►
From YouTube: Kubernetes Resource Management WG 20171213
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
A
All
right,
so
this
is
the
December
13th
meeting
of
the
resource
manager,
working
group
and
a
number
of
items
on
the
agenda.
I
believe
this.
She
put
a
number
of
these
there
if
I'm
not
mistaken.
So
in
the
interest
of
time,
let's
go
through
each
of
these
items,
one
by
one.
So
do
you
want
to
go
through
the
V
ones
and
work
plan
topic
fish,
yeah.
B
You
Balaji
awesome,
so
I
mean
I
had
a
few
thoughts
in
my
mind
but
like
before
I
go
on
about
them.
I
want
to
understand
what
is
it
that,
as
a
group,
we're
considering
achieving
in
the
one
pattern
time
frame?
I
just
want
us
to
be
cognizant
of
the
fact
that,
even
though
we
have
a
lot
more
people
in
the
community
now
focusing
on
this
area,
the
review
bandwidth
is
like
pretty
limited
on.
B
The
churn
rate
also
needs
to
be
limited,
so
we
need
to
like
figure
out
some
form
of
balance
between
graduating
existing
features
and
stabilizing
them
versus
adding
new
things,
and
so,
let's
explain
like
the
next
ten
or
fifteen
minutes
together,
identifying
priorities
for
all
of
us
and
then
seeing
how
we
can
how
we
can
sort
of
like
place
them
in
the
1:10
time
frame.
The.
A
B
I
plus
one
most
of
that
from
our
main
priority
now
is
to
is
to
graduate
device
plugins
to
beta
I'm.
Also,
they
continue
focusing
on
graduating
all
the
other
existing
experimental
features,
because
we
have
had
them
for
a
while
and
it's
I
think
it
gets
through
the
stack
and
that,
like
we,
have
a
few
experiment,
controls
and
so
on.
So
like
trying
spend
energy
on
graduating
them
would
be
the
most
priority
and,
secondly,
I
would
like
to
start
collecting
use
cases
for
better
resource.
B
Api
is
not
necessarily
like
finalize
the
design,
but
just
at
least
have
a
sketch
of
like
one
of
the
different
scenarios
that
we
are
trying
to
cover
and
maybe
have
like
a
few
high-level
ideas
that
are
circulated
in
the
community
and
that's
about
it.
I
guess
so
about
finalizing
or
it's
not
about
like
deciding
the
execution
plan
and
so
on.
D
E
D
A
Just
on
the
resource
class
front,
like
I'm
curious,
if
I
was
to
gauge
where
the
community
of
folks
here
are
like
or
from
my
perspective,
I
still
work
to
do
to
make
it
easy
to
operationalize
like
a
homogeneous
set
of
clusters
or
notes
that
some
of
the
advantages
around
resource
classes
and
getting
heterogeneous
support
are
at
least
more
forward.
Looking
for
me,
so,
like
is
something
that's
a
near
term.
Production
concern
for
folks
that,
having
it
be
anything
beyond
a
stretch,
goal
would
be
like
a
major
detriment.
B
B
D
D
Think,
like
a
whether
we
put
here
the
pupae
up,
you
too,
let's
try
to
make
some
progress
and
I
mean
I
feel
nobody
will
really
come
to
us
and
try
to
put
I.
Don't
think
that
will
be
the
situation
but
I
think
I
agree
with
Derek
that
we
may
want
to
come.
Put
some
time
put
more
time
on
this
ethnic
graduating.
The
existing
features
to
beta
I
think
it's
always
her
party,
but
like.
We
may
also
want
to
plan
to
put
some
time
on
this
effort
as
well
at
least
arm
I,
said
I.
B
B
D
B
So
yeah,
so
there's
one
more
thing
that
I
forgot
to
mention
I
would
also
like
to
see
in
a
one
time
time
time
anything
consumption
of
NPD,
a
cheap
used
in
a
really
easy
manner,
should
make
you,
unlike
other
local
cluster
scenarios,
without
useless
to
go
through
a
lot
of
me.
I,
don't
know
that's
a
common
interest
for
people,
but
like
at
least
with
the
cube
flow
project
that
we
announced
recently
I
mean
this
has
come
as
a
very
strong
signal
and
feedback
from
people
trying
it
out.
