►
From YouTube: Kubernetes Resource Management WG 20181024
Description
Meeting Agenda:
https://docs.google.com/document/d/1j3vrG6BgE0hUDs2e-1ZUegKN4W4Adb1B6oJ6j-4kyPU
A
Hello
Brett
so
welcome
everyone
to
the
October
24th
resource
management.
Working
group
meeting:
apologies
for
those
who
are
watching
this
and
video
Who
I
am
starting
the
meeting
a
little
bit
late,
and
so
thanks
for
those
who
stuck
around-
and
so
we
have
one
item
on
today's
agenda,
it's
my
own
Christian
feel
free
to
talk
to
your
cut.
B
B
It's
a
bug,
bug
about
our
own
fee
or
Cooper.
On
this
issue
about
AVX
and
the
idea
there
is
that
if
we
are
running
a
VX
/
clouds
that
are
mixed
with
this
irregular
workloads,
then
a
VX
will
actually
like
push
push
down
the
base
frequency
very
easily,
which
the
means
that
non
AVX
workloads
with
are
running
slower
and
the
problem
here
is
that
it
makes
it
where
it's
slower
and
makes
those
work
harder
to
kind
of
reason
about
the
workload
speed
when
this
ABX
workloads
are
running
surf
too.
B
B
So
the
idea
is
that
we
would
have
two
new
configuration
values,
the
couplet
which
would
repeat
this
best-effort
cpu
set
and
other
cpus
it,
and
that
would
mirror
this.
Best-Effort
port
would
be
limited
to
this
PE
cpu
set
antha
parent
granted.
What
would
be
running
on
the
other,
cpus
it
and
by
default
both
of
these
abuses
would
actually
be
or
cpus,
meaning
that
everything
would
run
everywhere
and
only
if
needed-
and
this
would
be
kind
of
system-
could
be
partitioned
in
the
and
the
smaller
smaller
pieces.
And
then.
B
B
Examples
and
first
variant
is
that
this
AVX
workloads
would
be
running
up
as
best-effort
workloads.
So
it
means
that
everything
else
would
have
a
cpu
resource
request
on
same,
and
so
they
would
be
running
on
some
on
some
on
this
other
pool
and
only
this
AVX,
which
would
run
on
this
best
effort
and
that
sort
of
the
kind
of
them.
B
B
But
we
say
here
that
it
can
negatively
affect
cluster
efficiency,
which
is
sort
of
true,
and
it
has
some
other
side
effects
like
most
likely.
People
don't
want
to
do
it.
We
heard
this
from
our
customers
that
they
are
not
not
sort
of
happy
about
this.
This
approach,
this
alias,
may
be
X
mitigation,
so
they
would
rather
rather
split
that
nodes
and
not
kind
of
split
CPUs
inside
the
nodes
and
not
the
nodes
themselves.
C
B
A
B
Yeah,
this
is
actually
a
really
really
good
question,
because
that
sort
of
comes
down
to
the
question
that
how
exactly
do
this
a
BX
bar
close-
affect
the
frequency.
The
base
frequency
after
for
this
kind
of
non-native,
expert
close
and
that
sort
of
depends
on
many
things.
For
instance,
I
am
stunted
it's
different
between
different
processors
and
so
on.
B
So
probably
there
will
be
some
tooling-
maybe
maybe
it's
not
done
by
us
on
us,
but
we
do
have
some
work
that
helps
sort
of
the
system
administrators
to
analyze
that
what
sort
of
heat
there
would
be,
for
instance,
if
you
think
of
the
kind
of
the
cleanest
separation
of
AVX
and
non
lyrics
for
a
club
stand,
they
are
running
on
the
different
sockets
to
hold
different
processors.
But
if
they're
running
on
the
same
processor,
then
what
can
we
have
to
expect
there?
What
kind
of
performance
changes
then?
That
sort
of
depends.
A
I
guess
the
what
you've
walked
thus
far.
It
seems
pretty
clear
at
the
low-level
side-
it's
just
when:
how
do
we
make
the
set
of
knobs
I
just
I?
Guess
that's
what
I'm
looking
at
on
my
side
me
reviewing
this
so.
