►
From YouTube: 2019-12-05 :: Ceph Performance Meeting
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
All
right
so
pull
requests
this
week.
Let's
see
I,
don't
I
didn't
see
anything
that
was
explicitly
new,
but
with
Thanksgiving.
This
not
super
surprising,
I,
guess
a
lot
of
stuff
close
though
so,
let's
see,
Patrick
had
a
PR
do
released
reheat
pages
after
trim
in
the
MDS,
similar
to
what
we
do
in
both
the
OSD
and
in
the
Mons.
A
A
A
A
So
that
was
a
the
one
that
we
talked
about
previously,
where
it
was
a
part
of
a
much
bigger
PR.
So
next
enhanced
OSD
Numa
affinity,
compatibility
that
merged
I,
don't
remember
everything
that
did
but
I
think
part
of
it
was
like
looking
at
bonding,
setups
and
properly
like
recognizing.
If
a
bond
was
on
the
same
Numa
notice,
the
rest
of
the
OSD
then
actually
sucked
the
affinities
right
and
everything
previously
didn't
do
that.
So
it
was
that,
amongst
other
things,
I
think
that
just
made
the
Numa
processing
code
a
little
bit
harder.
A
Let's
see,
oh
the
delete
range
stuff
merged,
so
that's
really
good.
That
was
a
PR
I
submitted,
but
was
really
actually
based
on
another
PR
that
had
previously
been
merged,
but
then
sort
of
got
overwritten
later
by
some
other
code.
So
that's
good.
We
need
to
be
very.
Very
careful
bug.
Delete
range,
it
turns
out
is
still
quite
impactful
when
it's
not
used
very
carefully.
A
See,
oh,
the
so
the
the
oh,
no
pinning
code
that
I
wrote
merged
with
a
fix
from
Josh
and
it
looked
like
it
was
working.
Fine.
It's
been
working
bastard
for
like
a
week
here
we
can
have
and
then
sage
hit
a
segfault
in
that
code,
but
unfortunately
didn't
have
a
core
dump.
So
I
think
the
plan
right
now
is
to
just
wait
and
see
if
we
hit
another
one
and
hopefully
get
a
core
dump.
A
If
one
of
those
is
yours,
I
guess
you
know,
go
back
and
reopen
if
you
care
about
it,
a
couple
that
were
updated,
the
I/o,
Ewing
io
engine
I,
think
Kiefer
ran
them
through
testing
and
it
had
some
failures
so
that
has
not
gotten
merged
things
updated.
This
collections
list
+
PR
of
his
that
was
kind
of
old
before
that,
but
he
needed
something.
I
didn't
look
closely
at
what
it
was.
A
My
refactor
do
OST
ops,
there's
been
no
new
work
on
that.
Only
I'm
trying
to
keep
it
in
the
back
of
my
mind
here.
I
think
the
way
I
want
to
approach
that
is
break
it
up
into
smaller
PRS.
It
was
really
bacon.
It
was
failing
in
kind
of
unexpected
ways
in
just
a
couple
of
areas.
I
think
so
probably
most
of
it
could
still
go
in
and
and
then
get
in
and
not
not
become
outdated.
A
Let's
see
this
PR
from
Russia
and
paying
needs
a
rebase
now
I
think
was
buggy.
It
wasn't
passing
testing.
So
then
these
more
work-
or
maybe
it
don't
know
I'm
sorry
I,
got
updated
but
hasn't
been
retested
I,
think
I
know
it
needs
another
rebase,
that's
it
and
then
Igor
is
blue,
store
version,
2
framework
or
intelligent
DB
space
usage
that
got
updated,
but
then
was
failing
testing.
So
it
means
looked
at
again
as
well.
A
A
A
C
A
C
D
E
D
A
Let's
see
otherwise,
we've
got
our
tuning
of
the
NBS
cache
memory
limit.
That's
gotten
more
update!
