
for
resource
classes.
B
Choose
surfaces
repeat
for
air
for
a
number
of
you
who
are
on
the
last
meeting,
but
you
know
a
level
of
indirection
between
of
what's
in
the
pods
back
and
what
is
in
the
node
capacity
to
allow
for
concrete
resources
to
be
aggregated
into
slightly
more
abstract
things
that
end
up
in
the
user.
Pods
back
and
resource
classes
is
mentioned
in
a
few
of
the
other
work
streams.
So
this
may
be
a
at
least
a
soft
dependency
and
a
number
of
the
other
items
that
that
we
have
here.
B
There
are
links
to
the
two
detailed
I
guess
what
design
we
have
is
is
in
those
Docs.
Probably
more
work
needed
on
all
of
these
ok
IM
to
was
the
extensible
device
support,
and
so
this
is
I
think
just
remembering
back
Nvidia
agreed
to
be
the
guinea
pig
to
implement
a
device
plug-in
and
also
help
out
with
the
implementation
of
the
delegation
side
on
the
cubelet.
B
But
the
main
idea
here
was
that
you
would
have
a
finely
shaped
API.
That's
intended
for
an
external
component
to
advertise
some
device
resources
and
part
of
that
information
includes
the
topology
sub
graph,
which
the
cubelet
would
add
to
its
local
node
hardware,
locality,
information
and
use
that
in
the
device
assignment
decisions,
and
it
would
also
be
responsible
for
initializing
tearing
down
and.
B
A
I
would
basically
describe
like
shut
up.
You
jump
in
yeah
with
the
I
said
for
this
one,
like
the
high
level
of
things
where
identified,
and
so
like
I
think
the
document
that
mission
myself
set
up
that
I
think
is
voiced
their
build
a
lot
on
a
lot
of
the
items
that
were
presented
in
the
originals
CDI
or
container
device
interface
with
puzzle,
but
add
some
additional
stuff.
A
That
was
not
there,
and
so
basically,
we
need
someone
to
go
and
take
the
next
step
and
build
a
proposal
that
kind
of
blends
the
two
but
yeah
right
now,
like
you,
don't
have
the
digital
framework
in
mind
for
how
like
the
key,
but
will
trust
individual
device
creators
and
stuff
like
that.
So
it's
somewhere.
C
I've
been
spending
some
time
here
in
the
money
like
thinking
over
the
basic
design
idea.
Oh,
we
came
into
the
session
interface,
I
mean
just
to
say:
someone
community
wants
to
go
and
like
Sabo
can
really
feel
free
to
do
so,
but
I
think
there's
the
thing
to
people
in
mind,
which
probably
has
already
been
told
here,
that
the
resource
class
API
is
a
probably
a
prerequisite
for
anything
you
wanna
do
with
devices,
so
that's
probably
going
to
have
a
higher
priority
relative
to
our
excellent
vegetable,
so
existing
in
media
integration.
B
B
B
So
the
third
item
here
is
some
enhancements
and
how
the
cubelet,
by
default,
treats
CPU
resources
and
I
think
the
concrete
goals
we
had
here
was
now
to
achieve
good
performance,
or
at
least
more
predictable
performance
when
compared
with
VM
based
orchestration
systems,
those
what
a
what
Derek
described.
As
you
know,
disappointing
bake-off
results
when
he
was
asked
to
compare
you
know:
M
based
containerization
versus
the
you
know
the
CFS
CPU,
sharing
that
that
it
implements
this
is
the
equivalent
of
a
default
and
qiblah.
B
So
the
idea
is
to
try
to
reduce
the
process
scheduler
latency
by
giving
out
exclusive
course
to
some
guaranteed
pods
based
on
how
the
resources
are
requested,
and
there
are
two
strategies
that
were
proposed.
One
is
called
the
dynamic
strategy
which
would
be
based
on
you
know:
feedback
from
the
president
for
the
for
the
pod
and
using
that
to
do
something
smart
with
CPU
sets,
and
then
the
aesthetic
strategy
would
be.
B
A
The
only
thing
I'll
just
add
is
like
the
dynamic
strategy
will
be
the
default
most
likely,
and
so
we
need
some
people
to
prototype
this
before
getting
a
more
informed
proposal
out
so
I
know,
at
least
it
right
hat
here
on
the
call
is
going
to
start
to
prototype
something
I
know.