B
Yeah,
of
course,
it's
the
kind
of
thing
sort
of
the
ideal
solution
could
would
be
that
you
could
actually
adjust
the
CPU
pores
without
having
to
restart
the
Cupid,
but
it's
still
sort
of
under
investigation
by
our
stud.
What
kind
of
restrictions
exactly
out
there
when
you
do
this,
because
I
mean
I'm
kind
of
the
simple
solution.
B
Is
that
if
you
want
to
adjust
the
CPU
set
sizes,
let's
say
then
you
just
drain
the
node
at
just
the
CPU
sets
and
then
sort
of
run
coordinate
again,
but
it
could
be
that
there
are
some
operations
or
some
adjustments
which
you
could
always
make
without
having
to
do
that,
and
which
would
mean
that
you
just
let
them
CPU
manager
kind
of
let's
get
out
the
pods
to
the
CDs.
And
then
this.
B
It's
it's
typically
like
varied
butter
clocks,
because
the
whole
thing
is
a
problem
only
when
you
have
both
a
VX
and
no
Navy
exper
cross
running.
At
the
same
time,
because
I
mean
a
VX
bar
plots
are
good
in
that
sense
that
they
actually
like
it's.
They
they
perform
really
well,
and
the
only
problem
comes
when
you
have.
This
kind
of
this
work
loves
running
like
mixed
in.
In
the
same
note,
and
but.
B
So
it
so
it
depends,
I
mean
I,
don't
have
any
kind
of
concrete
examples
of
what
kind
of
workloads
could
be
this
kind
of
mixed,
what
a
clutch
first,
there
would
be
like
two
pods,
because
for
in
order
to
have
this
this
core
adepts
that
class
is
there
like
Portland
level
things.
So,
if
needs
to
have
two
ports
when
running
a
B,
X
and
one
running
on
Navy
X,
which
would
need
to
be
then
skeleton
same
note,
that's
what
you
mean.
I
think
yeah,
but
but
another
case
is
just
that.
B
B
That's
right
so,
for
example,
you
you
might
require
the
Debbie
expert
clods.
If
you
knew,
if
you
actually
know
what
they
are,
because
this
kept
kind
of
presumes
that
there
is
that
this
is
kind
of
cooperative.
So
the
expert
clubs
actually
are
somehow
like
known
to
behavior
expert
clothes.
So
there
would
be
this
label
or
'tony
this
convention
that
they
are
just
resources
set
so
that
they
call
this
certain
classes.
A
B
It
the
problem
is,
or
the
real
problem,
isn't
a
mental
sort
of
different
types,
if
they're,
both
AV
experts
and
its
students,
sort
of
effect,
because
the
base
frequencies,
anyway
robot
Ontario,
X's
sort
of
just
getting
older
power.
We
are
talking
about
sometimes
about
this
kind
of
disdains
a
be
expert
cause
when
you
accelerate
executing
a
lot
of
these,
a
B
X,
maybe
X,
in
structures
on
the
same
CPU.
But
that's
that's
not
the
problem
that
this
can.
D
D
One
of
them
is
the
SK
SK
you
so
basically
what
type
of
CPU
running
on,
but
the
other
thing
is
the
what
kind
of
area
constructions
you
are
running
and
the
third
thing
which
affects
it
is
that
how
many
cores
within
the
package
are
running
AVX
instructions,
and
one
of
the
basic
ideas
of
this
cap
is
that
by
being
able
to
limit
the
maximum
number
of
cores
that
are
executing
AVX
instructions
within
a
node,
you
are
able
to
sort
of
limit
the
down
clocking
effect
to
a
accepted
worst
case.
For
that
node.
A
Like
there's
this
idea
that
the
operator
knows
which
workloads
are
using
ABI
X
versus
not
a
VX
and
they've
steered
their
pots
appropriately
to
take
that
into
account
and
and
then
we're
kind
of
contrasting
that,
with
this
like
well,
I
could
be
running
on
a
cloud
and
I'm
kind
of
black
box
and
I.
Don't
know
anything
about
my
and
user
workload,
but
I
pre
pre
configured.
