►
From YouTube: OpenZFS Developer Summit Part 2
Description
Boris Protopopov of Nexenta on multi-tiered storage. Channel Programs (Chris Siden & Max Grossman)
B
C
A
D
F
F
F
F
F
F
F
F
F
F
B
F
Are
notoriously
difficult
to
my
knowledge
on
you
know
hard
drives
because
they
introduce
all
sorts
of
undesirable
disk
movements
may
cause
in
the
worst
case.
It
may
cause
like
head
thrashing,
etcetera,
and
you
know,
elite
support
performance,
especially
for
data
access
patterns
that
are
random
and
and
those
data
patterns
that
kind
of,
unfortunately,
you
know
correlate
with
with
some
background
activities
that
you
know,
like
a
metadata,
intensive
scan,
a
lot
of
metadata,
etc.
F
F
So
so
the
the
existing
tiers
and
z
pool
are,
you
know,
quote
unquote
normal
tier
following
the
you
know,
the
name
of
the
metaslop
class
and-
and
there
is
a
there-
are
logs
and
there
are
in
our
interest
or
there's
cash.
So
you
know
if
you
type
z-pool
status
or
something
will
see
those
listed
separately,
and
I.
F
F
F
F
F
And
finally,
it
would
be
probably
necessary
to
introduce
some
tier,
specific
square
management
because,
like
you
know
at
this
point
in
time,
there's
only
one
quantum
quote:
I
don't
want
to
use
the
word
pool
a
collection
of
tears.
Right,
I
mean
collection
of
spirits
are
in
the
pool
and
and
if
they're,
if
they're
not
homogeneous,
then
obviously
there
needs
to
be
some
logic
that
understands
that.
Well,
you
know
if
this
is
an
ssd.
Let's
replace
you
know,
drive
from
the
ssd
here
with
that,
and
it
has
a
hard
drive.
F
So
the
challenges
in
in
the
multi-tiered
group
situation,
so
the
general
case
of
of
multi-tiering,
you
know
generally
kind
of
requires
data
migration.
There
are
some
use
cases,
though,
that
do
not
in
particular,
if
we
decide
that
we,
you
know,
we
want
to
place
something
in
that
here
and
it's
just
going
to
stay
there.
So
that's
the
same
form
so
and
the
you
know
in
my
mind,
there
are
three
three
separate
issues
there
with
migrating
stuff
on
zfs.
F
B
F
D
F
F
F
If
we
try
to
come
up
with
a
general
statement
which
includes,
as
I
might
have
mentioned
earlier,
something
that
does
not
require
migration
in
the
first
place,
of
course,
plus
something
that
does
not
mind
the
the
side
effect
of
breaking
the
the
sharing
and
the
the
big
opportunity,
of
course,
is
to
enable
the
a
general
purpose.
Payload
mobility,
which
which
requires.
G
F
B
F
G
F
Early
february,
so
it's
a
couple
months,
two
and
a
half
months,
no
three
months
right
now,
I'm
sorry!
My
arithmetic
gets
a
little
around
around
the
around
that
time
frame
and
besides,
even
I'm
not
announcing
the
release
here,
I'm
just
basically
telling
you
we're
going
to
open
source
and
that's
the
approximate
the
time
frame.
So
I
don't
you
know
all
right.
F
F
B
F
F
From
placing
metadata
on
something
that's
you
know
low,
latency
and
and
high
iops,
because
well
because
some
workloads
like
it
and
and
also
because
there
is
a
whole
bunch
of
activities
in
outside
that
relate
to
management
and
background
tasks.
That
sort
of
you
know
try
to
elect,
to
read
metadata
and.
B
F
F
F
F
Two
tiers
actually
like
l2,
auric
and
logs,
that
are
also
traditional,
ssd
based
and
nzfs.
We
need
to
manage
interaction
between
those,
in
particular
case
file,
to
work.
You
know
if
people
say
metadata
is
already
placed
into
something:
that's
as
fast
as
a
torque.
In
the
first
place,
it
would
seem
it's
unnecessary
or
wasteful
to
place
it
in
the
torque
as
well,
but
in
terms
of
options,
we
were
thinking
about
like
a
special,
only
placement
on
that
special
tier
dual
placement
or
placement.
F
F
F
So,
and
in
this
case
we
actually
use
another
feature
that
we
introduced.
That's
called
class
of
storage
and
the
cost
of
storage
is
basically
a
slight
generalization
of
of
properties
as
well.
So
at
this
point
in
time
you
know
in
xerops
you
can
only
tweak
a
tunable
which
is
which
is
global
for
the
whole,
for
all
the
tools
in
the
kernel
or
you
can
change
the
properties
that
are
for
pool
or
for
video.
F
What's
what
we've
added
we've
added?
You
know
all
these
other
properties
per
tier
of
recollection
of
virtual
collection
of
videos
plus,
you
know
a
set
of
properties
for
video
just
just
for
extra
flexibility,
so
things
like
q
depth,
for
instance-
or
you
know
how
good
this
device
is
for
reading.
