►
From YouTube: Keynote: Reperforming a Nobel Prize Discovery on Kubernetes - Ricardo Rocha & Lukas Heinrich
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
Keynote: Reperforming a Nobel Prize Discovery on Kubernetes - Ricardo Rocha, Computing Engineer & Lukas Heinrich, Physicist, CERN
Back in 2012, CERN announced one of its most important achievements, the discovery of the Higgs boson leading to the 2013 Nobel Prize in Physics. In this presentation, we will redo the data analysis that led to it, this time on top of Kubernetes, the new infrastructure stack growing in popularity in the laboratory. The analysis submission will be done via a jupyter notebook into a small cluster on our private cloud, and both the application and the cluster itself will automatically scale out to exhaust resources - we will detail our setup and deployment decisions on the way. Then we will show how the work being done in the Multicluster SIG helps us define a set of placement and scheduling policies to scale out to external clouds. The end result will be a physicist’s dream: a histogram with a spike that back in 2012 indicated the discovery of a new particle.
https://sched.co/MRyv
A
A
Cool
so
so
we'll
quickly
introduce
the
the
CERN
CERN.
So
CERN
is
the
European
Research
Organization
for
Nuclear
Research.
You
probably
heard
about
it.
It's
the
biggest
particle
physics
laboratory
in
the
world
and
we've
based
in
Geneva
Switzerland
to
switch
okay.
So
we
we
build
a
lot
of
experiments.
Our
main
main
goal
is
to
do
fundamental
research.
To
do
this,
we
build
very
large
experiments
and
we
try
to
answer
big
questions
about
the
universe.
So
we
want
to
know
what
is
matter
made
of
and
what
happened
just
after
the
Big
Bang.
A
What
was
the
state
of
matter
just
after
the
Big
Bang,
and
this
picture
here
is
about
one
of
the
other
things
we
want
to
try
to
see
is
what
is
antimatter
and
why
don't
we
see
antimatter
in
the
universe.
So
in
this
machine
here
one
of
the
big
experiments
we've
built,
which
is
done
just
down
the
road
from
our
offices.
We
actually
create
antimatter
and
we
try
to
keep
it
for
a
while
and
try
to
see
what
is
inside
of
it.
A
But
many
of
you,
because
this
is
a
computing
conference-
might
know
CERN
also
for
other
regions,
and
one
of
the
big
things
we've
done
in
the
center
was
the
creation
of
the
World
Wide
Web.
So
it
was
created
at
CERN
in
1989
by
Tim
berners-lee,
and
the
goal
of
the
World
Wide
Web
was
to
help
physicists
in
their
daily
life.
So
physicists
have
a
lot
of
needs
to
share
data
between
themselves
and
to
have
a
way
to
easily
access
this
data.
A
So
he
came
up
with
this
proposal
in
1989
and
here's
the
screenshot
of
the
actual
initial
proposal
and
you
can
see
a
lot
of
blocks
and
things
interconnected.
This
was
the
whole
point
of
the
endeavor
and
in
89
it
creatively
got
really
popular
inside
CERN,
but
by
93
the
center.
The
CERN
had
realized
that
this
could
be
think
to
be
used
by
many
more
people
than
just
just
the
CERN
physicists,
so
they
made
it
put
a
available
in
the
public
domain
and
of
course,
it
got
really
popular
well.
A
No,
that's
also
why
we
are
all
here
today
and
just
a
couple
of
weeks
ago
we
celebrated
the
30th
anniversary
of
the
World
Wide
Web.
You
have
a
picture
here.
We
had
a
big
party
and
there
was
broadcast
and
you
can
see
Tim
berners-lee
there
sitting
with
many
other
famous
people,
but
our
main
goal
is
still
to
do
particle
physics.
And
to
do
to
do
this,
we
need
a
very,
very
large
machine.
A
We
inject
them
using
initial
accelerators
that
you
can
see
smaller
circles
there,
so
we
first
start
speeding
them
up
and
increasing
the
energy
and
when
they
are
close
to
the
speed
of
light
very
close
to
the
speed
of
light,
and
they
are
already
in
the
big
accelerator
and
we
have
two
beams:
one
is
going
clockwise
and
the
other
one
is
going
counterclockwise
and
that
very
specific
point
we
make
them
collide
to
each
other,
and
we
built
this
big
detectors.
We
have
here
Alice,
Atlas,
LGB
and
CMS,
and
this
will
we'll
see
a
picture.
A
They
try
to
analyze
what
what
the
collisions
are
seeing
just
for
size
to
have
an
idea
of
the
size.
You
can
see
the
Geneva
Airport
just
there
on
the
bottom
right
and
then
you
can
see
that
it's
actually
a
quite
big
machine.