Well,
no,
as
and
I
guess
at
one
point
he
did
no
movement
in
the
last
two
weeks,
though,
and
there's
just
other
various
things
here,
they're
still
on
the
list,
but
haven't
been
a
bid
recently
so
uhm.
Alright,
that's
it
for
PRS
anything
I
missed
for
anybody,
hey.
G
A
A
F
A
What
do
you,
what
do
you
think
of
that
basically
saying
the
OSD
memory
target
below
to
give
Isis
typically
not
recommended
or
give
Isis
the
current
default?
The
kind
of
balance
memory
requirements,
an
OSD
performance
and
then
saying
it's
higher
than
four
gigabytes
may
improve
performance
when
there
are
many
small
objects
or
or
many
objects
who
are
you
know,
very
large
data
sets
being
processed?
Is
that
yeah
Dean
might
a
reasonable
statement
from.
A
H
A
A
H
I
The
octane,
persistent
Dems
or
whatever
oh
I,
have
heard
that
others
out
there,
like
other
customers
and
we've
been
talking
to
that.
They
have
tried
using
the
the
dims
or
rocks
TV
wall
and
DB,
but
their
initial
test
results
showed
that
yeah
basically
didn't
have
any
performance
benefits,
but
they
didn't
have
much.
H
K
H
C
C
Stories,
or
is
the
right
approach
or
a
so.
There
was
the
first
thing.
I
do
and
I
saw
brick
pleased
with
it,
but
but
I
good.
This
conversation
becoming
a
rathole
there's
a
lot
of
it's
a
longer
trivial.
Due
to
this,
the
same
thing
that
Marcus
mission
is
neither
sufficient
is
that
is,
is
also
also
why
we
can't
really
move
the
light
laterally
in
in
the
right
path
at
all
in
blue
store.
C
If
listeners
I
would
love
to
propose
someone
or
people
work
on
making
that
possible,
but
I
don't
hesitate,
a
big
lift,
I
mean
being
able
to
try
out
others
other
other
other
right.
I
just
have
a
simpler
right
path
for
rocks
to
be
but
or
alter
my
ways
to
try
other
things.
I
mean
these
would
be
effectively
new
stores,
but
but
the
compatible
with
or
with
the
traditional
OA
c
code
base.
No,
this
pleased
hardened.
A
J
D
H
J
F
L
That's
true,
yeah
excuse
me,
there's
two
kinds,
of
course
of
envied
imps
one
is
like
the
flash-based,
which
is
where
you're
right
cycles
have
come
into
play
and
then
the
other
one
is
what's
called
env
Tim
n,
where
it's
basically
just
like
a
regular
memory
module,
but
it's
it's.
Basically,
you
can
put
it
backs
up
to
flash
if
this
power
failure,
and
that
would
be
that
kind.
You.
L
F
H
L
L
L
L
C
H
Let's,
let's,
let's
I
saw
I'm
putting
together
a
document
literally
started
five
minutes
ago
and
realized
I
was
late
for
this
meeting.
Where
you
know,
as
we
talked
a
couple
of
weeks
ago,
Matt,
you
know,
put
other
sort
of
a
working
group
to
figure
out
how
did
to
take
advantage
of
stories.
You
write
had
I've
had
side
conversations,
Tom
Coughlin's
with
segi
from
Compton
steam.
You
know,
let's
I'm,
just
gonna
share
it
out
as
widely
as
possible
and,
let's
start
throwing
stuff
at
it
and
figure
out
what
we
could
do
so.
C
A
Okay,
Johnny
back
to
you
guys
what
are
the
things
that
we
need
to
do
or
worry
about
if
we
were
using,
not
the
the
Intel
ones
necessarily
or
you
know
some
other
constraints
that
we
need
to
think
about,
but
and
the
envied
inside,
whereas
you
know
looking
like
traditional
memory
with
you
know,
battery
back
to
flash.
Is
there
anything
there
that
that
we
need
to
think
critically
about
compared
to
how
we
do
it
now.
