►
From YouTube: Kubernetes SIG Scheduling meeting - 2018-08-09
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).
A
A
It
turns
out
that
it's
useful
for
those
predicates
that
are
more
computationally
expensive.
Some
of
the
predicates
that
we
have
worked
for
deep,
pretty
lightweight
in
a
way.
For
example,
we
have
pot
fits
hosts.
Pop
fits
holds
only
checks
if
the
given
node
name
is
equal
to
the
name
of
the
note.
So
it's
basically
a
simple
string
comparison
moving
some
some
of
the
practice
like
this
to
the
equivalents
cash
causes,
multiple
cash
flow
crops
that
actually
decrease
the
great
performance
instead
of
improving
it.
So
we
need
to
rethink
that.
A
We
need
to
change
the
equivalents
cash
to
only
cash,
some
of
the
predicates,
those
predictions
that
are
more
computationally
expensive.
An
example
is
interpret
affinity,
an
anti
affinity
when
we
tried
the
equivalent
cash
for
some
of
our
benchmarks
that
are
targeting
only
anti
affinity.
We
saw
16x
performance
improvement,
of
course,
in
our
benchmarks
or
cluster
size
is
pretty
large
and
we
were.
A
We
are
scheduling,
a
relatively
large
number
of
parts,
but
anyhow,
this
number
is
quite
significant,
so
we
believe
that
equivalents
cash
is
important
and
is
useful,
but
we
need
to
change
it
to
make
it
basically
predicate
based.
If
the
predicate
is
lightweight,
we
don't
need
to
cash
it
we
just
can
run
the
predicate
itself.
If
the
predicate
is
more
expensive,
we
can
use
the
equivalence.
A
The
benefit
of
this
approach
is
that
every
time
and
that
a
new
predicate
is
added
that
predicate
in
itself
should
should
be
actually
optimal,
should
get
optimized
for
running
with
equivalents
cash
or
not
using
the
equivalent
cash.
The
alternative,
which
is
like
equivalent
cash
knowing
about
each
predicate,
has
the
drawback.
It
has
a
drawback
that,
every
time
that
a
new
predicate
is
added,
we
need
to
go
and
make
changes
to
the
equivalent
cash
as
well
since
harry.
A
A
C
So
the
modification
in
kill
features
Dargo
seems
having
merged
today.
So
there
I.
Think
in
the
last
item
is
the
website
so
cost
promised
to
reveal
this
week.
So
when
that's
done,
we
can
kind
of
close.
This
and
I
will
contact
them
to
see
them
if
they
need
any
release,
notes
on
that,
then
we
can
cause.
This
sounds
great.
A
I
guess
class
is
not
present
today,
all
right,
I
don't
see
him.
So
with
respect
to
gang
scheduling,
gang
scheduling,
we
have
had
some
discussion,
but
the
community
and
the
PR
that
Klaus
has
written
for
the
proposal
and
we
think
that
we
are
almost
there.
We
have.
We
have
a
good
idea
of
how
we
want
to
create
the
API
for
for
gang
scheduling
and
then
the
next
phases
will
be
starting
implement.
Some
of
these.
A
We
think
that
we
are
gonna,
build
a
separate
scheduler
called
battery
scheduler
the
for
for
just
doing
gang
scheduling
and
by
batch
scheduling.
We
also
need
a
controller
that
takes
care
of
lifecycle,
events
of
gangs
or
pod
groups,
as
we
call
them
now
in
the
API.
So
the
controller
is
responsible
for
taking
action
when,
when
certain,
when
certain
properties
of
a
pod
group
is
violated,
for
example,
if
the
pod
group
requires
a
minimal
number
of
cards,
what
should
be?
A
What
should
we
do
if
the
number
drops
below
that
the
number
of
paths
drops
below
that
number
required
number?
So
events
like
that,
whether
we
should
wait
for
those
or
whether
we
should
kill
all
of
the
instances
in
order
all
of
these
decisions
were
take
action
taking
is
going
to
happen
by
by
that
controller.
A
So
let
me
see
there
has
been
a
couple
of
other
items
that
I
have
been
working
on.
One
one
is
actually
the
scheduling
framework
and
there
is
a
proposal
out
for
for
this
scheduling
framework.
I
already
told
you
guys
that
we
team
that
building
a
separate
scheduler
from
scratch
means
a
lot
of
effort
and
also
rolling
it
out
will
be
a
little
hard
given
that
there
are
so
many
users
that
are
relying
on
the
specific
properties
or
behavior
of
the
currency.
A
Scheduler
replacing
it
with
a
completely
different
scheduler,
although
it's
mostly
backward
compatible,
may
not
be
very
easy.
So,
as
a
result,
we've
decided
to
gradually
bring
the
ideas
of
the
scheduling
framework
to
our
tree
scheduler,
and
we
think
we
are
gonna
start
with
some
of
those
that
help
move
the
dynamic
resource
binding
that
are
currently
partly
in
extenders,
partly
in
the
scheduler
code.
