►
Description
Join us for Kubernetes Forums Seoul, Sydney, Bengaluru and Delhi - learn more at kubecon.io
Don't miss KubeCon + CloudNativeCon 2020 events in Amsterdam March 30 - April 2, Shanghai July 28-30 and Boston November 17-20! Learn more at kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects
NFD and My Adventure in the Cloud Native Project Jungle - Markus Lehtonen, Intel
At its core, Kubernetes has a somewhat limited understanding of system details during operation. Node Feature Discovery (NFD) is an add-on that increases Kubernetes platform awareness by detecting and advertising the hardware and software capabilities of the underlying system. This information can be used to facilitate intelligent scheduling of workloads in a Kubernetes cluster. This talk introduces NFD and its role in Kubernetes. It covers the technical details of NFD, including how the node capabilities are advertised and how they can be utilized in workload placement. The presentation also covers my journey in the Kubernetes community and assuming maintainership of the NFD project. The journey passes through collaboration with third parties, both internal and external.
https://sched.co/Nrn4
A
So
hello,
everyone
good
to
see
you
here.
Are
you
tired
just
trying
to
survive
through
the
very
last
session
of
today?
Well,
that
will
be
about
no
teacher
discovering
a
project
I've
been
maintaining
for
the
left
last
year,
or
so
so.
First,
the
agenda
of
them
or
the
talk.
First
I'll
talk
a
little
bit
by
bit
about
my
my
journey
in
the
humanities
landscape
that
will
be
fairly
short.
First
of
all,
I
don't
have
much
time
for
that,
and
second
I
don't
know.
A
If
there's
anything
particularly
interesting
is
that
then
I'll
show
the
recap:
the
resource
management.
Thank
you
burn.
It
is
all
that
works
and
how
anything
or
not
feature
discovery
pitch
into
that
picture.
Then
I'll
go
a
bit
more
deeper
into
note,
which
is
discovering
it's
yeah
penalty,
its
architecture,
details,
deployment.
A
And
lastly,
there
was
certainly
certainly
the
use
cases
for
NFD,
but
the
project
by
the
time
of
beginning
of
last
year,
I
entered
in
a
sort
of
stagnant,
stagnant,
hibernated
state,
with
no
no
much
new
contributions
coming
in
and
not
that
actively
maintained
anymore.
So
I've
decided
to
start
started.
Looking
that
submitting
PRS
and
soon
I
found
myself
as
the
main
maintainer
of
the
project,
and
this
has
been
a
fun
year,
what
what
learned
back
then
I
dropped
it
up
short.
A
Plan
for
future
plan
for
NFD
and
presented
that
to
signal
and
everybody
seemed
to
be
happy
with
that
and
continued
working,
and
we
were
able
to
Margaret
Margaret
NFD
from
the
deprecated
kubernetes
incubator
in
the
community
six
organization
into
safer
home,
and
now
we've
we've
been
able
to
get
some
new
contributors,
new
reviewers
for
the
project.
So
things
looks
pretty
good
now
and
I've
been
feeling
very
welcomed
by
the
by
the
community.
A
A
A
This
is
very
simplified.
You
have
a
Cuban
Redis
cluster
master
and
a
worker,
node
or
probably
master,
could
be
ruthless
with
or
control
plane.
I've
left
out
some
key
parts,
no
key
components
not
relevant
here,
like
controller
manager,
queue
proxy
and
so
on,
but
anyway
the
results.
Money
management
work
so
that
we
have
so-called
native
resources
like
CPU
memory
and
huge
pages,
and
they
are
detected
by
key
bullet
directly.
Then
we
have
extended
resources
with
GPU
FPGA.
Typically,
some
hardware
accelerators-
and
these
are
typically
managed
by
separate
device.
A
Plug-Ins
Cuba
left
advertises
the
resources
to
API
server
stores
them
in
their
CD
store
and
then
all
is
created,
for
example,
by
user,
with
a
cube
at
all.