B
B
G
B
Criminal
deployment
function.
It
is
not
meant
to
be
a
full-fledged,
fully
user-friendly
crystal
power.
So
then,
so,
let's
not
even
get
to
like
what
is
the
right
abstraction
for
expressing
that.
Maybe
this
is
not
the
right
form
for
that,
but
yeah
we
definitely.
We
have
to
identify
what's
the
least
common
denominator
for
focusing
on
energy.
A
E
Of
free
factoring
right,
just
just
designed
for
110,
okay,
I'm
I
romantically
get
done
with
everything
I
actually.
B
Maybe
we
first
invest
time
on
building
some
some
good
real-world
performance
benchmarks,
because
it's
synthetic
once
we
can
like
set
them
up
such
that
we
will
see
performance
differences,
but
does
it
actually
matter
to
use
this
and
and
that
maybe
we
can
give
them
specific
guidance
or
white
papers
on
how
to
cuz.
You
notice,
right
so
before
we
even
like
in
this
more
energy
in
like
identifying
more
of
these
features
and
setting
them
up
in
specific
but
specific
policies.
I
would
like
to
see
some
benchmarks
available
in
cuteness.
B
D
So
maybe
we
should
make
backup
some
standard.
A
performance
benchmark
is
a
critical
now
to
find
the
new
mom,
because
I
also
I
also
think
they
say
something.
Maybe
we
would
like
to
say
and
I
hope
the
benchmark
is
not
just
limited
to
new
MA,
although
maybe
Numa
will
be
the
first
feature
that
we
may
use
to
evaluate.
G
B
G
B
A
That's
what
I
was
trying
to
differentiate
so
Connor
said
one
thing
which
was
he
wanted
to
get
locality
decisions
made
relative
to
the
pinning
CPU
choice.
I
can't
tell
vish
if
you're
saying
you
don't
want
to
do
that
and
just
depend
on
Auto
numeral,
oh,
but
I
feel
like
we
do
need
to
get
that
aligned
yeah.
B
I'm
my
thought
process
is
typically
like
more
CPU
cores
than
the
external
PC.
A
hardware
devices
like
wait.
I
mean
choose
the
rest
based
on
that,
just
because
of
like
how
machines
are
typically
set
up.
But
again,
this
is
getting
into
the
details
where,
like
you,
don't
have
any
proper
data
to
verify
where
the
performance
level
is
today
and
what
the
day
latency
is
like.
I
think
anything.
The
folks
did
some
work
around
it's
one
or
something
which
tries
to
benchmark
this
thing.
I
would
like
to
see
that
become
part
of
the
equivalents
community.
B
B
Scaling
and
suggesting
are
trying
to
something
similar
so
just
see
if
we
can
have
have
like
single
abstraction
layer
for
running
these
kind
of
benchmarks,
because
otherwise
we'd
have
to
go
build
all
of
it,
which
is
probably
not
that
interesting
for
us.
So
maybe
we
need
to
like
engage
with
those
things
and
find
out
what
is
the
working
group
that
is
working
on
just
real
those
benchmarking,
and
maybe
we
have
a
few
tests
that
are
specific
to
our
features.
B
B
B
A
Discussion,
like
I,
didn't
hear
any
rampant
disagreement
of
let's
work
towards
graduating.
What
we've
had
thus
far
at
best
we're
kind
of
now
like
prejudging
any
any
design
thoughts
that
Connor
might
put
out
right,
which
is
a
bit
premature,
I
guess
so,
like
as
a
community
like
I'm,
not
saying
we
should
commit
to
doing
anything
with
Numa
got
a
code
standpoint,
110
I,
believe
you're,
saying
the
same.
I
think
it's
it's
fair
for
Connor
to
go.
A
B
E
I
We
did
try
to
kick
this
off
like
Jeremy
and
Christopher
and
I.
At
some
point
we
could
look
at
what
we
could
scribble
down
back
then
and
Mikey
would
hausky
was
also
a
part
of
that
you
pump.
We
could
have
the
gun
and
see
where
lifted.
B
G
Real
the
world
world
benchmarking
is
significant
effort.