It
spends
two
countries
France
in
Switzerland
and
inside
the
experiments.
If
we
try
to
see
things
like
this,
this
is
a
visualization
of
what
happens
inside
and
when
the
particles
collide
to
each
other.
We
try
to
make
what
we
could
call
a
pick
a
photo
of
what's
going
on.
A
Although
we
use
a
lot
of
hardware
for
this,
there
are
billions
of
collisions
happening
per
second
and
we
are
taking
for
something
that
would
be
40
million
pictures
a
second
here-
and
this
is
a
picture
of
the
CMS
detector.
The
CMS
stands
for
compact
muon
solenoid,
but
it's
not
compact
at
all.
It's
a
14
tonnes,
it's
written
there.
Even
so
it's
a
very
big
machine.
The
cavern
is
40
minutes
by
40
meters,
again
100
meters
underground.
A
We
have
four
of
these,
so
you
can
see
a
picture
here
of
the
circle
of
the
accelerator.
You
can
see
the
four
experiments
and
we
pull
fast
links,
Network
links
to
our
data
center,
so
the
white
building
there
is
the
CERN
data
center,
the
machine.
The
detectors
are
generating
in
orders
of
peda
petabytes
a
second,
and
we
can't
deal
with
this.
So
what
we
do
is
we
have
Hardware
filters
that
will
bring
this
down
quite
significantly.
A
Then
we
have
subsequent
software
filters
and
what
we
are
actually
storing
something
like
10
gigabytes,
a
second
in
the
data
center,
and
that
we
can
almost
manage
with
the
data
center.
We
have,
we
have
something
like
300,000
cores
and
we
can
store
storage.
We
have
like
400
petabytes
today,
but
it's
still
not
enough.
So
in
the
last
20
years
we
developed
what
we
call
the
grid
and
there
we
double
the
capacity
and
how
this
works
is
we
connected
200
different
centers
around
the
world,
universities
and
the
other
research
labs.
This
double
our
capacity.
A
If
you
look
into
it,
you
will
see
something
like
400
or
half
almost
half
a
million
jobs
running
concurrently,
and
we
are
actually
adding
70
new
petabytes
from
from
the
accelerator
every
year
that
we
have
to
store
and
analyze,
and
this
is
the
computing
and
then
I'll.
Let
Lucas
cover
the
physics
part.
Ok,.
B
So
what
we
came
here
to
cube
Khan
for
is
to
demonstrate
how
we
can
use
kubernetes
to
analyze
these
large
scale.
Scientific
data
sets
such
as
the
ones
that
we
produce
at
the
LHC,
and
so
we
also
prepared
a
little
live
demo
and
we're
gonna
try
to
rediscover
the
Higgs
boson
live
on
stage,
and
hopefully
they
will
work.
B
And
so
here,
on
the
right
hand,
side
you
can
see
the
announcement
of
the
discovery
of
the
Higgs
boson
by
both
the
Atlas
and
CMS
experiment,
and
also
the
champagne
that
we
drank
and
then
on
the
left-hand
side.
You
can
see
from
Swan
layer
and
Peter
Higgs
who,
in
the
next
year
in
2013,
we're
awarded
the
Nobel
Prize
in
Physics,
because
due
to
that
discovery
because
they
laid
the
theoretical
foundation
that
predicted
this
particle
to
kind
of
explain.
B
Why
we're
so
happy
that
we
found
this
Higgs
particle
I'll
need
to
explain
a
little
bit
about
the
Higgs
particle.
So
it's
part
of
what
we
call
the
standard
model
of
particle
physics.
It's
the
current
theory
that
we
use
to
compute
different
interactions
of
fundamental
particles,
and
it's
very
successful.
So
we
can
predict
the
outcome
of
particle
physical
experiment
to
many
many
decimal
places,
but
before
we
had
the
LHC
there
was
a
of
a
puzzle
because,
on
the
one
hand
we
had
this
incredibly
successful
theory
and
on
the
other
hand
the
same
theory
was
predicting.
B
The
existence
of
this
new
particle
called
the
Higgs
boson,
and
but
we
hadn't
seen
that
yet,
and
so
the
Higgs
boson
is
a
very
special
particle.
It
has
unique
properties,
it's
unlike
any
other
particle
in
the
standard
model,
and
the
existence
of
this
particle
signals
that
we
also
understand
why
particles
have
mass
and
not
everything
is
zipping
through
the
universe
at
the
speed
of
light.
So
obviously,
when
we
had
the
LHC,
we
went
looking
for
it
and
one
of
the
issues
when
looking
for
the
Higgs
boson
is
that
it's
a
very
short-lived.
