►
From YouTube: 2019-12-17 :: 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
Regarding
to
the
allies,
blue
store
appear
to
me.
I
think
it
would
be
helpful
if
you
can
commit
message
with
with
a
strategy
and
what
that,
what
a
change
you
made
instead
of
a
single
title,
that
would
be
helpful
for
the
reviewers.
B
I
have
sent
out
the
release
version
test
result,
so
it
seems
in
highway
workload
and
crimson.
Ost
have
a
better
performance,
but
when
the
workload
isn't
is
only
one
one
job
job
number
is
equal
to
one.
When
the
block
size
is
a
little
bigger.
B
So
when
I
build
the
release
version,
I
found
that
the
c,
the
c
star
c
star
used
use
the
default
allocator
for
the
back
version
and
use
the
c
star
allocator
for
the
release
version
and
when
it
used
a
c
star
allocator,
it
can't
allocate
the
pages.
B
Page
number
81
92,
so
it
will
give
a
error
for
the
in
this
this
function.
I
will
paste
here.
B
B
And
so
it
will
return
a
noun
pointer
for
here
so
couldn't
allocate
the
pages
for
the
blue
star.
It's
some
some
called
a
bit.
I
think
it's
called
the
bitmap
for
blue
fs,
something
like
that.
So
I
just
make
it
use
the
default
allocator
to
build
already.
B
I'm
not
sure
I
just
I
did
not
debug
on
it.
I
just
found
this
branch
the
condition
meet
for
the
further
when
the
aliens
when
the
blue
star
allocates
some.
Let
me
check.
A
Determine
regarding
to
the
performance
regression
degradation
of
when
the
the
block
size
is
greater
than
like
4k,
I
think
that
behavior
is,
is
expected
and
should
be
addressed
by
a
great
exchange
of
into
battlefield
factory.
D
Input
buffer
factory
system,
when
you,
when
you
use
the
posix
stack
of
c
star,
to
read
data
from
from
network
like
the
content
of
file
of
radius
object
during
during
write
operation,
then
you
do
that
a
bit
in
efficiently.
It's
because
sister
always
use
eight
kilobyte
chunk
to
interact
with
a
kernel
to
fetch
the
data
from
kernel
buffers
into
user
space
buffers,
and
then
it
does
another
copy
just
for
the
buffer
linearization,
it's
not
so
efficient.
D
D
Exactly
we
have
two
possibilities:
dpdk
user
space
drivers
which
are
absolutely
free
of
that
issue,
or
we
we
can
go,
we
can
push
for
the
buffer
factory.
The
code
is
actually
ready.
It
will
be
right
here,
political
pressures
on
in
the
industrial
community
to
get
this
feature
merged.
B
D
D
Yes,
but
but
to
do
that,
we
would
need
to
change
sister.
D
D
Have
the
no
no
classic
classic
is
using
machine.
No,
no,
no
classic
osd
is
absolutely
free
of
this
efficient
socket
reading
it
messenger
uses
completely
async
messenger
of
classical
osd
uses
completely
different
strategy
to
interact
with
gardener
when
it
comes
to
data
reading,
it's
three
of
the
issues,
so
it's
at
the
moment.
It's
not
april
to
april
comparison.
B
But
currently
a
single
messenger
used
a
pulsing
stack
right.
B
The
kernel
I
use
the
kernel
tcp
stack
and
they
have
the
copy,
they
need
the
column.
A
Outside
each
code,
re
try
to
read
from
a
kernel
like
for
4k
each
time,
but
but
in
case
of
classic
osd,
it's
done
in
for
such
such
kind
of
it
doesn't
have
have
such
kind
of
limitation.
Actually,
I
think
using
has
a
change
to
to
to
to
do
this
in
an
adaptive
way,
but
that,
but
it
turns
out
that
performance
improvement
is
not
sometimes
the
performance
improvement
is
not
visible,
so
that
that
change
was
put
it
down
by
me,
but
I
think
it
would
factor
in
profit.
B
B
Yeah,
so,
oh
so
that
is
the
the
performance
is
not
good
for
christmas.
I
see
when
the
block
says
it's
over
or
4k,
maybe
over
one
big
bath.
D
We
are
terrible
for
bigger
chunks.
B
Oh,
I
see
I
see
so
so
who
is
working
on
it
using
we'll
fix
it.
