►
From YouTube: 2017-FEB-01 :: 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
B
There's
there's
the
difficult
request
that
does
the
partial,
restarting
in
blue
store
marks
testing
it
unclear
whether
it's
really
helping
the
aio
reads.
Finally,
merge
semester:
Bassam
did
a
backward
to
crack
and
also
but
p
squash
into
one
Pat's
I
have
some
to
keep
it
separate,
but
I
put
that
in
the
next
cracking
point
relief.
B
B
C
B
B
D
I
didn't
have
a
whole
lot.
I
guess
I'm
here
to
talk
about
with
what
they
are--
that
merged
and
the
testing
that
we
done
so
far
makes
it.
Look
like
it's
much
faster
for
sequential,
read,
month,
object
oriented!
So
that's
good
as
well,
though
that's
doing
exactly
what
it
was
intended
to
give
based
on
pages
ER
the
whip
list.
Early
sharp!
Oh,
maybe
talk
about
that
a
little
bit
from
the
testing
that
we've
done
so
far,
but
the
good
news
is
that
looks
like
pretty
much
across
the
board
for
small
random
write.
D
Cp
usage
is
dramatically
lower.
Derby
the
performance
interview,
limited
situations
should
be
much
higher
from
just
kind
of
what
I've
seen
going
through
doing
lots
and
lots
of
tough.
The
things
a
little
surprising
is
that
it's
not
strictly
faster.
It
seems
like
in
some
cases
and
performances
better
in
case
of
the
performances
work
with
a
16km
Alex
eyes,
which
is
the
default
right
now
for
ssds.
It
seems
like
for
4k
random,
writes
about
the
same
for
16
k,
+
EK,
it
might
be
faster,
but
it's
not
totally
clear.
D
There's
there's
a
lot
of
tunable
that
you
just
have
to
kind
of
change
things
and
performance,
or
the
behavior
seems
to
be
any
kind
of
different
than
used
to
be,
but
not
necessarily
create
better.
One
thing:
that's
so
showing
up
a
lot
in
testing
is.
D
In
the
first
factory
stumps,
what
we
see
is
flocking
and
bitmap
alligator
and
I'm,
not
sure
if
that's
necessarily
a
bottleneck
right
now
or
not.
It's
just
kind
of
hardly
tell
just
based
on
for
stumps,
but
it
is
showing
up
a
lot
so
since
invites
a
completely
mindful
of,
in
any
event,
we're
still
testing.
All
of
this
page
also
has
another
whipped
loose
or
we
sharks
full
branch.
D
That
kind
of
revert
some
of
the
behavior
changes,
but
saw
the
new
code
to
play
so
for
texting
that
as
well
Parmalee
agency
more
more
and
out
some
more
okay.
What's
going
on
here,
I
think
I
was
going
to
try
to
replicate
mixes
for
synchronous
right
test
findings,
but
has
this
new
PR
that
he
just
submitted
the
people
for
mothers?
I,
don't
have
a
documented
in
here,
but
if
there
he
eat
your
submitted
I
think
this
morning
that
hopefully
make
investment
see
that
helps
fun.
D
That
way,
you
can
tell
until
we
can
get
a
chance
to
see
those
results
of
chief
astronaut
before
Dominion,
also
potentially.
Clearly
we
have
some
additional
ETO
break.
Has
me
to
do
the
tests.
I
did
maybe,
a
month
or
two
ago
did
not
look
at
the
differences
between
testing,
where
we
have
say
like
four
plus
three
verses,
6
plus
two,
and
with
the
way
that
he
/
lights
work.
There
could
be
potential
performances,
differences
between
those
different
kinds
of
of
tests
looking
at
at
different
values.
D
E
A
E
E
Has
this
right
performance
of
writing
who
is
cracking
with
racial
cut
short?
Last
month
we
have
calculated
a
little
de
facto
brothers
of
rough
TV
I
ox
all
day
on
the
organist
overall
performance
or
white,
so
we
I
want
to
share.