Common
line
to
object
is
created
in
HCV
scheduler
fix
that
up
and
Trust
to
find
a
node
that
can
satisfy
the
resource
requirements
of
the
pod.
If
there
are,
if
there
are
any
when,
after
finding
a
suitable
node,
it
schedules
the
pot
in
the
node
and
what
we
realize.
This,
of
course,
goes
with
a
cube
a
lot,
but
anyway
works
like
this
father
compute
resources.
A
Some
some
workloads
might
have
strict
requirements
for
certain
capabilities
for
them
to
be
able
to
run
run
on
the
node
and,
on
the
other
hand,
some
capabilities
might
provide
measurable
benefits
in
the
clock
workload
performance
and
how
all
this
gap
gap
between
the
like
traditional
computer
resources
and
note
features
or
capabilities
is
filled.
Currently,
that
is
the
answer
in
kubernetes
is
note,
feature
discovering
or
NFD.
A
It's
basically
running
one
one
instance
per
node
and
advertises
the
detective
capabilities
of
node
node
labels,
which
in
turn
can
then
be
used
for
the
user.
The
labels
can
be
used
by
the
user
to
aid
in
workload
scheduling
so
any
of
these
project,
sponsored
by
cig
note
and
it
lives
on
the
Kuban,
a
v6
organization.
A
Indeed,
problem
at
the
moment,
then
a
bit
deeper
into
any
of
the
architecture.
Simple,
simple
view
like
this:
the
worker
node
in
typical
deployment.
We
have
any
of
the
worker
running
on
the
worker
node,
which
does
the
actual
the
device
or
feature
discovery.
I
mean
if
the
master,
the
colleague
of
only
runs
on
the
master,
node
and
nfe
master,
is
the
one
responsible
for
actually
creating
them.
A
node
labels
and
and
annotations
so
in
other
words
it
is
responsible
for
updating
the
node
object
in
the
cluster.
A
Anybody
master
is
actually
essentially
a
JRPG
server.
It
listens
to
request
G
RPC
request
from
the
NFD
worker.
It
has
support
for
MPLS
to
dust
some
filtering
on
the
label
so
that
so
that
the
worker
nodes
kind
of
cannot
create
arbitrary
labels
and
only
on
the
labels
that
are
that
are
allowed,
and
it
has
a
node
update
part
which
basically
creates
labels
and
annotations
in
that
node.
And
if
you
worker,
it
is
a
bit
bit
more
bit
more
complex
software
still
quite
simple.
A
The
main
part
of
it
is
the
actual
feature
discovery
which
is
represented
in
the
yellow
box
in
the
middle.
So
the
teacher
discovery
is
divided
into
multiple
picture
sources,
one
by
hardware
type,
basically
and
some
other.
It
has
simple
configuration
handling,
then
G
RPC
clients
talking
to
the
inner
key
master
and
in
the
left,
the
lower
left
part.
We
have
groups
and
feature
files.
A
These
are
essentially
ways
to
customize
next
thing
in
a
fee
in
a
flexible
way,
I'm
from
there
and
the
discovery
part
basically
uses
simple,
also
simple:
kernel
interfaces
like
sea
surface
and
poor
campus
to
discover
the
capabilities
and
also
read
some
data
from
from
boot
and
et
Cie
from
the
post,
but
essentially
any
of
these
kind
of
read
only
read-only
components,
so
it
doesn't
do
any
modification
in
then.
In
the
note,
it
is
running
on
don't
reach
reach
for
more,
for
example,
then
about
feature
sources
basically
or
discovery
is
organized
into
set
of
feature
sources.
A
You
have
Hardware
show
sources
like
CPU,
I
will
a
menu
and
so
on,
and
then
some
other
kernel
system
and
one
power
source
and
local
is
basically
for
customization
and
the
labels
have
a
certain
scheme
or
pattern.