If
we
block
on
the
inclusion
of
that
we're
going
to
be
waiting
for
a
long
time,
potential
suggestion
is,
is
her
onyx
or
or
perf
kit
benchmark?
Does
a
lot
probably
be
a
lot
easier
to
integrate
and
a
lot
less
debate
about
what
gets
run
a
lot
simpler,
I
think
our
team,
even.
B
I
On
110,
but
there's
just
a
collection
of
benchmarks
that
exists
like
like
for
Cassandra,
like
the
Y
of
Y,
is
a
white
CSB
and
CPUs
bag
and
whatnot
well,
I
hope
grown
parallel
right,
because
some
things
won't
work
on.
Unless
we
have
no
awareness,
like
the
tooth
pages,
isn't
that
right?
Some
things
are
just
getting
on
me.
If
you
get
to
like
many
socket
systems,
will
won't
be
a
great
experience
on.
B
The
Google
side,
like
I,
some
of
the
teams
that
are
maintaining
from
a
benchmark
are
actually
add
benchmarks
with
tensorflow,
but
we
like
do
random
device
allocations
and
try
to
measure
standard
deviation,
so
I
mean
all
that
all
those
benchmarks,
I
would
expect
them
to
be
upstream.
So
I
I
would
consider
it
to
be
less
technical
work
to
get
that
running
in
a
pod,
and
then
it's
just
running
a
lot
of
them
and
like
storing
the
data
and
Prometheus,
or
something
like
that.
B
G
The
pircuit
bench
marker
was
brought
up
in
that
document.
I'm,
not
I'm,
not
against
that
at
all.
It's
just
the
the
fullness
of
that
document
was
well,
it
was
ambitious,
I
would
say
necessary.
Quite
honestly,
we
pitched
a
giant
proposal
to
sing
scale
like
a
year
and
a
half
ago.
That
was
a
real.
What
we
called
it
sink
and
it
just
didn't.
We
never
snapped
it.
So
I
was
glad
to
see
someone
else
in
the
community
bring
this
up
again
if
they
can
actually
staff
it
yeah.
B
I
agree,
I
definitely
needs
like
one
or
two
people
who
would
work
for
a
month
or
two
at
least
get
something
going,
because
once
we
have
what
I've
noticed
is
like
once
we
have
these
frameworks,
the
people
usually
like
end
up
adding
more
and
more
benchmarks
and
one
more
test.
It's
that
initial
inertia
of
like
building
that
framework.
That
is
what
is
typically
that's
the
thing
that
typically
takes
more
time.
Oh.
A
B
A
A
G
So
if
everyone
on
the
phone
was
there
than
or
or
attending
remotely
that
I
don't
need
to
go
into
it
again,
I
would
say
that
there
was.
There
was
good
a
lot
of
good
scheduling.
Talk
there
I
mean
the
there's
like
Cuba.
It
was
aq
arbitrator,
I
learned
about
that,
but
anyway,
I
can't
see
who's
on
the
call.
So
I
guess
I'll
just
assume
that
folks
have
a
good
sense
of
what's
discussed
there.
J
G
You
what
the
one
thing
was,
one
thing,
I
kind
of
heard
loud
and
clear-
was
capacity
management
capacity
planning.
That
was
the
one
thing
and
then
the
other
one
was
like,
although
it
was
probably
heavily
biased
or
the
gentleman
from
Intel
was
they're
talking
about
FPGAs
we
keep
hearing
about
that,
especially
since
Amazon
is
selling
them.
So
that
was
the
other
thing
that
came
up
requirements
wise
and
I.
Quite
honestly,
I
think
everybody
is
waiting
for
us
to
get
the
device
plug
in
and
user
story
around
GP
is
nailed
down.
B
There
was
one
more
thing
which
is
like
resource
sharing
across
namespaces,
especially
like,
if,
if
you
have
a
large
set
of
resources
and
if
you
have
different
people
like
holding
or
and
like
holy
aim
through
utilization,
I,
think
there
was
a
project
called
cube
arbitrator
but
like
trying
to
make
it
more
upstream
and
like
trying
to
make
it
more
relevant.