Yes,
if
you
think
that
coming
testing
for
you
to
understand
our
context,
that
can
be
different
to
your
contacts,
sure
every
right
very,
very
quickly.
This
can
take
us
five
winning.
F
E
Well,
yes,
but
as
here
in
the
left
side,
we
have
the
same
test.
Writing
one
megabyte
file
in
five.
Has
this
Drive
in
the
left
and
on
the
right
will
have
the
same
test,
but
only
what
about
using
for
drugs
VV,
an
external
disk
unless
an
sss,
DVDs,
ok-
and
here
there
is
a
period
of
time
that
is,
that
rabbits
all
the
time
to
hype
cycle
that
repeats
all
the
time
and
the
meaning
in
performance
on
the
left.
A
screen
is
when
the
little
compact
rondalee
conversation
takes
place.
E
Okay,
so
the
problem
is
that
while
they
learn
video
conversation
takes
place
in
the
husband
right.
We
are
using
iOS
of
this
for
writing
on
the
on
the
road
to
be
partying,
so
these
ions
are
decreasing
the
items
for
writing.
Okay.
So
that's
quite
the
overall
bandwidth
during
10
minutes
period,
because
we
have
a
lot
of
objects.
In
this
case,
a
shops
of
the
performance
with
10
minutes
decrease
a
lot
and
then,
after
the
letter,
G
of
competition
increased
because
there
are
no
more.
There
are
no
more
right
on
the
rug
v.
E
So
after
making
several
change
that
can
change
in
some
way
how
these
competitions
that
pays
the
bill,
that's
the
site
of
the
day.
There
is
5
for
the
velocity
of
these
conversations.
We
have
the
same.
I
mean
this
is
a
loop,
maybe
short,
maybe
longer,
but
it's
always
like
this,
but
we
are
trying
to
use
test
for
real-time
video
ingestion
and
video
a
great
okay.
So
we
need
to
well
as
in
certain
performance
whatever,
but
the
minimum
performance
will
be
one
and
the
left
case.
E
E
E
E
E
D
B
B
E
B
E
H
B
E
Yeah
yeah
yeah
yeah,
okay,
the
disguise
a
speed
test
with
all
the
discs
overloads
all
the
time,
I
mean
yes,
the
limit
the
limitation
is
clearly
tell
me.
The
limit
is
imported
by
the
by
the
disc
in
the
case
of
hard
disk
or
the
all.
The
five
discs
are
overloaded
by
yolks
all
the
time
and
in
the
case-
and
in
the
second
case,
the
only
disc
of
a
lot
like
rice
is
the
beta
Beast.
Well,
it
is
now
you
we
can
achieve.
20
megabytes
per
second,
at
least
is.
D
B
Fix
that
okay,
I
right
so
mark
anything.
This
is
just
suggest
to
me
that
we
should
probably
try
to
replicate
this
turn
doing
it
to
it,
but
for
+1
or
for
close
to
whatever
it
is
and
I
rate
a
sponge
one
meg
and
just
see
what
it
cute
it
does,
because
I
would
expect
to
get
a
lot
all
up
on
that
out
of
out
of
the
disks
me.
B
B
Kind
of
I'm
skeptical-
oh,
that's,
that's
actually
the
issue
because
the
those
just
hit
shipped
in
the
note
it
will
affect
the
amount
of
rock
three
metadata
that,
if
sir,
are
excused
on
an
SSD
not
going
to
change
their
still
only
getting
20
max
per
second,
it
should
be.
Oh
yeah.
D
B
F
E
Ok,
we
have
HDL
a
to
the
hundred.
Are
you
these
are
either
price
per.
Second,
in
our
case,
yeah
I
owed
is
around
100
and
is
constant.
Ok,
it
is
profitable
and
as
viz
one
are,
they
I
abdominal
is
on
the
ad
one
with
the
data
exists.
So
after
in
combat
season,
we
have
the
growth
files
right
into
it
here
here
with
the
free
flash.
Ok
before
the
next
completas
edo
competition
takes
place
here.
