►
From YouTube: SIG - Performance and scale 2022-08-18
Description
Meeting Notes:
https://docs.google.com/document/d/1d_b2o05FfBG37VwlC2Z1ZArnT9-_AEJoQTe7iKaQZ6I/edit#heading=h.tybh
A
Good
all
right,
these
are
already
looking
much
better.
Okay,
let's
look
at
these,
so
I
forget,
if
we
saw
this
last
time
and
doesn't
say
test
if
we
had
a
chance
to
compare
some
of
the
numbers
for
the
the
200
density
test,
we
load
both
of.
A
I
don't
know
what
it
saw.
Let's
see,
what
else
we
have
all
this
stuff
looks.
Good,
I
mean
this
is
like
looks
like
it's
totally
random
like
has
no
correlation
to
the
basically.
What
I'm
looking
for
is
like
think
like.
If
we
see
any
correlation
between
this
number
and
this
number
over
time
or
across
jobs,
then
then
something
we
have
to
either
question
like
if
it
makes
sense
or
if
it
like
something,
we
need
a
threshold.
So
I
don't
I
don't
know
let's.
This
is
this
seems.
B
Do
you
mean
to
me
put
that
in
the
in
the
document,
because
I'm
very
sure,
like
next
time
we
look
at
it
we'll
forget.
A
Okay,
this
is
600,
if
you
want.
Let
me
just
link
to
the
job
here.
B
So
I
had
a
question
these
matrix
value
right
at
least
to
get
and
edge
are
these?
Is
this
like
a
static
list,
or
do
we
generate
this
at
runtime.
A
We
generate
this
after
the
testing,
but
so
this
is
these:
are
these
come
from
prometheus?
We
generate
them
after
the
test
and
we
we
do
it
over.
We
grab
the
data
over
the
duration
of
the
test
so
like
we
basically
run,
we
have
like
checkpoints,
we
get
like
a
point
in
time
and
we
start
a
point
in
time
on
the
end
and
then
we
we
capture
the
data
over
this
period
of
time.
B
A
Oh
it,
it
only
depends
on
whether,
like
during
this
test,
if
that
api
is
called
okay
yeah,
so
we're
only
like
right,
we're
only
creating
and
deleting
vmis.
A
Okay,
so
these
are
like
I
mean
these
look
like
they
have
a
positive
correlation,
but
this
one's
a
little
weird
I
this
looks
like
it's
kind
of
random.
I
mean
it's
like
that.
Doesn't
so
like
our
the
what
what
we
usually
do
is
we
usually
relate
it
to
the
to
the
crate.
Pods
account,
that's
what's
like
the
most
accurate,
so
I
mean
that's
this
this.
This
looks
well,
okay,
this
one,
maybe,
but
these
two
for
sure.
A
A
A
A
All
right
what
else?
Oh
actually
so,
oh
wait
here.
We
actually
have.
I
forgot
we,
don't
we
don't
have
thresholds
in
this
test.
Let
me
check
the
other
one.
Let
me
pull
the
other
one
as
a
reference,
so
we
where's
our.
A
So
these
half
thresholds-
I
think
I
remember
correctly.
Actually
these
might
already
be.
A
I
have
them,
let's
just
compare
okay,
so
we
have
update
patch
okay,
so
the
patch
steps
hold.
So
this
is
a
hundred,
oh
yeah,
let's
see
like
so
between
these
two.
So
this
is
what
do
I
say,
patch
virtual
machine
instances,
patch
first
message
machine
since
it's
about
a
2-1.
A
This
is
less
than
a
2-1
because
it
should
be
100
right,
121,
okay,
so
you
know
roughly
100,
so
we
up
to
242.
So
it's
a
two
to
one
relationship
and
we
actually
have
less
so
okay,
we
do
have
a
threshold
here.
So
actually,
let's
see
this
two
to
one
in
the.
A
A
Okay-
I
don't
know
about
this
one.
This
isn't
really
so
this
this
idea
of,
like
one
to
one,
doesn't
really
hold
well,
I
mean
it's
like
one
to
one
plus
a
constant
or
something.
It's
like
tell
their
rates
updating
points,
yeah,
okay,.
C
A
Do
you
have,
please
add
it
to
the
agenda?
I
have
it
if
you
don't
here
I'll
share
it
in
the.
A
A
A
A
A
No,
no
I'm
wrong.
We
are
over
so
it's
over
ten
to
one
right
in
my
mouth
right.