J
J
L
A
A
J
A
F
I
would
encourage
you
guys
to
go
look
at
the
persistent
memory
api's
that
have
come
out
of
sneer.
They
basically
address
how
things
are
allocated.
There
is
still
transactional
issues,
because
updates
can
still
be
stuck
in
the
cache
at
a
time
of
a
power
failure,
and
you
can
have
inconsistent
structure
updates
and
things
of
that
nature
and
the
persistent
memory
api's
that
are
coming
out
of
Adam's
new
address.
All
of
that
right.
So
you
might
want
to
take
a
look
at
that
stuff.
F
There
are,
and
there
they're
available
in
Windows
and
Linux
they're
available
widely
available
because
they
had
to
do
that
to
get
adoption
started
and
regardless,
if
it's
an
NV,
NVRAM
type
of
thing,
with
DDR,
backed
with
super
capacitors
or
even
you
know
the
the
more
solid-state
stuff
from
Intel.
If
they're
in
the
DDR
path
on
the
DDR
bus,
they
have
the
same
issues
with
regard
to
flushing
and
though
Ruby.
L
J
F
C
Well,
this
came
up
and
the
conversation
in
spring
and
until
it
was
just
not
with
librarian
even
inch,
but
you
had
to
your
point
someone
you
value
stories.
I
mean
Intel
claimed
that
they
had
Ellen
to
be
running
on
PLP
mm
and
the
results
were
herb
and
concurs.
Just
talk
to
them
further
about
it
and
then
disappeared.
A
C
Okay
but
but
they've
been
you
know,
but
that
it's
it's,
but
it's
a
double
buffer
map
strategy
right,
no
right
ahead
log,
but
P
men
that
might
be
great
yeah
I.
B
H
Saying
P
D
cake
contains
p
m--
MKV
yeah
at
the
bottom
of
the
page.
Sure.
A
E
J
A
J
L
A
A
J
H
No
there's
some
money's,
no
object.
Is
there
another
reason
not
to
random
renderings
there?
They
should
be
just
as
indoor
you
know.
It
was
just
much
endorses
with
normal
DRM
right
the
battery
bad
things
right.
That's
what
that's!
What
I've?
That's?
What
I'm
in
my
head,
differentiating
between
P
James
and
every
dims
money.
C
C
I
L
H
L
H
L
L
H
H
L
H
Just
don't
want
to
make
sure
let's
not
make
assumptions
I
mean
I
did
exercises
more
than
10
years
ago,
where
I
could
hold
up
to
64
gigabyte
memory
boards,
with
literally
two
like
double
A's,
so
long
enough
to
push
them
push
the
memory
out
the
flash
quote
least
according
to
my
Rick
from
tick,
which
may
have
been
right
or
wrong.
But
you
know
I
was
you
know
doing
a
lot
of
that
kind
of
stuff
back
then
so
I'm.
H
J
C
A
J
C
A
J
J
L
L
J
L
H
A
L
A
L
L
Well,
maybe
maybe
you
know
you
need
to
come
back
to
this
subsequent
week,
but
I
just
was
just
trying
to
get
the
if
this
helped
the
discussion
at
all.
At
some
point,
I'd
like
to
talk
about
a
different
topic
like
we're
at
a
time
almost
so
bring
that
up
yet
listen.
I
L
Minutes
all
right
so
well,
but
the
topic
I
wanted
to
bring
up
and
maybe
it'll
be
a
short
one
is
Alex.
Calhoun
has
been
doing
some
testing
with
the
combination
of
the
auto
scale,
PG,
autoscaler
and
balancer,
and
he
got
some
really
I
thought
excellent
results
with
using
both
where
he
was
getting
damn
near
perfect,
balancing
across
OS
DS
and
we've
been
having
this
email
discussion
about.
Well,
how
important
is
that
for
performance,
and
we
need
to
do
more
measurements
of
that,
or
he
may
have
already
done
it.