The
price
of
the
puzzle
results
of
last
level
field
marketing.
What.
E
E
E
G
E
F
B
You
move
or
activity
on
to
the
under
the
SSD.
You
don't
you
don't
really
see
the
impact
on
the
right
to
put
my
list
is
pretty
small,
so
that's
that's
good.
At
least
it's
interesting.
Actually,
if
you
look
go
back,
I
think
two
slides
where
you
have
the
periodic
big
spike
in
Reed's
right
after
compaction
finishes.
That's
the
pressure,
that's
the
what
the
cash
and
validation
that
happens,
because
yet
the
after
compaction
finishes
you
and
validated
all
of
your
old
SST
files,
and
you
have
a
whole
bunch
of
mrs.
in
your
cash
can
to
read.
B
E
B
B
Those
blocks
on
the
block
device
and
what
we're
probably
gonna
see
here
is
that
when
compaction
happens
and
it
writes
out
all
those
new
SSDs,
it's
doing
it
using
the
buffered
stuff,
which
means
that
it'll
go
into
the
page
cache
and
then
the
reeds
right
after
that
we'll
hit
the
page
cache.
The
colonel
paid
cash
blue,
store
default,
buffered,
reads
and
default.
Buffered
rights
are
different,
so
for
the
rest
of
blue
store.
B
We
have
our
own
buffer
cache
in
blue
store
itself,
and
that
just
means
that
a
raid
0
sigh,
oh
that
comes
through
with
no
hints
at
all,
will
default
going
into
it's
our
buffer
cache.
Okay,
I
can
remember
if
those
are
on
or
off
by
default.
I
think
that
might
be
off.
Of
course,
if
you
hint
the
raid
us
to
buffer
or
cash,
something
than
it
will,
okay.
D
Rights,
okay,
so
blue
is
buffered,
I'll,
primarily
region
right
or
is
right
for
reason.
Rights
for
rock
CD
data
loose
or
default
buffered
right
is
the
default
behavior
that
that
you
do.
When
were
writing.
You
know
data
out,
I
really
but
yeah,
exactly
no
relevancy
each
option
but
may
be
overridden
by
him.
B
The
only
concern
is
that
in
if
the
right
is
on
by
default,
then
it's
usually
I
mean
usually
it's
not
needed
in
either
case
for
RVD.
It's
not
because
you
have
caching
happening
at
the
client
side
for
rgw.
It's
pretty
important,
at
least
in
the
reed
side,
because
you
have
multiple
gateways
that
might
be
getting
the
same
hot
objects
for
the
right.
It
might
be
a
good
idea
for
you
to
be
also
to
do
it
just
hope.
You
write
and
read
something
you'll
get
it
back
without
having
to
do
any
I/o,
but
I
presume.
D
B
B
D
A
B
Well:
okay!
Well
thanks
thanks
for
sharing
guys
Oh,
we'll
take
a
look,
because
that
doesn't
I
guess
it's
good
news
that
the
SSP
helps
but
I
understand
why
it's
only
20
megs
for
sex,
that
it
should
be
better
than
it.
B
E
Not
have
presentation
already
tuition,
I
I
would
like
to
send
you
this
afternoon
tomorrow
morning,
okay,
yeah
sure
sure,
whatever
they
have
in
time
yeah.
Thank
you
cause.
You
I
believe
that
the
problem
that
the
benchmark,
if
you
try
to
make
a
benchmark
during
the
first
minutes
or
the
first
hour
while
there
is
not
a
lot
of
information
than
the
rods
VV
database,
which
is
much
better
okay,
yeah,
so
that.
E
E
D
E
D
D
Just
emailed
verbosity
on
the
what
I
was
really
interested
in
there
was
the
Iraq
Stevie
data
on
compaction,
so
even
in
the
default
was
that
it
will
dump
out
all
the
state
about
compaction,
and
it
might
be
that
they
might,
they
might
say
something
interesting
about
homely
solve.
