►
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
Service Meshes: At What Cost? - Lee Calcote, Layer5 & Girish Ranganathan, SolarWinds
“What is the performance impact that a service mesh has?" "What overhead does being on the mesh incur?” By far, this is the most common questioned by engineers coming to with the value of functionality provided by a service mesh. Generally, this question goes unanswered. We will share methodology and results of performance testing research done in collaboration with a university, through the lens an open source service mesh benchmark tool - a tool used to provide a common benchmark across service meshes (their control planes, like Istio) and modern proxies (their data planes, like Envoy). Over 10 service meshes projects will be reviewed. In addition to performance, we’ll take an in-depth look at the landscape of service meshes, characterize and contrast their functionality as well as their data plane and control plane architectures.
https://sched.co/MPf7
A
A
Well,
we
were
so
so
my
name
is
Lee
co-presenters
gireesh.
We
were
here
at
cube
con
last
year,
giving
a
workshop
on
serviceman.
Who
was
at
that
workshop
anyone?
It
was
a
very
good.
Ok,
you
did
not
learn
your
lesson
earlier.
You
were
gonna
suffer
more
today
gave
a
lot
of
workshops
who
gave
give
them
at
docker
con
at
velocity
at
cube
con.
A
A
Some
of
those
questions
were
trying
to
pet
rock
the
various
service
meshes
that
are
available
that
are
out
there,
but
a
fair
number
of
questions
around
related
technologies,
which
makes
a
lot
of
sense
to
me
as
I
go
to
learn
a
new
concept
a
lot
of
times.
It's
very
helpful
to
me
when
someone
compares
it
to
something
I
already
understand:
I've
got
a
common
kinda
mental
framework
for
it.
In
this
case,
API
gateways
client
libraries
they
tend
to
do.
They
say
that
they
do
similar
things
and
they
do
some
similar
things.
A
This
is
not
that
talk.
Point
of
this
is
to
say
there
is
an
emerging
landscape.
Are
a
number
of
services
available
you've
got
choice
which
is
excellent.
It
also
brings
with
it
some
challenges,
as
you
go
to
try
to
figure
out
what
choice
to
make
and
what
to
adopt.
Those
challenges
include
the
two
most
commonly
asked
questions
from
our
workshops.
A
Each
of
the
questions
started
with
we
were
pretty
similar.
We
were
done.
The
workshop
people
were
absolutely
blown
away
mostly
by
the
presentation
and
the
corny
jokes
that
I
had
had,
but
they
also
said
for
the
most
part:
hey,
we
see
the
value
of
this
sort
of
Smashing.
We
want
to
try
it
out,
we're
ready
to
adopt
which
one
should
we
use?
A
How
should
we
get
started?
How
much
is
this
gonna
cost
us?
What's
the
performance
overhead
here,
you
don't
get,
you
know
very
rarely.
If
ever
does
do
you
get
something
for
free?
We
didn't
necessarily
have
fantastic
answers
on
behalf
of
the
projects
that
we
were
speaking
about
on
behalf
of
those
meshes.
Some
of
these
service
meshes
are
easier
to
install
than
others.
A
A
I
think
maybe
convinced
at
this
point
that
that's
you
know
an
actual
exact
same
comparison
is
probably
not
feasible,
but
we
can
get
close
enough
or
you
might
be
able
to
get
comfortable
and
close
enough
that
you
understand
what
that
cost
is,
and
you
weigh
it
against
the
value
that
you
derive
from
having
deployed
a
service
nice
we're
going
to
talk
about
a
tool
that
will
help
us
in
this
regard.
Before
we
talk
about
that,
I
want
to
acknowledge
that.
A
Not
only
has
the
landscape
been
evolving,
but
service
measures
themselves
have
been
I
suspect
some
of
us.
It
wouldn't
take
me
just
one
row
of
folks
here
in
the
audience
to
disagree
about
the
definition
of
a
service
mesh
potentially
and
what
makes
what
makes
one?
Let's
see
if
we
can
disagree
together
right
now,
you
can,
if
that
happens,
please
ignore
my
Twitter
handle
go,
go
talk
somewhere
else.
A
A
Please,
yes,
solidarity,
very
good,
okay!
So
some
of
these
terms-
and
even
if
you
don't
have
a
networking
background,
these
terms
are
kind
of
familiar
to
you
anyway,
as
you've.
