►
From YouTube: Kubernetes Resource Management WG 20170622
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
B
C
B
D
D
D
I
did
have
a
few
specific
things
to
talk
about
today
and
the
first
one
is
yesterday
I
and
tackled
the
problem
ways,
one
for
example,
cubic
crashing
or
the
device
again
crashing
and
I
want
to
share
my
accounts
about
that
and
see
what
you
thought
of
the
solution:
I've
employed
and
and
the
second
one
was
mostly
about
I.
Don't
know
if
anyone
here
had
some
time.
D
D
And
well,
it's
always
a
be
boots.
If
let
me
just
take
out
the
Vice
plug-in.
B
E
B
D
So
I
stopped
at
the
discovery
phase,
and
so,
let's
say
a
pod
record
requests
a
GPU
and
then
the
qiblah
crashes
or
the
device
plane
crashes.
And
what
happens
is
that
when
the
qubit
will
device
for
parts
again
and
then
we
stablish
connection,
we
needed
a
well
I
was
thinking
that
we
needed
a
way
to
say
to
the
light
plug
in
them.
There
are
some
device
that
needs
to
be
initialized
and
something
that
should
not
be
realized.
D
B
D
G
G
So,
just
just
wondering
all
these
crashes
movie
be
real
common
cases,
because
you're
actually
looking
at
you
know,
systems
which
are
pre
validated
right,
I
mean
this
thing.
Is
you
provided
in
there
bringing
of
this
in
production
systems?
For
expressing
here,
I
mean
this
is
more
of
developmental
system
problems?
It
all
the
scratches
and
all
that
well.
B
G
B
I
mean
from
a
prototype
standpoint-
maybe
I
you
don't
like
for
showing
a
demo.
You
don't
care
about
this,
but
for
like
it's
really
too
big
people's
the
beta's.
We
like
need
to
handle
all
these
corner
cases
right.
So
maybe
it's
unnecessary
to
fix
all
that
this
week,
but
it
is
still
talking
about
clay,
yeah
and
those
requirements
of
move.
The
device
plugin
is
expected
to
work
and
how
keep
that
it
expected
the
interface
level
s
plug-in
these
questions
have
to
be
a
case.
It
is
not
super
lightly.
D
So,
just
to
finish
on
that
I
have
been
a
person
and
so
say
a
point
using
a
GPU
is
running,
qubit
crashes
and
qubit
gets
up
again
and
the
con
terminates
and
then
after
the
pod
emanates
the
GPU
device
going
connects
itself
to
the
chelate
and
x-cubed
supposed
to
wait
until
the
GPU
device
program
connects
itself
to
and
tell
it
that
I
need
to
deallocate
the
GPU.
That
was
just
a
minute
terminated
or
is
there
or
do
we
expect
that?
Because
the
device
volume
just
could
be
connected,
then
we
don't
need
to
do
locate.
B
Having
said
that,
the
village,
plugin
and
I
also
mentioned
that
iDevice
plugins
have
to
be
creative
and
stateless,
ideally
because
writing
stateful
plugins
would
be
a
big
deal,
also
have
some
bugs.
So
you
want
to
centralize
all
this
logic
and
cubelet
that
cubelet
will
have
to
come
back
up
and
then
in
love.
B
It'll
have
you'll
have
to
have
like
a
loop
that
we
go
and
figure
out
what
devices
are
allocated
and
then
like
which
ones
are
actively
being
used
by
pause
and
then
like
find
out
the
ones
that
have
to
be
de-allocated
and
then
go
in
here
like
is-
and
this
is
standard
cubelet
does
for
a
lot
of
other
things.
So
this
will
only
new
with.
B
B
B
So
Cuba
would
have
a
loop
then,
like
that,
does
take
care
of
the
allegations.
It
basically
look
at
the
current
date
under
the
five
state
and
then
figure
out
the
disk
and
go
ahead
and
like
be
allocated
deep
use
or
device
that
are
not
expected
to
be
attached.
Rain
falls
then
it's
necessary
viewing
allocation.
B
It
can
trigger
this
clean
up
again,
just
to
see
if
any
you
need
devices
can
be
freed
up
for
a
new
part,
because
you
say
for
some
reason:
it
figures
out
that,
like
every
every
device
is
allocated
right,
then,
if
I
should
perform
a
trick
just
to
see
if
some
of
them
can
be
be
allocated,
and
so
it
could
could
do
such
a
thing.
Then,
if
there's
no
devices
exist,
then
it's
going
to
reject
the
pod.
B
D
What
I
was
thinking
was
just
adding
a
instead
of
looping
and
waiting
for
the
device
plug-in
to
reappear,
whereas
with
thinking
of
is
adding
a
call
just
after
the
in
it.
That
would
say-
and
we
can
soften
upon
that-
says
these
creepy
you
are
used
and
these
DP
you
are
not
used.
Can
you
do
locate
these
confuse
yeah.
B
H
B
B
D
D
Let
me
see
the
pot
requests
for
one
GPU
allocates
one
vpu
in
a
cod
and
then
say
everything
goes
wrong
at
this
point
you
play
crashes,
device,
plane
crashes
and
whichever
you
within,
whichever
whatever
order
everything
goes
on
at
this
point,
I
was
thinking
that
and
when
the
qubit
goes
up
again
and
I
say
the
user
says
I,
remove
this
pod
right
at
the
device
plug
and
still
hasn't,
registered
and
I.
