►
From YouTube: 2019-06-19 :: Crimson SeaStor OSD Weekly 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
C
A
So
to
me:
well,
wouldn't
be
joining
us
she's
still
on
vacation
last
week,
I'm
mainly
to
fix
the
flavor
and
give
the
true
test
drive.
We
see
be
cheaper
title
that
way
somehow
filter
to
reports
tell
us
updated
status
to
to
managers.
So
it's
a
third
to
the
test.
Suite
failed
to
wait,
wait
for
ring
and
active
PG.
Do
I
need
to
fix
that
first.
Otherwise,
I
need
to
disable
this.
This
check
just
for
crimson
test,
but
I'd
prefer
to
fix
it
on
the
conservative
side.
I
think
that's.
D
A
D
B
C
B
B
Finally,
we
have
merging
the
merging
of
adjacent
buffer
pointers
work
as
expected:
I
updated
the
PR
and
posted
also
a
new,
a
new
iteration
of
the
result
of
new
iteration
of
performance
comprising.
Basically,
there
is
small
improvement.
I
believe
it's
because
we
are
not
creating
temporary
buffer
pointers
over
and
over
ii
am
working
right
now
is
integration
with
integration.
The
performance
testing
with
CI
CD
I,
made
a
small
er
small
change.
That
I
believe
will
be
useful
for
larger
testing.
B
B
B
C
B
D
Is
so
exit
means?
Something
else
can
go
into
the
pipeline
there
exclusive
right.
D
D
This
is
a
yeah
me
too,
unless
there's
some
other
metaphor
that
works,
but
this
there
may
be
pipeline
phases
that
aren't
exclusive.
For
instance,
you
could
you
can
enforce
order
of
Exodus
order
of
entering
without
enforcing
exclusivity.
So
that's
why
I
chose
this
terminology
that
might
be
useful
later.
B
D
That's
right,
that's
right!
I
I
didn't
add
comments
to
the
other.
I
just
changed
the
name.
Did
you
feel
that
what
was
it
I
think
I
called
it
user
private
handle
or
something
do
you
feel,
but
that
was
better
yeah.
B
D
Okay,
it's
I,
don't
think,
there's
a
huge
hurry
unless
someone's
blocked
on
actually
people
may
be
blocked.
On,
from
my
perspective,
it's
essentially
done.
I
was
going
to
address
whatever
comments
come
up
of
course,
but
I
feel
like
it
kind
of,
does
the
job.
For
now
it
will
definitely
change
going
forward,
like
I
just
feel
like
this
is
some
as
much
as
I
maybe
get
out
of
the
current
or
the
state
of
encoder
until
we
started
the
wedding
tonight,
the
next
set
of
stuff,
so
the
question
really
is
given
the
last
week.
D
D
D
C
D
The
idea
I
could
I
suppose
have
combined
them,
but
I
felt
that
that
would
not
be
as
ideal.
We
don't
really
want
client
request
blocked
behind
period
of
us,
especially
on
it's
it's
unclear
in
general,
only
one
of
the
two
is
gonna
be
used
because
you
don't
get
client
requests
from
other
OS.
These
you
don't
get
period
for
question
clients,
so
in
that
sense
this
is
unnecessary.
We
could
have
two
different
OSD
session
or
whatever
I
called
it.
Os
do
whatever
I'd
stuck
on
the
connection.
C
D
C
D
D
D
But
but
mainly
been
trying
to
sort
of
get
a
handle
on
blue
stores,
latency
properties
this
week,
so
I'm
abundance.
That's
what
your
person
I
expect
this
week
to
address
any
comments
they
come
up
with
us.
A
C
Okay,
so
the
first
thing
is
the
perf,
the
the
perfe
movement
is
implemented,
but
it
hits
another
bug
off
of
pre-emptive
socket.
Closing
though
so
it's
after
it
runs
some
second,
it
will
close
the
connection
preemptively
and
there
will
be
some
bugs
so
I'm
writing
up
a
fix
to
address
this,
but
this
will
impact
the
performance
so
I'm
wondering
to
implement
another
version,
but
before
that
I
need
to
implement
socket
unit
tests.