No
doubt
heard
about
control
planes
quite
a
bit
a
lot
of
times,
as
we
talked
about
in
a
network
centric
discussion,
but
it,
but
certainly
that
concept
applies
too
well
to
kubernetes
as
an
example.
A
So
I
won't
explain
these,
because
I
expect
that
folks
pretty
much
know
what
those
are
in
the
land
in
the
sort
of
in
the
annals
of
computer
science.
When
folks
are
studying
tcp/ip
and
becoming
a
network
engineer
you
typically,
you
also
talk
about
a
management,
plane
and
I
suspect.
There's
a
need
here
for
management
plane
as
people
go
to
earnestly
adopt
surface
meshes
important
to
look
at
the
architectures
for
a
brief
moment
really.
The
point
here
isn't
necessarily
to
comprehend.
A
There
are
more
of
these
nerd
charts
at
a
resource.
I'll
point
you
to
later.
These
are
really
fun
to
spend
time
on
and
and
just
try
to
tweak
I'm
gonna
skip
some
other
ones,
but
there's
probably
maybe
by
my
account
about
ten
a
little
over
ten
service
wishes.
You've.
No
doubt
heard
of
the
two
that
you
saw.
This
is
amazing.
I
can
get
this
far
into
the
talk
and
not
actually
say
the
name
of
the
service
master.
So
I'm
surprised.
A
So
not
only
are
the
architectures
different
not
only
do
they,
the
features
and
functions
differ,
but
the
deployment
models
can
differ
as
well.
I,
don't
know
if
you
caught
some
snark
on
those
last
couple
of
slides.
Maybe
it's
good
that,
since
my
twittle
Twitter
handle
is
on
the
pages
that
I
didn't
get
snarky
about
my
definition
of
a
service
mesh
come
come.
Come
ask
me
I'll,
be
happy
to
tell
you
it
was
on
the
slides.
In
case
you
saw.
A
We're
not
gonna
actually
discuss
the
advantages
and
disadvantages
of
different
deployment
models
rather
highlight
just
a
couple
to
say
you
can
different
service
meshes
have
a
number
of
different
options
for
how
they
get
deployed.
This
particular
diagram
is
probably
an
easy
starting
point.
Something
of
a
starting
with
an
ingress
gateway
is
something
of
a
gateway
to
service
meshes,
it's
a
good
place.
To
start,
it's
good
place
to
modernize
your
existing
monoliths.
A
Some
proxy
per
node,
which
is
more
or
less
an
agent
a
daemon
set,
if
you
will
an
agent
per
node
advantages,
disadvantages
small
print
here
the
slides
are
available
later.
The
point
of
what
we're
going
to
discuss
today
isn't
so
much
the
particulars
of
of
each
of
these.
Rather
to
point
you
to
resources
that
can
explain
them
better.
We
want
to
talk
about
get
back
to
the.
What
does
it
cost
to
run
a
service
mesh
which
one
do
I
use
and
how
do
I
get
started?
We
want
to
get
back
to
those
free
book
hit.
A
A
Maybe
because
I
couldn't
think
have
anything
more
clever,
a
lot
of
folks,
think
I'm,
saying
mastery
I
think,
but
well
we'll
get
that
corrected.
It's
a
multi
mesh
management
plan,
it's
a
multi
mesh
performance,
benchmark
and
playground.
It's
also
a
google
Summer
of
Code
project,
which
is
kind
of
kind
of
neat.
We've
got
some
bright
young
individuals
helping
us
on
this
project.
We
also
have
a
couple
of
PhD
students
at
the
University
of
Austin
Texas,
which
is,
which
is
why
I'm
wrong
was
awesome.
Texas,
that's
why
she
guys
that
wasn't
obvious
from
my
accent
earlier.
A
A
A
Rather
simple,
it
does
measure
he
does
require
that
you
have
a
kubernetes
cluster.
If
you
do
you'll
point
measure
e
at
it
and
deploy
adapters
service
meshes
specific
adapters
right
now,
this
project
about
4
months
old.
It
has
three
adapters
for
East,
EO,
linker,
D
and
console
two.
Others
are
being
contributed
by
other
projects.
The
network
service
mesh,
which
is
a
CNC
F
sandbox
project,
as
well
as
octarine,
which
is
another
security,
centric
service
mash.
Those
adapters
are
forthcoming,
while
they're
individually
contributed.
