►
Description
KCD Italy 2021* - The video of the talks are now available!
Kubernetes has become the new application server of modern cloud-native applications.
Efficiency is one of the top benefits, thanks to containers' lean resource footprint and Kubernetes ability to pack them while ensuring performance isolation.
The reality can be different: many teams adopting Kubernetes see high infrastructure costs and performance issues. Why is that? Applications need to be carefully configured to achieve the touted benefits, but Kubernetes resource management is tricky and there are many microservices to be tuned.
In this talk, we analyze Kubernetes resource management concepts from a performance and reliability perspective. We then show how AI can help developers and SREs to achieve the right balance among low costs and optimal application performance and reliability, automating the sizing of container resource settings driven by your application SLOs.
🤓 SPEAKER: Stefano Doni, CTO @ Akamas
______________________
* KCD Italy 2021 is the first edition of a community-organized non-profit event that will take place on November 17th and 18th and will gather adopters and technologists from open source and cloud native communities for education, collaboration, and networking.
📯 Follow us to be notified about the upcoming MeetUps: https://community.cncf.io/kubernetes-cloud-native-milano/
A
Here
we
are
for
the
last
part
of
this
eataly
2021
crisis.
We
are
sadly
reaching
the
end
of
this
event
before
inviting
the
next
guests
on
stage
watch
the
chat,
because
there
is
the
form
with
the
latest
available
t-shirts.
There
are
still
very
few
t-shirts
available.
You
have
to
book
super
fast
with
the
form
that
arrives
in
chat
now,
I
invite
the
next
representative
of
the
corgan
heiser
christmas,
hello
christmas.
You
are
silent
here
we
are
hello,
yes,.
B
C
B
B
B
I
saw
a
great
participation
in
the
chat
and
the
next
talk
is
also
very,
very
interesting,
but
before
going
to
the
next
talk,
I
wanted
to
do
chiara
felt
a
little
king
cup
a
about
what
we
did
so
this
english
day
of
a
community
event
organized
for
the
italian
community.
Thank
you
all
for
being
here.
Thank
you
all
for
attending
the
event,
and
I
wanted
to
especially
thank
these
fantastic
organizers,
chiara,
starting
with
you.
Then
there
is
nicla.
There
is
angela
there
annalisa.
There
is
giulio.
There
is
celiac.
Then
there
is
fabrizio
adriano
paolo.
B
B
A
Fantastic
and
congratulations
to
you
too,
that
yesterday,
you
did
a
super
job
and
tracing
and
definitely
paved
the
way
for
today's
work.
Christmas
tell
us
briefly
who
you
are
so
also
who
saw
you
yesterday,
maybe
you
weren't
there
yesterday
and
today
the
two
words
given
and
then
present
the
last
one
speaker
how
sad
you
are
to
present
the
last
one
and
the
event
is
concluding.
B
But
in
fact
that
is,,
thank
goodness,
we
have
already
anticipated
that
we
continue
with
the
gubernatis
meetups.
So
we
continue
the
valley
of
sales
of
this
clone
of
vital
this
cube
n
d,
italy,
fantastic,
so
I'm,
sad
that
let
's
end
it
happy
also
because
we
continue
to
do
a
little
introduction
about
myself
for
the
other
sessions.
My
name
is:
natal
won,
I'm
really
kate
in
edited
robben
shift.
I
also
wrote
a
book
recently
this
modernizing
the
flight
java.
B
If
you
want
to
download
it,
you
can
find
it
on
developers
ed
it's
from
the
point
com.
I
deal
with
the
activities
of
evangelists,
bernardis
and
clowns
and
ting,
and
I
am
very
very
happy
to
be
here
and
help
in
the
organization
of
this.
What
about
eataly,
which
is
going
great?
Oh,
please,
don't
forget
to
tweet,
let
it
be
that
he
is
italy
lendle
that
risitaly
tweet
share
to
the
screenshots.
Let
us
know
what
you
think
well.
B
You
use
the
form
to
get
the
shirt
they
are
selling
fast,
so
stefano
and
cto
adac
hamas
is
obsessed
with
performance,
optimization
and
leads
the
autonomous
performance
optimization
project.
The
artificial
intelligence
of
akamas
expert
speaker
has
obtained
several
talks
over
the
years
and
won
the
computer
measurement
in
2015
group
best
paper
award
for
his
java
application.
