►
From YouTube: 2017-MAY-18 :: Ceph Performance Weekly
Description
Weekly collaboration call of all community members working on Ceph performance.
http://ceph.com/performance
For full notes and video recording archive visit:
http://pad.ceph.com/p/performance_weekly
A
B
B
B
C
B
Team,
okay,
alright!
So
this
first
one
came
in
one
type,
one
for
one
that
basically
puts
all
of
the
key
transaction
work.
It
will
work
you
cores
right
now,
it's
being
done!
Synchronously
I'm
surprised
that
it
helped
I,
guess
but
I'm
also
concerned.
This
isn't
the
right
direction
to
go
with
blue
store,
it
sort
of
end
up
the
opposite
of
what
the
RTC
run
to
completion
branch
is
trying
to
do
where
we
try
to
do
as
much
work
as
we
can
seriously
in
for
it.
B
Any
contact
switches
at
all
I
think
is
going
to
make
more
sense
for
fast
storage
like
in
gaming.
In
fact,
we
in
many
cases
will
probably
able
to
actually
commit
synchronously
without
doing
any
contact
switch
in
the
UST
portion.
So
I
think
that's,
probably
not
the
direction
we
want
to
go
in,
but
mark
did
some
tests
on
August
we're
going
to
talk
about
that.
The
kV
Vanisher,
so
I
think
we'll
probably
go
down
that
road
for
now,
we'll
see
this
one
from
how
my
on
async
messenger
and
that
reduces
right,
lock,
contention.
B
I
think
this
is
motivated
by
the
profile
that
marked
it
last
week.
So
that's
encouraging
I
can
refute
included
any
performance
results
in
there
did
he
do
look.
He
posted
some
perf
gainers,
but
I
didn't
have
any
context.
I
can't
tell
if
it
actually
made
any
difference,
but
hopefully
it
did
so.
We
should
test
that.
B
Interestingly,
it
removes
the
synchronous
encode
path,
so
there
was
an
option,
an
async
messenger.
That
would
actually,
when
you
call
send
message
it
which
encodes
message
right
then-
and
we
ended
up
not
doing
that
and
instead
doing
the
encode
work
in
the
worker
thread,
because
it
improved
performance
latency
so
on
I
need
just
remove
that
entirely.
B
So,
but
just
fine
I
guess
we're
just
not
going
back.
If
you
could
supply
the
right
choice,
simplify
things.
Let's
see,
there's
a
rock
CB
thing
that
supports
sliced
parts,
and
this
is
just
avoiding
a
memory
copy.
When
we
had
multiple
buffers
that
are
being
passed
into
box
to
be
there's
an
interface
that
lets,
you
actually
do
that,
instead
of
copying
into
one
contiguous
buffer.
So
that's
that's
good.
B
B
Let's
see,
there's
some
stuff
that
merged
reasoning,
TCC,
intrinsics
or
byte
swap
I,
think
that
is
slightly
faster.
In
some
cases.
Let's
see,
there's
a
optimization
manager
and
the
blue
store.
One
I
put
the
blue
FS
in
the
middle
of
a
hard
disk
device
that
got
merged.
We
didn't
really
see
a
difference,
but
intuitively
it
should
be
better.
So
that's
what
every
other
file
system
do.
It
does
also.
It
puts
the
journal
manila
disc,
since
on
average
the
stick
should
be
shorter.
B
There
are
a
couple
other
ones.
There's
the
blue,
FS
sync
right,
one
that
includes
a
bug
fix
for
sync
right
option
that
we
definitely
need
to
merge
and
then
also
changes
the
option
to
true
which
doesn't
seem
to
make
any
difference
that
we
can
measure
so
I
think
maybe
I'll
just
drop
that
and
leave
it
asynchronous
until
we
show
positive
evidence,
I
guess
that's
on
my
plate.
Let
me
cut
out
my
list
a.
B
There's
a
pull
request
that
simplifies
the
locking
after
more
discussion,
we
decide-
that's
probably
slower,
so
we're
not
going
to
set,
certainly
not
in
its
current
form.
There's
lots
of
discussing
around
the
CRC
calculation
of
zero
buffers.
I
haven't
been
following
that
discussion,
so
I'm
not
sure
what
status,
and
maybe
someone
can
chime
in
anybody,
know
offhand
where
that's
at.