Interestingly,
if
you
were,
was
it
Monday?
Was
it
yesterday
yesterday?
A
Thank
you
very
good.
She
watched
Gabe
from
Microsoft
up
on
stage.
He
talked
about
SMI
service,
mash,
the
interface
I.
Imagine
many
of
us
are
still
puzzling
over
what
that
means
to
us
what
that
is
and
what
it
means
to
us
and
true
to
what
I'd
said
before
about
at
least
for
me
learning
by
relating
to
something
I
already
know,
or
at
least
I'm
already
familiar
with.
If
you
think
about
C
and
I,
the
container
network
interface
CSI,
the
container
storage
interface
CRI
I
contain
a
runtime
interface.
A
Smi,
unfortunately,
is
not
a
C.
Try
to
fix
that
service
mesh
interface,
even
without
explaining
what
it
is.
My
hope
is
that
sort
of
settles
in
a
little
bit
very
young
specification
does
have
partnership
and
support
from
a
number
of
the
players
in
the
ecosystem,
including
this
open-source
tool
mystery,
which
is
compatible
with
this
very
young
spec.
This
spec
is
well
I,
already
said
it
young
on
this.
A
There's
been
some
a
number
of
people
who
have
put
out
yep
I
feel
I'm,
just
I'm,
just
I'm
just
chewing
the
fat
until
you're
up
here
to
present
there's
a
number
of
recent
reports
actually
I'm,
one
that
was
just
done.
A
study
commissioned
by
linker
D
to
a
group
of
bright
folks
at
kin
Volk
who
who's
seen
this
the
the
thing
I'm
talking
about.
It
just
happened
this
last
weekend
right
in
time
for
cube
con
ok
anyway,
so
the
net
of
it
is.
There
are
a
number
of
people
putting
out.
A
A
One
of
when
I
say
that
if
you
want
to
test,
if
you
want
to
compare
how
does
one
serve
as
measure
perform
with
the
next,
we
probably
need
to
run
it
in
a
very
similarly
resourced
environment,
similar
traffic
it
used
to
be
inside
of
a
vacuum.
It
needs
to
be
probably
the
same.
App
same
rate,
same
sustain
rate
different.
You
know,
there
are
many
different
variables
to
account
for,
and
different
messages
are
gonna
perform
differently.
A
B
Hey
guys,
my
name
is
garish
when
he
introduced
I
wasn't
on
the
stage.
I
was
taking
a
seat
with
you
guys
so
I'm
I'm
gonna
actually
showcase
my
Sri
and
of
course
it's
it's
something
like
you
know,
risky
to
actually
doing
this
kind
of
an
environment,
because
you
know
we
I
have
a
lot
of
network
issues.
Obviously
it's
not
something
new
that
you
guys
don't
know
about.
So
hopefully,
I've
tried
my
best
to
pray
as
much
as
I
can
to
the
demo
gods
and
I've
come
here.
So
you
know
I'm
gonna,
give
it
a
try.
B
B
So
in
order
to
make
people's
life
easier,
I
mean,
if
you
just
want
to
try
my
Sri.
What
we
have
done
is.
We
have
actually
put
together
a
CLI,
a
simple
CLI,
which
I'm
going
to
actually
showcase
as
a
starting
point
which
actually
helps
you
get
started
with
my
Sri,
but
before
I
actually
start
typing
more,
can
you
guys
see
my
screen?
Well,
no,
okay,
good,
good,
good!
Okay!
B
B
So
there
is
a
start/stop
update
for
updating
the
images
logs
for
getting
the
logs
from
all
the
containers
and
a
status
to
actually
report
on
the
status
and
a
cleanup
like
you
know,
which
actually,
if,
for
example
like
you
know,
if
there
are
some
Network
issues
like
I,
went
downloading
the
configuration,
you
can
actually
run
cleanup
and
then
start
it
again
to
a
fresh
start,
but
in
my
case,
like
I,
already
have
the
configurations
ready.
Now.
This
is
a
very
simple
script.
B
B
I'll
just
go
back
to
my
screen
for
a
second
and
you
can
see
like
you
know,
those
are
all
the
services
that
are
running
in
memory
right
now,
so
you
have
a
question:
okay,
gotcha,
so
you
can
see
like
you
know,
Lee
already
introduced
the
architecture
of
mastery,
so
there
is
a
mystery
app
and,
of
course,
like
you
know,
we
have
all
the
adapters.