Some
members
with
me
from
Intel
have
expressed
an
interest
as
well,
so
it'd
be
good.
If
people
are
interesting
to
go
ahead,
and
so
if
they
could
speak
us
and
we
can
surely
collaborate
to
build
a
it
could
outcome
for
all,
but
yeah.
B
That
would
be
great
I
know
some
some
of
the
members
of
our
group
started
just
by
seeing
what
they
could
do
in
terms
of
refactoring.
The
external
isolator
work
to
be
a
kind
of
internal
delegate
that
we
talked
about
during
the
meeting
yep
to
change
it
into
or
just
a
CPU
manager
interface
inside,
be
a
container
manager
package.
A
E
I'm
also
looking
in
the
container
manager
and
I
mean
the
cpu
set.
A/C
group
is
kind
of
strange.
It's
got
a
CPU
under
core
occlusive
are
pretty,
but
we
do.
We
don't
want
to
set
that
because
that's
my
cause
nightmare
scenarios
with
races
and
things
like
that.
If
the
what
it
does
is
it
you
know,
you're
defining
the
CPU
set
for
your
the
cpu
set
CPUs
for
your
CPU
set
and
if
you
said
exclusive
well,
one,
if
you
can
use
in
your
CPU
set,
are
included
in
a
sibling.
E
If
you
set
set
fails,
if
it's
not,
then
you
you're
set
exclusive
equals.
It
succeeds
and
then
those
course
can't
be
assigned
to
CPU
sentences.
Link
CPU
sets
so
yeah.
It's
kind
of
complicated,
but
I
think
what
we
want
to
do
is
basically
make
cores
explicitly
exclusive
by
just
removing
them
from
all
the
sibling
sets
and
not
setting
CPU
under
chore
exclusive,
but
yeah
that
it's
pretty
it's.
E
A
A
killer
thing
to
call
it
if,
like
a
code,
passed
I,
think
for
either
strategy
should
not
be
like
diametrically
different
like
they
have
very
similar
problems
like
they
both
need
to
schedule
a
CPU.
They
both
need
to
really
succeed.
You
they
both
you
to
potentially
know
how
fragmented
they
are
so
I
think
the
hope
is
a
free
de
compasso
like
we
can
get
a
base
primitive
that
works
from
a
code
standpoint.
The
things
not
look
at
the
fragmented
right.
So
yes,.
B
B
E
F
B
A
A
I
think
the
thing
we
need
to
iron
out
on
this
is
I
think
in
their
prototypes
that
are
out
there.
The
well.
Let
me
single
out
purpose
in
if
we
choose
to
map
huge
pages
via
resource
classes
and
have
an
indirection
we'd
like
to
just
say
this
is
the
size
of
a
huge
page.
I
want
independence
of
the
actual
page
size.
So
that's
awkward
to
say,
okay,
but
if
you
just
say,
I
want
to
gig
a
few
pages
and
you
don't
care
if
it's
2
megabyte
or
some
other
alternate
architecture.
A
C
A
C
A
Yeah
other
way,
I
agree
that
so
I
guess
we
need
to
that's
why
the
one
edge
case
on
the
heat-ray,
stuff
and
I
think
the
other
edge
case
was
just
ensuring
that
memory
counting
is
working
as
we
expected
so
to
be
able
to
to
verify
that
the
mem
stage
group
is
not
double
counting.
Php
function
outside
of
food
scale,
but
yeah
generally,
we
should
be
close
to
beyond
I
get
a
final
proposal
on
the
stuff.
B
Cool
okay,
so
second,
the
last
one
here
is
Numa.
So
the
intention
is
for
the
cubelets.
You
understand
the
node
topology,
so
it'll
do
some
sort
of
I
guess
we
talked
about
an
initial
round
of
introspection
into
the
Ellie's
via
CPU,
socket
topology,
and
then,
when
device
plug-ins
advertise
their
own
resources,
those
those
graphs
would
be
appended
to
the
the
known
topology
of
the
node
and
they
would
all
be
considered
when
making
locals
of
resource
allocation
decisions
on
the
Cuban
side.