B
B
So
in
this
demo
we'll
try
to
search
so
the
Higgs
can
decay
in
many
ways
and
we'll
try
to
find
it
and
what's
called
the
golden
decay
channel
of
the
Higgs,
where
the
Higgs
decays
into
electrons
and
muons,
and
so
this
is
the
plot
that
we
had
during
the
publication
when
we
discovered
the
Higgs
particle
and
will
try
to
re-create
this
plot
using
kubernetes.
Now
so
Ricardo
will
work
on
submitting
yeah
the
workload.
So
we
have
a
couple
of
clusters
set
up
that
Ricardo
will
explain
later
on
and
I'll
also
set
up
the
visualization.
B
Okay,
so
I
have
so
all
the
results
that
we're
going
to
talk
about
are
being
aggregated
are
being
aggregated
while
the
results
are
coming
in.
So
we
are
starting
to
notebook
that
dust
up
aggregation
and,
as
you
can
see
here,
the
plot
is
empty
right
now.
So
it's
not
like
the
plot
that
we
just
saw
on
the
slide
before,
but
hopefully,
while
we're
talking,
these
results
will
come
streaming
in
and
then
we'll
actually
see
the
same
result.
Yeah.
B
A
Here
so
the
the
submission
is
happening.
Can
you
increase
I?
It's
right,
the
submission
is
happening
and
what
we
are
doing.
We
we
are
tuning
into
the
event
API
of
kubernetes
and
we
are
seeing
the
events.
The
jobs
we
have
something
like
I
will
go
into
detail
of
how
many
things
we
are
submitting.
But
here
we
are
listening
into
the
job
API
and
you
can
see
the
different
states
that
the
jobs
are
in
polling
means
they
are
staging
in
the
data
and
then
we
already
have
a
couple
of
running
and
eventually
they
will
start
succeeding.
A
So
this
takes
a
bit.
So
while
we
wait
for
this,
I'll
give
a
brief
description
of
what
the
demo
is
really
about
yep.
So
in
the
data
set,
we
are
now
analyzing,
it's
the
Higgs
data
set
and
it's
70
terabytes
of
data.
Now
this
includes
a
lot
of
different
pieces.
We
have
simulation
data
and
we
have
then
the
real
data
and
the
goal
is
that
we
will
try
to
process
this
70
terabytes.
While
we
are
standing
here
and
we
have
to
speed
up
the
demo
a
bit.
A
A
While
this
how
this
is
happening
so
actually
and
then
this
25,000
files
show
a
limitation
that
we
have
in
this
in
this
demo,
which
is,
we
can
only
paralyze
up
to
25,000
parallel
jobs,
because
at
the
time
with
the
software's
from
2010,
it
was
single
core.
So
the
resources
we
are
using
it
at
CERN
we
have
a
cluster
with
a
SEF
for
storage.
We
have
a
70
terabyte
data
set
there.
A
Then
we
have
a
kubernetes
cluster
that
we
deploy
in
our
internal
infrastructure
and
we
push
the
data
to
Redis
and
then
from
Redis
the
results
to
the
Redis
and
from
rarities
we
can
do
the
visualization.
The
problem
is
that
CERN
has
a
production
infrastructure
and
even
during
upgrade
time
like
now,
we
are
always
trying
to
max
out
the
resource
usage.
A
So
when
we
went-
and
we
asked
for
for
25,000
cores
for
a
demo-
we
did
not
get
there,
so
we
started
asking
around
and
we
then
met
some
really
nice
people
from
Google
and
we
have
the
CERN
open,
lab
collaboration,
which
is
the
the
goal
of
this
collaboration
is
to
as
work
with
industry
partners.
So
Google
is
one
of
these
members
of
this
collaboration,
so
we
did
the
same,
we'll
just
replicate
it
the
same
setup
but
in
the
Google
Cloud.
A
So,
instead
of
set
for
and
there's
three
we
have
GCS,
we
have
the
70
terabyte
data
set
there.
Then
we
have
a
large
cluster
of
managed
by
gke,
it's
actually
deployed
in
a
single
region,
but
we
split
it
in
three
different
versions,
which
gives
us
really
a
really
nice
way
to
to
spread
the
load.
We
deployed
the
25,000
kubernetes
job
stare.
We
push
the
job
results
to
the
equivalent
of
Redis,
the
cloud
Google
Cloud
map
memory
store,
and
then
we
visualize
in
Jupiter
and
there's
two
things
I
would
like
to
highlight
here.
A
One
is
that
we
saw
in
the
previous
watch
that
the
the
scalability
and
of
kubernetes,
because
we
can
really
submit
all
these
jobs
very
quickly
and
they
go
to
scheduling
and
running
quite
quickly
and
then
the
second
one
is
that
by
developing
the
application
is
in
the
kubernetes
api.