A
A
A
A
The
reason
that
we
decided
that
we
are
gonna
do
this
is
to
improve
performance
basic,
so
in
especially
in
larger
clusters,
we
probably
don't
need
to
go
through
all
the
nodes
once
we
find
a
certain
number
of
nodes
that
are
there
enough
for
running
a
part,
and
we
can
start
scoring
only
the
set
of
nodes
that
are
found.
We
see
some
performance
improvement
from
this.
It
has
been
in
the
ballpark
of
like
30
percent
in
larger
clusters
in
smaller
clusters.
A
Actually,
this
feature
is
not
gonna
have
any
major
impact
and
it's
not
gonna
be
actually
be
active
in
a
way.
So
in
a
small,
since
we
are
you're
expecting
to
score
a
certain
number
of
parts
at
minimum,
a
certain
number
of
nodes
minimum
in
a
small
cluster
that
minimum
is
oftentimes
less
than
the
size
of
the
cost.
The
minimum
is
larger
than
the
size
of
the
cluster
and,
as
a
result,
we
were
gonna
end
up
scoring
all
the
nodes.
A
C
A
Correct
yes,
so,
yes,
that
is
very
true
I.
Actually
there
was
some
discussions
about
this
on
undertand,
a
PR,
the
gist
of
it
is
that
the
most
optimal
node
for
running
a
particular
for
any
particular
part
is
not
necessarily
very
stable
choice
anyway.
So
let's
say
that,
for
example,
you're
trying
to
spread
pods
among
nodes
and
you're
trying
to
balance
the
amount
of
resource
usage
among
nodes.
You
may
go
and
find
the
perfect
node
in
the
cluster
that
gives
you
the
best
balance
of
resource
utilization,
one
second
after
the
node
is
scheduled.
A
Another
part
on
a
different
note
terminates,
and
suddenly
a
lot
of
resources
become
available
on
that,
not
so
very
soon
that
note
that
you
chose
a
second
before
no
longer
is
the
most
optimal
node
or
the
optimal
node
for
running
the
part.
So
this
is
in
practice
not
necessarily
very
useful,
at
least
in
larger
clusters.
This
is
where
there
is
a
lot
of
churn
picking
the
most
optimal
or
the
optimal.
No,
it
is
not
necessarily
worth
a
lot
of
extra
activity
and
a
lot
of
extra
computation
does
it
make
sense.
A
B
I'm
thinking
of
it
more
like
just
power
usage
overall,
if
they
have
a
large
cluster,
that's
half
and
use
if
you
keep
half
the
notes
busy
and
half
of
them
idle.
You've
got
a
better
power
situation,
now
failure
but
looks
with
that.
But
but
what
are
the
larger
metrics
that
that's
going
for
other
than
being
faster?
Yes,.
A
That
is
true,
that's
valid,
although
we
are
intentionally
actually
in
order
to
provide
a
better
spreading
in
clusters,
we
actually
go
through
different
failure
domains
instead
of
just
no
today,
our
schedule
actually
starts
looking
at
all
the
nodes
in
a
in
a
cluster
in
a
sort
of
like
a
very
linear
manner.
It
just
goes
through
their
list
of
all
the
row,
all
the
nodes
and
scores
and
checks.
The
feasibility
of
all
when
this
new
feature
is,
is
merged.
The
way
that
we
are
looking
at
the
nodes
becomes
slightly
different
or
may
be
fundamentally
different
Jeff.
A
So
what
we
do
is
that
we
actually
go
through
different
zones
and
different
regions,
but
it's
good,
it's,
let's
say
different
zones
that
that
makes
things
easier,
because
a
differential,
different
zones
in
different
regions
are
still
different
zones.
So,
basically
we
go
through
different
zones
and
we
pick
one
one
node
from
let's
say:
zone
one
we
evaluate
it
and
then
the
next
one
is
chosen
from
the
next
zone
and
the
third
one
is
chosen
from
your
third
zone.
If
there
are
multiple
zones
across
it,
it's
not
just
a
linear
scan
anymore
and
once.
B
A
Yeah
for
for
power
optimization,
we
actually
can
you
know
their
priority
function.
That,
of
course,
you
just
said:
that's
right,
so
I
I
guess
for
that
they're
for
the
purpose
that
you
have
in
mind.
Our
best
priority
function
is
most
used,
so
we
have
my
Yoda
function
that
tries
to
actually
maximize
the
usage
of
resources
and
I
know
it
before
going
to
another
node
and
so
on.
So
basically,
that's
that's
something
that
can
be
used
to
optimize
power
usage.
D
For
the
counts,
gathering
I
think
I
update
the
document
or
calling
over
to
our
little
meeting
last
week
and
I.
Think
in
the
last
wait
await
put
a
lot
of
discussion
about
a
cannibal
transaction
of
the
whole
group
and
I
checked
the
slurm
as
I
pan
in
the
in
the
SEO
firm
also
do
similar
things
to
the
automation
controller,
but
the
sing
one
different
is
that
slurm
pay
sounds
a
in
top-level
mm-hmm.
