►
From YouTube: 2019-12-12 :: 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,
there's
still
a
cool
folks
in
core
that
might
come
in,
but
we
can
just
get
started
here
all
right.
Let's
see,
I
only
saw
one
new
PR
this
week
and
that
was
from
heat
regarding
changing
some
of
our
documentation
to
be
slightly
more
updated
than
a
used
to
be.
This
is
just
regarding
recommendations
for
CPUs
and
memory,
the
OSC
side.
A
A
Let's
see
next
on
the
list,
we've
got
a
bunch
of
stuff
closed.
The
big
one
here.
Well,
two
big
ones
here
were
the
I/o
Ewing
support
chief
who
merged
that
a
couple
days
ago
and
then
also
Igor's
work
on
I
guess,
a
framework
for
letting
you
more
intelligently
use
the
database
space
in
blue
store,
so
not
not
doing
kind
of
what
we
currently
do
where
once
you
start
compacting
stuff
to
the
block
device
instead
of
the
database
device.
I
just
can't
sticks
around
there
and
forever
and
doesn't
ever
move
back.
A
So
that's
good
I
think
that
maybe
has
a
peer
based
on
that.
That
actually
does
aren't
stuff.
But
I,
don't
think,
there's
been
any
progress
on
that,
but
maybe
there
will
be
now,
though,
that's
good,
the
VL
APR
that
is
being
closed
in
favor
of
potentially
future
PR
doing
the
same
thing,
but
I
think
he
was
running
into
some
bugs
during
testing
influence,
I
guess
properly
test
it
before
submitting
a
new
PR.
So
hopefully
we
see
that
again
soon,
I've
got
an
OP
track
or
PR
that
got
closed
by
this
Dale
bot.
A
They
kind
of
get
rid
of
the
mixed
way
that
we
kind
of
categorized
stuff.
So
the
gist
of
it
is
that
we
could
do
the
whole
thing
better,
but
it'll
be
more
work
and
needs
to
be
done
a
little
bit
differently
than
that
PR
did
it
up
here
does
include
kind
of
an
interesting
optimization
for
how
we
store
and
process
events
too,
but
maybe
that
should
be
handled
as
a
for
PR
as
well.
A
A
Think
he's
been
updated
and
probably
needs
to
be
retested
again,
and
then
this
one
from
Eric
that
does
better
filtering
in
CLS
for
bucket
listing
it
does
fill
basically
just
reduces
the
amount
of
data
that
needs
to
be
sent
over
the
wire
by
pre-filtering
it
in
the
OSD
I
believe
he
has
now
has
that
in
testing.
So
that's
that's!
Basically
it
for
stuff.
That's
been
updated
and
a
whole
other
interesting
stuff.
Here
this
week,
I
see
any
a
pianist
and
he
PRS
people
wanna
talk
about.
A
A
So
the
in
memory
as
well
and
I
tried
to
look
at
a
couple
of
different
configurations
that
we
know
improve
performance
or
become
the
base
configuration
at
least
in
this
gamma
test.
One
is
to
increase
the
memory
target
so
that
you're
not
doing
any
kind
of
reads
of
map
entries
during
writes
and
then
the
second
I
guess
isn't
known,
but
it
was
suspected
to
possibly
improve
throughput
when
we
have
PG
log.
No
longer
writing
stuff
out.
Smaller
bright,
have
log
buffers
and
rocks
TB
so
that
hopefully
we
we
have
less
work
to
do
there.
A
And
so
the
just
of
it
is
that
it's
kind
of
an
incremental
improvement
each
time
being
the
OMAP
rights
out
of
wraps
DB,
helps
and
then
totally
gain
rid
of
peih-gee
log
rights
and
trims
in
general
helps
even
more.
The
improvement
was
not
quite
as
dramatic
as
it
could
have
been.
I
mean
it's
decent.
It's
like
I
think
around
20
percent.
If
I
remember
right,
when
looking
at
kind
of
the
optimized.
A
But
the
big
thing
that
I
thought
was
very
interesting
is
the
reduction
in
tail
latency.
It's
pretty
dramatic.
We
go
from
like
above
60
milliseconds
many
90%,
the
90th
percentile
found
to
like
20
to
25
kind
of,
depending
on
which,
which
configuration
you're
looking
at
exactly
so
that
was
that
was
really
nice
to
see
a
kind
of
a
pretty
big
effect
on
a
big
cluster
I.