Like
those
other
discussions
going
on.
D
B
B
G
G
You
know
where
this
started
guys
was
the
where
the
heck
do
we
set
Cube
reserved
and
what
do
we
set
system
reserve
to
and
we
started
looking
into
it's
really
difficult
to
answer
this
at
cluster
scale
right
now:
we've
given
the
existing
metrics
so
on
our
side,
well,
I'm,
not
exactly
sure
where
the
changes
will
go,
but
that's
how
I
was
attempting
to
like
address
this.
It's
a
big
concern
internally
as
well.
So.
D
G
A
So
personally,
I
think
this
discussion
should
probably,
if
we're
talking
about
like
reservations
on
a
note
and
stuff,
that's
probably
something
we
should
drive
through
signal,
more
proper
and,
as
you
know,
Jeremy.
Our
our
observations
are
that
that
couples
a
lot
of
things
like
we've,
we've
had
large
numbers
services,
destroy
our
CPUs
as
well.
So
were
there
any
other
topics
we
wanted
to
come
out
of
from
the
Keep
Calm
meetings,
or
do
we
want
to
go
and
is
someone
here
that
want
to
talk
through
the
cluster
scope,
resources.
A
B
A
Do
not
hear
an
author,
so
I
would
say:
let's
move
on
to
the
next
topic
here,
which
was
what
an
interest
in
talking
through
Moriches.
Can
we
so
at
the
end
of
1:9,
we've
had
the
very
good
discussion
around
like
graduating
plugins,
my
vision.
If
you
want
to
give
an
update
now
on
110
and
how
we
can
get
over
the
hump
that
we
fell
in.
B
B
B
B
B
A
G
So
the
that's!
How
I
spent
my
week
last
week
for
the
most
part
talking
with
different
people
and
if
you're
subscribed
to
sig
network
you'd,
have
seen
me,
propose
a
new
working
group
to
sort
this
out
amongst
the
vendors
and
when,
in
those
meetings,
I
made
sure
they
understood
that
there's
an
awesome
event,
rural
intersects
between
the
device
plug-in
and
whatever
they
come
up
with
initially
using
CR
DS.
G
So
we
we
actually
committed
some
engineering
resources
inside
Red
Hat
to
to
work
and
build
prototypes
in
that
area,
along
with
several
of
the
other
vendors,
including
Intel
and
that'll,
be
if
you
track
sig
network
you'll
see
it's
coming
up
there
and
one
thing
we
can
do
when
that
kicks
off
on
December
21st
is
bring
up.
The
concerns
mentioned
now
around
what
to
do
in
110,
I,
just
don't
know
if
they're
going
to
have
solid
direction
in
the
timeframe
where
we're
looking
to
potentially
graduate
things.
So,
but
it's
worth
mentioning
to
that.
D
G
D
Clarify
the
the
what
it
cause
we
want
to
get
from
the
I
mean
I
know
it's
definitely
helped
us
evaluate
the
API
design
with
different
types
of
device
or
device
tracking
implementations,
but
I
also
want
to
don't
want
to
depend
on
the
like,
because
it
since
we
don't
really
have
all
resources.
Human
resources
allocated.
I
know
the
device
plug
implementation.
Right
now
seems
like
no
one
has
confirmed
they
will
actually
in
110
timeframe,
they
will
be
able
to
report
like
make
available
a
best
tracking
presentation.
D
I
just
want
to
set
clear
goal
like
what
we
want
to
get
from
the
different
team.
Experimentations
I
think
if
we
just
want
to
hear
a
feedback
and
I
think
they
already
have
some
faith
bags
from
early
developers
on
different
device
practice,
and
maybe
we
can
just
summarize
those
feedbacks
and
then
maybe
we
can
just
like
ask
her
on
the
folks
to
say
like
why
that
they
think
the
feature
is
ready
to
graduate
to
beta.
Instead
of
really
like,
depending
on
the
real
implementation
you
use
the
in
production
or
whatever
I.
A
D
I
see
it
I
also
feel
the
same
way
like
you
from
a
just
a
whole
to
feature
at
the
alpha
stage.
