►
From YouTube: 2020-02-27 :: Ceph Performance Meeting
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
A
B
B
B
B
B
A
A
A
There
were
two
peers
that
closed
one
from
Radek
for
a
buffer
list,
a
small
bufferless
change.
I
think
I
think
originally
were
maybe
going
to
wait
until
octopus
for
that,
but
keeping
merged
it.
They
must
been
fine
and
then
Igor,
an
old
PR
of
yours
for
changing
the
auto
tuning.
The
Murray,
auto
tuning
for
blue
store,
got
closed
by
stale
bot
and
that
PR
was
actually
I
think
made
before
we
moved
the
everything
into
the
priority
cash
camp
into
its
own
separate
names
face
and
everything.
So
it's
not
super
relevant
at
this
point
anyway.
A
D
A
A
And
then
same
thing
with
your
memory
reduction,
your
PR,
it
looks
like
they
just
got
released
as
well,
okay,
and
then
that
that's
basically
it
haven't
had
movement
on
anything
else.
This
week
that
I
can
see,
although
I
think
folks
are
still
diligently
working
on
various
things,
the
atom
still
working
on
some
things,
with
his
rocks
TV
showing
PR
other
atom,
as
still
has
the
objection
sustained
out
there
and
all
other
things
add
many.
Any
news
on
Ahmad.
C
A
Then,
probably
my
stuff
for
the
the
blue
store
cash
to
avoid
double
cash
rights
for
or
double
caching
of,
oh
no,
it's
will
probably
be
reintroduced
and
after
the
Brock's
to
be
cash,
sharding
is
merged.
And
after
that
is
done,
then
I
can
go
back
and
try
to
get
the
the
age
based
bidding
code
kind
of
free
reintroduced
and
appear
me
for
that
and
hopefully
merged
so
kind
of
a
chain
of
things
dependencies
here.
But
hopefully
we
can
start
reading
progress
on
them
again.
A
A
post
octopus,
alright,
so
discussion,
topics
for
this
week,
I
just
have
two
quick
ones-
were
continuing
to
work
on
IO
500,
testing
of
stuff
S
on
officinalis
notes
were
I.
Think
now,
somewhere
around
twelfth
place,
maybe
even
getting
close
to
11th
place
on
the
10
no
challenge
list.
We've
seen
a
couple
of
improvements
there,
that's
with
two
extra
plication.
We
could
do
even
better
if
we
just
drop
down
to
1x,
which
I
think
some
of
those
scores
are
doing
with
other
distributed
file
systems.
So
we're
not
doing
bad.
A
You
could
do
a
lot
better
than
I.
Think
I
was
able
to
track
down
the
poor
balancing
we
were
seeing
in
the
heart,
that's
where
they
write
lots
of
files
from
many
clients
to
a
single
directory.
That
was
due
to
the
balancer
attempting
to
do
exports,
but
failing
the
can
there,
we
are
right
lock
on
both
the
file
lock
and
the
nest
lock
and
so
I've
been
trying
to
discuss
with
Patrick
and
the
rest
of
the
the
self-esteem.
A
What
exactly
that
means
how
to
get
around
it.
I
had
a
first
attempt,
Kevin
naive
attempt
at
at
just
experimenting
with
it
that
did
a
sleep
and
then
would
retry
grabbing
the
lock.
It
sounds
like
that's
probably
kind
of
totally
irrelevant,
because
even
when
you
sleep
and
retry
it
if
a
a
cap
has
been
revoked
as
the
MDS
can't
process
it,
or
at
least
that's
the
the
understanding
or
the
impression
I'm
getting
though
we'd
have
to
do
more
than
just
sleeping
there,
I
guess,
but
frankly,
I
don't
totally
understand.
A
Let's
see,
I
did
as
a
matter
of
testing.
All
this
see
fairly
high,
lock
contention
within
the
MDS
between
the
different
messenger
threads
and
the
other
ten
more
primary
threads
in
the
MDS
for
doing
work
and
based
on
the
fuse
testing
from
last
week.
That
was
horribly
slow,
I'm
kind
of
looking
at
potentially
writing
loops
of
us
back
end
for
MD
test
in
I/o.
Are
they
make
it
pretty
easy
to
do
so?
A
There's
actually
an
existing
Rados,
back-end
I
didn't
even
know
about,
but
it
can't
be
used
for
MB
test,
because
in
crate
directories
would
you
searches
or
anything
it's
just
basically
for
writing.
You
know
files
as
objects
or
I,
or
essentially
so
it'd,
be
very
interesting,
I
think
to
see
how
loops
ffs
did
without
fuse
attached
to
it,
if
it's
faster,
if
it's
not
how
it
compares
to
both
the
other
existing
Brennan's.
So
anyway,
that
was
kind
of
all
the
update
and
400-500
any
or
thought
some
neva.
A
The
good
news
here
is
that
David
Stockman
has
a
PR
that
we
think
will
help
pretty
significantly.
It
sounds
like
it
avoids
calling
dump
OSD
stats
or
dump
PG
stats
when
any
either
when
only
one
of
those
things
is
required.
I
guess
previously,
it
would
try
to
grab
everything
all
the
time.
So
hopefully
that
helps
quite
a
bit
but
we'll
see.
So
that's
all
I
have
any
questions
on
manager,
stuff.
D
A
D
E
D
D
F
A
D
D
Well,
there
are
some
test
cases
for
different
rides,
just
relevant
for
locator,
but
nothing
that
is
specific
to
hybrid
alligator
or
for
the
hollows
G,
just
just
unit
testing,
but
speaking
about
hyper
locator.
It's
the
safest,
leagues,
I
think
because
it's
well
I
actually
brings
in
some
modifications
to
the
yellow
locator,
but
well.
D
D
A
D
A
Cool
yeah,
if
you
I
mean
even
even
as
is
I,
guess
you
know,
if
it's
not
the
default,
we
could
probably
just
merge
it
it
to
me,
I
mean
it's
a
slower
on
the
reads,
but
if
we
you
know,
if
there's
a
subsequent
PR,
that
excited
might
be
fine,
especially
if
it's
not
default
be
nice.
If
we
figure
that
out
first,
but
Josh
would
do
you
think
that
you'd
rather
get
it
in
now
for
more
testing
or
or
you
know,
wait
until
that's
result.