B
Anyway,
yeah,
whenever
that's,
whenever
we
decided
ready,
it's
fine
to
converge
it,
but
I,
don't
I
haven't
been
following
the
discussion,
so
I'm,
not
sure,
and
then
I
had
a
pull
request
for
blue
restore
that
tries
to
change
it
so
that
the
blobs
I
get
marked
on
shareable
or
chaired
and
unmute
immutable.
As
a
result
of
the
cloning
for
the
rollback
for
richer
coding,
one
knows
that
Roebuck
information
get
tired.
B
Then
it
tries
to
on
unshare
the
blob
so
that
it
can
be
modified
again,
basically
well,
but
it
failed
in
QA,
so
I
need
to
debug
it
and
fix
it,
and
that's
all
that
that
stuff,
let's
see,
there's
other
stuff,
there's
1396,
which
I
think
is
pretty
encouraging
when
it
last
time
I
put
it
through
QA,
it
I
saw
much
a
sailor,
so
I
couldn't
explain
it
so
I
pulled
it
out.
I
need
to
go
back
and
retest
it
and
see.
If
it
is
that
pull
request,
fault
or
not,
my
suspect,
to
deuce.
B
See
I
said:
discard
method,
I'm,
hesitant
to
merge
that
until
we
actually
have
some
evidence
that
it
helps
captive,
throttle
I'm,
not
sure.
If
we're
going
to
merge
that
or
not
we
have
sort
of
a
really
simple
throttle
in
there
now
I
think
wait.
Where
do
we
end
up
word
we
land
on
the
throttle.
Do
you
remember
mark?
Are
we
doing?
Let's
see,
I
think
we
just
tuned
it
for
hard
disk
on
SSD,
right,
yeah.
A
A
B
A
B
Okay,
maybe
we
just
double
check
but
I
think
I
think
we're
mostly
okay.
There
yeah
it
might
be
worth
coming
back
a
little
bit
later
and
looking
at
it.
If
the
adopted
throttle
makes
any
further
improvement,
but
I
kind
of
doubt
it
is,
but
modeling
doesn't
really
improve.
It
can
only
hurt
mostly
it's
just
a
matter
of
managing
the
queued
up,
so
that
the
QSF
will
work
better
mm-hmm.
D
B
Makes
it
kind
of
hard
to
test
I
guess,
but
so
it
is,
let's
see
it's
a
pull
request
that
changes
the
compressed
block
cache,
basically
a
separate
tunable
for
the
cache
or
compressed
blocks.
If
you
happen
to
disable
or
any
little
compression
which
we
don't
by
default,
so
I
think
we
can
merge
that
in
case
somebody
changes
I,
don't
know
in
case
somebody
in
case
somebody
enables
compression
in
markscopy,
then
we
can
do
it
and.
A
B
A
Of
the
things
that
really
is
weird,
which
I
don't
understand-
and
maybe
it's
just
maybe
I
think
maybe
they're
going
to
change
it,
but
it
doesn't
seem
like
the
d-block
cache
and
the
compressed
block
cache
actually
help
you
in
compaction,
like
you're,
not
doing
reads
from
cache
or
always
reads
from
the
SST
files,
and
now
that
we're
doing
like
the
read
ahead.
It's
way
way
better
is
not
nearly
as
much
of
a
big
deal
as
it
used
to
be.
But
it
looks
like
we're
never
hitting
hitting
the
cache
yeah.
B
B
B
B
B
Yeah.
Another
scale
stuff
here
still
hoping
that
rate
of
sloth
can
resurrect
the
stuff
that
he's
doing
with
the
in
boost
or
a
while
back
they
still
finishing
at
the
Rios
gateway
pieces.
I.
Think
the
blue.
If
that's
right,
cash
is
still
buggy
I
think
it's
not
really
critical
for
a
luminous,
because
you
can
just
turn
on
the
option
to
use
the
operating
systems
paid
cash
so
not
really
needed
just
like
to
get
away
from
that
eventually
and
I
guess:
that's
it!
B
But
I
think
that
needs
well.
They
need
to
rebased.
It
was
mostly
complete
as
far
as
it
one
at
least
I.
Think.