L
I've
got
it
just
hasn't,
presented
the
data,
but
the
main
thing
I
was
concerned
about,
is
you
know
in
his
tests
he's
starting
out
with
everything
empty,
the
cluster
is
empty
and
then
he
sets
it
up
the
way
he
wants
it,
though.
The
the
PG
autoscaler
basically
can
instantaneously
come
to
the
right
AG
numbers,
because
you
tell
it
with
the
target
size
ratio
of
what
you
want,
but
in
the
real
world
you're
gonna
I
already
have
data
in
the
system.
I
J
Yes,
the
bouncer
will
only
make
a
run
when
the
cluster
is
healthy
and
it
only
makes
us
very
small
changes
at
a
time
there.
So
they're
kind
of
the
ideas
that
they've
worked
together
in
that
way,
so
that
they're,
if
they're
catching
PD,
counts
the
you're
going
through
some
recovery,
and
so
the
balancer
wouldn't
run
at
that
point.
So.
L
G
You
could
always
go
into
a
Chand
right
like
once.
The
autoscaler
stops
working
and
you
have
the
balancer
kicking
in.
That
might
not
be
the
best
scenario,
but
with
the
way
the
origins
the
default
configuration
for
the
the
balancer
is
right.
Now
it
there
is
a
tunable
when
it
starts
making
those
changes.
It
makes
like
ten
changes
every
few
seconds
if
we
change
those
to
not
happen
as
frequently,
but
then
it
might
not
be
as
big
of
a
problem.
A
A
The
PG
count
like
you
know
bit
by
bit
at
a
time
and
maybe
you're
growing
it
fast
enough
there
that
the
the
balancer
doing
a
whole
lot
of
work,
because
your
your
do
it
again
again
quickly
enough
that
it
doesn't
do
a
lot.
What
what,
when
you
get
out
to
having
enough
data
in
it
that
you
have
a
big,
a
fairly
big
jump
between
them?
You
could
still
have
the
balance
or
like
filling
in
that
gap
when
it's
still
healthy,
doing
a
bunch
of
balancing
work,
but
you're
still
growing.
A
So
then
you
just
kind
of
like
you
know,
do
a
scaling
event
which
blows
everything
away.
It
changes
everything
around.
Then
you
balance
for
a
while.
Then
you
blow
it
all
the
way
you
know
scale
hitter.
You
know
again
so
now
you've
got
a
totally
new
one.
You
rebalance
do
it
for
a
while.
You
know
you
could
end
up
like
constantly
in
a
state
of
either.
J
Think
it's
kind
of
a
they're
gonna
dependent
triggers,
though
I'm
the
PK
autoscaler
is
a
based
on
the
capacity
of
the
cluster
and
the
balancer
is
based
on
the
PD
distribution.
So
if
your
capacity
is
stable,
that
I
bounce
or
what
ya
scalar
will
stop
changing
and
in
the
balance
were
a
lot
of
time
to
stabilize
things
in
the
distribution
wise.
A
What,
if
you
have
like
a
filled
workload
where
you're,
just
like
filling
in
a
bunch
of
stuff
like
you're,
you're,
constantly
writing,
like
you
know,
logs
or
something
into
into
the
storage,
won't
that
continuously
trigger
the
autoscaler,
as
you
add
more
stuff,
and
continue
to
fill
up
the
disk?
Yes,.
J
But
it's
a
it's:
it's
not
gonna
happen
like
continuously
it's
a
fairly
conservative
and
when
it
makes
changes.
So
it's
only
when
I
think
it's
like
a
three
times
threshold
errs
and
clean.
So
when
you
have
like
three
times
as
much
data
as
what
make
sense
for
the
current
PG
count
is
when
it
makes
it
makes
a
big
change
in
the
increases,
then
repeat,
ease.