You
saw
over
time
the
percentage
of
time
that
was
stalled,
how
many
esse
files
were
in
levels?
You
know
all
these
kinds
of
things
that
that
might
tell
us
something
interesting
about
what
rock
CD
is
doing.
E
G
D
D
The
effects
of
compaction
become
much
worse.
The
disks
that
we
have
the
test
in
our
lab
at
Red,
Hat
I,
think
there
are
only
one
or
two
care
by
gift,
so
you
don't
have
at
my
face.
They,
the
beliefs
there
is
that
that
Zetas
scale
their
key
value
store
that
they've
been
working
on
intervene
to
do
so
will
do
much
better.
In
these
cases,
we
we
haven't
been
able
to
test
that,
but
tom
mack
steps
seem
to
indicate
that
that
it
is
much
better
with
very
large
discs
like
this.
E
Yeah
lately
I
would
is
kinetic,
is
very
efficient
because
it
this
is
for
recording
video
so
for
our
our
clients
is
very
important
to
make
use
of
the
food
storage
because
they
are
stirring
the
channel
for
several
days
or
weeks
or
even
months
and
bov
catalĂ .
So
the
college
is
very
important
of
the
day
they
all
plan
to
use
at
the
maximum
storage
capacity.
The
disc
so
aware
our
benchmark
should
be
when
the
disc
is
filled
with
data,
so
this
is
the
only
test
that
is
valid
sheriff.
E
D
Might
have
to
be
interesting
to
know
if
the
Rashard
lean
patch,
that
page
will
helps
you
guys
at
all.
The
idea
behind
it
is
that
we
right
now
we're
doing
way
too
much
work
whenever
we
Rashard
work
we're
both
basically
going
through
and
recomputing
all
of
the
shards,
even
the
ones
that
don't
change
when
you're
charting
happens.
If
we
could
be
interesting
to
know
whether
or
not
you're
seeing
yet
a
Quebecer.
E
What
we
also
have
tested
I
mean
this-
is
we
have
been
characterizing
that
the
bottleneck
of
where
our
realist
in
areas
with
this
test?
Okay-
with
this
a
small
test
with
like
this?
But
the
problem
in
fact,
is
that
in
the
real
cluster
with
600
disks,
what
happens
is
that
the
overall
performance
is
good,
but
the
rest
of
their
time
and
every
time
that
we
with
some
reason,
maybe
11
hundred
percent
of
their
risk,
takes
several
seconds.
And
this
is
a
consequence
of
the
disk
being
overloaded
by
these
complications
care.
Okay,.
D
H
D
My
understanding
is
that
they
have
been
primarily
focusing
on
the
nbme
case,
but
I
know
that
Alan
mentioned
that
they've
been
thinking
about
spinning,
just
as
they've
been
working
on
it
and
he
was
speculating.
Maybe
a
couple
of
weeks
ago
that
he
thought
for
lunch
thing.
That's
actually
better
than
rock
TV
does
I,
don't
know
that
anyone's
actually
tested
it.
But
that
was
kind
of
the
thought
that
when
you
have
very
large
ofes,
like
you
know,
six
terabyte
dis
for
an
detailed
like
this,
that
the
overhead
caused
by
rocks
BB
compaction
becomes
so
great.
H
E
Thanks
thanks,
we
have
these
guys.
What
happens
is
that
very
sophistical
part
of
the
dish
that
is
allocated
for
rock
to
delete
the
partition
size
physically
at
back
of
the
disc
and
the
different
parts
of
that
is,
is
allocated
for
we
are
video
or
real
data
and
when
the
competition
is
tend
to
write,
the
heat
of
the
dish
have
to
move
because
it's
grating
on
both
places
on
both
physical
separated
parts
of
the
disc.
At
the
same
time.
Okay,
that
is
not
it.
E
D
D
D
G
Only
the
two
sort
of
know
each
separate
issue
here:
one
is
that
with
polar
places,
Wow
and
DB,
you
know
during
compaction.