Optimization
skills,
stefano
vaj.
B
D
So,
thank
you
all
for
being
with
us
today
welcome
to
my
talk
on
using
machine
learning
to
actually
automatically
optimize
costs
and
efficiency
and
the
reliability
of
applications
on
gubernatis.
This
is
the
agenda.
Therefore,
we
will
make
a
rapid
introduction
to
cuba
metis,
especially
aimed
at
the
challenges
that
the
organization
today
who
are
equipping
it
are
encountering,
above
all,
on
the
issues
of
optimizing
performance
and
reliability.
Costs
which
are
three
very
connected
dimensions
we
will
go,
can
then
understand
what
are
the
roots
of
these
problems
a
little
bit.
D
Then
we
will
do
a
little,
who
is
not
an
exploration
of
what
are
the
mechanisms
that,
as
metis
adopts
for
the
management
of
the
resources
of
the
containers
of
the
applications
after
which
I
will
limit
it?
What
is
the
new
approach
that
we
have
introduced?
Adac
hamas,
with
the
use
of
machine
learning
to
automate
the
optimization
itself?
We
will
then
see
how
this
approach
is
applied
in
a
real
case,
and
we
will
conclude
with
some
takeaways
a
few
words,
even
if
you
have
already
made
an
introduction
thanks.
D
I
am
precisely
stefano
two
websites,
akamas
h,
but
if
an
italian
company
and
was
born
within
the
furniture
group,
where
we
have
been
doing
performance,
optimization
capacity
planning
testing
for
more
than
twenty
years,
then
let's
start
with
a
quick
interruption
cover
netease,
even
if
we
are
in
a
dedicated
event
with
veneti,
if
in
fact,
I
like
to
start
with
a
story
of
how
kuber
netease
was
born,
which
I
found
interesting.
In
fact,
I
think
it
is
not
entirely
known,
obviously
governed
part
within
google.
It
is
an
answer.
D
D
Ling
with
tenderness
you,
the
system
at
the
beginning,
was
obviously
a
trade
secret
of
google
was
called
borg,
and
it
is
interesting
how,
at
a
certain
point,
a
group
of
engineers
within
google
decided
to
make
it
open
source,
and
the
interesting
aspect
is
this
reaction,
which,
at
the
beginning,
let's
say,
was
a
bit
adverse
in
effects
within
google.
So
you
want
to
make
open
source
a
project
that
actually
has
our
greatest
competitive
advantage.
D
Fortunately,
the
team
found
us
on
several
occasions
and
in
the
end,
got
it
in
light
by
the
organization
of
google
itself
and
therefore,
today
we
have
this
big
open
source
project
that
we
are
all
working
on,
which
is
definitely
extremely
interesting.
How
we
boo
to
so
many
benefits
here
in
this
topi.
I
want
to
dwell
mainly
on
the
issues
of
performance
efficiency.
Among
the
main
benefits
it
clearly
happens.
D
However,
let's
say,
despite
these
promises,
what
can
be
observed
is
that,
in
fact,
the
netease
cover
courses
are
growing,
that
you
see
a
serve
made
this
year,
which
basically
demonstrates
this
aspect
by
collecting
feedback
and
also
from
mature
organizations
that
adopt
this
technology.
The
main
question
that
comes
is
why
these
costs,
in
fact
the
changes
are
growing
so
much
clearly,
we
have.
The
main
reason
tends
to
be
that
it
is
having
more
and
more
adoption.
Therefore,
there
is
a
mature
increase
in
pervasiveness
with
an
increase
in
use
of
the
platform.
D
D
In
fact,
a
little
bit
even
unexpected,
if
you
want
from
a
certain
point
of
view
here,
are
also
many
videos
available
to
taste
with
this
cube.
Metis
philos
also
interesting
torres
that
a
site
entirely
dedicated
to
bankruptcy
stories,
launched
by
an
engineer
from
zalando,
where
all
the
companies
report,
on
the
one
hand,
the
performance
practices
and
established
issues
that
are
connected
to
them.
So
once
again,
the
question
is
given
that
the
platform
makes
available
these
these
possibilities
of
obtaining
performing
systems
and
coefficients
on
costs
and
stable
on
the
other
hand,
in
fact,.