So
it's
like
it's.
It
should
be
six
one.
One,
zero
yeah
we're
six,
two
four
zero,
so
we
are
over
so
we're
actually
over
threshold
here,
that's
interesting!
So
I
guess
what's
happening
is
like
so
looking
at
the
this
is
this
is
interesting
data,
so
we
have
on
a
smaller
scale
of
a
hundred
we're
getting.
We
don't
have
to
have
as
many
updates
as
the
larger
scale.
A
We
do
so
I'm
guessing
we're
going
to
get
so
we're
here.
This
is
like
a
reflection
of
some
like
some
400
errors
that
were
hitting
some
conflicts,
some
405s
or
something.
A
That's
interesting,
so
it's
roughly,
I
mean
so
it
this
was
okay,
so
the
ratio
is
about
six
six
out
of
ten
or
six
times
the
is.
There
was
a
ratio
so
okay,
so
it
like
slowly
climbs
to
about
this-
is
about
it's
like
10
and
a
half
here
or
10,
and
a
quarter
or
something
maybe
less
10.1
or
something
it's
interesting.
A
A
A
These
are
good
to
add
all
right.
Let's
see
what
else
we
have.
Okay,
the
vmi
creation
times
all
right:
let's
do
a
little
comparison
here
of
our
smaller
scale:
okay,
much
faster,
okay,
all
right
many
times!
This
is
this
is
interesting.
Okay,
so
we're
at
600
and
we're.
So
the
idea
with
this
test
is
we're
creating
them.
I
don't
think,
there's
any
weight.
I
think
we
create
them
like
as
fast
as
we
possibly
can
so.
The
this
is
how
long
it's
going
to
take
for
these
600.
A
A
A
B
Right
do
we
have
any
data
on
so
I
know
that,
after
creation
to
running
there
are
like
a
bunch
of
kubernetes
things
that
happen
like
binding
pvc,
and
you
know
picking
the
right,
node
and
stuff
like
that.
So
is
there
any
kind
of
data
point
on
where
that
time
is
being
spent
while
going
to
running
state.
A
Yeah,
it's
a
good
question.
Well,
this
is
where
it
would
be
good
to
gather
all
of
the
the
phase
transitions
to
see
like
to
get
a
measurement
of
like
exactly
the
breakdown
was
what
the
breakdown
was.
I
think
maybe
we
can
get.
I
think,
there's
a
dashboard
and
marcelo
has
the
dashboard
for
this
somewhere
and
found
out
that
I
think
it's
public
I'll
have
to
dig
it
up,
or
maybe
I
have
it
is
this
it.
I
think
this
is
it.
A
And
we
should
be
able
to
see
the
different
these
transitions.
Okay,
so
let
me
see
so
this
is
from
what
is
this?
I
think
it
was
just
well.
I
can't
see
the
thing
8
17.,
so
it's
just
yesterday
or
just
yeah
yeah
just
yesterday,
so
I
think
it's
this
one.
So
this
looks
like
start
recreation
time.
Okay,
so
let's
zoom
in
here
looks
like
a
big
number.
A
A
A
Okay,
I
don't
know
how
to
look
at
how
the
this
is
generated,
but
the
what
I
believe
it
is
it's
the
time
it
took
to
transition
from
scheduling
to
scheduled
okay,
but
these
are
telling
these
are
too
different,
though,
like
so
this
is,
I
don't
know
what
this
one
is.
This
might
be
the
time
since
the
vmi
was
created.
I
think
so
it
running.
A
Does
that
make
sense,
because
it
sort
of
should
be
additive
here,
right,
yeah,
so
green
should
be
on
top.
So
what
about
the
orange
yeah?
So
then
this
is
okay,
so
this
is
additive,
so
okay,
that
that
makes
sense.
So
this
is
the
time
since
it
was
scheduled
all
right.
This
is
a
good.
It's
actually
a
really
good
depiction,
zoom
in
on
this
one,
so
the
okay,
so
vmi
was
created,
and
the
time
since
then
was
to
I
don't
know
if
it's
to
reach
scheduled
phase
or
if
it's
like
to
transition
out
of
scheduled
phase.
B
B
Like
it
seems
to
me
that
at
five
minutes
like
if
you
look
at
the
middle
line
like
right
with
when
there
are
two
running
green
ones,
up
to
five
minutes,
it's
in
scheduling
like
scheduled
state
and
then
from
there
on
up
to
six
minutes,
it's
in
running
state
that,
like
that's,
how
I
am
interpreting,
I
don't
know
if
I'm
right.