Let's
say
in
the
in
the
mirror:
kind
of
configuration
could
be
controlled
by
video,
specific
properties
so,
and
we
will
leverage
the
video
specific
properties
for
basically
attaching
the
labels
to
make
the
spare
replacement
here.
F
B
F
B
F
We
can
configure
these
tiers
per
data
set,
so
it
can
be
used
as
a
log
device
called
slog
and
the
meta
device
and
and
the
right
cache,
f4
and
all
right.
So
all
these
things
are
kind
of
inclusive
in
our
current
implementation.
So
if
it's
a
meta
device,
it's
also
used
as
a
log.
Unless
there
is
a
service
log,
you
know,
in
which
case
it
kind
of
takes
off
it
takes
to
learn.
So
let's
say
the
customer
doesn't
like
doesn't.
F
F
F
Also
so
the
way
the
way
we're
thinking
about
it
is
that
there
will
be
a
background
task
that
will
be
sensing.
How
intense
the
workload
is.
You
know
how
much
data
is
being
written
by
the
user
and
when
that
amount
of
data
drops
basically
based
on
the
telemetry,
that's
available
in
the
right
throttle
line
framework
right
when
that,
whether
the
amount
drops
you
know,
basically,
we
start
moving
off
the
data.
F
B
F
F
The
applications
are
going
to
experience
performance
drops,
so
we
have
also
like
high
and
low
watermarks
they're,
quite
events
to
alert
people
of
the
fact
that
they
need
to
add
some
more
stories
there.
So
you
know
going
back
to
the
to
the
metadata.
I
mean
normally
it's
it's
fairly.
It's
not
very
hard
to
kind
of
provision,
a
metadata
appropriately
to
not
overflow,
with
the
exception
of
the
case
on
the
applications,
maybe
because
duplication
is
also
metadata.
F
And
so
this
this
basically,
what
I've
mentioned
about
what
we're
working
on
is
basically
either
this
one.
It's
around
this,
this
pretty
much
fairly
complete
piece
of
work
that
it
deals
with
the
met
with
this
storage
pool
as
a
place.
That
is
a
place
to
call
it
metadata.
Android
is
right,
cache
and
the
associated
work
with
the
cluster
storage
and
video
specific
properties
yeah,
and
this
is
the
future
work
and,
as
I
mentioned
previously
in
my
window,
like
two
big
categories
here
that
need
to
be
addressed.
F
Let's
say
in
case
of
duplication
in
the
general
case
of
crosstee
or
payload
migration,
and
the
first
case
I
think,
is
it's
a
little
easier
than
the
general
purpose,
because
you
know
it
could
be
addressed
with
a
or
replaced
like
from
the
high
level
standpoint.
It
would
seem
that
it
could
be
addressed
with
transient
dt
class
because,
as
you
might
know,
the
current
duplication
framework
in
zfs
uses
different
classes
for
the
duplication
table
at
this
point
in
time,
they're
basically
broken
down
based
on
the
frequency
on
the
reference
count
of
a
block.
F
To
kind
of
pull
back
and
look
up
things
in
multiple
application
classes,
so
if
we
were
to
you
know,
think
about
migrating
stuff
and
introduce
something
to
introduce
a
transient
duplication
table,
we
could
move
things
on
the
assumption
that
it
will
change
the
data
path
in
the
read
path.
That
says:
well,
if
we
can't
find
something
this
block,
we
need
to
go
and
look
at
this
transit
application.
F
But,
but
only
when
there
is
a
special
global
proflux
as
we're
transitioning
stuff,
that
sort
of
thing-
and
this
is
a
very
real
idea,
just
like
you
know.
Obviously
you
know
there's
not
enough
detail
to
have
a
yeah,
but
it
looks
like
it
would
work.
Yeah
I
mean,
generally
speaking,
it
would
seem
that
it's
it's
easy
a
little
bit
then
dealing
with
the
block
pointer
right.
So
so
we
will
try
that
first
and
if
it
doesn't
work
see
if
that's,
but
you
know
at
what
point
I
read
is
something
that's
very
intriguing.
B
F
I
don't
know,
I
don't
think
it's
a
significant
change.
I
mean
in
cfs
there's
already
a
framework
for
going
through
and
classifying
the
the
type
of
functions
of
the
right.
So
you
need
it's.
You
know
there
is
a
the
function's
name
is
dmu
rate
policy
that
basically
gets
the
block
pointer
and
returns
like
cio
flags
is
here
properties
or
something
that's
called
in
nci
all
right,
and
that
basically
says
what
it
is.
F
G
D
So
can
you
just
kind
of
overview
like
what
you
mentioned,
that
you're
working
on
open
sourcing
some
of
this
stuff?
What
of
what
specific
features
are
done
and
you're
ready
to
work
on
open
sourcing
it
versus
like
future
ideas
here
so
like
what?
What
do
we
have
to
look
forward
to
specifically
like
what
new
properties,
what
new
settings.
F
F
It's
the
different
stories,
but
it's
here,
but
I
mean,
if
you
guys,
have
a
better
idea
as
to
what
to
call
it.