C
D
There
are
two
there
are
two
waiting
appears:
one
is
for
our
theft,
slash
sister
ripple,
which
brings
the
interface
extensions.
Another
one
is
for
crimson
to
to
ma
to
make
it
use
to
let
it
use
the
new
api.
D
Likely
we'll
need
to
do
well.
During
the
the
sister
summit,
I've
talked
with
avi
on
the
input
buffer
on
the
input
backer
factory
concept.
He's
fine
to
to
reiterate:
we
can.
We
can
start
even
even
now.
D
Still
the
best
approach
is
to
that
is
to
decide
on
our
side.
What's
the
priority
between
between
driver
among
drivers?
Do
we
want
to
use
dpdk
or
posix
stack
will
be
our
first
choice
when
it
comes
to
drivers.
If
posix
stack
is
first
choice,
we
need
to
fix
the
performance
very
soon
much.
E
Sooner,
I
think
we've
decided
to
go
with
project
stack
for
now.
I,
by
the
way,
I
saw
a
response
by
one
of
us
db,
guys
regarding
a
native
stack,
and
he
specifically
said.
Yes,
there
are
problems
there
and
we
are
not
sure
we
are
not
testing
it
and
which
we're
just
making
sure
that
it
more
or
less
works,
but
there
are,
there
are
problems
there.
There
are
bugs.
If
you're
going
to
use
it,
you
better
debug
it
as
someone
else.
They
ask
that
question
in
the
group.
I.
D
See
I
see,
on
the
other
hand,
we
linking
made
the
testing
the
performance
comparison
between
native
and
posix.
It
turns
that
the
ipc
it
turns
out
that
the
apc
factor
for
for
native
is
significantly
higher.
It's
not
just
10
percent
or
something
that
is
it's
much
higher
around
40
or
maybe
more
so
it
still
might
be
worth
the
burden
it
into
the
side.
D
D
B
Okay,
so
I
will
take
a
vacation
next
two
weeks
for
the
christmas
holiday,
maybe
escape
at
the
the
weekly
meeting
for
next
two
weeks.
A
E
D
Just
maybe
one
thing
from
from
the
previous
week:
it
is
the
fix
for
a
generator
that
allows
for
in
for
integration
with
sister
the
weave.
Another
news
is
on
sister
site:
all
patches
are
merged.
D
Yesterday,
I've
sent
a
pr
to
to
pick
to
let
pick
them
up
we're
already
using
updated
sister
right
now,
I'm
working
on
two
bugs
related
to
car
related
to
the
the
raiders
that
came
from
technology,
but
I
will
be
able
to
to
switch
back
to
to
the
exclusive
lock
of
rbdson.
D
I
plan
to
work
specifically
on
the
notify
class
of
the
osd
slash
watch.cc.
That's
me.
E
E
I
will
handle
those
if
anybody
else
would
like
to
go
over
this
pr.
I
appreciate
that
and
maybe
get
it
to
merge
this
week,
hopefully
and
other
than
that
two
more
issues
one.
What
is
this?
I
had
a
discussion
with
a
teammate
regarding
his
here.
Some
some
ideas
regarding
our
store
objects,
our
object
classes.
D
It's
the
support
is
good
enough
to
to
work
to
hand
to
handle
rbd
some
parts
of
of
the
object
class
api
are
stopped,
but
the
general
mechanism
machinery
is
there.
So
if,
if
another
sub
classes
needs
to
be
parted,
it
may
be
just
they
might
just
implement
some.
We
will
need
to
implement
some
parts
of
the
object
class
api
that
is
missed.
E
E
E
F
You
say
something,
sir
sorry,
I
was
asking
you
if
I'm
up.
F
Well,
anyway,
that'll
talk
about.
I
tried
to
respond
to
you
guys
comments
on
the
document.
There
isn't
any
code
yet
to
help
with.
I
got
sidetracked
working
on
balanced
reads
in
the
last
week,
but
I
hope
to
get
into
that
more
in
the
next
couple
of
weeks.
I
am
off
the
next
two
weeks,
so
I
won't
be
at
the.
I
won't
be
at
this
meeting
next,
two
tuesdays,
but
I'll
be
back
up
after
that.
That's
it
for
me.
E
So
should
we
should
we
answer
those
two
in
next
meetings.