D
We
realize
that
organizations
adopt
juventus,
which
is
actually
not
so
easy
to
obtain,.
Therefore,
from
this
first
part
of
the
talk,
let's
understand:
why,
therefore,,
which
ones
are
in
fact
resource
management
mechanisms
which,,
as
you
put
it,,
stop
in
place.
So
let's
review
the
fundamentals
a
little
bit,
because
there
are
some
aspects
in
my
opinion
that
are
counterintuitive
or
in
any
case
even
new
compared
to
a
previous
world
of
virtualization,
if
you
want
to
which
we
have
always
been
used
to.
D
So,
let's
see
4
facts,
fundamental
four
concepts,
plus
concept
is
that
of
resource
requests
request.
So
when
a
developer
clearly
has
to
define
a
bit
of
a
deployment,
let's
say
the
application
on
gubernatis
can
define
resource
requests.
So
the
requests
are
in
fact,
from
the
amount
of
resources
that
a
container.
In
fact,
then
it
is
guaranteed
at
the
level
of
by
cover
nettis.
So
this
example
becomes
a
bit
like
the
developer
is
asking
for
two
cpu
and
2gb
of
theron
pod
memory,
and
therefore,
why
is
this
pod
spent?
D
of
available
resources,
if
then,
as
in
the
example,
let's
say,
a
bit
of
the
same
amount
of
cpu,
should
arrive
in
a
second
pod
for
two
other
cpu,
as
light
is
the
obstacle
was
on
the
same
way
if
a
bit
dc
queen
were
to
arrive
in
this
case?
A
third
pod
that
assuming
we
have
this
node
in
the
cluster?
In
fact,
this
way
at
this
point
it
has
a
capacity,
was
full
in
the
sense
that
all
four
cpu's
are
started,
so
this
pod
would
be
there
would
be
waiting.
D
So
this
already
gives
us
a
first,
let's
say,
insight,
a
first,
a
very
interesting
first
lesson
or
that
how
metis
manages
the
capacity
of
infrastructural
resources.
Obviously
this
is
true.
We
more
than
understand
that
on
the
cloud
on
the
basis
of
the
requests
we
ask,
there
is
no
over
commit
nor
on
these
requests.
D
There
is
also
a
consequence
of
this
aspect,
which
is
due
to
the
fact
that
this
capacity
management
comes
precisely
on
the
requests
here
on
the
location
without
actually
looking
at
the
actual
use.
Therefore,
a
fairly
typical
effect,
which
what
we
notice
organizations
struggling
as
metis
is
clusters,
let's
say,
become
full
capacity
capacity
level,
but
actual
utilization
level
very
low.
So
let
's
say
there
is
potential
level
of
wasting
ease,
also
in
wasting
infrastructure,
and
this
is
one
of
the
main
reasons
why
we
find
ourselves
having
these
high
costs.
D
The
second
aspect
is,
inter
alia,
another
interesting
fact
which
instead
has
to
do
with
the
performance
and
stability
of
the
applications.
Therefore,
the
requests
we
have
seen
serve
to
guarantee
resources
a
container,
but
in
fact
it
can
console
you
more
as
much
as
more
up
to
another
concept,
which
are
the
rhesus
limits.
Therefore,
limits
that
can
be
specified
beyond
which
we
say
consumption
cannot
not
go
substantially,
depending
on
the
fact
that
we
have
cpu
memory
of
type
of
resource.
Different
things
happen
when
our
application
tries
to
go
beyond
the
limits.
D
So
if
it's
about
cpu,
let's
say
how
cpu
is
managed
as
a
so-called
compressible
resource,
so
in
fact
it
is
artificially
reduced
overboard.
The
two
cpu
we
requested
in
the
example
previous.
Obviously,
if
this
change,
we
have
performance
problems,
because
the
application
of
artificially
slowed
down
different
instead,
the
case
in
which
these
limits
are
met.
As
regards
memory
resource
here,
things
happen
a
little
more
drastic
in
the
sense
that
it
is
connected
with
you,
it
is
directly
killed.
Then
this
was
also
a
bit
of
a
novelty
compared
to
what
we
were
used
to.
D
D
The
third
fact
is:
this
is
a
little
less
known.