Think
as
you'd
expect,
the
amount
of
time
we
spend
in
compaction
goes
down.
A
A
Let's
see
what
else
we
don't
see,
or
at
least
I
didn't
see
a
real,
dramatic
reduction
in
CPU
usage,
which
is
a
little
interesting
because
when
we
tested
them
store
on
a
year
maybe
two
years
ago,
something
like
that.
It
didn't
have
a
whole
lot
of
improvement
on
performance,
but
we
saw
big
reductions.
A
If
you
use
F,
though,
here
we're
not
we're
spending
more
time
doing
other
stuff
in
blue
store,
whereas
no
store
is,
you
know,
apparently
not
doing
nearly
as
much
so
PG
log
was
probably
how
much
bigger
fraction
of
the
time
spent
do
see
have
a
way
you'd
expect
to
see
in
terms
of
my
apps
per
core
improvement,
just
based
on
the
fact
that
the
core
usage
was
around
the
same,
but
throughput
was
higher.
So
that's
that's
basically
it
for
now
I'm
going
to
also
I'm
in
the
process
of
rerunning.
These
tests
on
our
officinalis
nose.
A
I
did
notice
that
throughput
on
those
notes
was
lower
than
I,
expect,
I
think
something
merged
in
mastering
the
last
month.
That
is
hurting
us
a
little
bit
so
I,
probably
him
get
be
busy
doing
a
little
bit
of
bisection
to
see
if
I
can
figure
out
what
that
is,
or
was
real,
but
then
I'll
run
some
of
these
on
the
new
nodes
and
also
probably
some
ohio
st
tests
just
to
see
if
that
reduction
in
latency
is
helping
significantly
when
you
have
replication.
A
B
B
C
I
think
you
find
it
Adam
and
confirm
in
attorney
that
you
know
in
any
of
these
areas
and
probably
we're
irrelevant
reduction
in
compaction
is
intrinsically
relevant.
There's
no
benefit
performance
coming
over
time
right
as
soon
as
you
don't
pay
those
but
I.
Think
then
yeah
I,
don't
think
I,
don't
think
you
can
make
everybody.
C
Hopefully
you
pull
you
to
fully
capture
the
benefit
of
HIV
and
the
point
changes
biting
off
the
off
the
top
line
necessarily,
but,
as
you
remove
remove
things
from
the
stack
you
collect
them
later,
I
think
you
prove
the
benefit
of
this,
even
if
it
hasn't
max.
Even
if
you
haven't
in
a
massive
payoff
and
don't
you
no
no
and
all
the
performance
measure
issue,
you
have
all
the
show.
That's
worth
doing.
This
yeah.
A
B
A
Another
thing
I
need
to
go
back
and
figure
out
what
the
bisect
is
for
some
reason
now
and
master
I
can
very
easily
stall
the
OSD
when
trying
to
do
a
wall,
clock
profile
it
just
like
totally
breaks
and
then
doesn't
doesn't
respond
to
requests
anymore.
So
it's
harder
to
get.
But
let
me
see
if
I
can
say
something
you
might
take
a
little
bit
so
feel
free.
It.
A
A
A
A
A
How
do
you
think
we
should
approach
the
no
oh
map
piece
of
this
rather
than
you
know?
The
no
PG
log
has
that's
fun,
but
you
know
it's
not
very
realistic,
but
the
old
map
piece.
Maybe
we
can,
can
we
reasonably
move
this
over
to
some
other
kind
of
blue
store?
You
know
on
this
format.
Is
that
something
we
should
think
about
right
now,
like
a
new
interface
for
that
earth,
using
the
same
interface
and
just
like
hijacking
it
yeah.
A
B
D
D
It's
harder,
but
either
way
this
is
the
way
PT
logs
worked
before
2012
work,
so.
D
B
D
B
A
C
C
B
D
Yeah,
but
in
the
common
case,
where
knows
DS
just
went
down
briefly
and
came
right
back
up
for
the
trim
operation
you're,
just
removing
a
couple
of
them
Matthews
and
that's
a
common
case.
So
I
don't
think
you
want
to
customize
that
case
by
making
good
rewrite
the
whole
block.
D
You
look
if
you
have
three
OS
GS
and
you
take
primary
down
the
primary.
Almost
always
the
divergent
entries
right
does,
the
other
two:
don't
don't
don't
have
them.