►
From YouTube: 2019-02-28 :: 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,
I
think
sage
is
gonna,
make
it
I,
don't
know
that
we'll
have
a
whole
lot
to
talk
about
today.
We
I
think
everyone's
super
focused
on
Nautilus.
At
the
moment,
though,
but
as
many
experimental
performance
related
things
coming
in,
but
again
we
can
go
over
to
the
stuff
of
skiing.
Up
did
sage.
A
A
All
right.
We've
got
two
merged
PRS
this
week,
radix
a
pen
buffer
removal
from
buffer
list
looks
like
he
food
merged
that
and
then
also
oh
I.
Sorry,
this
one
just
closed.
It
didn't
merge.
This
was
a
sea
star,
charred
lru
thing,
I'm,
not
sure.
Why
must
have
been
they
decided
to
go
to
another
direction
or
something
all
right.
Let's
see,
there
are
a
couple
that
updated
its
macing
messenger
stuff,
a.
A
B
A
Fun
times:
well,
good
luck!
Let's
see
what
batch
handle
send
message.
Yeah
I
haven't
looked
at
that
at
all.
Apparently
there
oh
I,
remember
that
was
the
one
where
there's
differences
of
opinion
regarding
the
the
performance
benefits
of
this,
so
they're
gonna
do
more
testing
and
and
luck
specifically,
it
may
hurt
larger
io
sizes,
and
then
there
was
a
question
about
whether
or
not
for
smaller
Iowa
sizes.
If
it
actually
helps
or
not.
A
Alright,
there's
one
here
for
reducing
the
number
of
blue.
If
a
space
allocation,
that's
a
little
older
I,
don't
know
I,
don't
know
that
much
about
it
actually.
A
A
It
makes
the
cash
balancer
much
happier
when
it
doesn't
have
these
kind
of,
like
you
know,
short
periodic
spiky
behavior
happening
in
the
background
with
memory
usage
so
that
it
doesn't
seem
to
actually
slow
anything
down
either.
It's
not
really
that
much
worse.
It
looks
like
doing
a
short
trim
on
right,
rather
than
a
long
trim
kind
of
on
a
regular
interval,
so
I
think
I.
Think
it's
gonna
be
a
good
change.
I,
don't
know
other
stuff
going
on
I,
don't
think
so
yeah,
that's
it
I.
A
C
C
A
A
Like
hard
drives,
they
probably
didn't
have
as
much,
because
we
had
a
bunch
of
users
that
were
complaining
that
the
OSD
was
using
too
much
memory.
So
I
think
the
idea
was
that
they
had
hard
drives.
We
should
probably
meant
that
they
were
older,
but
then
we
just
default
to
use
less,
but
it
I
I
think
in
retrospect
it
was
kind
of
a
bad
idea,
because
people
got
really
confused
about
it,
like
everyone
or
a
lot
of
people
were
setting
the
now.
A
A
Or
the
the
global
one
and
they
had
the
other
one
I'll
reset
somewhere
and
they
got
themselves
confused
about
whether
or
not
it
which
one
was
taking
precedence
right.
So
they
had
said
it
one
way
in
one
place
in
another
way
in
another
place,
and
then
it
didn't
match
what
they
thought.
It
was
going
to
be
I.
A
So
I
think
at
least
to
me
it
seems
like
in
retrospect
we
we
may
be
trying
to
be
a
little
overly
clever.
A
There's
a
legitimate
thing
here
on
when
you
have
an
SSD
or
an
nvme
drive
most
likely,
you
don't
have
anything
faster
for
oxy
be
like
most
people
are
sticking.
If
they're
going
to
put
the
block
device
on
nvme,
I
haven't
seen
too
many
people,
yet
that
are
using
something
like
obtain
or
something
super
fast
for
the
database
and
when
that's
the
case
hitting
the
cash
4x
like
metadata
reads
so
so
like
if
you're
gonna
do
a
new
write,
you
actually
end
up
reading
some
stuff
out
of
Rhapsody
be
doing
that
from
the
blue
store.
A
Cache
is
much
faster
than
reading
from
Rocky's
block
cache
or
from
disk,
though
having
all
the
o
nodes.
Cached
is
is
really
really
nice
if
you're
on
a
hard
drive,
a
lot
of
people
end
up
running
with,
like
the
block
device
on
a
hard
drive,
but
then
put
a
Rox
tbe
the
database
in
the
right
head,
logon,
flash
and
reading
data
metadata
out
of
rocks
TB
in
that
scenario,
isn't
really
a
big
deal,
because
your
your
ball
might
buy
the
hard
drive
anyway.
A
No
problem
in
general
I
think
we
have
more
work
to
do
in
this
area
right
now
we
end
up
double
cashing
in
a
lot
of
scenarios:
the
own
ODEs,
both
in
in
the
racks
to
be
cash
and
in
the
blue
store.
Oh
no
I
think
down
the
road.
Assuming
that
we're
we
don't
end
up
just
moving
totally
over
to
C,
store
and
C
star.
A
If
this
is
kind
of
like
a
long-term
performance
thing
with
blue
store
and
nvme
drives,
you
probably
want
to
have
some
kind
of
mechanism
by
which
we
don't
end
up
double
cashing
that
there
may
be
various
approaches
we
could
take.
But
it's
it's
kind
of
unfortunate
right
now
that
that
happens,
because
they
can,
you
know,
hurts
us
in
in
various
scenarios.