►
From YouTube: 2020-05-20 :: 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
So,
let's
start
I'm
posting
the
the
url
to
the
to
the
pad
in
the
nutella
window
and
the
last
week
I've
been
working
on
the
easy
implementation
based
on
cmga
branch,
which
is
the
working
progress
but
visualizes,
it's
quite
completely
able
to
support
offers.
What
I
need
to
to
to
improvement
on
the
tree,
I'm
still
working
on
the
and
merge
support.
I
and
also
I've
been
doing
some
good
reviews
and
that's
me
redick.
B
C
Yeah,
I
I'm
still
implementing
the
b3
node
and
so
first
I
have
to
get
understand
the
lba
3
implementation
and
try
to
figure
out
how
the
code
can
be
reusable.
D
D
I
think
the
implementation
itself,
however,
does
a
lot
of
fancy
footwork
regarding
the
physical
addresses,
which
I
think
won't
be
as
useful
up
to
you,
though,
whatever
you
think
you
can
factor
out
usefully
or
cleanly
will
be
good.
C
So
sam
said
that
y'all
code
layout.h,
is
the
bad
it's
a
good
foundation,
but.
D
Well,
if,
if
you've
had
a
look
at
my
code,
you'll
notice,
I
did
a
thing
where
I
just
kind
of
wrote
an
iterator
interface
over
the
raw
byte
ranges
in
the
node.
I
don't
actually
know
whether
that's
a
particularly
popular
or
common
way
of
doing
it,
and
I
was
going
to
take
a
look
at
kifu
at
the
code
from
absil
kifu
added
to
see
whether
that's
cleaner.
E
A
D
I
think
you
were
probably
just
using
the
cpu
usage,
including
any
busy
waiting,
but
the
thing
is
for
crimson.
It's
only
running
on
one
core,
so
we're
doing
is
we're
overestimating,
cpu
usage
for
crimson
and
even
under
those
conditions,
it's
still
faster
than
classic
right,
because
classic
uses
several
cores.
D
A
But
but
I
think
if
you
want
to
get
a
better
number
or
pre
more
precise
number,
probably
you
can
take
a
look
at
the
engine's
reply.
I'm
not
sure
you
are
you're
in
in
that
loop.
I
will
I
fold
the
mail
to
you
and
she
mentioned
that
syllabus
three
db
guys
are
using
a
different
metric
for
mirroring
the
cpu
usage,
which
is
so.
B
B
Indirection
here
all
you
want
the
primary
metric.
I
believe
you
want
to
expose
is
cycles
per
operation
cycles
per
op,
however,
ask
as
c
star
does
a
busy
waiting
that
cycles
per
up
the
number
of
cycles
burned
per
single
operation
could
be
actually
higher
if
crimson
isn't
fully
saturated,
so
kyla
guys,
as
far
as
understood,
made
some
kind
of
internal
metric
that
allows
how
the
how
often,
how
frequented
the
busy
loop
really
is.
B
So
if
you
can
get
access
to
it,
you
can,
and
if
you
see
it's
it's
too
frequented,
then
you
can
add
yet
another
instance
of
rather's
bench
that
way.
Finally,
you
go
you
get
you
get
the
proper,
the
proper,
the
variable
cyclist
drop,
but,
to
be
honest,
there
might
be
much
easier
approach.
B
I
was
usually
sticking
to
just
adding
another
radius.
Bench
instance
till
I
till
the
total
total
number
of
the
total
iops
starts,
starting
because
started
being
constant,
which
means
cruising
saturation
of
crimson.
B
A
E
B
Yeah
we
had
similar
in
the
very
similar
issue
on
on
insert
zero
seven.
It
was
okay
for
crimson,
it
was
popping
up
as
well,
but
it
was
terribly.
It
was
terrible
for
rocks
some
low-level
benchmarking
conference
db
and
anyway,
when
we
we
finished
with,
we
ended
with
just
turning
off
all
the
power
saving,
mem
power
management
features
in
cpu.
D
D
F
I
have
covered
here
which
is
based
on
my
experience
with
arms,
not
with
86s,
but
when,
if
you're
tweaking
the
frequency,
then
you
still
have
some
guards
some
a
guarantee
against
the
burning
gaps
totally,
but
most
of
the
thermal
guards
are
off,
which
means
that
there
is
a
chance.
You
will
just
burn
down
your
a
cpu.
It's
not
something
that
I
would
do.
B
Yep
mark
mark
nelson
was
discussing
that
with
intel
many
times
actually,
maybe
preferred.
