►
From YouTube: Ceph Performance Meeting 2022-08-04
Description
Join us weekly for the Ceph Performance meeting: https://ceph.io/en/community/meetups
Ceph website: https://ceph.io
Ceph blog: https://ceph.io/en/news/blog/
Contribute to Ceph: https://ceph.io/en/developers/contribute/
What is Ceph: https://ceph.io/en/discover/
A
All
right,
well,
maybe
it'll,
be
a
small
crowd
today,
not
very
many
people
here,
but
that's
okay,
there's,
maybe
only
a
couple
things
to
discuss.
So
maybe
this
will
be
a
short
meeting
all
right.
Well,
let's
get
started
okay,
so
I
saw
one
new
pr
this
week
and
that
was
from
you
ronin.
Do
you
want
to
talk
about
your
your
get
prefixed
pr.
B
Well,
the
performance
is
it's
a
bit
an
exaggeration,
it's
just
one
function
that
is
improved
by
15
percent
in
calculating
some
strings,
not
no,
no
impact
on.
I
don't
envision
any
impact
on
the
as
a
whole,
I'm
not
sure.
Okay,
maybe
I
should
have
removed
this.
A
Stack
nah,
I
wouldn't
don't
don't
worry
about
it.
It's
good
is
I
like
to
have
more
just
because
then,
if,
if,
if
we
see
some
kind
of
regression
somewhere
later
on,
if
it
might
impact
it,
that's
good
to
know.
B
A
Well,
good
good,
even
even
small
wins
are
our
wins,
so
it's
it's
good,
okay,
so
next
that
was
the
only
new
one
I
saw-
and
I
did
not
see
anything
that
had
closed
this
week
as
far
as
I
could
tell,
but
there
were
a
couple
that
were
updated.
A
So
there's
been
some
discussion
on
the
this
kind
of
work
in
progress,
pr
that
I've
got
for
rex
tv
store
to
try
to
kind
of
redo
some
of
the
rock
cp
settings
adam,
and
I
had
a
really
good
conversation
earlier
this
week
about
it.
He
expressed
some
concern
about
how
drastically
we'd
be
changing
the
some
of
the
settings
that
we've
had
for
so
long,
and
I
I
I
think
it's
valid
to
be
worried
about
that.
A
One
thing
he
was
particularly
concerned
about
is
the
read
amplification
impact,
which
it
is
it's
quite
possible
possible
that
some
of
the
new
settings
that
I
was
looking
at
may
make
re-amplification
worse,
but
at
the
benefit
of
making
rights
faster
and
maybe
actually
helping
a
little
bit
on
right
amplification,
although
that's
that's
still
kind
of
it,
it's
probably
close
either
way.
But
in
any
event,
you
know
all
this
tuning
is
is
really
a
balancing
act.
So
he's
gonna
look
at
a
couple
of
situations
he's
concerned
about.
A
We
did
see
some
evidence
that
rgw
bucket
index
listing
may
be
a
little
bit
slower,
so
that
may
also
be
some
evidence
that
the
m's
concerned
about
that
is
correct,
and
it's
just
a
question
of
how
do
we?
What
do
we
want
the
defaults
to
be?
Do
we
want
to
kind
of
favor
making
rights
faster,
possibly
at
the
expense
of
making
reads
a
little
slower?
How
much
right
amplification
are
we
willing
to
take
on
if
any
over?
A
What
we
have
now
another
thing
that
factors
into
this
somewhat,
especially
for
rgw,
is
compression.
It
looks
like
we
can
greatly
improve
right
amplification
if
we
compress
it's
probably
specifically,
the
o,
node
column,
families,
oh
and
and
kevin
offshoot
of
that
we
were
talking
about
whether
or
not
we
should
start
looking
at
per
column.
Family
default
settings
rather
than
global
settings
across
all
of
rocks
tv,
and
maybe
that
we
want
to
treat
different
calm
families
differently,
which
is
really
complicated.
A
A
He
also
brought
up
the
idea
that
we
could
maybe
look
at
submitting
a
full
request
against
roxdb
to
add
the
ability
to.