F
A
C
Yeah
last
week,
I'm
trying
to
understand
enough
background
knowledge
and
try
to
review
systole
design,
and
I
think
I
will
put
more
comments
into
that
document
and
an
interesting
finding
is
that
f2fs
has
two
separate
levels:
the
sections
and
the
zones
for
gc
to
support
like
stream
asset
stream,
interface,
ssd
and
the
set
associative
ftls.
So
I'm
trying
to
understand
if
it
is
necessary
also
for
us.
F
Yeah
those
name
collisions
were
why
I
tried
to
use
stream.
I
was
trying
to
avoid
hitting
any
of
the
other
terms
that
have
been
used,
but
I
don't
really
care
except
I
don't
I
don't
want
to
use
zone,
but
other
than
that
doesn't
matter
to
me.
Okay,
so
we
had
two
opinions.
Actually
one
was
not
stream
and
the
other
one
was
stream
quick
show
of
hands
who
was
okay
with
stream.
F
D
Well,
stream
suggests
to
me
that
it's
contig
and
it
is
contiguous
so.
F
F
A
Because
the
reason
I
don't
want,
like
stream,
is
that
it's
it
only
matters
when,
when
we
write
to
a
stream
like
like
this,
the
live
stream
in
c
plus,
when
they
we
want
to
read
from
it,
we
we
need
to
materialize
it
using
like
dot
str
to
to
grab
the
like
a
snapshot
of
it,
but
it
actually
does
not
apply
to.
In
our
case.
In
our
case,
we
can
just
randomly
access
into
a
stream
where,
whenever
the
right
right
pointer
is
after
the
the
yeah.
F
That's
all
true,
but
I've
worked
in
systems
that
refer
to
a
thing
like
that,
as
a
stream.
Also,
honestly,
all
of
these
words
have
probably
been
used
for
everything,
at
least
once
I
kind
of
I
I
kind
of
lean
towards
segment,
I
think
segments
a
little
better,
but
I
don't
feel
that
that
strongly
does
anyone
else
want
to
want
to
take
a
swing.
F
A
F
F
F
F
Yeah,
it's,
including
precisely
it's
it's
an
input
to
the
algorithm
we
use
to
choose,
which
stream
to
garbage
collect
for
some
streams.
We
probably
want
to
wait
as
long
as
we
can
before
garbage
collecting
right,
but
if
a
stream
is
almost
exclusively
journal
entries-
and
we
know
that
they're
obsolete
great-
that's
a
great
one
to
collect,
because
it's
almost
empty
right,
we
have
we
have.
We
have
to
do
almost
no,
no
work.
F
There's
a
lot
more
detail
that
needs
to
go
into
that
I,
or
rather
implementing
it,
will
turn
up
trickiness
that
I
didn't
think
of,
but
that's
the
gist
of
my
plan.
My
plan
is
that
that
structure
should
be
trivial
to
keep
in
in
memory
and
we
ought
to
be
able
to
keep
it
updated
with
delta's.
We're
already
writing
writing
down.
E
I
have
a
misunderstanding,
probably
regarding
the
blockage
and
in
the
segment
in
the
first
paragraph
in
the
blocker
segment,
you
state
that
every
block
is
in
one
of
two
states
and
I'm
not
trying
to
stand
dirty
status.
If
a
block
is
dirty
shouldn't,
it
be
combined
with
all
current
deltas
in
the
cache.
E
E
What
what
I'm
trying
to
say,
if
you
have
a
dirty
block
in
the
like
you
said
the
must
be
fully
present
in
the
cache
okay
yep.
So
if,
if
you
have
a
dirty
block
in
the
cache,
isn't
it
up
to
date
with
all
relevant
date,
deltas
already
applied
to
it?
It.
F
F
C
C
B
A
How
many
miles
will
be
around
next
week?
I
think
the
mandarin
one
won't
be
able
to
be
around
about
you
to
me
and
now
an
ancient.
C
A
D
Another
question
maybe
just
quick
note
about
abner
abner's
work.
Last
week
he
made
the
cycles
per
up
in
in
cbt's
raiders
bench
operational.
D
Now
now
we
are
now
we
are
designing
the
way
how
to
do
we,
how
to
extend
sub-effort
metrics,
also
in
more
benchmarks,
especially
rbd.
E
D
Abnormal
is
working
at
the
moment
on
a
document
enlisting,
possibly
a
possible
possible
options,
possible
approaches
to
tackle,
to
tackle
the
to
tackle
the
the
generalization
of
matrix
in
cbd.