Probably
what
we
really
need
to
do
is
try
that
out
on
either
an
NV
dim
or
an
nvme
for
the
journal
and
just
demonstrate
that
it
actually
helps,
and
if
that's
the
case,
then
we
should
merge
it,
but
until
we
actually
show
that
it
helps
it
has
all
that
I
complexity
to
the
to
those
paths
and
there's
a
bit
of
risk
associated
with
it.
So.
B
Think
we
want
to
do
this
eventually.
I
think
there
are
a
bunch
of
places
where
we
could
benefit
from
using
slots
containers,
because
we
have
containers
that
are
almost
always
small,
like
em
yeah
sets
and
maps
that
only
have
a
few
items
yeah,
but
it
needs
it
needs
some
like
careful.
It
needs
to
be
actually
going
to
be
cleaned
up.
The
code
needs
to
be
like
carefully
gone
over
by
somebody
who
is
savvy
with
all
the
C++
11
language
features
and
make
sure
it's
sort
of
the
best
we
can
make
it.
B
B
A
A
But
in
when,
when
we
talk
about
this
profile
later
on,
here,
there's
there's
a
lot
of
big
things
that
are
showing
up.
But
what
was
some
of
the
stuff?
That's
not
really
showing
up,
and
this
is
his
dealings
all
that
kinda
stuff,
which
I
kits
kind
of
scattered
all
over
but
I
kind
of
wonder
if
maybe
it's
bigger
than
it
looks
like
when
you
just
stare
at
one
of
these
things.
Yeah.
B
A
Sure
so,
beyond
what
I
have
here,
I
mean
not
that
much
to
talk
about
yeah,
I,
guess,
Igor
kind
of
reminded
me
that
he
was
seeing
really
big
improvements
with
this
a
couple
weeks,
maybe
a
month
ago,
I've
been
meaning
to
look
at
it
and
just
haven't
yet.
But
finally,
this
weekend
was
able
to
and
yeah
for,
for,
K
random
writes
on
one
OS
d,
using
a
client
running.
If
I
able
to
live
RVD
back-end,
it
was
pretty
dramatic.
A
I
was
like
if
we
went
from
like
25k
I
ops
on
an
nvme
up
to
kind
of
fluctuating
between
about
thirty
to
forty
K.
So
it
was
a
nice
nice
bonus,
nice
bump
I
did
do
a
wall
clock
profile
while
testing
it
and
yeah
we're
we're
spreading
work
out
much
better
than
we
did
previously,
but
both
the
beast
or
K
V
final
thread
and
beast
or
K
vsync
thread
are
quite
busy.
A
They're,
not
they're,
not
as
bad
as
it
used
to
be
like
I
said,
but
there's
a
lot
of
work
going
on
so
yeah
there's
a
profile
there
for
anyone
that
wants
to
look
at
it.
I
kind
of
tried
to
go
through
and
look
at
stuff
that
seemed
jump
out
at
me
that
look
kind
of
interesting
see
still.
A
A
A
A
A
A
That
it
all
been
good
news,
that's
spending
a
fair
motifs
time
there,
this
box,
TV
table
key
comparator,
is
now
really
big.
Mm-Hmm.
A
B
A
Otherwise,
bufferless
rebuilds
and
rebuild
wine-
this
rebuild
blind
size
and
memory
is
showing
up
there's
other
places
where
it
does
too,
but
that.
B
E
Hiya
Nick,
here
hey
so
just
to
absorb
smart
last
week
about
the
deferred
right,
stuff
and
so
I
just
gave
it
never
protest
with
the
latest
master
and
it
seems
pretty
much
to
be
the
same
so
performance
now
as
file
store
and
the
disk
definitely
sort
of
in
the
same
equation
as
before.
So
that
looks
like
it's
sort
of
working
up,
so
I
got.
E
At
me,
hopefully,
I'm
tweeting
on
some
more
MDM
ease
for
our
next
batter
nodes.
There's
SSD
drought
at
the
moment
southern
and
when
they're
going
to
turn
up,
but
I
should
be
able
to
run
that
on
our
cobweb
for
the
guards
production.
So
if
they
can
hit
the
same
extreme
low
latency
we're
getting
so
the
tests
I've
done
so
far
just
been
on
test
box,
but
yeah
I
thought
I'd.
Just
put
that
back.
Awesome
cool.
F
Hello
stage
we
are
handsome
from
Alibaba,
so
we
submit
a
poor
request
for
executor
to
transaction
or
synchrony
in
some
sense
retro,
as
the
object
store
there.