B
So
I
guess
the
one
hard
design
constraint
that
we
decided
during
the
meeting
was
that
the
node
topology
will
not
be
advertised
to
the
to
the
scheduler.
So
guess
where
we'll
just
deal
with
counting
resources
and
the
the
cubelet
will
be
concerned
with
its
mapping
the
allocated
number
of
resources
and
concrete
resources
once
the
plan
is
assigned
to
the
to
the
node.
A
So
I
don't
know
this
if
you
wanna,
add
more
color
here
to
this
I
think,
generally
speaking
to
in
order
to
go
deeper
on
Numa,
we
need
to
get
basically
the
grass
modeled
in
the
qiblah
there's
a
lot
of
prereqs.
We
need
to
land
before
them
and
I
think
we'll
learn
some
things
along
the
way
around
like
how
we
should
treat
relative
good
so
which
resources
which
should
give
better
weights
and
I
want
to
be
surprised
if,
in
a
few
months
time
we
have
a
bunch
of
like
like
scoring,
policies
might
be
taja.
A
G
B
Okay
and
then
the
as
a
less
topic
that
we
talked
about
was
performance
measurement
and
the
main
idea
is
to
provide
some
way
that
we
could
witness
performance
regressions
that
happen
specifically
with
node
local
policies
and
also
to
compare
proposed
proposed
isolation
strategies
against
each
other
in
different
situations.
Hopefully,
one
common
commonly
deployed
workloads
in
the
cloud
so
there's
one
example
linked
here,
which
is
a
project
that
it
happened
at
Intel
and
it's
mainly
about
allowing
the
user
to
program,
experiments
and
each
experiment
consists
of
a
bunch
of
different.
B
You
know
configuration
values
that
get
unfolded
into
a
matrix
and
then
what
you
end
up
with
is
a
big
graph
of
sensitivity,
profiles
or
specifically,
a
high-priority
workload
versus
some
co-located
agressors.
So
you
end
up
with
a
graph
like
this,
and
the
main
point
was
not
to
you
know,
push
this
project
specifically
as
a
as
a
solution,
but
just
use
it
as
inspiration
for
know
what
kind
of
data
we
might
want
when
we
are
making
decisions
about
the
ocean
cooler
policies,
because
in
I
know
a
bunch
of
the
people
on
the
collar.
B
A
Wondering
this
is
like
how
we
can
take
like
these
performance
measurement
data
and
then
like
make
something
more
intelligent,
so
that
we
get
like
affinity
and
entice
any
scheduling
roles
assigned
like
intelligibly,
so
I
don't
know
how
to
spot
about
that
much
Connor
but,
like
generally
speaking,
I
know
we
spend
all
the
time
talking
about
like
initial
resources
and
potentially
a
thing
in
the
cube
project.
We're
like
based
on
historical
usage.
You
can
see.
A
B
C
A
Got
to
read
through
your
stuff,
more
a
little
more
but
I
feel
like
we
might
get
a
large
smell
of
like
decent
data
and
it'd
be
good
like
it's
unclear
to
me
that
we
need
new
new
code
to
like
responsive.
You
won't
like
how
do
you
know
where
it's
safe
to
schedule,
stand
angle,
huntable,
lose
affinity
or
anti
affinity
to
hit
your
desired
outcomes.
B
C
A
B
A
A
D
B
D
G
G
A
We
can
move
on
to
the
second
topic
on
the
agenda,
so
Dennis
I
don't
know
if
you're
there.
Thank
you.
There
I
think
you
wanted
to
get
a
rundown
on
your
thoughts
behind
the
container
device
interface
proposal,
and
maybe
you
can
walk
us
through
that
now
and
then
we
can
give
you
some
of
our
perspective
back
for
items
that
we
might
have
touched
on
at
the
face
of
it.
F
A
I
I,
don't
I,
don't
think
the
mechanism
or
the
trend
for
it
is
defined
it's
more
of
a
quality.
What
are
the?
What
are
the
use
case?
Requirements
I,
think
we
kind
of
nailed
down
set
of
use
cases
you're
looking
for
and
then
what
are
the
high
level
like
design
goals?
You
need
to
achieve
I,
think
we
nail
down,
but
they
actual
mechanics
I
think
are
still
to
be
determined.