Performance
suffers,
oh
yeah,
maybe
there's
someone
that
we've
done
to
minimize
that,
but
I
think
that's
always
going
to
be
an
issue
because
you
that
that
happen
and
if
you're,
not
farming,
that
out
the
NFC
is
gonna
suffer
I.
Think
there's
also.
D
Yeah,
you
very
well
could
be
rate
one.
What
in
question
Nick
using
that
pages
PR
this
morning,
it's.
D
G
D
D
We
have
too
much
metadata
going
in
to
watch
TV.
It's
at
least
with
the
way
the
rock
CD
works
were
seeing
a
lot
of
rare
amplification,
we're
seeing
a
lot
of
read
and
write
traffic
in
rocks
Phoebe,
so
the
the
recharging
patch
that
page
did
I
think
is
is
good,
hopefully,
because
we
don't
resubmit.
A
lot
of
unnecessary
shards
tube
is
the
database
that
haven't
changed
at
all.
D
It
seems
like
maybe
Rex
media's
doing
more
force
based
on
this,
but
in
a
lot
of
cases,
performances,
no
differences,
that's
all
surprising,
but
if
it
were
possible,
that's
been
the
limited
testing
they
have
done
on
it
looming
a
teasing,
the
full
text-
and
there
may
be
other
interests
positively
need
to
work
through
so
I
mean
I'm,
very
glad
that
people
are
looking
at
this.
D
You
know
it's
kind
of
in
the
past
and
test
I've
been
running
in
SS
and
discus
and
running,
and
we
we
found
issues
that
were
looking
through
but
having
having
community
test,
but
this
is
really
really
good
for
us.
So
thank
you,
Nick
and
thank
you.
Everyone
to
nokia
what's
been
looking
at.
This
is
literally
going
to
help
us
so
and
I'm
quite
glad
to
see
you
yeah.
G
All
right
sort
of
thing
is
resolved
by
can
actually
start.
Looking
at
we've
got
like
a
semi
production
class
so
that
you
know,
if
ever
got
lost.
It
would
be
annoying
rather
than
a
disaster,
so
I
might
start
looking
to
actually
start
running
some
loose
Dorothy's
annette
to
give
it
a
bit
all
the
real
life
success.
Okay,.
D
Okay
Nick
another
thing
that
could
be
interesting
to
do
during
your
test.
If
you
don't
see
any
improvement
with
sages
PR
would
be
if
we
could
try
to
get
a
better
idea
of
what
the
ofc
is
doing
during
those
tests
that
that
their
purpose
one
option
but
I've
been
kind
of
a
unhappy
with
perfect
isn't.
It
seems
like
a
missing
big
thing
that
that
are
taking
up
time
and
I'm,
seeing
things
that
may
or
may
not
actually
be
relevant
in
the
results.
D
One
thing
that
could
be
done
would
be
to
use.
Gdb
do
grab
stack
traces.
You
know
a
bunch
of
different
samples
of
fact
raises
as
though
tests
are
going
on
and
just
see
what
the
osu
sending
this
time
doing,
while
while
you're
your
tests
are
ready,
we
might
be
able
to
get
an
idea.
There's
my
glock
contention
there
something
else.
You
two
unknowns
in
this
house.
The.
G
Market
see
initiative
from
zoar
a
more
performant
OS
performance
monitoring
sort
of
side
of
things.
It
was
very
similar
server.
Nakia
thing
is
that
the
disks
were
the
bottleneck.
They
even
I
was
only
doing
a
single
really
bright
at
4k.
It
was
like
every
single
right
it
was
requiring.
That
did
to
be
the
spinning
disks
to
acknowledge
before
that
was
passed
back
to
the
client.
So,
even
though
you've
got
the
wow
and
DB
on
FSC,
the
actual
data
part
was
require
a
knack.
G
D
Ok,
yeah,
let's
see
if
this
helps
run
and
if
it
does
great,
if
it
doesn't-
and
you
might
have
some
work
to
do
see,
what's
going
on
both
get
somewhere.
You
good
plan.