A
A
J
Is
if
you
know
that
you're
gonna
be
using
a
lot
of
data
for
this
particular
pool,
you
can
use
the
target,
a
ratio
for
that
pool
that
the
balancer
doesn't
have
to
keep
changing
as
un-quit
data.
It
just
goes
to
write
directly
to
that
that
and
and
your
hope.
A
L
J
It's
kind
of
tricky,
though
cuz
and
when
you're
you
don't
necessarily
know
account
which
rules
like
what
at
what
point
you
decide:
okay,
I'm
gonna
use,
but
I'll
put
a
bunch
of
kikis
on
these
pools
like
when
you're,
initially
creating
the
pool.
You
probably
don't
want
to
say
this
pool
is
gonna,
be
using
all
of
my
IP
d
space.
Why
not
be
then
creating
a
bunch
of
other
poles
directly
after
that.
A
L
H
J
A
And
in
that
situation,
to
where,
if
you
already
have
a
bunch
of
data
in
the
cluster
and
a
bunch
of
peas
in
the
cluster,
that's
exactly
the
kind
of
memory
or
PG
pressure.
That
would
tell
you
that
your
new
pool,
maybe
shouldn't
be
very
big,
but
you
shouldn't
necessarily
give
it
lots
of
you.
Jeez
yeah,.
J
C
All
depends
right
if
you
turn
on
multi
site
and
other
things
to
do
logging
you
get
tons
of
stuff
in
the
log
pool,
etc.
Sure,
okay,
so
you
can't
really
predict,
there's
also
an
evils
and
stupid
pool,
but
called
multi
fault
apart.
I'm
sorry
was
it
called
extra
data
pool?
That's
that's
not
as
much
of
multiplier
stuff,
and
so,
if
you
do,
if
you
don't
do
multiply
upload
workload,
you
won't
see
anything
in
there
about
there
less.
You
will.
J
A
J
H
The
interest
of
time
and
Matt's
question
can
I
ask
a
non
sequitur
question
so,
with
with
straight
replicas
like
OSDs,
are
doing
well,
client,
ops
and
replica
ops
like
you're,
doing
client
ops
for
things
your
primary
on
and
replicas
for
things
that
you're
not
primary
on
brightest
in
but
easy
doesn't
work
that
way.
Does
it
or
does
it
still
kind.
H
J
H
C
H
C
J
H
Might
like
a
let's
say
you
were
just
kidding:
let's
say
you
had
no
client
I/o
and
you're
an
OSD,
that's
but
you're.
Actually
you
know
you're
getting
hammered
like
you're
everybody's
replica.
Is
that
possible?
You
know
if
that
happened,
then
you
know,
there's
still
would
kind
of
not
work.
I
mean
I.
Think
it's
worth
trying
and
we're
thinking
about,
but
I'm
I
think
you
know
it
yet
either.
C
But
I
think
you
might
spray,
but
I
think
but
I
think
but
I
think
the
thing
thing
one
by
applying
flow
control
at
the
low
priority
queue
makes
a
ton
of
sense
and
it
would.
It
would
tend
to
be
effective,
I'm,
not
sure
I'm,
not
sure
it
rules
on
off
I'm,
not
sure
it
goes
all
the
possible
where
the
system
can
overload,
but
it
I
think
it
helps
a
lot.
It.
C
H
J
J
J
H
But
but
yeah
anyway,
all
right
well
I,
don't
take
anywhere
near
ten
minutes
all
right.
It's
far
too
early
I
would
say,
but
I
knew
this
was
important.
We're.
L
A
C
Right
read
description,
guys
thanks
or
bring
up
the
premium
stuff,
I
think
I,
think
I,
think
I
think
I
think
that
I
think
coming
up
with
a
with
the
media
to
medium
term
and
longer
term
buckets
for
things.
I
think
I'm,
not
gonna,
argue
against
the
perfect,
but
making
it
an
Emmy.
The
good
only
I
want
to
see
it's
come
for
things.
We
can
do
soon,
think
things
who
can
do
later
and
excellently
yep.