A
My
nodes
with
these
separate
CPU
set
policies
like
I,
guess,
I'm
kind
of
wondering,
like
the
consumer
of
this
feature,
is
a
user
who,
like
I'm,
still
hung
up
a
little
Don.
Why
I
would
not
just
have
dedicated
node
pools
for
one
class
of
workload
versus
another
and
what
the
major
like,
if
I'm
sensitive
to
this
issue,
with
the
major
workload
requirement
itself,
is
that
requires
that
you
were
two
pots
to
be
co-resident
like
one
that
is,
and
it's
not
using
any.
B
Access
them
I
made
a
splitting
of
the
cluster
into
this
AVX
nodes
and
no
networks,
not
it.
It
works,
works
too.
So
no
problem
about
that,
but
it's
just
like
sort
of
an
maybe
bit
easy
solution
for,
because
you
can
actually
use
that
for
any
other
things
so
for,
for
example,
but
I
have
extended
resources
for
let's
say
GPUs.
If
you
can
touch
us
say
that
these
are
the
GPU
nodes
antes
at
the
non
GPU
nodes.
Well,
I
have.
A
An
extended
resource
for
GPU
is
useful,
so
you
can.
They
have
a
count
of
how
many
GPUs
are
scheduled
in
your
cluster
and
like
the
presence
of
a
GPU
extended
resource
on
the
node
directs
schedule
link
to
that
node.
So,
whereas
this
it's
not
quite
the
same,
it's
more
like
I
have
a
very
performance,
sensitive
workload
that
takes
advantage
of
a
B
X
or
is
adverse
to
the
usage
of
a
B.
X
and
I
want
to
steer
that
workload
to
a
particular
node
pool.
That
is
not
impacted
by
that
noisy
neighbor
effect
you're
describing.
D
D
D
You
do
not
need
to
set
aside,
for
instance,
time
slots
during
the
day
when
you
can
schedule
your
AVX
workloads,
but
somehow
you
want
to
prevent
them
from
severely
affecting
the
performance,
sensitive
ones.
Then
one
option
is
that
you
are
limiting
how
large
number,
of
course
within
a
node,
can
concurrently
execute
AVX
workloads,
and
by
that
you
are
indirectly
limiting
the
base
frequency
downscaling
effect
of
the
noisy,
neighbor
avx-512.
B
You
have
two
spare
cycles.
One
and
one
other
possibility
is
that
if
you
have
like
a
specialized
cluster,
for
instance
like
a
single
node
cluster
or
other
cost
after
consists
of
like
special
nodes,
which
need
to
be
performing
many
things
because
they
happen
to
be
somehow
like
special
send
you,
then
you
actually
could
run
a
be
expert
close
on
those
notes
to
if
needed.
Well,
it
would
not
affect
the
other
other
workloads
there.
So
much.
A
A
Partitioning
within
the
node
and
the
CPU
manager
used
cases
we've
explored
previously
were
pretty
clear
to
me
on
who
the
beneficiaries
were
like
they
were,
whether
it
was
DP,
DK
or
some
other
class
of
applications
like
that.
That
was
that
is
clear,
I
guess
and
Ike.
But
in
this
particular
case
the
who
the
who
is
not
able
to
work
the
run
a
workload
best
on
Korea
as
it
exists
today
and
in
what
use
case
environment
with
this
issue
with
ABX
is
not
as
clear
to
main
I
apologize
cause
like
I
said.
B
Yeah
but
that's
that's
a
fair
point
and
I
think
for
many
users,
just
sort
of
like
splitting
up
the
cluster
would
work.
Just
fine,
so
I
definitely
agree
with
that,
but
I
maintain
that
they
are
these
kind
of
these
special
cases
where
it
actually
would
be
like
beneficial
to
have
this
yeah.
A
So
that's
what
I'm
looking
for
is
like
what
is
the
special
case
where
the
were
the
cost
outweighs
the
benefit.
I
mean
where
the
benefits
outweigh
the
cost.
I
guess
is
it?
Is
it
like
an
edge
appliance
grannies
in
a
box
type
case
where
you
have
very
particular
intimate
knowledge
like
that's
what
that's?
