►
From YouTube: 2018-08-16 Ceph Performance Weekly
Description
B
A
A
B
B
B
If
I'm
not
terribly
concerned
about
it,
I
think
what
we
should
do
as
we
do
the
messenger
to
implementation.
We
should
be
really
careful
there
about
what
calls
we
do
and
when
and
where,
and
in
that
case,
but
because
I
have
a
feeling
that
the
current
thinking
is
that
the
message
signatures
are
kind
of
going
to
go
away
with
messenger
to
you,
but
you'll
either
just
be
fast,
and
this
won't
be
an
issue
or
you'll
have
also
been
doing
encryption,
in
which
case
it'll
be
slower.
B
A
A
B
After
he's
gonna
review
that
that
seems
to
be
important
when
the
amended
a
two
servers
have
lots
and
lots
of
memory,
and
this
OST
shard
Duan
commits
one
done:
paying
has
stuck
a
little
bit
because
the
technology
tests
show
Villiers
left
and
right
and
he
isn't
able
to
reproduce
or
understand
what
that's
why
it's
happening.
We
can
run
it
through
again,
I
guess,
the
last
time
I
did.
It
was
just
like
50%
failed
or
something
so
quite
sure.
What
the
deal
is
there,
but
I
think
so.
B
B
A
Really
well
at
least
I
just
can't
put
some
update
stuff
in
there,
but
maybe
the
only
kind
of
interesting
thing
I'll
mention
with
my
stuff
is
that
there
is.
There
is
a
cost
to
switching
over
from
one
kind
of
like
favoring
one
kind
of
cash
to
another
kind
of
cash.
The
allocator
I
think
they.
It
keeps
like
groups
of
objects
in
a
particular
like
allocation
group.
You
know
for
similarly
sized
things
or
whatever,
and
it
looks
like
when
you
switch
from
like
predominantly
own
owed
to
predominantly
maybe
like
block
cash
stuff
in
memory.
A
A
C
B
B
C
Have
a
branch
that
adds
some
caching
to
the
entitiy
ad
Artie:
it
would
require
a
magic
fix
for
clearing
the
instances
of
that
class.
First,
however,
it's
my
github
it
it
should
provide
some
benefits.
Definitely
we
want
eradicate
all
the
overhead,
because
there
are
also
other
another.
There
are
multiple
sources
of
of
the
overhead.
C
A
B
A
C
A
B
I
would
be
I
would
hesitate
to
change
the
level
because
for
all
history,
sending
all
the
one
shows
you
as
messages
and
that's
usually
what
you're
asked
for.
So
if
we
want
those
not
to
show
by
default,
I
think
setting
the
level
to
zero
is
the
way
to
to
make
them
go
away,
but
but
again
there.
They
are
the
thing
that
has
been
useful
for
establishing
context
when
people
run
into
problems.
B
C
B
Yeah,
the
problem
is
that
the
operator
Aero
operators
are
context.
Sensitive,
aren't
see,
bug
little
sensitive,
they're
already
relatively
concise,
like
the
only
ones
that
really
matter
are
that
are
going
to
dominate
our
post
Diop
in
and
out.
Basically,
it's
the
off
the
most.
You
have
to
fly,
I
guess
some
sub
ops
but
I
think
they're,
already
relatively
concise.
B
B
C
I
will
focus
on
profiling
in
death
without
that
effects
I
mean
in
the
CBT
world
to
see
whether
the
numbers
match
what
is
expected,
maybe
maybe
trying
to
fight
some
of
the
overhead,
cashing
sorry
to
go
hungry.
It
just
adds
extra
complexity
anti-d.
I
just
took
a
look
even
quite
detailed
on
entitled
ERT
and
it's
mutating
it's
supposed
to
mutate,
so
we
will
need
to
also
take
care
of
coherency
between
actual
values.
C
B
C
C
B
Yeah
I
think
focusing
the
the
performance
testing
on
cases
where
office
turned
off,
though,
makes
sense,
because
that's
that's
where
we're
gonna
be
with
much
needed
food
and
hopefully,
hopefully
we'll
get
the
prototype
or
the
initial
draft
of
its
in.
Is
your
vacation
everybody's
out,
but
the
Ricardos
working
on
the
initial
pass
of
messenger
to
that
just
splits
out
the
protocol
enough
and
has
just
enough
implemented
so
that
it'll
actually
take
the
third
path
and
then
we
can
go
from
there.
C
B
Something
so
I
think
the
semantic
value
those
logs
is
the
same.
Regardless
of
what
the
underlying
protocol
is.
It's
still
the
same
messages
that
are
being
passed
across.
It's
just
the
wire
on
the
wire
protocol.
This
treatment,
so
I,
wouldn't
I,
wouldn't
want
to
map
the
duty
that
tied
it
to
you
together.
B
A
Page
going
back
to
debugging
and
disabling
debugging,
this
was
this
is
pretty
much
a
common
thing
in
the
lustre
world.
When
I
managed
lustre
storage
it
was
it
was.
You
know
known
that
if
you
care
about
performance,
just
disabled,
debugging,
I,
don't
I,
don't
think
it's
terrible
to
tell
users.
You
know.
C
B
C
A
Guess,
from
my
perspective,
it's
it's
worthwhile
to
make
it
better.
I
mean
it's,
it's
not
a
bad
use
of
time,
but
you
know
we're.
Never
gonna
get
it
to
be
the
same
speed.
So
at
some
point
we
have
to
just
say:
hey,
you
know,
disable
it.
If
you
care
yeah
yeah,
there's
a
lot
of
other
things.
We
could
work
on.