In
fact,
it
is
a
research
that
we
have
also
done
within
acam
or
what
happens
when
we
have
these
cpu
limits
on
the
cpu
that
are
set
and
when
we
say
so,
our
applications
begin
to
approach
this
limit
from
the
diagram
we
have
seen
before.
It
would
seem
that
the
application
is
slowed
down
in
some
way
only
when
your
level
of
use
reaches
the
maximum
level
of
our
limit.
In
reality,
something
much
more
happens,
much
more
impactful
or
these
limits.
D
D
One
of
the
reasons
why
this
happens
is
that
suddenly,
let's
say
access
to
cpu
is
artificially
slowed
down
by
cover
netease,
or
rather
the
system
by
the
kernel
mechanisms
of
underlying
linux,
on
which
we
enter
right
now.
But
if
you
are
interested,
you
can
contact
me.
I
am
absolutely
available
to
deepen.
Are
the
aspects
that
can,,
let's
say:
so,
can
impact
the
performance.
Some
companies
a
year
ago
tried
to
see
what
happens
by
completely
removing
the
limits.
D
Here
you
see
an
example
of
quite
famous
of
buffer,
an
american
company,
where
you
see,
let's
say
in
this
graph,
the
latency
of
some
micro
services
at
some
point
in
the
middle
of
the
graph,
cpu
nimitz
have
been
removed,
notice
how
there
has
been
a
drastic
drop
behind.
It
is
a
factor
of
22
factor,
22
of
the
latency,
simply
removing
the
limits
without
changing
the
application
between
structures.
Obviously,
this
brings
the
question,
but
then
the
ris
limits
we
have
to
use
anymore.
D
Actually,
no
because
rhesus
of
miss,
we
remember
the
net
cover
benefits
on
one
of
the
mechanisms
for
which
it
allows
gubernatis
to
limit
the
consumption
of
applications.
If
we
remove
them,
we
could
have
applications
that
suddenly,
due
to
a
problem
with
a
mou,
wrong
release
or
a
bug,
can
consume
all
our
resources,
in
spite
of
other,
perhaps
business
critical
applications.
In
fact,
even
google
itself
recommends
it
as
like
best
practices
to
avoid
established
ixus.
So
this
leads
us
to
say
that
optimizing
free
applications
is
actually
a
challenge.
D
We
have
the
resources
for
the
quests
that
determine
they
can
determine
the
exaggerated
costs
limits
that
we
have
to
use
them.
However,
if
we
put
them
too
tightly,
obviously
they
are
not
chosen
for
each
single
micro
service.
In
fact,
we
have
even
a
little
counterintuitive
performance,
but
in
reality,
then,
if
we
broaden
our
objective
a
bit
typically
a
micro
services
application,
we
have
many
micro
services.
Our
objective
typically
is
to
guarantee
gravity,
reliability
and
performance
of
an
entire
application,
which
can
typically
be
built
from
dozens
of
micro
services.
D
15
are,
let's
say,
extremely
a
very
large
number
of
parameters
inside.
Let's
say
this
is
the
application
cover
netease
requests
limits
are
also
themes,
auto
scaling,
then,
by
the
way,
within
an
application
that
contains
itself.
There
are
also
the
gvm
runtime,
for
example,
about
700
configuration
parameters
itself.
So
this
is
kind
of
the
challenge
facing
us.
Today
is
a
modern,
claudette
application
so
that
we
say
that,
typically
of
being
the
performance
engineering,
the
developers
are
faced
with
to
ensure
reliability.
Therefore,
an
extremely
complex
challenge.
D
Given
this,
this
large
number
of
parameters
was
the
dynamism
of
these
environments.
Therefore,
to
face
this
challenge
at
da
tamas,
we
imagined
a
totally
different
approach
where
actually
the
machine
learning
comes
into
action
at
home's
approach
is,
in
fact,
based
on
four
big
pills:
four
big
innovative
elements.
First
of
all,
an
approach
must
be
of
this
type.
To
solve
this
problem
correctly.
It
must
be
food
stack.
What
does
it
mean
that
it
doesn't
in
reality?
It
is
enough
just
to
look
at
an
infrastructure
layer,
for
example
my
resources,
the
crash
limits
of
my
containers.