A
A
B
A
A
A
Okay,
I'm
not
sure
I
don't
know,
I
don't
know
what's
supposed
to
be
in
here-
maybe
it'll
load
in
a
few
minutes.
Okay,
so
well!
I
I
think,
though,
what
would
be
so
this
is
good.
I
mean
this
is
like
this
is
kind
of
good.
It
goes
a
long
lines
of
what
you
know.
Marcelo's
observed
where,
like
as
we
increase
the
number
of
of
create
requests,
which
it
looks
like
that
happens
until
let's
see
when's
the
last
scheduling.
Okay
here
it
is
so
50
to
56..
A
So
it
looks
like
over
five
minutes
as
we
do
the
creating
yeah
okay
and
then
we
stop
and
it
basically
takes
well.
Let
me
get
rid
of
these,
so
I
can
let's
just
do
running
okay,
so
then
it
stops
and
then
we
go
the
remaining.
I
wish
I
had
a
count.
A
A
Okay-
well,
I
I
don't
know,
let's
we'll
have
to,
I
think,
like
the
the
thing
from
this
is
like
I'm
really
interested
to
see
like
like
these
numbers.
The
yellows
here
is
like
the
thing
that
we
want
to
explore
more.
I
think
it's
like
it's
somewhere
in
kubernetes
in
the
way
that
it's
going
through
and
scheduling
these
things
like.
A
Maybe
we
can
find
some
more
data
some
way
we
can
make
something
more
efficient
there
or
or
something
because
yeah,
because
our
this
is
just
it
seems
like
it's
just
going
to
keep
increasing
and
as
we
increase
the
numbers
number
of
vmi's
just
increases.
I
I
I
don't
think
this
happens
with
pods.
I
don't
know
if
anyone's
familiar,
but
I
I
don't
know
I
don't
think
so.
A
B
Yeah
ryan,
can
you
talk
a
little
bit
about
what
the
the
cluster
resources
look
like,
where
this
test
is
running.
A
A
B
Like
I'm,
like
I'm
talking
in
context
of
we
were
discussing
somewhere,
what
would
be
like
good
things
to
share
when,
when
going
to
v1,
this
comes
to
my
mind,.
A
Yeah
the
thing
is,
though,
with
with
this
this
count
right.
This
is
a
hundred.
This
is
150.
like
I
would
still
expect.
I
mean
we
should
be
able
to
handle
this
right
like
we
shouldn't
be
like
it
is,
it
is
clearly
there's
a
relationship
between
the
female
account
and
the
performance.
B
A
A
Yeah,
okay
well,
but
we'll
just
leave
it
here
I'll.
Let
me
find
this
thing
and
let's
do
a
study
on
this,
because
I
I
I
don't
know
this-
is
this
isn't
well
just
out?
Let's
just
outlaw
that,
let's
just
rule
out
that
idea
and
then
we
can,
I
think
the
next
step
after
that
is
like
we
completely
see
that
there's
a
slowdown
in
the
scheduling
time.
A
We're
gonna
need
to
do
some
further
investigation
either.
Metrics
like
I
think
the
best
thing
to
go
with
this
is
that
I
think
we're
gonna
need
some
sort
of
metrics
that
appear
just
like
this
in
a
few
areas
that
are
targeted
like
what
we
think
are
slow
and
I
think
maybe
pvcs
or
something
with
kubernetes
or
signing
something
with
the
networks,
signing
networks
or
something
we
should
try
and
find
and
attach
those
things.
B
Yeah
quick
question:
so
if
you
go
to
the
dashboard
I
see
in
on
the
top
that
there
is
a
so
after
the
general
it's
keyboard.
Is
there
a
section
that
is
related
to
kubernetes,
like
on
on
the
top
top
bar.
A
Oh,
this
is
the
last
hour
where's
my
I
lost
my
time
frame,
so
let
me
see.
B
Yeah
yeah,
I
was
hoping
if
we
had
like
pvc
or
some
kind
of
other
things
in
there.
B
Oh
no,
this
this
is
just
space
utilization.
It's
yes!
Maybe
if
you
can
go
to
controller
manager
that
that's
what
processes
the.
B
A
Okay,
well
I'll
try
to
explore
yeah,
I
don't
I
don't
know
yeah.
I
think
we
can
do
some.