Think
what
vishwas
saying
is
that
we
are
going
to
have
a
background
process.
B
G
B
D
D
B
D
B
So
the
thing
is
devices
which
are
are
also
going
to
have
topology,
and
so
it
also
look
like
programming,
see
groups
and
figuring
out,
like
your
location
books,
and
you
know
what
so
forth.
So
the
ideal
ideal
place
like
device
allocation
would
happen
would
be
as
part
of
the
container
manager.
So
like
this.
B
So
all
I
coming
down
from
the
container
monitor
and
the
runtime
is
just
like
passing
it
through
through
the
CRA
and
the
five
beers
of
I
am
so
have
this
grander
vision
that
I
don't
know
if
this
is
like
I'm
thinking
of
this
Hardware
plug-in
interface,
similar
to
the
container
storage
interface,
and
that,
like
I,
don't
reason,
white
cubelet
should
be
the
only
one
using
this
interface
in
theory,
like
doctors
really
know
like.
If
rocket
is
considered
as
like
a
darker
thing,
then
leap
could
use
that,
like
other
runtimes
could
also
in
theory
uses.
B
B
D
G
So
a
wish
to
this
summarize
what
I
said:
are
you
proposing
this
specifically
for
device,
lugging,
property
injection,
so
whatever
specific
devices
you
want
to
have
110
any
way
of
thinking
those
containers
so.
B
I'm,
imagining
you're
exactly
imagining
a
gold
library
of
the
people
which
could
expose
an
API.
That
would
be
the
basic
methods
that
are
available
as
part
of
the
G
RPC
server.
And
then
that
would
then
like,
like
all
the
interactions
with
like
two
more
plugins,
and
then
any
consumer
would
be
a
cubelet
or
a
test
or
like
some
other
authorization.
B
H
D
B
D
G
G
So
one
part
is
essentially,
you
know
installing
the
kernel
driver,
so
that
is
the
user
space
drivers,
but
both
these
operations
seem
more
natural.
You
know
during
the
initiative,
then
the
device
plug-in
is
started.
So
basically,
when
you
start
the
diff
I
mean
spawn
the
device
plug-in.
That's
when
you
do
these
operations
or
when
you
stop
the
device.
Plugin
is
when
you
do
these
operations,
not
you
know
typically
on
the
fly,
but
that
said,
can
you
look
at
some
models
where
I
mean,
of
course,
to
install
these
drivers?
We
need
root
permission.
G
Basically,
my
thought
process
goes.
If
you
can
kind
of
get
the
install
done
as
soon
as
the
device
plug-in
starts
and
some
method
before
actually
doing
anything
else,
you
know
as
root,
and
you
know,
then
the
rest
of
the
device
plug-in
could
be
running
pretty
much
as
non-root
brake
because
it
will
be
then
looking
at
kind
of
operating
things
like
failure,
detection
or
you
know,
monitoring,
etc,
I.
So.
G
What
I
meant
was
a
wish
I'm
trying
to
explore
is
any
possibility
where
you
don't
even
start
the
disk
enough,
and
this
can
happen
elsewhere
it
basically
the
driver,
install
assuming
I,
mean
what
I'm
seeing
is.
This
is
basically
when
you
start
the
device
plugin
or
stop
the
device
that
in
the
first
operation
right.
So
if
you
can
get
it
then
before
actually
exactly
like
it
said,
starting
the
device
plug
in
the
process,
then
tip
of
the
device
plug
in
then
we
can
potentially
is
not
the
device
again.
B
I
B
H
I
B
No
I
mean
I
assume
what
you
mean
is
that,
like
every
user
pod
would
install
its
own
drivers,
I
still
think
that
that's
an
Oracle
and
I
still
think
that
it's
probably
better
to
like
use
some
other
scheduling
primitives
to
achieve
multi-tenancy
in
the
in
that
regard,
because
a
lot
of
resources
are
being
spent
and
installing
it
I'm
not
saying
that
it's
it's
not
a.
Why
of
the
solution
and
I'm
missing.
It
is
not
a
it's,
not
okay,
with
making
that
the
default
option
for
for
every
user.
I
B
You
know
the
compile
is
not
once
classified,
because
the
node
version
keeps
changing
during
the
lifetime
of
the
cluster
so
and
then
the
the
kernel
version
is
also
changing
as
part
of
that,
so
we
until
I
keep
compiling
and
then
the
image
the
image
has
to
be
smart
enough
to
download
the
right
set
of
drivers
for
each
version
and
so
forth
in
theory,
it's
possible.
Can
we
like
consider
that,
as
like
step
two
rather
than
step
one?
I
I
D
I
Yeah
I
mean
go
like
I
was
just
thinking
of
generic
machinery
that
could
be
used
for
any
module,
and
you
know
it's
the
fact
that
those
drivers
are
more
difficult
and
the
other
shouldn't
matter
to
cube.
We
should
push
that
difficulty.
You
know
away
from
the
platform
like
there's.
Other
drivers,
I
have
similar
concerns,
but
like
there
should
be
a
HelloWorld
kernel
module
and
a
similar
workflow
for
every
other
module,
and
you
know
this
driver
does
XYZ
when
it
needs
to.
B
B
J
J
A
B
G
Yep
and
then
I
just
wanted
to
have
a
word
with
you
on
the
code
as
VD
okay,
so
we
can.