D
It
is
very
important
to
also
look
at
what
is
inside
the
container,
therefore
runtimes
such
as
gvm
or
even
go,
rather
than
also
the
application
parameters
of
the
dell
'application
or
even
going
below
actually
also
how
configured
what
kind
of
nodes
or,
in
the
cluster,
the
second
central
aspect.
We
have
all
these
parameters.
Actually,
human
brain
is
not
well
equipped
to
solve
the
problem.
D
Even
the
performance
aspects
can't
make
it
because,
actually
beyond
to
the
ability
to
reason
when
we
have
hundreds
of
parameters,
we
really
have
billions
of
combinations,
and
this
is
where
machine
learning
is
central
to
efficiently
explore.
This
configuration
space
find
the
optimal
configuration
that
gives
us,
but
if
performance
at
minimum
cost
gaulle
east-
and
it
is
another
fundamental
aspect,
that
is,
with
this
type
of
approach,
we
can
indicate
to
machine
learning.
What
is
the
objective
of
optimization
here?
It
is
not
a
generic
machine
learning
that
we
set
up-
let's
say
ours
contained.
D
We
can
actually
indicate
the
objective,
which
can
be
a
specific
business
metric
example
I
want
to
increase
payments
according
to
my
application,
and
machine
learning
works
behind
the
scenes
so
that
all
contains,
for
example,
are
correctly
configured
for
this
type
of
goal,
and
this
type
of
goal,
among
other
things,
can
also
include
constraints,
and
we
will
see
it
later
in
applied
case
study.
How
last
pillar
and
central
was
there
automatic?
What
does
it
mean
by
tick?
D
So
let's
say
what
is
my
goal
of
optimization
our
vision
that
a
button
is
pressed
and
the
platform
let's
say,
takes
charge
of
the
optimization
that
can
proceed
in
a
totally
automated
way
quickly.
How
this
optimization
process
works
at
a
high
level
guided
by
the
masci
days,
then
we
see
the
applied
case.
Basically,
we
have
five
phases
of
work.
The
first
phase
you
see
on
the
left.
There
is
conceptual
akamas
conceptualized
on
the
right.
There
is
a
cabernet
system,
cestistically
an
application
built
by
n
micro
services.
D
Therefore,
in
the
first
phase
of
the
cycle-
and
we
apply
the
configuration,
therefore,
as
configurations,
we
tend
the
parameters
that
we
are
going
to
touch,
for
example,
of
its
requests
in
limits
the
parameters
of
ours,
of
simplification
so
on
we
can
do
this
through,
for
example,
the
use
of
the
governed
plans.
The
same
second
step
is
to
apply
work.
What
does
it
want
say?
We
have
to
understand
if
the
configuration
that
we
have
applied
actually
good
when
the
application
is
under
stress
in
a
typical,
let's
say
way
to
do
this.
D
This
second
stage
is
to
integrate
consultants,
their
heads
typically
to
size,
the
containers,
one
of
the
ways
is,
we
make
use
of
los
test
in
staging
environment
to
simulate
what
happens
in
production,
and
then
we
use
we
integrate
with
these
tools
to
automate
for
the
launch
of
the
test
without
having
to
rewrite
reinvent
the
wheel
third
step,
we
can't
find
them
system,
so
we
get
m
of
performances
course
from
the
system.
We
are
optimizing.
D
Once
again,
we
don't
reinvent
the
blank
we
employ
agents
with
existing
monitoring
tools
or
prometheus
very
common
case
in
cover
netease
lacks
enterprise
like
dynatrace
fourth
phase.
Once
we
have
collected
the
metrics,
let's
say
so:
let's
analyze
the
goal,
the
goal
could
be,
for
example,
how
many
mine
business
metric
that
we
indicated
before,
or
even
a
ratio
between
two
meters
which,,
as
we
will
see
in
the
example,,
is
once
we
have
defined
the
goal.
Therefore,
the
goodness
the
score
of
this
experiment.
D
How
good
is
this
configuration
at
which
point
5
machine
learning,
in
fact
to
take
the
decision
to
figure
out
what
is
the
next
configuration
to
apply
and
this
and
this
process
restarts
from
phase
1?
So
this
is
a
bit
high
level
and
if
we
go
to
see
with
a
practical
realworld
case
like
this
optimization,
let's
say
it
works
in
one
case
in
one
context
of
a
cover,
next
application,
friends
or
services.