What
I'm
just
hung
up
a
little
bit
on
us
understanding?
What
the
special
cases.
B
I
can't
give
you
any
kind
of
kind
of
this
concrete,
concrete,
like
clusters,
that
this
is
now
definitely
like,
something
somebody
who
will
benefit
from
this,
but
there
are,
but
we
have,
we
have
understood,
understood
factory.
There
are
these
these
people,
who
would
actually
want
to
want
to
use
this
okay.
A
B
I
think
I
think
this
was
sort
of
them
kind
of
the
beef.
Then
we
have
this
implementation.
Details
and
I
have
a
small
proof
of
concept.
Implementing
this,
but
I
think
this
is
a
kind
of
sort
of
secondary
I.
Think
the
first
question
that
I'm
hoping
to
have
this
kind
of
these
comments
is
that
if
this,
this
gap
is
something
with
that
could
actually
like
at
some
part,
be
be
accepted
and
and
what
kind
of
changes
or
other
ideas
must
come
like
regarding
its
whole
ABX
air
leaks
like
mitigation
issue,
the.
A
B
A
A
Actually
setting
it,
whereas
typically
at
sunset
and
setting
it
based
on
the
sum
of
first
of
all
and
guaranteed
pods
on
that
nodes
request
getting
subtracted
from
when
a
best-effort
could
use,
and
so
if
one
of
the
things
you
were
showing
here
was
basically
having
a
cpu
set
that
was
fixed.
That
would
just
run
best-effort
pods,
but
I
guess
I
was
wondering
as
if.
A
Rather
than
fixed
that
CPU
said
you
could
make
that
CPU
set
just
further.
Restricting
but
I
guess
that's
kind
of
what
the
static
CPU
manager
already
does
in
practice
by
just
moving
non-guaranteed
workloads
to
to
those
other
cords.
And
so,
if
I
was
to
compare
like
a
user
who's
running
the
static
CPU
manager
or
not.
A
D
So
usually
so
again,
if
you
look
at
the
details
in
some
cases
it
might
matter
which
exact
course
I
guess,
because
it's
so
complex
but
I,
don't
think
that
would
be
a
significant,
so
I,
don't
think
you
could
come
up
with
a
workload
or
a
configuration
where
it
would
show
significant
performance
difference
with
two
setups,
which
only
differ
on
which
particular
course
within
the
same
package
you
are
running.
The
avx-512
would
end
up
significantly
differently,
affecting
the
other
known
AVX
workloads
on
the
cpu.
A
Right
or
just
like
turn
on
a
static,
CPU
manager
policy
inside
cube,
reserved,
really
high,
and
then
the
impact
would
that
be
would
be
you're.
Guaranteed
workloads
would
run
in
the
integral
cpu
course
that
you
want
to
write
for
either
I
guarantee
for
all
your
best
effort.
Workloads
would
Ronen
that's
lust
bucket,
and
would
you
not
have
the
same
basic
effect
as
pre,
reserving
particular
CPU
steps
for
best
effort.
B
It's
true
that
if
I
remember
correct
the
static
policy
it
only
like
it
or
it
doesn't
sort
of
at
least
every
subset
from
the
really
from
any
from
the
whole
CPU
said,
but
it
just
is
something
that
they
make
sure
that
that
it
must
be
nonzero
sort
of
so
that
this
this
other
pots
can
also
like
executed
there.
So
it
actually
might
might
work
needs
are.
D
A
Okay,
well,
I,
guess
that
that's
the
set
of
feedback
I
would
have
looking
at
this
initially,
and
maybe
some
of
that
feedback
was
good
and
useful,
and
maybe
other
is
less
so.
But
I
want
to
thank
you
for
presenting
today
and
I'll
Jess
were
running
late
on
starting
the
meeting,
and
we
can
bring
this
topic
again
up
and
subscript
meeting
where
we
can
hopefully
get
some
discussion
and
other
folks
giving
your
feedback
after
watching
the
video.
So
thank
you,
everyone
so
anyhow,
going
to
cover
for
today.
Otherwise,
I'll
give
back
during.