We
may
ask
him
on
on
on
today's
radius
sink.
B
I
don't
only
did
it,
but
I
can
recall
that
he
was
mentioning
that
instrument
at
the
time.
F
B
F
B
Not
sure
when
I
understood
you
do
you
mean
observing,
observing
the
internal,
easy
weight
related
metric
and
use
it
to
add
further
radius
benchmark
radius
bench
instances
or
somehow
the
reef.
The
number
of
that
somehow
the
reef,
the
cycles
pair
of
on
the
on
t
on
the
on
the
basis
of
this
metric.
F
Yeah
sorry,
I
wasn't
clear
what
I
was
suggesting
is
two
things:
first,
just
to
eventually
publish
the
numbers
as
part
of
the
cbt
results
of
test
or
whatever
publish
the
number
of
the
either
loops,
and
the
second
thing
is
trying
to
calculate
trying
to
find
out.
If
this
is
a
good
and
now,
if
it's
a
number
we
can
use
instead
of
just
if
the
numbers
match.
B
A
A
B
C
So
current
telemetrics
only
can
value
the
the
cpu
used
in
the
c
star
right
in
our
code.
There
is
no
the
metric,
and
so
it
is
not
accurate
for
our
permission,
and
so
all
your
guests
suggest
to
use
this
saturated
to
trigger
the
saturated,
osd
performance
use,
use
more
or
client
side
to
saturated
ost,
saturated
the
cpu.
A
I
think
so
it
depends
on
what
what
you
expect
from
the
results.
If
you
are
expecting
a
saturated,
saturated
load
and
push
crimson
to
its
limits,
then
we
we
probably
could
use
the
tools
and
to
figure
out
what
does
right
laws
to
to
add
it
on
top
of
the
crimson
sd
and
the
classic
quality.
If
you
wanted
to
to
do
a
apple
to
apple
compression.
E
B
All
you
need
to
do
is
to
tune
is
to
tune
the
yaml
files
in
the
test,
cbt
directory
in
the
main
cell
repo
at
the
moment.
Well,
if
you
will
correctly,
we
we
can
configure
okay,
there
are
two
places.
First
is
about
the
number
of
of
clients,
I
mean
raido's
benches,
and
this
is
configurable
in
the
yaml
files,
but
there
is
also
an
information
about
cluster
layout.
B
C
B
We,
it
will
be
it
yes,
cbt
can
consume
fio.
B
And
avnom
added
the
the
generalized
types
of
metric,
so
it
should
be
able
to
cover
to
cover
to
quite
easily
integrate
with
fio,
but
this
will
still
require
some
some
code
changes
but
but
pretty
small,
I
believe
yeah.
So
cbts
can
you
use
for
sure
what
is
for
sure
dbt
can
use
fio
crimson
just
suffered
by
very
basic
fio.
A
Redick,
I
think
probably
I
could
copy.
You
include
you
in
the
loop
where,
where
jimmy
was
asking
for
help
regarding
to
the
performance
test,.
B
C
A
B
A
A
F
Yeah,
I
almost
forgot
my
research
still
doing
scrubbing,
but
actually
mostly
testing
and
reading
code,
to
see
what
was
done.
E
F
F
D
To
be
clear,
there
are
no
hard
and
fast
rules
about
what
libraries
we
permit.
Not
if
you
have
a
reason
to
use
a
library,
then
we'll
talk
about
it
like
that's,
that's
just
how
it
goes.
It's
not
like
there's
a
specific
white
list,
it's
just
when
we
think
there's
a
good
reason
to
do
it.
We
discuss
it
and
then
we
make
a
choice.
D
I
got
sidetracked
with
some
other
stuff
about
the
or
about
due
duplication.
Last
couple
of
days
earlier
last
week,
I
have
merging
working
with
the
lpa
tree,
so
I
need
to
do
a
bunch
of
cleanup
and
documentation
and
I
need
to
do
a
more
thorough
review
of
keith's
code
and
I
think
I'm
going
to
do
that.
Basically
by
trying
to
use
the
layout.h
implementation
in
lba,
it's
just
easier.
It's
a
cleaner,
simpler
implementation.
It
doesn't
actually
change
very
much
code.
D
It
doesn't
do
anything
splitting
emerging,
obviously
it'll
just
replace
the
iterator
interface.
I
was
using
that's
it
for
me,
I'm
going
through
the
crimson
backfill
stuff,
and
I
think
I
owe
reviews
on
a
few
other
things
too.