We
started
by
testing
it
at
cern,
and
then
we
didn't
have
enough
resources,
and
then
we
went
to
the
public
cloud
and
without
changes
the
application
just
runs.
So
this
is
a
like
major
features
for
our
community.
B
We'll
go
back
to
the
visualization,
hopefully
some
results
should
be
coming
in.
Okay.
There
you
go
okay,
there's
stuff
happening,
so
you
see
the
plot
updating
in
the
clock
and
so
what's
happening
now.
So
as
you
see,
as
you
see
it,
the
plot
was
empty
before
and
now
it's
filling
up
slowly
and
you
see
the
black
dots
slowly
inching
up
and
the
black
dots
are
actually
the
data
that
we've
observed
at
the
LHC
like
actual
collisions
right
in
the
color
histogram.
So
the
blue
light
blue
histogram
in
the
red
outline
histogram.
B
These
are
the
computations
that
we
have
based
on
our
theoretical
understanding
of
the
standard
model.
So
this
is
the
expectation,
the
prediction
that
this
animal
gives
us
for
what
the
data
should
look
like
and,
as
you
can
see
in
the
middle
around
the
value
of
125,
you
see
that
there's
a
small
bump
forming
that
is
not
covered
by
this
blue
histogram.
So
everywhere
else,
the
black
dots
pretty
much
agree
with
the
blue
histogram,
but
in
the
middle
there
you
have
a
small
bump,
and
this
is
actually
the
existence
of
the
Higgs
particle.
B
So
this
proves
its
existence,
because
this
plot
will
be
actually
see.
So
what
we
do
in
these
twenty
five
thousand
jobs.
We
download
event
data
and
we
go
through
these
collisions
and
we
select
events
that
have
electrons
and
muons
in
them
because,
as
I
said,
the
Higgs
is
decaying
into
electrons
and
muons,
and
sometimes
we
pick
also
events,
obviously
where
there's
no
original
Higgs
boson.
But
sometimes
you
also
pick
up
events
that
actually
have
the
Higgs
boson.
And
so
you
see
the
peak
is
forming
pretty
nicely
right
so
yeah.
B
B
B
Right,
okay:
here
we
go
so
there
you
go
so
I'll
zoom
in,
and
so
this
is
pretty
astonishing
right.
So
we
had
25,000
jobs,
pulling
data,
and
so
you
can
see
the
total
data
rate
that
we're
pulling
from
Google
storage
was
almost
200
gigabytes.
A
second
right.
This
is
multiple
terabytes
per
second,
and
so
we
were
pretty
amazed
when
we
saw
this,
and
so
this
actually
allowed
us
to
run
this
demo
in
real
time.
B
So
basically,
so
this
is
using
open
data
that
we
are
making
accessible
as
CERN.
So
many
of
the
experiments
at
CERN
have
an
open
data
policy,
and
so
in
this
case
we
use
the
open
data
from
the
CMS
experiment
and
but
for
these
large
scientific
applications
or
experiments,
it
has
always
been
a
little
bit
of
an
issue.
While
we
can
make
our
data
available
and
dump
petabytes
of
data,
it
remained
an
open
question.
B
How
can
people
that
don't
have
a
data
center
like
we
have
actually
realistically
run
these
types
of
applications
on
their
own
right,
and
so
how
can
they
actually
make
use
of
this
data
and
analyze
it
and
using
this
public
cloud
infrastructure?
It
actually
turns
out
that
anybody
can
try
to
rediscover
the
Higgs
boson
within
a
couple
of
minutes.
Ok,
so
I
also
wanted
to
talk
a
little
bit
about
software,
archiving
and
reproducibility.
B
So
not
only
the
software,
so
we
make
the
data
available,
but
also
the
soft
and
so
the
software
runtime,
so
that
we
are
only
depending
on
a
very
stable
interface
layer,
which
is
the
UNIX
kernel,
so
next
yeah
right
so
yeah,
so
containers
basically
give
us
reproducibility
in
space-time
right.
So
we
can
not
only
run
an
application
in
the
different
data
center
today,
but
we
can
also
run
it
in
the
future
and
maybe
reuse
an
application
later
on.
B
So
in
my
opinion,
this
has
been
a
complete
game-changer
for
open
science,
at
least
for
these
large-scale
scientific
collaborations.
Ok,
so
if
you
want
to
learn
more,
maybe
try
it
out
yourself.
You
can
also
go
to
the
certain
open
data
portal
and
learn
more
about
what
experiments
actually
make
data
available,
and
yes,
so
we
actually
want
to
just
thank
the
entire
Cooper
News
community.
This
has
been
great
for
us
yeah.