►
From YouTube: 2016-SEP-28 :: Ceph Performance Weekly
Description
Weekly collaboration call of all community members working on Ceph performance.
For full notes and video recording archive visit:
http://pad.ceph.com/p/performance_weekly
A
A
Let's
see,
there's
this
first,
one
about
garbage,
collect
for
partially
overlap.
Blobs
I'm
curious
to
see
how
much
that
helps,
though
that's
basically
blowing
down
reads:
overtime
and
skid
that
we've
got
that
in
there
else
is
in
here,
Oh
sages.
This
isn't
blue
star,
but
this
is
a
sages
ER
for
less
frequently
updating
certain
PG
and
protease
stats
I.
A
So
originally
we
had
been
count
looking
at
just
bar
and
encoding
the
whole
thing-
and
you
know
it
could
be
a
kind
of
brute
force,
easy
fix,
but
I
think
this
is
actually
going
to
be
better.
The
conditional
checks
is
doing
here,
probably
lower
overhead.
Then
the
encoding
was
so
yeah
that
looks
good
I'm
excited
to
see
how
much
that
helps.
A
A
Speaking
of
that,
one
that
one's
kind
of
going
through
testing
right
now,
it
looks
like,
in
the
very
most
recent
version
of
that
pull
request.
We
may
be
seeing
a
regression
specifically
with
aged
random,
reads
or
random
reason
on
an
aged
system,
so
we're
kind
of
trying
to
track
that
down
is
a
little
bit
annoying,
but
hopefully
we'll
we'll
figure
that
out
soon.
A
A
A
So
let's
see
well
do
we
have
here.
I
guess
I've
already
had
talked
in
discussion
topics
about
some
of
these
bull
requests
and
things
ice
age.
Is
there
anything
you
wanted
to
add
to
any
of
that.
A
B
C
A
B
Looks
like
you're
cutting
up
yeah,
I
mean,
I
think
this
this
seems
is
independent
of
them,
using
the
the
slab
allocation
for
like
small
vector
and
slab
vector
and
so
on.
B
C
B
Yeah,
we
lost
them,
yeah
all
right,
Oliver
I,
to
come
back
to
this
yeah
I,
don't
know
if
it's
possible
just
instantiate
an
existing,
a
standard,
allocator
st
alligator
in
class,
or
if
you
want
to
use
one
of
the
boost
ones
and
try
to
optimize
it
for
type
I,
don't
know
but
I
think
simple
to
start,
because
I
think
the
biggest
thing
I
want
is
just
the
memory
accounting
but
I
can
make
the
the
backpressure
stuff
trim.
Cash
is
based
on
actual
memory
used
instead
of
these
hard
to
configure
counts.
A
A
A
B
B
Do
is
change
the
Ceph
configurable,
and
so
they
are
defined
in
terms
of
a
fraction
of
the
total
cache
size
that
we
allocate
the
rocks
TV
so
that
there's
one
knob
that
somebody
says
I
want
the
Oh,
Stevie
use,
one
gigabyte
of
RAM
or
whatever,
and
then
a
moo
barn
to
figure
out
how
much
to
go
drugs
to
be,
and
so
on,
yeah,
but
think
I
can
come
later
run
for
now.
It's
all
just
sort
of
crammed
into
that
rocks
to
be
tunable
string,
yeah.
A
B
Answer:
yeah:
no
thanks,
so
I'm,
probably
the
hardest
thing
that
the
OST
is
going
to
do
it's
going
to
the
most
sensitive
is
that
it
has
these
PG
logs
that
are
a
bunch
of
omap.
The
inserts
are
spread
out
over
time
in
different
parts
of
the
namespace,
and
if
you
have
a
postit
that
goes
down
and
then
comes
back
up,
if
you
do
backfill,
you
have
to
read
these
logs
in
no
now
I.
Take
it
back
that
that
happens
on
startup,
so
it
should
be
okay,
yeah,
I.
C
B
Needs
it
I
mean
everything,
even
though
is
the
outside
of
the
blue
store
needs.
It
like
this
has
always
been
annoying
figuring
out
how
much
memory
the
OC
is
going
to
use
and
tuning
it
to
use
that
much
memory
is,
is
a
black
art,
so
maybe
we
solve
the
general
alligator
problem
and
we'll
we'll
be
able
to
clean
up
a
lot
of
this
stuff
is
actually
the
MDS
is
pretty
bad
name.
Yes,
okay.
A
B
Anyway,
I
think
that
next
step
is
to
continue
the
discussion
on
the
list.
I,
don't
think
we
have
anything,
we
need
to
go
through
right
now.
Sure,
let's
see
the
fast
info
stuff
is
just
queued
up
for
testing
I.
Don't
think,
there's
anything
else
to
do
there.
I
could
do
some
performance
tests.
I
guess
I
see
how
much
it
helps
on
I'll
store
and
on
blue
store.
It
should
reduce
the
metadata
load
quite
a
bit,
but
you
notice
what
the
performance
impact
there
is
just
apply.
Some
testing
did
you.
There.
B
A
B
As
long
as
I've
had
actually
I'm,
let
me
frame
work
than
theirs
I'm
going
to
keep
rebasing
that
branch
as
I
fix
things
so
and
I
pull
it
up
to
wait
a
semester.
So
I
would
probably
I'm
gonna
make
sure
it's
bisect
able
and
once
I
do
that
then
I
think
we
could
retest
on
that
yeah
figure
out
figure
out
where
we're
we're
broke.