It'll
be
me
know
we
may
never
get
enough
feedback
and
I
know
it's
a
little
commitment
that
we
to
support
Romania
to
support
the
different
versions
API,
but
I
just
want
to
say
like
it.
So
from
other
people
who
have
implemented
if
ur
enter,
who
have
tried
to
implement
different
packets
if
they
say
some
some
feature,
I'm
working,
it
had
to
be
found
for
beta
I
would
like
to
say
the
particular
feature.
D
A
Me
the
question
is,
like
other
other
devices,
that
users
in
production
or
vendors
in
this
space
are
looking
to
support
in
any
timeframe
outside
of
GPUs,
and
if
that's
not
the
case,
is
it
realistic
to
expect
that
we
can
meet
the
goal
of
demonstrating
that
device?
Plugins
could
support
more
than
one
device.
A
You
know
at
Red
Hat
we
tried
to
push
forward
the
solar
flare
information,
because
that
would
be
something
that
would
be
compelling
to
our
user
community,
but
obviously
that
had
met
resistance
and
right
now
you
know
correct
me
if
I'm
wrong,
but
I'm
not
aware
of
any
strong
right
now
pushing
us
to
get
like
FPGA
support
in
the
next
three
months
right.
So
there
is
a
major
user
need
for
accelerated
workload.
Types
though
right,
so
the
best
I
can
think.
A
A
Guess
how
firm
is
the
commitment
is
that,
like
I
mean
there's
a
lot
of
ways
we
can
handle
this
right?
But
it's
like
it's
if
we
find
ourselves
at
the
end
of
1-10
without
a
clear
alternative
device
plug
into
like
to
be
I'm,
not
sure
if
we
continue
to
hold
right,
because
at
some
point
like
where
there's
actual
demand,
the
need
like
that
seems
to
be
what
the
community
is
directed
or
realize.
Then
right,
I.
G
G
A
J
B
K
K
D
We
can
discuss
them
in
more
details
like
and
what's
the
right
way
to
send
those
parameters
between
occupied
and
the
device.
Rocky
I
just
feel
like
a
sending
container
idea.
Our
idea,
our
path
cycle
may
may
open
the
tower
too
wide,
but
maybe
we
can't
discuss
the
particular
use
cases
and
say
by
the
we
can't
come
up
with
some
solution.
A
That
was
the
last
topic
on
today's
agenda.
The
only
other
topic
I
want
to
have
is
I.
Do
not
think
we
should
hold
any
more
meetings
for
the
remainder
of
the
year.
So
just
a
heads
up
on
that
and
then
another
topic,
I
wanna
discuss
was
have
people
think
through
a
little
bit
is
right.
Now
we
have
a
weekly
cadence
for
this
meeting,
I
think
given
what
we've
discussed
right
now
for
what
the
community
interest
is
around
110
and
where
our
focus
was
would
lie
like
if
we
were
to
move
to
a
biweekly
meeting.
A
Would
there
be
objections
to
that
or
you
know,
given
a
new
year,
we
have
some
time
to
think.
If
you
want
to
do,
anything's
I
mean
anything
differently.
So
if
people
have
suggestions
on
how
we
could
run
the
group
better
or
more
efficiently
or
anything
in
that
regard,
please
just
reach
out
to
both
me
and
fish,
but
let's
use
the
new
year
to
do
things
better.
If
possible,
yeah.
G
B
People
are
collaborating
on
like
specific
specific
items
in
an
agenda
then
they
can
always
have
their
own
meetings.
The
only
request
from
Brian
has
been
that,
like
those
meetings
should
also
be
documented,
so
it's
just
if
you're
having
those
dialogues
meetings
is
just
like
documented
in
our
community
page.
F
Balaji
pundits,
very
quick
and
I
actually
talked
to
this
corner
about
this
before
yeah
about
the
topology
of
achieves
you
use
them,
you
are
using,
so
this
is
about.
Actually
we
are
planning
to
move
or
the
Microsoft
research,
machine,
learning
or
cloud
to
kubernetes.
It's
very
big
plans,
but
every
machine
learning
training
job
actually,
depending
on
this
kind
of
cheap.