Well,
partly
you
know
the
the
complexity
here
is
in
that
we
we
do
want
to
make
this
apparently
a
special
purpose
right,
so
the
the
best
and
in
the
most
general
use
case,
is
to
define
the
tier
in
on
the
pool
level.
It's
like
it
is
done
for
other
tiers
in
the
zippo
and
then
use
data
set
level
properties
to
actually
define
how
every
data
set
uses
it.
F
D
F
That's
work
in
progress
right
now.
We
just
need
to
it's
it's
sort
of
available
raw,
not
from
the
necessary
from
the
point
of
stability,
but
from
the
point
of
what
to
do
with
it
and
how
to
configure
it
and
how
to
get
things
going
that
sort
of
thing,
but,
but
you
know
to
address
the
the
question
of
you-
know
how
it
could
be
done.
I
mean
it
really
is
not.
F
I
mean
keeping
in
mind
that
the
equals
is
a
copy
and
grade
system
and
under
the
assumption
that
there
is
already
you
know
this
general
framework
leveraging
the
new
raid
policy
that
basically
lets
you
control
where
to
place
a
particular
block.
Then
all
you
need
to
do
is
just
look.
I'm
gonna
read
it
read
the
blog
mark.
It
dirty
mark
that
block
dirty
and
want
to
basically
tell
this.
If
that's
where.
F
D
F
D
It
sounds
like
the
the
first
code
that
you're
working
on
putting
back
would
be
basically,
the
like
class
of
storage,
so
you'd
be
able
to
configure
special
or
fast
devices,
and
then
say
this
type
of
data
goes
on
to
the
fast
devices
versus
the
slow
devices
and
it
just
like
it
just
stays
where
it
lives
forever.
That's
right,
yeah.
B
F
Well,
because
if
you
I
mean
because
here,
if
you
don't
do
so,
then
if
you
import
it
and
basically
start
messing
around
with
it,
you
cannot
violate
all
the
constraints.
I
mean
it
will
be
readable,
etc
and
writeable,
and
this
is
this-
is
a
problem
because
then
the
metadata
that
changes
will
go
to
the
wrong
place.
F
Three
values
for
this
profit
here
you
know
this
one
one
value
is
zero.
Those
critical
data
set
only
uses
it
for
the
slope
device
and
there
is
a
meta
property
that
basically
says
this.
Particular
data
set
uses
a
special
tier
for
metadata
and
zero,
and
then
there
is
right
cache,
which
means
that
particular
and
the
right
cache.
Let's
say
we
don't
we're
not
ready
to
operate
because
they're
still
publishing
things
like
understanding
the
department's
implications,
etc.
F
F
D
H
B
F
H
F
I
mean,
but
but
but
so
we've
been
kind
of
a
little
worried
about
also
about
you
know
the
the
failure
characteristics
of
ssd
is
etc.
Given
the
fact
that
this,
you
know,
this
thing
costs
metadata,
it's
a
little
unnerving
to
you
know
if
you
lose
it.
Basically,
the
whole
pool
is
gone
as
far
as
I
can
tell
so.
You
know
it's
probably
like
a
durational
year
to
have
some
sort
of
backups
etc.
B
G
F
H
B
H
D
D
H
I
D
F
Yeah,
the
you
know
the
automark,
I
guess
I
mean
I
agree
with
you
that
there's
a
lot
of
commonalities
and
we
thought
about
it.
But
you
know
our
reservation
about
using
altora,
for
this
purpose
is
that
it's
hard
to
control
what
stays
there
for
work
and
what
goes
out
so
because
it's
just
it's
got
to
be
pushed
out
at
some
point.
Right.
Do.
I
F
J
F
Now,
yes,
well
also
loop
was
kind
of
important
here
because
you
know
ddts
are
you
know
the
biggest
consumer
of
the
of
this
metadata
space
on
this
device,
and
so
what
we
added
also
who
added
some
tweaks
to
control
the
number
of
data
copies
for
the
edt
and
the
reason
we
did
it
is
because
ddt
is
actually
rebuildable
right.
It's
not
really
it's
it's
redundant.
B
F
You
really
lose
it
so
so
you
can
tweak
the
number
of
jitter
pointers
for
the
dtu
with
the
level
property
and
then
you
know
you
can
kind
of
have
easier
time.
It's
not
the
so.
The
full
solution
is
kind
of
like
a
half
measure,
but
it's
it's
fairly
effective
because
in
fact
ddt
is
the
biggest
concern
of
the
space
and,
as
I
mentioned,
it's
a
you
know
more
or
less
we're
done
in
the
debate
in
the
first
place.
F
D
D
This
is
like
the
mega
version
of
the
lightning
talk,
I
guess
for
an
xenta,
but
for
other,
maybe
we
can
get
one
person
from
each
other
company,
that's
here
to
just
talk
about
how
like
what
product
you're
creating
with
zfs
and
like
any
cool
stuff
that
you've
done
that
you're
also
looking
to
push
back
and
put
into
to
open
source.
At
some
point,
it's
cool
thanks.