F
A
F
So
yeah
well
I
I
just
tried
to
mimic
CNI
The
Container
network
interface
because
I
thought
well,
it
works
already,
and
it
has
some
elements
that
are
similar,
so
I
yeah
just
did
that
for
device
discovery,
I
implemented
that
in
the
proposal
after
your
suggestions-
and
it
also
has
well
device
discovery,
location
and
de-allocation
and
communication
in
this
proposal.
In
this
version
of
the
proposal
happens,
the
command
go
through
environment
variables,
while
the
output
goes
through
Jason's
who
send
it
out.
D
Let
you
know
this
is
Daniel
is
through
sick
network
and
also
a
CNI
maintainer
were
sort
of
moving
away
from
environment
variables
for
some
of
the
stuff,
because
we
thought
they
weren't
quite
as
flexible
and
so
instead
of
environment
variables.
Removing
to
push
that
information
through
the
standard
input
Jason
configuration
instead.
That.
F
B
F
A
A
C
We
did
talk
about
the
scenario
that
I
like
as
a
vendor.
You
don't
really
care
about
giving
a
plug
in,
but
then
you
just
follow
right
by
you
and
if
they
recall
correctly,
general
agreement
and
all
that
the
different
thing
here,
that's
going
to
be
the
plug
in.
Thank
you
professional,
binary
analysis.
You
can
have
one
more
level
of
indirection,
look
like
you're,
basically
having
a
proxy
the
proxy
some
any
API
that
you
have.
That
could
be
an
executable,
for
example,
and
then
it
clocks
it
over
to
on
the
GRDC
and
points
in
Cuba.
C
A
But
I
guess
the
other
things
that
I
was
thinking
about.
Vishna
I
agree
with
your
foot.
There
was
that
and
I'd
like
to
spend
all
the
time
discussing
how
to
get
kernel
drivers
as
well
as
your
space
drivers,
for
that
might
be
as
though
there's
in
your
device,
and
then
we
spent
a
fair
amount
of
time
discussing
like
in
the
allocate
step.
A
If,
if
that
was
going
to
be
used
and
so
I
feel
like,
we
got
a
lot
of
pushback
on
the
GPU
side
that
they
did
not
want
to
use
the
allocate
some
way
of
exposing
container
I'm
wanting
to
do
it
with
1c
hooks,
but
I
wasn't
sure
like
Dennis.
If
there
were
particular
devices
you
were
looking
at
that
might
have
similar
challenges
that
might
help
us.
F
A
Are
you
interested
in
pushing
the
device
proposal
board
further,
like?
Are
you
agreeable
to
maybe
looking
at
to
the
document
that
we
put
out
and
then
digesting
those
and
then
potentially
iterating
on
your
existing
proposal,
and
maybe
we
could
use
that
as
the
as
the
pr2
rally
against
or
I
guess,
what
I'm
trying
to
figure
out
is
like
what
how
long-term
engaged
are
you
looking
to
be
in
the
project
to
push
the
teaching
board.
F
Well,
I'm
I
already
read
most
of
the
documents
from
the
face-to-face
meeting
and
yeah.
If
that's
still
an
option
for
you
to
propose
to
follow
this,
this
proposal
sure
I
can
I
can
go
through
that
and
iterate
with
you,
but
I.
Don't
think
it
makes
much
sense
to
to
go
to
different
routes.
That's
why
I
would
like
to
coordinate
with
you
on
like
one
one
proposal,
yeah.
A
A
A
lot
of
detail
and
I
think
the
the
I
think
one
of
the
things
that
came
up
there
is
we
wanted
to
cubelet
to
be
able
to
do
a
list
and
a
watch
on
individual
device,
plugin
providers,
rather
than
pulling
so
that
we
can
respond
more
quickly
when
we
know
a
device
failed
and
so
I
think
I'm.
Just
if
this
is
a
topic
that
you're,
particularly
interested
in
you
want
to
evolve.
A
A
To
both
prototype
to
inform
the
puzzle
and
write
the
proposal,
I
think
what
core
nation
standpoint.
If
we
assume
that
the
concept
of
a
resource
costs
will
exist,
then,
like
I,
assume
dish
myself
and
maybe
we're
know
from
Nvidia
or
probably
the
ones
to
coordinate
live
but
like
I
know
personally,
myself,
I
won't
have
cycles
to
work
on
a
proposal
for
the
next
few
weeks
and
I
assume
dishes
in
a
similar
boat.