C
The
C
star
code
is
continue
evolving
right.
Yes,
so
with
this
socket
level
test
we
can
find
out
if
the
code
behavior
is
different
because
they
are,
there
are
a
lot
of
things.
I
found
that
it
is
not
documented
how
to
use
the
interface,
though
it
might
be
changed
in
the
future,
and
we
can
find
it
at
the
first
place.
Using
this.
C
B
D
C
D
B
D
D
Not
in
this
case,
this
is
that
code
hasn't
been
modified
and
I,
don't
think
it's
worth
going.
I
can
change
ago.
Just
now,
but
that's
just
using
the
perfectly
normal
blog
standard,
Princeton
people
sports
exception,
propagation
stuff.
So
when
you
throw
an
exception,
the
the
trampoline
that
spawned
that
panda
catches
it
does
some
stuff
and
propagates
it
down
to
the
future
chin
as
appropriate.
D
D
A
A
A
A
D
A
B
C
B
C
I
think
we
can
first
do
sound
test
performance
test.
Maybe
get
some
data
out
of
the
two
different
implementations
and
then.
A
C
C
C
So
do
you
think
of
the
two
different
ways
to
do
that?
I
think
your
your
implementation
is
as
conceived,
has
more
control
of
how
the
memory
is
allocated
the
other
agreed
and
but
is
a
little
simpler
and
and
continue
to
use
the
existing
implementations
and
maybe
add
adaptive
prefetching
to
improve
the
buffer
sizes,
but
still
maybe
need
more
copies,
but
I
don't
know
how
the
performance
will
look
like
or
how.
C
B
So
we
can
the
mem
coffee
between
carbon
and
user
space
with
all
you
can
always
put
a
one
huge
buffer
and
that,
but
for
DP
DK
we
will
need
to
avoid
the
to
avoid
fixing
fixing
with
transforming
the
scattereth
output
of
network
stack.
We
will
need
to
to
avoid
the
BEM
copyrighted
to
it.
We
will
to
use,
would
need
to
branch
we
would
need
to
have.
We
will
need
to
avoid
using
till
it
exactly
based
on
some
kind
of
happy
detection.
If
it's.
B
If
it's
look
like
if
it's
a
DP
decay
stock,
then
don't
care
about
alignment
and
go
with
read.
Otherwise,
we
go
with
exactly
second
and
specify
the
alignment
I'm
I'm,
feeling
that
it's
it's
affecting
well,
it's
controlled
to
the
abstraction
pistol,
put
a
lot
about
it,
but
just
take
a
look:
they
try
to
abstract
the
way
how
application
interact
with
network
stacks
I.
C
Think
that's
it's!
It's
only
our
user
case
that
doesn't
that
can
allow
that
using
VB
DK
the
data
can
be
fragmented
and
we
we
can
just
use,
read
fragmented
because
we
don't
require.
It
is
not
a
requirement
so
really
exactly
I
I
think
it
is
a
requirement
to
to
make
the
memory
continue.
Continuous
continuously.
This
is
a
requirement,
so
it
even
for
DVD
K,
if,
if
it
is
required
to
be
continuous,
I
should
be
increment
like
that.
So.
D
B
Rich
well,
I
got
impression
that
read
fragmented
is,
is
a
concept
from
our
abstraction
of
input
of
of
sea
stars
socket
in
sister
in
input
stream
system.
Recent
fragmented
there
is
read,
there
is
consumed
and
receive
it
exactly.
I
can
I
can
recall
all
those
three
methods
right,
strap
to
the
payload
extraction
from
socket.
C
C
My
impression
is,
is:
is
ceased
our
input
stream
up
that
sea
star
when
we
read
read
from
sister
network
and
sister,
provide
to
two
kinds
of
interfaces
to
us.
One
is
read
and
another
is
read
exactly
and
the
read
allows
us
yes
us
to
to
allows
the
buffer
to
be
fragmented
and
from
the
network
duck
and
exactly
reading
exactly
means.