So
so,
right
now,
as
part
of
this
deployment,
you
can
see
the
memory
console
nessie
linker,
the
mesh
ratio,
so
those
are
the
adapters
that
are
actually
being
deployed
by
default.
B
The
other
one
is
a
playground
component,
which
is
where,
like
a
knight,
actually
will
let
you
play
with,
saw
this
meshes
and
the
third
one
is
actually
the
result
section.
So
we
also
give
you
the
capability
to
actually
purchase
the
results
and
compare
them
so
which
I'll
walk,
walk
you
guys
like
in
the
next
minute.
B
So
as
a
first
step,
you
know
I'm
now
connected
to
my
Cherie.
The
the
first
step,
like
you
know,
would
be
to
actually
configure
the
playground
components.
So,
obviously,
as
of
today,
like
you
know,
since
most
of
the
meshes
that
is
publicly
known
work
with
kubernetes,
we
have
started
to
base
it
off
of
kubernetes.
Eventually
we
we
could
like
in
a
move
away
from
it.
So
the
first
thing
is
I
pick
a
configuration
file.
You
know
it
can
be
a
queue
config
file,
but
it
has
to
be
a
queue
of
valid
cube,
config
file.
B
We
have
some
validations
in
place,
but
obviously
so,
once
a
valid
cube,
config
file
is
presented,
you
hit,
submit
it's
validated
and
once
it's
accepted
like
you
know,
you
will
actually
see
the
indication
that
the
cube
config
is
accepted.
The
next
thing
is
the
adapters.
Also,
to
note
is
that,
like
you
know,
there
is
an
asterisks
right
next
to
the
first
two
sections
like
notice,
the
kubernetes
and
the
adaptors
Griffin.
So
we
also
have
an
integration
with
Gravano
which
I'll
briefly
showcase,
but
the
first
two
are
mandatory.
B
A
All
right,
so
let
me
since
you're,
not
Mike.
Let
me
I'll
speak
to
this
just
to
get
both
of
these
hands
on
the
keyboard,
because
we
want
to
show
you
some
results
and
we
got
late
started
the
here.
You
can
see
that
each
of
the
adapters
have
some
common
functionality,
hopefully,
as
we
move
forward
with
surface
mesh
interface,
that
that
new
specification
that
will
help
with
that
common
functionality
right
now,
because
not
every
surface
mesh
has
made
the
same,
you
will
find
different
capabilities
in
each
adapter.
A
This
is
also
in
part
based
on
the
surface
messages
that
have
contributed
those
adapters,
they've
written
more
things
versus
the
next
one,
and
so
with
that,
one
of
the
things
that
people
are
able
to
do
is
apply
a
custom
configuration.
So
if
you
have
a
yamo
configuration
that
you'd
like
to
push
to
your
mesh,
you
could
do
that
through
the
tool.
A
Once
you've
gotten
your
you've
gone
through,
this
configuration
you've
connect
up.
Your
adapters,
you
can
measure,
will
facilitate
deploying
sample
apps
each
of
the
surface
meshes
come
with
come
with
their
own
sample,
app
to
show
off
what
a
mesh
doesn't
help.
You
understand
it
measure
e
helps
you
deploy
cross,
deploy
those
sample
apps
across
different
meshes,
so
you
could
have
a
kind
of
an
apples
to
apples
comparison
as
you
go
to
run
performance
tests.
B
B
You
can
pick
and
choose
the
panels
from
a
board
and
once
you
click
Add
like
you
can
actually
see
the
graph
on
our
charts
integrator
at
the
bottom
to
just
validate,
you
can
actually
expand
and
see
the
charts
and,
of
course,
like
we
have
we've
tried
to
give
the
same
behavior
like
you
know,
with
the
graph
on
of
filters.
Now,
once
you
have
configured
Ravana
once
you
go,
then
you
go
to
the
load
testing
page.
So
here
you
can
actually
perform
the
load
test
now,
given
the
time
I'm
not
going
to
show
it.
B
Actually
so,
essentially,
what
we're
trying
to
do
is
like
provide
a
single
pane
of
glass,
where
you
can
actually
see
the
client-side
metrics
and
the
server-side
metrics
and,
of
course,
let
you
know
also
the
last,
but
not
the
least,
is
the
results
section.
