►
From YouTube: 2020-03-05 :: 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
A
A
A
That
should
allow
us
to
have
more
parallelism
during
writes
and
be
able
to
improve
performance
when
writing
to
a
single
bucket
with
lots
of
shards
and
lots
of
points
all
right.
So
that's
basically
it
as
far
as
when
I
saw
that
had
really
moved.
I
didn't
make
it
quite
through
the
whole
new,
no
movement
list
there,
but
don't
expect
a
whole
lot
of
other
stuff.
Probably
got
looked
at
this
week,
so
any
PRS
I
missed
from
anyone
or
any
updates.
A
All
right,
so
the
only
thing
I
really
have
for
this
week
myself
is
I,
have
been
continuing
to
work
on
testing
sefa
fests
with
IO
500
and
trying
to
get
our
scores
up
so
that
we
have
a
good
result
that
we
can
give
to
them,
and
one
of
the
issues
that
I
had
run
into
previously
was
that
the
kernel
client
was
showing
slow,
read
performance
for,
like
large
sized
reads
and
I
couldn't
really
figure
out.
What
was
going
on
can't
talk
to
some
of
the
guys
about
it.
A
A
So
it's
at
least
in
that
specific
case,
quite
a
bit
faster,
it's
quite
a
bit
slower,
seemingly
so
far
for
file
operations,
so
creating
lots
of
files
deleting
files,
static
files,
pretty
much
everything
we'll
see.
What
happens
why
increase
the
number
of
MDS
is,
but
it
does
appear
to
be
lower
for
that
kind
of
thing
right
now
the
code
is
closing
the
connection.
A
Every
time
a
file
is
closed
and
I'm,
not
sure
if
that's
actually
the
behavior
I'm
supposed
to
be
doing
if
I'm,
if
that's
correct,
I
am
kind
of
going
based
on
the
notes.
I
saw
in
lips
ffs
Ted
H,
so
anyone
that
knows
that
if
you
could,
let
me
know
that
would
be
fantastic
but
anyway,
this
is
these.
Are
the
early
results
I'm
seeing?
It
needs
a
pad
here
so
kind
of
interesting?
A
B
B
You'd
expect
that
it
doesn't,
it
doesn't
need
to
go
through
the
you
know:
Colonel
use
your
space
layer
and
the
reason
that
fuses
slow
is
not
because
the
user
space
code
is
slow.
It's
because
the
fuse
like
operations
layer
requires
a
lot
of
serialization
and
bundling
and
kernel
the
outer
space
halls
and
stuff
so
yeah.
So.
A
A
A
A
A
We
can
play
tricks,
then
we're
like
technically
I.
Don't
think
we're
supposed
to
do
like
create
the
directories
upfront
for
directory
pinning,
if
we
do
it
as
part
of
our
lib
stuff,
has
back-end
MD
test
in
IOR
I
think
we
can
just
like
set
the
whatever
we
create
the
directories.
You
can
just
set
the
the
outsiders
right
there
ourselves
and
it's
fine.
So
we
could
like
implement
our
own
like
custom
round
robin
pinning
right
in
the.
A
C
A
A
Remember,
I
was
also
going
to
do
this
testing
for
rgw,
but
for
one
of
her
reasons,
maybe
I
did
do
it.
I
just
put
them
in
here
or
maybe
I
didn't
get
around
to
it,
but
but
this
is
at
least
four
little
bar
biddies.
So
you
know
it's
something.