You
comment
that
so
first
aired
its
kind
of
or
disagree
with
your
direction
of
optimization
basically
is
adding
more
contacts
which
are
more
threads,
but
we
saw
some
problems
like
in
companies
work.
It
has
a
long
function,
has
business
doing
a
lot
of
work
and
the
teacher
lock
is
held
a
while
doing
this.
If
there
are
on
some
other
target
on
same
PG,
basically
they
are
blocked.
F
F
F
B
B
E
B
B
B
I'm
not
I'm,
not
sure
this
is
the
time
to
do
it.
I
guess:
okay,
it's
I
mean
it's
encouraging
that
this
this
works
for
sequential
a
is
and
that
so
I
got
helpful
data
point
to
make
sure
that
we,
what
we
build
is
going
to
be
can
accommodate
that,
but
I
don't
want
to
I,
don't
want
to
complicate
the
walking
right
now.
F
Okay,
so
thank
you
so
I
have
another
question
so
basically
in
our
profiling
performance
for
finding,
we
find
that
for
brew
store
the
I/o
pad
the
metadata
synchronize
on
that
Nathan
seat
much
higher
than
accurate.
The
data
I,
oh
wait,
so
do
you
think,
is
it
better
to
either
as
submit
the
rocks
DB
sync
in
the
individual
workers
thread,
instead
of
just
using
one
case
rates
to
do
a
bad
thing.
B
B
So
this
this
that's
roughly
what
this
polar
Crescent
tries
to
do
so
it
makes
it
so
that
you
can
either
queue
the
transaction
synchronously
or
you
can
actually
commit
it
to
currency
where
you
wait
for
the
idea
to
complete
from
the
worker
thread.
There
are
a
whole
bunch
of
sort
of
restrictions
on
when
it's
safe
to
do
that,
because
you
have
to
preserve
the
ordering
those
iOS
and
commits
within
your
sequencer,
but
all
that
infrastructure
is
in
that
pull
request.
So
I
think
it's,
I
think,
that's
roughly
the
right
way
to
go.
B
There
are
a
few
commits
towards
the
end
that
try
to
also
make
it
so
that
you
can
do
to
I/o
completion
callback
synchronously
also,
those
aren't
very
well
tested,
I'm,
not
sure
if
that's
sort
of
the
right
way
to
go.
So
you
could
just
ignore
those
for
the
time
being
and
focus
just
on
the
front
half
of
the.
So
the
that's
very.
B
Done
yeah
l
came
for
you
take
a
look.
Minx
is
based
on
work
that
I'm,
like
somebody
else
submitted
like
I,
can
remember
his
name:
Boyd
bag.
I
think
this
is
a
username
and
so
I
took
those
patches,
yeah
he's
comforting
on
the
PO
requests
and
and
putting
them
up,
but
it
needs
them.
So
this
is
actually
I'm
into
this
a
bit
earlier.
I
starting
to
mark
I.
B
Think
we
need
to
if
we
can
rebase
this
and
demonstrate
that
it
actually
is
a
improve
things
when
you
have
a
MV,
meanie
or
need
beat
in
for
a
journal
and
I
think
it's
worth
emerging
at
least
the
first
part
of
this
PR,
but
we
should
to
sort
of
do
that.
Testing
to
make
sure
it
actually
is
is
making
a
difference,
because
it
does
that
a
little
bit
of
complexity.
G
E
B
G
G
B
B
G
G
Sort
of
looking
at
this
perfectly,
whatever
aggregate
and
hold
on
functions,
exclude
function.
Good
people,
oh.
B
G
B
B
The
bakken
we
merged
the
greatest
block
in
patch
and
Adam
I,
just
disabled
us
all.
You
had
it
cute
kids.
He
make
you
turn
it
on
you
and
I
yeah
I,
don't
know.
I
was
harassing
I'm
I
can
refer
as
Casey
O'riley
to
update
that
just
a
document.
That's
just
like
a
how
to
use
it
because
I
when
I
fold
it
are
since
they
didn't
work
so
I'm
not
entirely
sure
what
stash
is
there
George.
B
B
Cool,
yet
there
is
so
the
rate
of
block
in
one
merge,
there's
a
and
as
the
second
patch
that
adds
all
the
instrumentation
to
our
BD.