First,
there's
a
standard
label,
namespace
feature
node
gates
the
tile
and
then
name
of
the
feature.
Sauce
then
name
of
the
feature
possible
additional
attribute
and
then
value,
and
this
buddy
folder.
A
A
Crypto
accelerator
from
refraction
and
so
on,
and
on
arm
that
uses
HW
captain
from
internal
hardware.
Multi-Threading
is
basically
telling
if
the
technologies
such
as
Intel,
hyper-threading
or
AMD
SNT
is
enabled.
This
might
be
interesting
for
workloads
that
are
into
the
work
or
interested
more
more
in
performance.
A
Predictability
versus
maximum
performance
and
P
state
turbo
is
basically
into
turbo
boost,
and
it's
similarly
like
max
per
versus
predictive
predictability
of
performance.
Then
we
have
power
source
power
Peter,
it's
basically
detects.
Currently,
if
the
int
of
late
speech
select
technology
has
been
enabled,
it's
basically
can
be
thought
of
as
a
permanent
turbo
boosting
of
some
CPU
CPU
cores
in
the
package,
and
then
oddity
is
for
cache
and
main
memory.
A
Monitoring
and
and
allocation
from
kernel
we
detect
kernel,
config
options.
It's
fun
agreeable,
which
options
are
advertised
by
default.
We
detect
the
idle
taken
preempt
options
which
can
be
interesting
for
computationally
heavy
or
some
real-time
per
clause.
For
example,
SEO
Linux
is
detective
that
is
enforcing,
is
enabled,
and
then
we
did
the
kernel
version
and
that
that
is
also
well.
There
are
some
special
work
clubs
which
might
require
some
new,
your
cranial
features
or
some
new
kernel
features
provide
miserable
benefits,
benefits
for
the
workload
from
system.
A
So
this
is
significant
or
maybe
significant
for
some
college
email
optimized
we're
closed
for
them.
On
the
other
hand,
forward
loss,
guitar,
no,
that's
well
behaving
in
Numa,
Numa,
environment
and
the
envy
is
the
reason.
No
one
non-volatile
run
technology-
and
this
is
this
might
be
interesting
for,
for
example,
in
some
database
deployments
that
need
local
volatile
run
from
PCI.
We
detect
the
presence
of
PCI
devices
by
default.
A
We
detect
GPU
and
accelerator
cards
is
configurable
and
basically
the
main
use
case
that
I
know
for
this
is
our
device
plug
in
so
that
device
plugins
can
be
deployed
on
them
on
the
nodes
in
the
cluster.
That
really
have
the
hardware
that
the
Divas
plugins
is
managing
from
network.
We
detect
single
route
by
over
generalization,
capable
and
configured
network
interface
cards,
and
this
can
be
relevant,
for
example,
in
for
a
network
function,
virtualization
workloads
but
mobile
today,
station's
firewalls
load
balancers.
A
Basically,
these
for
first
user
space
packet
processing
we
take
I
Oh
you
this
present
and
enabled-
and
this
basically
direct
hardware
access
in
virtualized
and
might
be,
might
be
significant
for
IO
heavy
workloads
when
using
hypervisor
based
container
runtime,
such
as
Papa
from
storage.
We
detect
presence
of
non
rotational
disk,
so
this
basically
sort
of
indication
that
first
first
local
storage
is
available
on
the
node.
A
A
So
if
it
outputs,
simple
mobile,
that
is
translated
into
a
label
my
source
as
mobile,
it
was
true
and
default
in
of
the
label
nationís.
Similarly,
we
can
give
a
value
for
the
Bourdon
feature
which
slash
we
can
override
the
name
of
them,
I,
never
the
features.
Also,
in
this
case
our
add
source
is
used
in
instead
of
instead
of
my
source
in
the
feature
label.
A
We
can
also
give
value
to
this,
and
if
the
end
of
the
master
has
been
configured
correctly,
you
can
also
using
a
custom
name
label
namespace
for
for
your
feature,
for
example,
some
Wender
vendor
specific
labels,
so
in
this
case,
creates
a
label
in
them.