F
You
topology,
for
example,
the
in
being
their
link,
for
example,
and
we
have
paper
and
benchmark
to
prove
that
this
kind
of
acceleration
can
can
improve
the
training
performance
to
very
high
level,
but
so
the
blog
here
we
see
her.
Actually
in
or
alignment
we
cannot
be
device
playing
at
all
because
plot
device,
the
device
called
me
only
can
say:
I
want
to
teach.
You
is,
for
example,
but
what
we
want
it
like.
We
want
to
use
two
GPUs,
which
is
which
are
connected
by
a
media
link,
so
its
cuddly
is
kind
of
different
requirement.
F
So,
although
we
are
using,
Chrome
is
one
canary,
but
we
cannot
you
device
plug
in
what
we
are
doing
today,
that
we
have
a
CR
ashame,
it's
very
ugly
because
we
have
to
move.
We
have
a
copy
most
of
the
logic
in
debugging
and
writes
the
meeting
Stalinists.
They
are
ashamed
to
do
that.
So
what
I
want
to
discuss
today
is
actually
the
topology
stuff
can
be
handled.
Most
of
the
case.
Indicators
can
be
hand
on
handle
being
the
resource
in
the
resource
class
proposal.
F
G
B
You
can
have
this
conversation
again
after
New,
Year's
or
email,
something
because
it
goes
like
TL
DR.
Is
that,
like
we
decided
not
to
deal
with
topology
at
the
cluster
level
and
do
it
at
the
node
level
like
have
cubelets
sort
of
handle
a
graph
and
do
some
scheduling
but
like
if
you
read
through
that
loads,
you
sort
of
understand
the
rationale
behind
that.
B
A
Class
proposal
did
come
out
of
that
meeting
right
and
it
was
prototyped.
So
if
you,
if
you're
saying
initially
the
ideas
is
germinating,
something
that
you
would
get
value
out
of
I
guess
that's
good
feedback.
But
if
you
haven't
had
a
chance
to
read
through
those
notes,
Harry
I'd
happy
to
point
you
to
them.
Just
I
would
ask
that
you.
Actually
you
should
yeah
it's
amazing
how
many
requests
I
get
to
be
able
to
get
access
to
those,
because
people
don't
want
to
join
criminate
staff,
but
that's
not
an
issue.
Yes,.
D
And
I
also
want
to
understand
a
highway
if
you
have
some
like,
because
I
think
all
right
now,
I
want
to
understand.
Like
a
are
you
interested
in
doing
this
yeah
prime
or
you
are
trying
to
use
something
cloud
pride
environment
because
I
think
there
are
certain
ways
to
work
around
the
problem,
even
I.
Right
now
like
say
you
can
create
some
cracker
with
the
particular
our
diagram.
A
group
node.
F
F
B
B
D
I
think
I
want
to
learn
like
the
particular
pin
PI.
So
the
current
way
of
doing
this
I
can
say
like
a
possibility.
It's
my
issue
and
the
MA
automation
required
in
future.
It
could
also
be
value,
but
I
would
also
would
like
to
hear
y'all
just
as
a
given
that
you
have
a
good
experience
on
this
and
yeah
just
ask
dick
the
PIMCO.
You
hype,
mm-hmm.
E
Speaking
of
following
up
anyway,
keep
con
a
few
people
mentioned
next
iteration
of
an
in-person
meeting.
I
was
wondering
if
it
if
people
were
interested
in
that
next
year
and.
A
No
I'm
just
trying
to
think
what
the
conference
states
were.
I
thought
I
knew
it
started
with
an
M
and
I
wasn't
sure
if
it
was
earlier
and
then
I
I
know,
there's
a
open
source.
Leadership
Conference
in
March,
but
I
would
be
interested
in
having
a
get-together
again.
I
thought
that
was
really
constructive
is
one.
A
Enough
of
a
breadth
of
topics,
vision,
I,
myself,
I
think
he
and
I
can
vouch
that
we
had
to
do
like
multiple
weeks
of
preparatory
work
for
that
face-to-face
that
was
rather
exhausting.
So
it
takes
time
to
coordinate
those
to
be
as
productive
again
so
I'm
in
favor
of
them,
but
we
just
need
to
make
sure
that
they
they
have
enough
lead
time
and
topic
space.