Therefore,
the
application
that
we
have
taken
as
an
example
is
enough.
D
Also,
quite
well
known,
the
google
online
boutique
application
that
a
netiv
cloud
application
made
by
google
made
up
of
about
ten
micro
services
that
precisely
it
simulates
an
ecommerce
shop
and
let's
say
it
consists
of
a
variety
of
technologies.
It's
actually
not
just
let's
say
here's
bernet
hicks
with
run
time
with
applications,
gol
js
java,
python,
also
a
regis,
and
so
on
the
interesting
thing
it
includes
a
cancer
generator.
We
tell
you.
D
Thus
it
also
allows
us
to
automate
the
testing
phase
very
easily,
precisely
because
it
is
already
included
in
the
project
itself
in
which
we
integrated
with
akamas,
to
carry
out
this
optimization,
guided
by
automatic
machine
learning.
Here
you
see
a
little
bit
of
overall
architecture.
Therefore,
here
veneti
10
applications
errata
point
consists
of
10
micro
services.
Actually,
there
is
also
a
redis
database,
so
they
make
11
micro
services
in
11
optimized
containers.
The
parameters
we
touched
are
22.
D
You
see
at
the
top
right
then
cpu
and
meet
memories
of
each
of
each
of
the
eleven
of
the
eleven
containers.
In
fact,
we
get
the
market
oriented
in
monitoring.
In
this
case,
we
leverage
prometheus,
which
will
give
us
metrics
that
tend
to
use
the
resources
and
the
stir,
which
also
gives
us
the
service
performance
metrics
at
the
level
of
each
single
micro
service.
Then
we
make
use
of
locus
and
lo
generated
already
included,
let's
say
in
the
google
open
source
project
to
submit
a
realistic
load,
determined
amount
of
time
towards
this
application.
D
D
This
is
the
optimization,
but
what
is
really
the
goal
of
these
optimizations
as
told
at
home
goldin,
where
we
can
speci
tar
is
the
goal
we
are
interested
in.
This
is
the
goal
we
have.
Let's
say
that
I
chose
for
this
case,
for
example,
so
the
optimization
here
aimed
to
have
a
service
that
was
more
cost
efficient,
therefore
increasing
the
throughput,
but
at
the
same
time
also
decreasing
the
cost.
D
Let's
say
to
run
this
service,
this
application
on
the
cloud,
so,
let's
say
in
dacca
massa.
We
have
given
a
goal
composed
of
these
two
meters
that,
on
the
one
hand,
we
have
the
throughput
income
in
saks
transaction
measured
by
on
the
front
end
so
effectively
the
transactions
successfully
disbursed,
divided
by
the
cost
la
the
cost.
To
be
noted
here.
It
is
interesting
because
we
can
calculate
the
course
of
these
all
with
you,
for
example,
by
leveraging
the
pricing
price
that
found
clouds
expose
us.
D
D
So
we
have
each
a
configuration
that
h,
but
xvi
ne
has
decided
that
that
says
the
stars,
the
size
jones
of
merges
these
boxes
to
an
associated
course
that
we
obtain
from
the
provider
and
to
a
certain
volume
of
performance
that
we
obtain
by
measuring
the
stress
from
the
stress
test
with
locust,.
The
evident
possibility
that
we
have
missing
tari
constraints
is
also
very
interesting,.
So
what
it
means
are
constraints
on
optimization
that
we
can
indicate
on
some
other
properties.
D
We
want
us
to
say
the
configurations
that
h
but
suggest
you
to
explore,
therefore
satisfy
this
typical
case.
Typically,
we
want
the
service
not
to
worsen
its
reliability,
so
we
don't
want,
for
example,
failures
beyond
a
certain
threshold
and
also
typically,
we
don't
want
that-
has
very
high
response
time
peaks.
So
we
have
also
made
explicit
a
constraint
on
response
time
in
the
front
end
of
500
milliseconds
2
seconds
lower
than
this
is
the
optimization
design
where
we
can
effectively
express
declarative
mode.
D
What
we
want
in
terms
of
performance
liability
of
our
application,
after
which
we
actually
tremble
play,
is
all
the
optimization
can
come
totally
automatic
maya
by
atlas.
Let's
see
the
results
that
have
been
achieved,
so
this
graph,
you
see
dots
that
actually
identify
the
performance
of
the
configurations
obtained
by
camas,
the
first
red
dot.