They
already
have
the
ability
to
look
at
how
many
tombstones
you've
iterated
over
versus
other
key
value
pairs
and
issuing
a
compaction.
If
you
see
too
many
tombstones
on
iteration,
what
they
don't
have
is
the
ability
to
execute
min
table
flushes.
A
If
you
see
lots
of
tombstones
in
a
round
table,
so
adam's
thought
is
that
we
might
be
able
to
extend
that
in
rocks
db
itself,
rather
than
like
tracking
deletes
on
our
end.
And
then
you
know
at
first
I
was
kind
of
thinking.
Well,
I
don't
know
if
you
want
some
anything
to
rock
cp.
It's
not
worked
out
so
well
in
the
past,
sometimes,
but
it's
still
a
good
idea.
Potentially
it
could
be
maybe
better
than
than
modifying
our
own
roxdb
key
value
interface
to
like
track
deletes
ourselves.
A
So
anyway,
lots
of
lots
of
kind
of
interesting
discussion
around
that
we'll
see
where
it
all
goes
all
right.
So
next
igor's
pr
for
getting
rid
of
statifes
updates.
I
think
that
got
more
fixes
all
and
then
this
pr
for
rgw
bench.
This
atomic
writer
benchmarking
dang
reviewed
that
and
there's
been
some
discussion
further
discussion
on
it,
and
that
was
all
I
saw.
Did
I
miss
anything
from
anybody.
A
All
right:
well,
then,
let's
see
I
already
kind
of
talked
about
this
forex
tv
tracking
deletes
and
alternate
approaches.
There
was
one
other
topic
casey
at
the
the
the
leads
meeting
mentioned
that
they
want
to
make
a
change
to
enabling
boost
val
grind
by
default
and
also
turning
off
the
system
boost
by
default.
So
basically,
this
change
would
mean
that
we
build
valgrind
ourselves
instead
of
using
the
system.
A
Better
sorry
build
boost
ourselves
rather
than
the
system
boost
and
that
we,
when
we
build
it,
we
enable
valgrind
support,
and
the
concern
here
is
that
it
might
have
a
performance
impact
mark.
Cogan
yesterday
did
some
testing
with
rgw
and
said
he
he
did
see
maybe
a
little
impact,
but
it
was
within
the
noise,
so
it
it.
You
know
it's
not
clear.
If
there
there
might
not
even
have
been
any
impact,
but
if
there
was
it
was
pretty
minimal.
A
A
Apparently,
there's
the
way
that
this
works.
It
just
results
in
a
little
bit
of
extra
cpu
overhead
for
certain
parts
of
valgrind
or
started
sorry,
certain
person
boost
so
oh
hey,
casey
just
came
so
casey.
I
forgot
what
was
it
in
the
the
boost?
Valgrind
change,
that
where
you
were
seeing
a
potentially
slower
performance
or
more
overhead,
it
was
something
with
context.
C
Hey
mark,
so
that's
we're
just
trying
to
evaluate,
in
general,
linking
against
boost,
builds
with
the
valgrind
flag,
enabled
the
the
place
where
that
flag,
fixes
rgw
is
related
to
the
boost
context,
library.
A
C
A
C
A
Yeah,
that's
kind
of
what
I'm
guessing
too.
I
was
just
mentioning
the
yeah,
the
testing
that
mark
hogan
did
and
then
that
I'll
I'll,
just
quick,
spin
up
a
test
cluster
with
with
the
the
parent
version
from
maine
and
then
that
that
I
guess
it's
not
even
a
commit
so
just
turn
it
on
and
off
and
see
what
happens.
A
And
I
on
that
note
I
think
that's
that's
all
I
have
today
folks
so
I'll
open
it
up.
Does
anyone
have
anything
they
would
like
to
talk
about
this
week?.
A
All
right,
then,
have
a
great
week,
everybody
and
see
you
next
week.
A
You
you
want
to
just
keep
staying
in
here,
or
do
you
want
to
do
a
separate
room.