We
can
do
some
some
looking
at
it.
I
don't
know
this
is
yeah
and
the
thing
that
I
would
say,
though,
like
marcel
has
said
previously
like
that
that
this
was
expected
and,
like
we've,
we
sort
of
come
across
this
in
a
few
different
ways
like
when
he
talked
about.
A
You
know
increasing
the
amount
of
threats
and
how
to
improve
performance,
so
they're
it's
sort
of
expected,
but
I
mean
still,
though
we
don't
want
to
well.
Actually
maybe
that's
the
question
like.
I
wonder
if
this
has
the
marcelo's
change
in
the
number
of
I
for
what
he
changed.
It
was.
The,
I
think,
was
the
rate
limiting
on
the
client.
A
The
number
of
active
or
open
requests,
I
think,
is
what
he
increased.
I
I
don't
know
what
version
of
kubert
this
is
actually
running
and
if
it
has
those
changes,
because
I
don't
think
this
gets
redeployed,
so
perhaps
it's
missing
that
and
maybe
that's
what
we're
observing
is
the
same
phenomenon
that
he
did.
A
Okay,
all
right
another
open
question:
all
right:
I'm
gonna
ask
this
one
to
marcelo
I'll,
send
him
a
message
and
see
if
he
see
if
we
can,
I
can
shed
some
light
on
this.
Okay.
I
think
so
I
think
we're
done
on
looking
at
these.
I
think
this
is
good.
This
is
doing
remaining
stable.
A
This
is
really
good
information.
I
think
for
a
comparison,
and,
oh
so,
here's,
okay,
the
100
density
test,
yeah,
I'm
still
within
the
or
much
closer
to
the
the
other
density
test.
Interesting
yeah.
I
think
this
is
the
one
that's
going
to
be
interesting,
we'll
look
at
the
see!
These
are
the
numbers
that
I
think
we
want
to
bring
down,
but
we
should
be
able
to.
I
think,
since
we
have
these
two,
this
will
actually
be
valuable
to
see.
A
Okay,
all
right!
That
was
good!
All
right.
Let's,
let's
go
to
the
next
topic,
so
andre
you
want
to
talk
about
yeah.
You
got
a
topic.
C
We
are
creating
a
stress
test
on
our
solution
for
hundred
thousand
concrete
windows,
11
rdp
sessions.
We
are
able
to
do
the
over
the
creation
of
the
vms
and
everything.
This
is
fine.
C
To
see
the
remote
desktop
session
actually
because
we
expose
them
over
in
web
browser,
is
there
any
tool
that
came
to
mind
to
be
able
to
open
hundred
thousand
web
browsers
one
for
each
session
and
see
every
rdp
session?
C
A
A
There
might
be
something
in
the
test
that
that
do
this.
I
think
I've
seen
this
before.
Like
do
you,
do
you
actually
need
to
open
a
browser
in
a
session,
or
can
you
like
just
do
it
over,
but.
A
Okay,
let
me
see
what
so,
where
it
look
is
in
this
in
the
rdp
tests
that
are
in.
A
A
Yeah,
I'm
not
sure
where
they
are
the
like
when
that's
what
I
would
look
for,
though,
because.
C
Yeah
well,
I
passed
this
video
to
my
developers
anyway.
If
you
can
send
me
later,
this
can
be
very
helpful.
A
Yeah
I
yeah,
so
what
I
was
hoping
for
is
like
I,
I
pretty
sure,
there's
some
code
that
we
have
that
in
the
test.
That
is
like
doing
something
interacting
with
rdp
or
and
just
like
doing
some
interactive
testing
there.
If
not,
then
I
mean
we
might
just
be
out
of
luck,
but
that
that's
like
at
least
how
like
some
of
the
like,
there's
already
some
like
like
there.
I
know
there's
some
code
like
that
interacts
with,
like
a
serial
console
like
we'll,
do
some
interactive
stuff
on
the
on
the
vm.
A
But
that's
what
I
was
asking
you
about
the
browser
like
if
it
needs
to
be
actually
going
and
simulating
something
in
the
browser
if
it
could
just
be
going
through
using
a
client
and
and
accessing
it,
and.
C
What
we
are
trying
to
build?
Yeah,
okay,
okay,
okay,
our
solution,
but
if
someone
already
have
done
something
can
be
very
helpful.
Okay,
then
we
don't
need
to
start
from
scratch.
A
C
C
If
you
go
here,
there
is
this
file
here.