A
I'm
not
I'm
not
expecting
anything
in
q1,
seven
right,
I
think
the
future
concept
has
has
been
passed
on
that
I'm
more
just
wanting
to.
If
people
have
time
and
cycles
to
dedicate
and
think
on
this
and
I
understand
where
we
have
to
brought
a
group
felt
like
any
proposal
should
go
then
like
I,
don't
discourage
people
from
from
iterating
on
my
proposal.
A
C
D
C
C
So
it's
saying
that,
like
oftentimes,
what
happens
is
like
this
amazing
work
happening,
but
they
don't
happening
together
and
they're,
not
part
of
the
roadmap
and
so
oftentimes.
Like
those
those
amazing
patches
proposes
designs.
They
don't
get
the
right
kind
of
attention,
so
I
want
to
avoid
that
with
all
this
one
and
that's
why
I
was
trying
to
set
the
expectation
straight
and
that
it's
probably
unrealistic
to
get
a
working
implementation.
Even
if
it's
alpha
in
the
next
release.
C
Not
sure
I
know
I,
don't
think
that
I
was
in
the
column
like
today.
Here's
an
intern,
so
I
don't
know
there
how
long
he
will
be
immediately.
If
general
is
not
responding,
then
I
can
probably
help
you,
because
it's
something
that
I
tend
to
focus
on,
but
enough
I
will
be
a
1990
contributing
to
it.
Okay,.
F
A
B
No
at
the
end
of
the
at
the
at
the
end
of
the
face-to-face,
we
put
some
kind
of
bring
leaders
on
to
the
topics
just
to
get
a
roadmap
together.
I
was
wondering
if
we
wanted
to
have
like
a
consistent
format
for
what
the
roadmap
should
look
like
for
each
one
or
if
there
are
some
examples
that
we
should
follow
when.
C
I
mean
basically,
we
we
listed
all
the
six
items.
The
six
major
items
I'd
be
description.
Then
we
tried
to
like
identify
some
basic
supports
or
like
high-level
owners,
would
then
have
some
form
of
a
roadmap
and
like
only
we
wanted
to
bring
all
those
features
together
and
I,
like
figure
out
an
overall
roadmap
that
makes
final
year
or
something
because
you
don't
wanna,
do
all
that
together
and
so
wondering
like.
If
I
can
merge
all
that
together
and
then
like
have
a
clear,
long-term
roadmap,
even
if
the
long
term
is
just
for
a
year.
C
You're
like
begin
the
process,
you
try
to
identify
some
owners
for
each
of
those
major
topics.
I
guess
a
peach.
Others
will
try
to
come
up
with
some
form
for
detail
work
plan,
even
though
it
may
not
be
accurate.
But
it's
supposed
to
give
some
idea
on
like
how
long
the
next
one
is
going
to
take
and
depending
on
that,
we
can
in
stagger
or
we
can-
or
we
can
do
things
in
Parliament.
So.
A
B
B
C
So
those
are
just
a
syllabus
for
now,
because
you
just
thought
that
we
could,
you
could
make
you
know
I
think
so
we
decided
you
need.
Okay,.
D
A
A
At
this
list,
the
only
thing
I
would
add
is
like
what
we
discussed
here,
which
was
we're
very
motivated
to
do
something
in
the
cpu
States
in
the
near
to
mid
term,
and
so
stuff
has
been
tasked
with
that
and
so
upon.
A
stuffing
Connor
are
communicating,
I'm
happy
and
then
I
will
go
through
the
rest
of
us
afterwards
to
more
detail
like
on
the
context
I
for
resource
passes,
fish,
that's
really
get
to
a
two
part
problem.
A
That's
like
a
scheduler
problem
and
a
node
problem,
and
so
I
want
to
sure
like
who
we
assumed
was
maybe
going
to
go
and
prototype
the
concept
to
ensure
like
even
basic,
counting
work
and
I.
There
are
potentially
people
on
my
team
that
I
could
sign
up
to
do
that,
but
that
would
just
be
verifying
that
the
thing
will
actually
work
and
not
a
so.
The
final
thing:
it's
not.