We
need
the
power
to
be
continuous,
though
it
doesn't
mean
that
it's
just
a
user
requirement
doesn't
care.
What's
the
the
network.
B
It's
it's
a
hint
because
well
at
that,
at
the
messenger
or
level.
Basically,
we
won't
never.
Okay.
We
don't
want
to
know
about
specific
requirement
of
not
only
above
objector
of
particular
device
driver
of
object
star
it's
very
far,
it's
put
very
far
away
from
from
from
messenger
and
fixing
the
memorial
out
based
on
requirement
of
very
very
far
and
tithing.
Well,
it's
a.
C
B
Take
a
look
just
take:
let's
take
the
AI,
oh
right
of
Brewster's
kernel
device
as
an
example
terminal
device-
AI.
Oh
right,
it's
we
are
thinking
about
about
the
alignment
I
be
only,
but
only
because
of
it
because
in
Brewster
we
are
doing
as
we
are
using
a
I/o
and
direct
IO
of
Linux.
It
puts
some
extra
requirements
on
the
alignment
of
memory,
but
just
take
a
look.
We
started
with
the
read
exactly
second
proposal.
B
We
started
thinking
that
it
must
be
one
huge,
contiguous
segments,
just
one
piece
of
memory:
that's
one
chunk
of
memory,
but
it's
not
true.
The
a
directly
strict
requirement
of
the
of
the
a
I/o
right
in
Brewster
is
that
your
that
you
have
lived
up
to
a
restricted,
limited
number
of
segments
in
your
scatter
data
list
and
those
segments
needs
to
be
aligned.
It's
weaker
required,
but
at
the
level
of
messenger
for
for
I
believe
for
long
time.
B
C
A
C
B
B
In
messenger
we
have
the
same
the
same
the
same
minute.
There
are
some
short
headers
like
32
bytes
that
need
that
must
be
contiguous,
but
even
if
the
we,
if
we
even
if
we
got
if
we,
if
we
good
buffer,
at
least
if
we
got
a
scatter
cutter
from
Network,
we
would
make
our
coffee
to
get
single
chunk
to
make
it
flattened
on
our
own,
because
we
want
to
reinterpret
cast
reinterpret
cast
to
a
pointer
to
some
pockets
track
and
I
had
sister
Fox
have
had
familiar
issues
and
that
and
then
they
introduced
some.
B
They
introduced
their
it
exactly
as
some
kind
higher
level
utility
some
kind
of
helper.
Yes
and
I
bet
I
all
the
reading
huge
chunks.
It's
not
the
use
case
of
it
all.
It's
not
useless
for
it,
because
it's
terribly
in
affiliate
each
implement
in
the
unstable
way.
If
it
needs
to
go
if
it
needs
to
to
work
on
adjacent
and
on
separated
chunks
till
it
works
on
single
junk,
it
works
on
small
small
till
it
searched
small
requests.
B
They
are
not
using
the
acceleration,
they
are
just
making
recursion
and
it's
not
tired,
we're
optimizing.
There
is
no
optimization
of
the
distribution.
It
makes
that
it
makes
that
each
recursion
recursive
call
it's
reflected
in
our
separated
stack
frame.
It's
terribly
inefficient
if
you
are
using,
if
you
use
it
really
big
things
for
small
things
like
headers,
it's
very
good.
B
C
C
B
B
To
cases
the
primary
one
I
don't
think
that
somebody
that
bad,
that
somebody
will
use
tippity
K
without
a
spirit,
it's
very
unlikely
to
me.
If
you
are
get.
If
you
are
accessing
your
network,
a
very
likely,
you
will
have
also
a
speedy
K.
That's
my
assumption
and
NPD
they're
low
level
and
via
me,
driver
of
aesthetic
a
is
is
good.
With
is
fine
with
handling
with
not
having
the
allied
the
second
flight.
There
is
no
single
word
about
in
the
documentation
about
they
need
to
have
those
stealth
segments
aligned
to
4k.