C
C
C
Can
you
explain
a
little
bit
about
that?
The
difference
why
the
linux
version
is
not
on
on
14
and
they
are
working,
unfortunately,
for
for,
for
instance,
and
vmware
and
windows
hypervisor.
A
I
I
don't
know,
I
don't
know
why
it's
what
the
I
don't
know
what
the
the
difference
like,
I
don't
know
like
why?
Why
it's
on?
Why
it's
not
on
14.,
I
I
don't
know.
A
There
should
be,
there
should
be
a
way
to
reach
out
through
the
like
the
like
there.
There
are
like
the
customer
facing
people
who
are
usually
interfaced
with
that
would
have
this
information.
A
I
don't
know
I
like
they're.
They
usually
publish
this
stuff
somewhere.
I
don't
know
where
it
is
I've.
Never
I've
never
done
it
before
I've
never
reached
out
to
some
of
those
folks.
So
I
don't
know
one
of
the
usual
channels
that
people
do
to
reach
out
to
the
I
don't
know
whatever
like
the
whatever,
like,
whatever
you
do
for
like
customer
support,
I
think
is
where
you
want
to
go
through
and
they
yeah
where
you
look
for.
C
A
Yeah,
I
I
forwarded
your
your
like
your
ass
on
to
someone
that
I
thought
would
help,
but
I
haven't,
I
didn't
hear
back
so
I
I
don't
know
like
it's,
it
might
not
be.
It
might
not
be
helpful
for
like
because
I
don't
know
I'm
not.
I
don't
interface
a
lot
with
some
of
those,
the
the
customer
reps.
So
it's
like
it's
hard
for
me
to
say
like
if
I'm
talking
to
the
right
people
or
not,
I
don't
know
so.
That's
why
I
would
say:
go
through
the
you
know.
Whatever
your
vendor
is.
C
Yes,
the
vendor
download
the
drivers
and
enable
them
on
on
them
websites
to
be
downloaded,
but
there
there
is
always
a
let's
say,
sometimes
big
gap.
Six
months
between
the
latest
drivers,
I
can
download
from
nvidia
and
the
drivers
that
I'm
I'm
able
to
download
from
from
google.
For
you
understand,
and
sometimes
I
already
find
huge
performancy
issues
without
the
latest
and
greatest
you
understand.
A
C
Anyway,
thank
you
so
much
for
your
time.
If
anyone
has
an
idea,
I
appreciate
that,
before
the
stress
test
on
on
100
000.
A
A
Yeah
look
look
in
here,
and
this
is
where
the
functional
tests
are.
I
don't
know
where,
specifically
like
all
the
serial
console
tests
are
in
here,
like
it'll.
Take
me
a
few
minutes
to
find
it,
but
it's
in
there
is
where
I
I
know
there
is
because
I've
seen
them
so
like
in
here.
There's
there'll
be
some.
A
You
know,
look
for.
I
don't
know
something
with
cloud
in
it
or
something.
Maybe
you
just
need
to
search
with
your
serial
console
and
or
vnc,
or
something
and
and
find
something
then
give
you
some
beautiful
use
of
resources
in
there.
Wonderful.
C
It's
a
nice,
it's
better
than
start
from
scratch.
A
C
A
C
Like
I
have
a
request
to
you,
andre
like
if
you,
if
you
are
doing
this
this
test,
could
you
maybe
share
like
some
results
afterwards,
like
I'm,
not
sure
what?
What
exactly
is?
Is
this
test
etc?
But
I
feel
it
would
be
nice
to
to
get
some
some
input
on
on
maybe
some
metrics
that
that
you,
that
you
gather
maybe
something
something
we
could
improve
later
as
well.
C
Hundred
thousand
concurrent
users
mean
10
cluster
with
10
000
users
in
each
cluster.
For
you
understand,
okay,
probably
we
we
have
when
we
have
the
results.
We
say
something
here
or
on
the
regular
on
wednesday's
meetings
from
covert
okay,
okay,
yeah
yeah.
That
would
be
great.
Thank
you.
Wonderful,
no
problem
at
all.
A
Cool
okay,
all
right
thanks,
andre!
Okay,
let's
see
I,
those
are
the
only
items
we
have
today.
So
there's
nothing
else
and
that'll
minutes
early.
C
A
Oh,
yes,
sure
so
what
happens
is
here
I'll,
just
stop
sharing.
So
what
happened.