So
every
time
like
you
know,
a
load
test
is
run,
the
results
are
persisted
and
you
can
always
come
back
and
see
your
results
like,
for
example,
so
that
was
one
run
and
you
can
see
another
run
side-by-side
or
you
can
actually,
you
know,
use
some
comparison
features
we
have
in
place.
B
A
A
They've
come
out
with
some
initial
results,
there's
yet
to
be
an
academic
paper
published,
and
that's
really
what
needs
to
happen
here.
There's
a
lot
of
details
that
needed
that
folks
need
to
grok
and
take
into
context
in
this
particular
environment.
The
the
results
that
are
being
shown
are
from
a
single
node,
kubernetes
cluster
running
Ubuntu
running.
A
Is
it
24
gigs
of
ram
and
eight
core
single
CPU
system?
You
know,
there's
a
bunch
of
a
specification
and
details
that
that
will
be
dead.
Aren't
in
these
slides
that
will
be
shared
I.
Think
the
thing
to
take
away
here
is,
as
you
understand,
that
different
service
meshes
they're
sidecar,
the
service
proxy,
that
they
use.
As
you
go
to
understand
that
different
meshes
sidecars
take
up
more
memory.
Maybe
then
the
next
one
does.
Please
also
understand
that
they're
performing
different
functionality.
A
A
These
simple
slides,
don't
do
that
that
justice,
they
are
eye-catching
and
sure,
do
make
for
a
great
tweet.
Be
careful
as
you
do
that,
please,
the
full
report
and
spec
and
kind
of
the
the
publishing
of
that
will
happen.
This
is
taking
analyzing
the
sidecar
again,
the
the
data
plane
saying
how
much
CPU
is
burned
by
that
sidecar
in
these
environments
by
these
different.
So
these
this
is
the
identical
environment.
This
is
the
same
application.
This
is
the
same
request:
r/a
the
same
load
being
placed
on
each
of
these.
A
A
Well,
I
hesitate
to
do
this.
I'm
almost
compelled
to
one
of
the
things
that
needs
to
happen
here.
I
think
that
will
be
insightful
for
folks
is
again
where's.
The
third
time
I'm
gonna
make
this
analogy.
Is
that
I
learned
by
comparing
the
things
that
I
already
understand
if
I
think
I
understand,
container
orchestrators,
pretty
well
Nomad
mesas
marathon
kubernetes
swarm
I've
laid
them
out
on
a
spectrum
before
saying
some
are
easier
to
stand
up
than
the
others.
Some
are
more
powerful
than
the
others.
A
I
think
a
spectrum
like
that
and
something
of
a
quick
comparison
like
that
may
help.
Folks,
we
didn't
put
it
in
here,
I'm,
not
going
to
say
what
my
perspective
is
there
just
yet.
So,
hopefully
that
was
memory
yep.
This
is
CPU
again
context
behind
this
needs
to
be
shared.
Given
the
time
there
are
other
results
that
we've
begun
to
publish.
A
What
I
will
say
is
that,
yes,
we've
got
our,
we
might
be
at
time.
What
I'll
say
is.
Thank
you
to
our
contributors.
A
number
of
these
are
the
service
messages
projects
themselves.
Some
of
these
are
just
folks
who
need
an
outlet
to
express
their
passion.
So
thank
you
for
that.
The
last
thing
I
says,
if
you're
interested
in
the
results
when
they
already
or
if
you're
interested
in
just
learning
more
about
mushy
things,
layer,
5,
dot
IO
is
a
service
mission
community.
It's
intended
to
be
a
repository
of
information
about
mission.
A
Things
you'll
also
find
some
other
free
workshops
that
we've
given
the
workshops
we've
given
that
these
past
dr.
Khan
cube,
Khan's,
velocities
they're
there,
the
for
you
book,
there's
a
the
surface
mesh
landscape
and
then
measure
e
itself,
give
it
a
shot.
Arguably
I
think
there's.
Arguably
it's
maybe
the
easiest
way
to
try
out
a
surface
mesh.
A
Well,
that's
arguable!
What
isn't
arguable
is
it
is,
will
soon
be
the
thing.
The
tool
that
lets
you
try
out
the
most
surface
meshes
I,
guess
not
important,
but
please
come
see
us
if
you're
interested
in
more
the
details
do
subscribe.
The
students
that
are
working
on
this
will
would
love
your
feedback.