►
From YouTube: Improving Resilver: operational results by Saso Kiselkov
Description
From OpenZFS Developer Summit 2017 (day 2)
A
Since
then,
we
even
progressed
a
little
bit
but
forget
is
that
or
the
one
person
you're?
Probably
that's
left
that
doesn't
know
how
to
algorithm
is
good.
I've
got
some
slides.
We
saw
from
last
year,
which
are
which
should
probably
explain
that
I'm,
basically
going
to
add
on
to
what
Todd
was
talking
about
in
his
to
talk
about
the
talk
about
the
3ds,
Pro,
premature
and
so
I'm,
going
to
take
a
little
bit
of
editorially.
Basically,
the
way
that
scrubbing
and
restoring
worse
is
that
it
reverses
the
data
hierarchy.
A
A
So
and
basically,
work
for
traversing
any
kind
of
like
metadata
processing
and
try
to
beat
that
back
what
I
was
trying
to
work
out
a
good
way
to
get
Rhys
little
performance
up.
So
the
problem
is
Bobo.
Grease
over
is
logical,
block
order
in
terms
of
the
way
that
the
data
is
arranged.
We
simply
said
it's
essentially
just
a
sec
marshal
pottery.
It
reads:
the
file
goes
on
to
the
next
month.
A
Reads:
the
file
goes
on
the
next
one,
and
just
that
simply
in
the
way
that
the
file
is
laid
out
so
essentially,
the
buffers
are
sort
of
laid
laid
out.
Logically,
is
the
way
that
generates
I/o
placed
on
the
other
complete
generate
some
more
I/o,
it's
enough
to
compete,
and
so
on.
It's
important.
That
is
perfect
if
you
have
a
mostly
sequential
run
layout.
So
on
initial
writing,
the
lot
of
the
order
that
the
blocks
appear
in
in
the
tremendous
layout
is
approximately
the
way
that
look
like
on
this.
A
So
you
just
work
issue
them
all.
In
order,
it'll
be
fantastic,
run,
really
well,
you'll
get
a
performance,
and
unfortunately
many
worklists.
Don't
look
like
that.
Many
pools.
Don't
look
like
that,
especially
after
the
globally
use,
because,
as
you
be
bright,
be
especially
when
you
running
things
like
dailies
and
me
ends
those
tend
to
rewrite
individual
pieces
of
your
bottle
and
though
and
the
pieces
then
don't
actually
get
forced
out
into
the
place
they
need
to
copy
over
to
some
other
place,
but
what
you
get
are.
A
But
what
you
have
with
is
that
if
you
go
in
order
in
logical
order
to
the
lower
part,
is
you
can
end
up
essentially
hitting
all
kinds
of
like
strange
paths?
And
so
this
ends
up
seeing
is
just
a
whole
bunch
of
out
border
lDA's
coming
into
it.
So
this
will
go
ahead
and
start
a
small
part.
The
other
then
little
jump
back
and
forth
totally
demolished
he
or
read
through
with
on
your
rubbery
silver.
A
So
you
can
easily
end
up
with
a
hard
drive
that
the
only
question
why
you
would
be
able
to
do
a
hundred
plus
megabytes.
A
second
you'll
be
like
ten,
which
is
kind
of
bad.
So
the
way
that
we
saw,
that
was
by
inducing
a
reordering
in
memory
we
ordered
hue.
That
does
its
best
to
basically
pull
the
items
back
into
order.
So
rather
than
me
going
straight
from
the
scrubber
of
any
silver
traversal
of
the
metadata.
A
A
I'm,
okay,
anyway,
so
the
so
the
way
we've
done
it
is
we
put
in
a
reorder
queue
that
completely
decouples
the
meditative
rehearsal
from
the
actual
reads
being
issued
for
the
data
blocks
to
disks,
and
this
is
pretty
much
the
stamping
code
about
a
year
ago,
as
I
said,
it
was
a
pile
of
rather
experimental
stuff
and
we,
since
then,
we've
run
into
a
bunch
of
trouble.
Initially
I
was
completely
confident.
A
A
Of
course,
something
you
got
to
have
resolved,
and
so
the
progress
that
we've
done
since
last
summer
ZFS
was
that
by
February
we
had
shipped
that
a
change
in
our
necks
and
store
that
over
lease
and
anything
past
that
that
basically
uses
that
as
new
default
algorithm,
there
is
a
way
to
go
back
to
the
old
one
you
really
really
really
want
to,
but
there's
basically
no
that
so.
The
other
thing
is
that
we
were
able
to
preserve
this
formality,
there's
no
basically
penalty
for
moving
back
and
forth
between
the
older
volition,
a
new
one.
A
We
also
involved
madam
George
in
some
of
the
design
work,
so
it
wasn't
white
paper
to
you,
but
I
think
we
got
most
extreme
now
and
we
also
got
some
help
over
from
Tom
and
now
who
were
crazy
enough
to
expose
her
still
autumn
boiling
and
yeah.
They
got
to
work
on
it
and
they
put
basically,
he
ran
into
some
performance
troubles
and
then
that's
where
the
prefetcher
changed
idea
came
out
of
is
that
he
saw
some
terrible
performance
on
some
Ludington
and
just
have
to
resolve
that.
A
So
that's
an
awesome,
awesome
piece
of
work
that
just
came
out
of
some
crazy
guy
who
just
wanted
to
go
ahead
and
experiment.
So
let's
talk
about
a
little
bit
about
cornice
numbers.
This
is
the
slide
that
showed
the
last
year
because
weirdly
big
big
promises
and
of
course
you
won't
be
only
slide
that
joke,
but
this
was
sort
of
the
Tom
Bend
of
the
performance
Delta.
So
we
were
getting
like
crazy
multipliers
of
on
really
really
badly
randomized
pools.
The
food
algorithm
was
like
just
light
years
ahead
and
so.
A
Is
that
something
that
we
serve
those
two
months,
you
in
a
different
color,
we're
so
slightly
more
realistic
numbers
that
we
got
ultimately
once
we
run,
ran
this
introduction
and
ran
this
on
systems
that
haven't
actually
used,
and
we
got
a
little
bit
closer
like
to
3x
ruffle
in
terms
of
improvement,
because,
realistically
the
deployment
deployed
systems
are
never
quite
as
badly
randomized.
As
you
know,
for
me,
there's
another
and
we
got
a
whole
bunch
of
customer
systems,
didn't
really
bother
us
now
the
detail
them
there.
A
A
This
is
basically
about
all
the
I/o
that
would
go
into
Z
pool
rather
sorry,
generally
from
the
functionally
silvering
the
blossom
and
they
which
ones
are
roughly
close
together,
clumps
and
so
because
they're,
those
he
reasonably
sequential
and
which
ones
are
so
huntress,
everybody's,
completely
say
Marshall,
in
which
case
the
new
implementation
is
completely
or
almost
totally
identical
to
the
Ultimate
Edition
and,
as
you
basically
get
further
down
the
proportionate
sequentiality.
So
you
get
more
randomized
stuff
on
the
pool.
The
old
scrub.
A
Some
of
them
know
there's
systems
that
have
is
some
didn't.
Have
it
so
I
think
most
most
of
these
systems
did
have
this
work,
although
some
of
them
babies
so
there's
a
there's,
a
bunch
of
experimental
work
that
was
done.
40
minutes
changes
over
so
a
few
T's
were
gathered
for
those
hope.
These
are
some
older
numbers
and
I
know
that
we
gathered
during
the
early
course
of
the
year,
which
I
know
including
Tom
changes.