Sorry
blue
on
the
left
identifies
the
performance,,
let's
say,
the
cost
efficiency
in
reality
of
the
beslan
configuration,
I.e.
D
the
initial
configuration
before
starting
the
optimization,
could
be
the
one
you
have
of
production
that
you
have
determined
by
making
the
marble
at
the
tuning,
all
the
other
pontines
the
green
ones
represent.
Let's
say
on
this
graph
course:
efficiency
level,
therefore,
always
our
aim,
our
goal
and
throughput
divided
by
the
cloud
cost
of
all
the
queues.
The
configurations
that
akamas
has
tried.
Therefore,
each
dot
is
all
the
execution
of
the
five
phases
that
we
have
seen
before,
where
ankama
sapri
to
apply.
D
The
eleven
mixi
more
memory
and
did
the
tester
has
determined
its
goodness.
The
interesting
aspect
is
that
a
77
percent
improvement
in
cost
efficiency
has
actually
been
obtained.
So
if
first
in
the
based
configuration
used
to
be
0
29
transactions
per
second
per
dollar
per
month
now
we
may
have
10
52
transactions
per
second
per
dollar
per
month.
So
an
efficiency
issue,
including
economics,
has
been
auto
optimized
by
comes
by
virtue
of
this
nature.
D
At
the
gold
river
also
note
very
limited
time,
let's
say
almost
resolve
compared
to
what
it
would
have
taken
to
do
this
activity.
How
do
you
manually
count
all
these
parameters,
so
in
villette
the
best
configuration
was
reached
in
about
35
experiments
so
very
quickly,
by
virtue
of
the
machine
learning
ability
to
learn,
despite
the
'high
number
of
parameters
very
quickly.
Where
are
my
best
space
exploration
areas?
D
The
typical
question
at
this
point
is
okay,
but
what
is
the
configuration
you
found
from
canas?
How
how
is
it
possible
effects?
How
was
it
possible
to
have
this
result
so
significant?
Here
you
see
a
representation,
the
most
important
parameters,
10
plants,
more
important
of
the
22-
that
we
have
optimized
for
each
container
here.
I
remember:
the
goal
is
to
optimize
the
cpu
and
memory
the
limits.
D
Therefore,
a
totally
new
approach,
totally
different
compared
to
the
current
approaches
and
also
interesting
to
go
and
see
by
applying
this
bes
configuration
what
changes
in
terms
of
performance
of
my
service
here,
which
is
what
typically
interests
me
here.
You
see
two
graphs
but
compare
where
we
are
going
to
compare
the
performances
in
the
same
graph
over
time,
this
time
of
a
single
configuration
of
the
best,
so
the
best
configuration
we
buy
it
against
the
baseline,
which
is
once
again
the
starting
configuration
here.
D
D
This
predictive
point
of
view,
therefore
not
exposed
and
very
low,
very
constant,
which
typically
also
important,
from
the
point
of
view
site
reliability
having
a
rapid
and
constant
system,
not
a
dancer
in
terms
of
response
times,
which
evidently
was
due
to
a
wrong
configuration
of
the
containers.
I
conclude,
therefore,
the
takeaways
we
have
are
those
of
three
substantially.
D
Therefore,
central
messages
are
actually
as
read
great
hall
platform
that
was
created
made,
I
think,
by
google
for
running
applications.
Micro
services,
a
number
of
annoyances
functionality.
That
c
makes
life
much
easier.
However,
optimization
remains
a
very
complex
problem
by
virtue
of
the
fact
that
we
have
a
lot
of
parameters
that
influence
each
other
and
the
current
approach
is
still
a
manually,
try
and
try
again
approach
that
takes
a
lot
of
time
and
requires
a
lot
of
resources.
The
gamut
jaw
is
revolutionizing,
inter
industry.
D
D
I
conclude
with
one
like
this
typically
a
frequently
asked
question:
if
you
often
want
performance
engineers
at
the
end
of
the
confluence,
they
ask
us:
it
is
possible
to
try
cras.
Very
simply.
We
have
a
free
trial,
you
can
go
to
our
site
and
you
can
download
it
and
apply
it
in
fact,
that
produce
a
little
even
this
scenario
that
I
have
shown
you,
but
also
others,
and
then
try
it
on
your
applications
or
finished
thanks.