That's
my
name.
Space
namespace
feature
pulse
are
similar
to
books,
but
they
are
simply
pass
that
enemy
tries
to
read
and
foil.
Contents
is
directly
directly
turned
into
labels,
and
the
format
of
the
feature
part
is
essentially
the
same
as
the
expected
standard
output
output
from
books.
A
A
Few
examples
of
NFD
deployment.
We
basically
provide
three
three
example
deployments
in
the
NFD.
Repo
first
is
what
demos
at
los
demás
at
configuration,
where
NFD
master
runs
as
a
demon
set
on
the
master
nodes
in
the
cluster
and
workers
run
as
FEMA
sets
in
the
worker
nodes
in
the
cluster,
and
in
this
configuration
service,
discovery
is
done
every
by
default,
every
60
seconds
and
when
new
nodes
are
added
to
the
cluster,
of
course,
automatically
labeled.
A
A
These
are
the
same
body
as
a
demon
said
in
in
every
worker
node,
then
about
work
loads
of
these
labels
or
workload
development.
Other
labels
can
be
used
basically,
two
methods.
First,
one
is
note
selector,
it's
simple:
we
want
this
label
to
have
this
certain
value
and
this
example.
In
this
case,
we
want
turbo-boost
to
be
enabled
for
the
workload
to
be
run
a
little
bit
more
involved
and
more
flexible.
Then
method
is
then
note
affinity.
A
A
Please
set
up
labels,
then
operators
and
then
list
the
values
of
a,
for
example,
in
not
in
operators
like
greater
than
less
than,
and
then
you
can
also
specify
soft
requirements
that
prepared
during
scheduling
and
ignorant
during
execution.
These
are
the
scheduler
trusted
tries
to
tries
to
fulfill
these
requirements,
but
they
are
not
hard
requirements
in
the
sense
and
here's
a
like
complete
example
of
a
pod
using
node
affinity,
so
affinity
and
node
affinity,
then
some
hard
requirements
in
this
case
kernel
version
make
an
internal
major
version
must
be
greater
than
3
and
kernel.
A
A
A
A
No.23.
Okay,
now
we
said
see
that
GPU
Intel
GPU
extended
resource
was
was
created
created
on
the
note.
Well,
that's
it
I'm
running
out
of
time.
So
let's
get
forward
so
short
about
the
future
of
the
project.
Yet
trying
to
get
any
of
the
operators
soon
upstream
red
happens.
I
think
the
Red
Hat
has
been
working
on
that
ton
in
downstream
and
trying
to
get
that,
obviously
trying
to
enable
multi
arch
build,
builds
in
the
near
future.
A
There
are
some
usability
and
configurability
improvements,
but
I'm
planning
to
have
then
possibly
supporting
or
extended
resources
in
NFA,
but
we
have
to
haven't
discussed
it
in
this
in
signal
yet,
and
we
are
they'd
like
to
have
a
logo
for
the
project
make
these
presentations
a
bit
more
appealing.
So
please
contribute
if
your
graphical,
if
your
graphical
skills
are
better
than
y,
know
the
other
guys
in
the
project
and
then
the
user
boring
uses
examples
of
mentation
improvements.
A
So
if
you
want
to
help
or
are
interested
in
the
project,
feature
request
request
bugs
bugs
are
what
reports
are
very
much
welcomed
as
duchies
and
also
you
can
tell
us
your
your
story.
How
are
you
are
using
unity
or
would
like
to
use
end
of
day,
and
I'm
planning
that
paulo
we
should.
We
should
try
to
collect
the
usage
examples
somewhere
in
the
NFP
repo
anymore,
more
organized
way,
so
that
would
be
useful
board
of
people,
Steven,
Leary,
actual
XO
eustacia's,
so
here's
one
small,
an
address
or
github
URL
URL
for
the
project.