Also,
that's
that
cleaned
up
version
of
the
original
that
a
Jason
just
posted
I
think
earlier
this
week,
so
that
should
pile
in
Charlie
but
yeah
I
think
that
the
doc
needs
to
be
updated
because
I
don't
know
the
knot,
works.
B
C
So
the
question
I
had
was
for
Peter
about
the
improved
CRC
calculation
for
zero
buffers
on
line
48
at
a
etherpad
I
know,
we've
been
discussing
the
crossover
point
for
very,
very
small
buffers
where
it
doesn't
make
sense
to
use
the
fast
zero
optimization,
but
other
than
that
it
seems
like
he's
incorporated
most
of
the
comments.
I
just
wondered
what
what
remains
to
be
done
on
that
request
and
if
it's
related
to
the
crossover
point,
then
I
can
try
to
help
with
that
or
do
two
more
testing.
C
C
Being
a
bit
stubborn
about
changing
that
one
point
and
I
guess
my
point
was
the
rest
of
the
code
looks
like
a
significant
improvement
and
if
somebody
changes
that
one
line
then
Peter
would
get
I
mean
I
agree
with
Peter
about
the
crossover
point
and
I
tried
to
data
in
the
TR
sets
back
that
up.
I!
Don't
know
why
Adam
doesn't
agree,
but
but
I
get
the
best,
a
one-liner
that
that's
wrong.
We
could
always
change
it
later,
but
they.
C
B
Oh
and
it
should
be
2
K,
you
think
right
got
it
yeah,
okay,
I'll
just
mark
this
needs
QA.
We
can
just
make
sure
it
goes.
It
makes
it
to
the
sweet
if
I
fine
to
do
it.
This
way,
anyway,
for
two
weeks
we'll
get
good
good
coverage
on
both
sides
of
the
threshold,
but
yeah
we
can.
We
can
switch
that
to
the
factor
by
the
way:
I
yeah
I've,
no
opinion
so
threshold,
okay,.
B
C
Yes,
both
so
that
one
is
I
marked
it
do
not
merge
this
morning,
because
I
need
to
do
a
little
more
on
it,
but
in
general,
so
on
x86.
My
findings
were
that
sorry,
when
we
back
up
when
I
first
put
up
the
pull
request,
the
only
reason
I
was
putting
it
up
was
because
Adams
code
falls
back
to
the
optimized
assembly
in
this
one
corner
case,
and
then
his
code
was
faster
than
than
all
the
other
versions
on
x86
and
PowerPC.
C
I
was
really
only
putting
it
up
for
that
one
corner
case
as
a
fault
I
reckon,
and
then,
when
Peter
pointed
out
this
problem
about
it,
does
a
bit
reversal
on
the
way
in
and
on
the
way
out
and
the
the
code
for
the
bit
reversal
was
kind
of
brain-dead
and
not
a
very
optimized
loop,
and
it
was
turns
out
really
slowing
things
down,
and
so,
when
I
replaced
that
with
some
code
that
was
already
in
another
part
of
set.
That
does
almost
the
same
exact
thing.
B
C
Yeah
sorry
I
I
mean
he
got
there
first
and
all
of
that,
and
so
having
emerged
first
is
certainly
fine.
I,
just
gonna,
say:
I
put
a
lot
of
side
side
things
up
on
the
repo
that
I
linked
from
the
pull
request
of
trying
to
back
up
my
findings
with
data
I.
Don't
know
if
anyone's
Peters
look
at
that
or
if
that's
too
much
but.
D
C
B
C
B
C
Right
so
the
J
eraser
code
is
not
optimized
for
PowerPC
and
I.
Don't
think
it
will
be
anytime
soon.
I
think
we
were
looking
to
have
I
to
L,
be
the
vehicle
for
for
power
P.
So
no
that's
right.
Okay,
let's
cook
a
lot
of
that
letters
that
one
is
taking
a
bit
longer
and
I
has
a
bit
more
technical
debt
to
it.
So
I,
don't
I,
don't
think
you'll
be
seeing
that
within
the
next
month
or
so
yeah.
B
B
There
is
another
library
that
is
called
GF,
erasure,
I,
think
it
is,
that's
that's
a
candidate
to
replace
Jerri
sure
that
might
be
a
place,
but
I
yeah.
That's
a
little
bit
early,
so
I'll
share
more
when
it
that
develops
bit
more
okay,
okay,.