B
B
Most
cases
but
I'm
entirely
sure
it
always
need
to
be
to
be
made.
That
way.
Requirement
we
saw
in
the
documentation
is
that
as
a
single
chunk
of
the
scatter
gutter,
you
are
passing
to
nvme
driver,
it
must
be
physically
concluded,
which
means,
if
you
are
using
the
Ambo
I
bet
we
are
the
emphasize
in
the
PD
in
the
PDK,
is
around
half
of
page
and
I
will
be
very
surprised.
It's
not
a
line
to
page
boundary.
B
B
Of
course
we
are
not.
We
don't
want
to
force
people
to
use
a
speedy
case
not
at
all,
but
even
if
this
would
stay
even
if
somebody
would
go
with
DP
DK
and
for
networking
at
Cardinal
for
storage.
In
that
way,
we
will
be
doing
basically
the
same
words.
The
bluster
Cardinal.
The
kernel
device
is
doing
right
now,
after
least
got
buffer.
It
has
a
special
facilities
for
pool
filling
requirements
of
Linux
direct
of
nukes
direct
IO.
B
D
A
B
C
B
B
C
C
B
But
it
would
that's.
The
problem.
I
meant
a
lot
of
you
committee
in
the
interface
good
and
forth
you
it
would
well.
This
approach
would
enforce
you
use
different
different
methods,
different
language,
basically
of
interacting
with
network
stuck
basing
on
on
some
kind
of
I,
believe
a
private
implementation
detail
of
sister.
You
would
need
to
expose
to
application.
Any
sister
will
need
to
expose
with
me
to
tell
just
unwrapped
9bk
and
do
we
need
to
tell
application,
which
particular
networks
that
is
being
in
use
and
the
application
would
need
to
branch
on
Tom.
A
B
B
B
Isn't
it
defaulting
to
positive
stock
or
even
well,
even
if
it's,
even
if
it's
I
believe
this
behavior
actually
will
be
a
private
implementation,
detail
your
sister
and
still
exposing
it
the
occasion?
It's
it
just
breaks
the
installation
of
responsibilities,
I
I,
don't
think
the
idea
of
using
it
the
same
instead
of
just
saying
this
is
its
yeah.
B
C
A
B
B
A
As
an
argument
to
a
common
argument,
or
as
an
option,
our
settings
to
sister
and
enable
it
this
part,
if
we
the
fails
to
do
so,
it
fails.
So
it's
an
otter.
In
a
note,
it
is
not
internal
to
sister,
it's
interface
to
the
users.
Use
user
of
a
sister
I
would
I
would
strongly
suggest
to
you
to
take
a
look
at
I'd,
react
at
OCC
and
a
search
our
score
using
skip
decay
to
see
how
it's
used
and
how
its
initial
initiated.
A
B
B
B
B
Undeniable,
its
you
are
baby.
You
are
trying
to
make
this
decision
in
messenger,
yes,
requiring
knowledge
that
is
available
in
not
even
in
objects
turn
in
object,
stores,
30
driver,
it's
very
far
away
and
just
a
cloak.
It's
just
the
AI.
Oh
right,
it's
a
perfect
example
of
it.
You
put
it
exactly
using
exactly
you
would
put
more.
You
will
put
click
the
requirements
on
network
side
that
are
absolutely
necessary
to
handle
to
handle
the
case.
B
B
Light
at
4k
and
Bravo,
it's
just
one
driver.
Even
now,
even
in
Bluestar
we
have
three
of
them.
One
is
SP
decay,
one
is
prepping
them.
One
is
an
is
for
Linux
a
IO,
it
would.
If
we
would,
if
we
would
want
to
make
messenger
a
good
choice,
we
would
need
to
Inc.
We
will
need
to
interconnect.
We
will
need
to
have
an
interface
between
objector
and
well
particular
driver
of
red,
star
and
passenger.
C
A
B
Believe,
that's
the
reason
why
we
have
that
fault.