A
D
D
D
A
D
Well,
first
of
all,
the
2
answers
primarily
and
some
services
could
be
oversized.
In
fact,
let's
say
so.
There
may
be
services
that
have
an
excessive
limit
and
therefore,
let's
say
I'm,
not
the
botnet
on
the
bottleneck.
The
other
interesting
aspect,
maybe
a
little
more
counter
intuitive
and
that
in
fact,
also
the
behavior.
We
must
always
consider
the
slush
of
a
container.
D
There
is
a
runtime,
for
example,
the
java
virtual
machine
or
even
the
go,
manages,
for
example,
the
memory
in
a
dynamic
way,
so
also
how
many,
and
so,
if
we
give
it
is
very
important
to
the
point
of
safe
performance,
sometimes
from
excessive
resources.
It
turns
out
to
be
counterproductive
precisely
for
the
management
mechanisms.
I
will
give
a
classic
example:
gvm
scoring
fewer
resources,
it
could
have
more
ease
in
memory
management,
also
change
mechanisms
of
gardez
collector,
another
counter-,
intuitive
aspect
that
is
usually
discovered.
D
It
is
not
discovered
because
it
is
not
seen
without
us,
knowing
it
on
the
basis
of
how
we
configure
the
container.
Therefore,
this
sea
of
​​,
what
we
have
inside
the
count
is
not.
It
can
involve
a
different
way
than
even
limits.
This
is
also
one
of
the
typically
unobservable
aspects,
but
which
fall
within
the
total
complexity
of
this
type
of
work,
but
which
also
gives
us
benefits.
If
we
succeed
on.
A
D
Ok,
we
discussed
this
I
think
I
have
an
exact
backup
site
in
kamata
punto.
It
was
created
with
a
concept
of
fusta
platform,
as
we
said,
the
first
central
kbc,
so
in
reality
we
can
go
both
above
and
below.
If
we
want,
with
respect
to
the
container
that
we
have
shown
today
as
an
example,
so
it
is
very
interesting
to
go
inside
the
runtime,
but
also
the
framework
here.
If
we
think
of
database
spark
big
data
applications.
D
Here
too,
we
have
hundreds
of
parameters,
but
also,
and
often
application
level
exist
parameters
and
that
they
too
configure
themselves
once
in
a
strange
way
compared
to
what
we
do
on
the
pod.
So
you
can
typically
go
up
to
the
application
level,
but
also
installment
down,
for
example,
one
of
the
interesting
aspects,
especially
on
the
cloud
and
the
choice
of
type
of
distances.
Therefore,
also
this
has
an
impact
on
performance,
feel
end
and
on
costs.
So
let's
say
the
whole
stack
is
typically
addressable.
D
D
Other
typical
question:
we
are
working
on
this
too.
So
the
example
we
showed
typical
mode
of
use
of
acam
in
production
environments
where,
however,
typically
organizations
already
have
stress
tests
for
the
most
important
applications
reproducing
business
scenarios
and
therefore
this
is
an
easy
way
to
attack
the
existing
single
system.
D
A
D
A
A
A
A
We
will
upload
all
the
speeches
there
and,
with
the
help
of
some
slides,
I
would
say
that
we
can
go
towards
the
final
greetings,
so
o
I
really
want
to
thank
time
that
you
take
it
upon
yourself
to
thank
all
our
speakers
of
this
day,
but
yesterday's
anticolli
also
edoardo,
who
was
a
really
high
level
presenter
yesterday,
I
thank
all
the
partners
who
they
endured.
This
event
had
great
resonance,
therefore,
project
new
life,
binary,
ethical
regis,
red
cubes,
fury,
storage,
sticker
miur,
thank
the
sponsors
and
co
organizers.
A
There
is
classic
skira
technique
platform
le
dat,
yes,
gap,
far
factories
and
vmware.
Have
you
seen
them
these
days
getting
on
the
stage
and
presenting
all
the
speakers?
Each
brought
a
contribution-
and
it
was
this
mind,
was
the
result
of
an
organization
that
lasted
a
very
long
time,
and
that
saw
us
very
busy,
and
we
are
speaking
on
behalf
of
all
very,
very
happy
with
how
it
went.