So
I'm
going
to
work
on
reviews
as
well.
E
Last
week
I
will
still
spend
my
spending
my
most
of
my
time,
refactoring
the
clean
shutdown
and
this
week.
I
think
I
can
concentrate
much
more
on
the
recovery
related
code
as
well.
For
me,.
A
E
I'm
trying
to
understand
the
concurrency
requirement
at
the
tree
level
and
I
was
looking
at
possible
solutions,
but,
as
the
discussion
goes
goes
on,
it
seems
to
me
this
requirement
is
currently
relaxed
later,
I'm
majorly
reviewing
kufu's
note
code
and
the
clean
shutdown
pr.
That's
all
from
me.
A
Thank
you
by
the
way
I
just
wanted
to
be
clear:
I'm
not
trying
to
persuade
you
into
to
abandon
your
your
your
your
working
progress.
I
just
want
to
to
to
emphasize
that,
so
it
will
be.
It
will
be
hard
because
it's
it's
currently
the
just
the
the
the
no
not
without
asynchronized
support.
It's
like
three
thousand
line
of
code,
so.
E
A
E
Think
we
are
currently
working
on
different
directions.
You
you
are
on
the
solution
and
I'm
trying
to
work
out
how
to
do
exactly
from
what
samuel's
proposal.
E
E
D
D
E
D
A
D
Or
what
I
call
it
the
owned
manager.
I
think
I
meant
it
to
be
an
interface,
if
I
recall
correctly
so.
D
D
The
point
is
that
there
may
be
other
implementations
later
first
instance
for
persistent
memory.
It
might
not
look
literally
anything
like
a
b3.
There
might
not
be
a
tree,
it
might
be
simply
a
statically
allocated
subset
of
persistent
memory
with
direct
updates.
That's
that
that's
the
intention.
There.
E
A
D
A
By
the
way
we
we
have
also
have
an
offline
discussion
regarding
to
the
integrating
the
student-based
cio
cd.
A
D
A
So
my
I
asked
my
my
my
suggestion
still
holds.
I
think
it's
it's
the
less
intrusive
change
and
the
most
with
the
least
risk.
E
B
A
B
I
can
recall
that
the
old
make
check
bot
for
arms
we're
using
two
different
two
different
arrangements.
One
was
on
ubuntu
one
on
one
on
centers
or
something
like
that.
They
were
differentiate
anyway.
A
A
The
performance,
the
other
test,
notes
per
running
performance
test.
I
send
us
eight
if
I
recall
correctly
they're
using
the
identical
I
didn't
do:
hardware
setting
and
software
setting.
B
Okay,
I
see,
but
maybe
it's
not
maybe
it's
not
there.
Maybe
it's
not
a
requirement.
Maybe
we
could
differentiate
like
in
the
case
of
alt
arm
64
check
and
then,
if
we
could
label
if
we
could
somehow
label
which
one
is
which
then
the
same
point
about
being
about
about.
D
B
D
D
D
B
Okay,
we
can
switch
pretty
easy.
I
believe
the
question
will
be
what
next?
Okay
there?
Let's
suppose,
let's
suppose
it's
operational
and
crimson
is
getting
more
and
more
audience
more
and
more.
Somebody
make
a
change,
and
this
change
fails
on
silang
silang,
isn't
so
popular
most
peop
most
folks.
B
My
impression
is
that
most
folks
go
with
gcc
so
we'll
be.
We
will
start
blocking
people
and
asking
them
to
tune
themselves
to
to
synagogue.
Maybe
we
should
we
should.
Maybe
we
really
should
have
some
automation
for
contributors.
D
A
D
A
D
D
E
F
E
D
F
E
D
D
F
D
E
B
And
this
is
this:
this
also
affects
ceiling
as
a
whole.
I
bet
we
don't
want
to
claim
we
official,
we
officially
support
silang
bills,
otherwise
we
will
need
to
get
into
technology
testing.
D
Would
be
like
we
can
go
as
far
as
making
a
point
of
for
one
thing
it.
It
is
useful
to
build
with
it.
It
captured
a
bunch
of
lambda
default
things
that
sometimes
we
notice
in
review,
but
we
usually
don't
and
that's
that's
worth
the
price
of
admission
on
its
own,
so
we
don't
have
to
go
past
just
using
it
locally
and
checking
with
the
test.
I
think
that
gets
us,
like
all
the
bank
for
the
buck
we
needed
and
we
get
to
continue
using
gcc,
which
usually
has
better
performance.