While
we
have
the
alignment
built-in
protocol,
it's
some
kind
of
hint,
that
this
particular
that
this
particular
mem
that
this
particular
payload,
it's
our
data
and
date,
it's
nice,
not
it's
nice
to
have
the
data
properly
aligned
because
it's
usually
money
storage
drivers
are
like
to
have
data
line.
Okay,
but
it's
only
nice
to
have.
If
the
network
stack
is
not
if
the
network
stack
is
not
able
it's
if
it's
not
it's
not
efficient
or
if
it
required
some
extra
work
at
the
level
of
network
stock.
B
A
B
This
R
here
for
K
aligned
database,
however
DP
UK,
would
need
to
do
extra
memory
copy
to
fulfill
this
this
hint,
so
it
ignores
it,
it
passes
it
passes,
it
produces
a
scatter.
Gutter
is
with
you
and
it
passes
it
to
object,
store
uses
pick
like
be.
The
objector
will
won't
need
to
do
anything
to
fix
the
incoming
batteries
because
it
will
be
speedily
cool.
B
B
However,
the
mem
copy
will
need
to
do
the
extra
work,
regardless
of
of
the
of
the
stack.
All
we
are
doing
is
basically
postponing
is
basically
doing
its
postponing
the
mem
copy
till
they
reach.
Inc
data
reached
the
distance
driver,
which
has
the
best
knowledge
about
the
the
intended
shape
of
the
input
storage.
If
you
would,
if
we
would
have
done
our
our
extra
work,
the
rectifying
of
input
of
network
data
at
the
level
of
of
the
network
stack,
we
will
move
on
it.
B
B
B
A
Yes,
I
think
I
think
I
tried
to
to
understand
the
radix.
The
opinion
I
think
is
the
whole
idea
is
to
try
to
encapsulate
the
difference
between
a
connect
to
the
couple
that
needs
picking
over
the
object,
store
and
and
what
the
messenger
can
offer.
So
the
idea
is
the
truth,
this
time
to
make
a
decision
later
and
let
the
and
let
the
object
started
to
to
to
to
to
what
is
a
good
for
for
for
it.
A
So
what's
the
lesson
you
can
do
is
to
to,
instead
of
to
make
the
decision
on
behalf
of
the
object
store
to
to
change
them,
whether
you
can
help
with
with
the
minimum
efforts?
What's
the
minimal
effort
we
can,
we
can
take
the
fragment
and
alignment
aligned
on
line
segments.
That's
what
we
can
do,
what
we
can
have
with
with
the
least
the
least
effort
and
the
way
we
the
best
without.
B
B
B
B
B
B
B
Well,
I
can
recall
it.
I
proposed
a
problem:
preserving
the
4k
hint
with
the
with
the
discovery
catalyst
just
for
this
encountering
and
an
encryption
it
would
make.
It
will
prefer
the
DP
decay
case.
It
will
mean
that
if
we
are
doing
with
the
encryption
and
the
artistic
without
qat,
or
something
like
like
that,
we
would
need.
We
will
use
this
transfer
matter
to
to
get
the
4k
alignment.
B
In
such
case,
the
target
the
destination
buffer
for
decryption
will
be
4k
elect
because
there
is
very
little
electricity.
No
difference
in.
There
is
no
difference
in
aligning
in
allocating
aligned
buffer
burst,
allocating
unaligned
buffer.
It
can
be
neglected,
the
cost
of
doing
that
can
be
neglected
and,
as
a
result,
we
are
getting
the
nice
4k
Imin
that
can
be
useful
for
for
for
object
or
drivers.
We
are
getting
it
for
free,
just
like
in
the
case
of
POSIX
tack.
We.
C
B
C
Yes,
because
we
have
the
hint
it's
just
hint
and
if
that
doesn't
implement
it
such
hint.
So
that's
why
we
need
not
at
the
moment.
Yes,
so
that's
why
we
need
to
propose
our
own
implementations
and
the
previously.
Why
I'm
implementing
my
really
exactly
true
because
I
think
that's
a
requirement
is
not
that's.
Not
a
hint
I
see.