Cuz
our
there
are
a
few
things.
I
have
changed,
but
I
don't
really
understand
why
they
would
have
affected
it.
Yeah.
A
Me
either
hits.
It
seems
to
be
fairly
clear,
though,
because
it's
I've
gone
through
like
seven
or
eight
different,
full
tests,
different
points,
and
it
seems
like
I'm,
seeing
pretty
consistently
a1
range
in
another
range
of
results
that
don't
really
overlap,
so
you
know
there's
their
variation
within
each
one,
but
they
they
seem
to
be
distant
clusters.
So.
A
A
B
It's
I
think
it's
because
the
least
bit
either
you
have
a
thread
for
a
connection
that
calls
in
the
fastest
patch.
So
it's
little
bit
slow.
It
only
affects
that
connection.
Now
it's
a
pool
of
threads,
that's
handling
all
of
the
reads
and
writes
off
the
socket,
and
so
in
fast
dispatch
slows
down
and
kind
of
slows
down
the
whole
little
thing:
okay,
I'm
guessing
I'm.
Guessing
that's!
Why
that's
for
me
high
level,
hand-wavy,
I
guess,
but
but
yeah
I
think
it
may
even
our
suicide.
There
are
some
some
blocks.
A
B
D
B
A
A
We
as
long
as
you
have
like
128,
it
seems
like
a
lot
of
times.
Simple
messenger
should
be
faster
than
nascent
messenger.
They
might
not
consistently
hold
all
the
time,
but
it
looks
like
at
least
with
JD
malik
and
which
is
and
gives
this
similar
results
that
you
see
with
a
fair
cash.
In
that
case,
that's
where
we
see
simple
higher
than
a
sink,
though
it's
it's
kind
of
like
once.
The
memory
allocator
is
out
of
the
way.
However,
you
achieve
that
then
simples
faster
than
a
sink
yeah.
B
A
Other
the
other
thing
is
my
my
customers
with
blue
store
in
his
was
this
file
store,
so
it
might
be
that
there's,
maybe
either
blue
stores
faster
and
we're
hitting
this
ball
next
posture
or
maybe
blue
stories
just
making
it
worse.
Somehow,
I
don't
know
yeah,
maybe.
B
A
D
A
One
one
question
I
had
for
you
sage,
speaking
of
perf,
is
in
those
traces
that
you
were
getting
recently.
Did
you
see
anything
related
to
the
bitmap
alligator,
except
those
still
almost
nothing.
B
Well,
I
still
don't
trust
my
traces
because
they
look
completely
different
from
yours
and
they
show
no
no
detail.
My
only
theory
was
that
maybe
my
compiler
version
is
more
aggressively
in
lining
and
so
I
don't
see
a
lot
of
the
tail
ends
of
the
call
chain,
but
I
don't
know
it's
really
weird
that
I
when
I
look
at
your
profile,
I'm
like
oh
I,
can
fix
this
in
this
net.
Then
I
look
at
mine.
There's
like
it
doesn't
there's
nothing
to
do.
B
I,
don't
know,
I,
think,
let's,
let's
get
the
encoder
stuff
sorted
out
and
then
I
spend
more
time
on
that
again.
I
have
a
cup
okay,
we
gotta
fix
first,
but
yeah.
C
B
B
About
getting
crackin
thing
hanging
out
grade
tests
and
getting
other
stuff
cleaned
up
to
merge,
yeah
I,
don't
want
to
get,
is
exciting
stuff
in
yeah.
B
A
B
C
C
There's
some
other
kind
of
low
level
stuff,
but
you
know
you
know
I
think
the
a
lot
of
it
resolves
around
being
able
to
save
the
container
code
that
we've
got
I.
Think
once
once
you
sort
of
treat
that
as
a
prerequisite,
you
know
then
you're
you're
dealing
with
an
STL
allocator
and.
A
C
Know-
and
you
know
they're
not
that
hard
fundamentally.
B
B
C
C
C
B
C
C
B
B
B
C
B
C
B
C
Reason
have
that,
so
you
know
what
I
was
going
to
use
is
about.
All
I
care
about
is
some
identifiers
with
some
affinity
to
the
cpu,
so
what
I
was
going
to
propose
is
the
basically
the
the
base
of
the
stack
frame,
which
is
what
I
think
the
P
thread.
I
forget
it's
a
P
friend
ID
or
the
P
threatening,
or
something
like
that.
But
it's
a
cheap
thing
that
you
can
get.
C
C
B
C
Know
as
you
only
do
that
in
frequently,
then
it's
not
too
bad.
You
know,
I
was
figuring
32
or
64,
but
you
know
I
think
that
I
I
think
that's
probably
plenty
of
disambiguation.
It
probably
only
really
needs
to
be
like
maybe
two
or
three
times
the
core
count
that
you
know
that
this
OSD
is
scheduled
against.
You
know
what
it
really
says.
Is
you
probably
want
to
partition
your
cores
against
your
OS
DS
yeah
yeah.
B
B
B
C
C
B
C
C
A
C
Looking
at,
like
you
know,
an
atomic
increment
after
some
shifting
and
masking
to
pick
one,
you
know
which
you
know
with
some
affinity
to
the
current
core.
You
know,
I'm
guessing
you
know
the
net
it
out
should
be.
You
know
a
dozen
instructions
or
something
like
that
times.
Whatever
the
effectiveness
of
the
sharding
is
on
the
cache
miss
that's.