D
Yes,
Apache
in
urban
areas
to
the
automation
pull
back
code,
almost
the
Impala
level,
a
single.
This
is
the
major
difference
and
I
also
checked
this
with
our
product.
We
are
word
to
the
similar
scenes,
so
we
also
do
the
to
the
cannibal
admission
in
job
level,
so
the
soulful
bad
salesman-
they
don't
have
such
kind
of
easier
but
for
corner
has
because
we
don't
have
a
concept
of
top
right,
so
you
can
kind.
A
D
D
I
mean
this
concept
is
in
Kremlin
has
led
to
the
Mosul
Cynthia
Nicole
partying
practices,
and
they
almost
do
everything
based
on
draw.
This
is
a
kind
of
major
differential
budget
single.
We
need
to
have
the
stubborn
proposal
or
some
solution
to
for
for
this
candle
for
for
this
kind
of
easy:
how
to
resolve
that?
Okay,.
A
A
E
A
E
D
I
think
they
don't
have
the
use
of
different
term.
We
have
a
plugin
named
the
Java
sublimation
crabby.
So
when
cha
was
submitted
to
the
system,
you
can
accept
or
reject
or
even
modify
the
job
when
they
sublimate
or
to
modify
its
a
map
to
the
Crillon
here.
So
there's
gonna
be
when
they
create
all
when
they
when
they
update
when
the
object
is
update,
so
they
don't
use
some
different.
The
slum
use
different
term,
but
I
think
that
something
is
a
similar.
Yes,.
C
D
E
D
You
talk
to
all
right,
okay,
you
know
you
were
you'd.
Do
the
patch
ourselves
and
piss
our
metals
I
think
because
they
also
I
use
it
talk
with
you
bruh
how
to
build
another
batch.
So
time
appears
on
kubernetes,
yes,
I
single
there
I
think
there
I
give
the
talk
to
them.
There
I
think
their
major
concerns
about
the
performance
performance.
D
A
E
A
B
E
F
What
we
think
someone
from
Alibaba
I,
don't
remember
his
name,
but
he
has
mentioned
they
are
using
the
variant
of
gang
scheduling
and
default
queue.
Vanity
scheduler
could
not
escape
I.
Think
it's
hong
kong
or
someone
from
Alibaba
and
it
all
day
it's
around
like
two
nk+
notes
or
15
k
plus
note
scale
that
they
are
able
to
use
and
they're
using
some
variant
of
gang
scheduling,
I'm,
not
sure,
but
he
has
posted
a
link
on
one
of
the
github
issues
that
we
have
created
for
gang
skating.
Okay,.
A
A
B
A
E
E
D
Estimate
there
may
be
different
form
from
the
concept
because
I
was
with
you
is
the
mall
about
for
the
to
share
salsa
so
for
them
of
hi,
probably
kill
ya
and
how
to
share
with
salsa
and
the
different
attendants
such
kind
of
things.
Gotta
battle
for
the
gun,
scheduling,
I,
think
even
with
how
queue
in
ESL
aromas
that
kind
of
concept.
We
also
need
to
handle-
maybe
transaction-
or
maybe
we
don't
have
transaction
eventually.
D
A
Yeah
there
are
so
this
is
I.
Don't
think
this
is
something
that
we
can
really
resolve
and
final
solution
and
this
Rodney
something
that
we
should
actually
follow-up.
This
is
one
of
the
pieces
that
is
left
to
be
decided
might
think.
The
API
part
of
the
gang
scheduling
is
kind
of
getting
there
yeah
like
creating
book
yeah.
D
A
E
A
D
Feature
and
it's
not
gonna
be
math,
yes,
I
think
we
can
follow
the
the
our
fire
feature
before
for
them
hallway,
cable
on
occasion,
key
right,
yeah.
And
yes,
when
we
have
some
solution
or
some
proposal
about
this
reduction,
we
can
introduce
another.
So
the
unity
opposed
when
we
upgraded
to
the
beta
Ogier
yeah.
E
E
This
is
the
best
option
we
could
think
of
now
and
then
like
a
prototype
for
that
and
then
start
solving
the
higher
layer
problem,
which
is
like
defining
a
controller
or
like
extending
the
existing
job
controller
to
actually
have
in
the
lifecycle,
semantics
that
we
and
and
then
like,
seeing
whether
the
whole
system
makes
sense
for
a
real
application
right.
So.
A
E
E
A
A
C
C
A
A
Yes
in
is
is
gonna
update
that
proposal
that
is
already
out
and
with
this
new
changes
to
the
design
and
changes
are
not
there.
Yet
what
we
are
hoping
to
see
the
proposal
updated
in
the
next
couple
of
days.
Is
there
anything
else
since
we
were
running
all
the
time?
Is
there
anything
else?
Questions
comments
about
any
other
features
that
you
guys
are
working
on.