►
From YouTube: CDS Infernalis (Day 2.2) -- LMDB backend for Ceph
Description
Videos from Ceph Developer Summit: Infernalis (Day 2.2)
04 March 2015
https://wiki.ceph.com/Planning/CDS/Infernalis_(Mar_2015)
A
B
A
B
A
I
think
it
was
someone
else
that
headed
the
blueprint
but
yeah
same
crew.
It's.
A
C
Yeah
yeah
yeah
our
RP
vs
test
issues.
The
early
there's
a
mean
shop
with
with
karan
equality,
with
implementations
such
as
a
large
larger
by
the
amplification
and
for
Roxy,
be,
are
never
too
busy.
They
needed
to
write,
they
need
to
write
log,
approve
it
a
priority
to
the
valley
disqus.
Oh,
the
the
read
and
write
and
professional
he's
a
very
large
arm
arm,
DB
the
single
our
store
without
a
locker
which
wich
tricked
a
memory,
I'll
single
single
dress,
so
I
think
I
dt.
C
The
man
and
the
more
the
code,
the
code,
the
code
tasca
for
this
PP
we
are
maybe
improve-
include
three
parties.
First,
a
user
implemented
a
arm
DB
star,
a
classic
in
her
in
her
from
who
led
be
which
we
are
implemented.
The
basic
basic
operation
such
as
gets,
delete
and
put
the
other
is
the
other,
is
a
transaction
cement
occur
and
the
another
a
director
implementation.
C
Sex
second
part,
is
the
you
integrated,
the
RMD,
be
other
sub
modo
say,
for
which
we
are.
We
are,
as
a
supporter,
auto,
configure
and
auto
auto
maker
for
the
RMD,
be
though,
sir,
the
third
parties,
maybe
we
need
refine
the
quality,
be
interfaced
to
integrated
some
advance
equality,
be
API,
something
like
had
something
like
a
for
Roxy
be.
There
is
a
lot
old
at
the
ones
ApS
such
as
much
offers
much
our
operation.
C
B
D
D
Is
possible
for
the
right
party,
the
only
like
20
out
70
seventeen
percent
over
the
falconer,
so
we
already
see
that
the
SSD
it's
a
pure
acid,
reset
half
and
the
SSD,
usually
things
around
90
percenter.
The
the
major
issue,
I
think,
is
because
there
r
sm
the
contacts.
Oh
yeah
there,
the
ratchet
I'm
vacating
like
27
or
28
times
it
means
a
client.
Is
your
one
out.
There
will
be
28
on
the
OSD
sad,
so
it
looks
like
rmvb.
B
B
We
definitely
I
think
we
definitely
want
to
try
level
to
be
or
sorry
l,
md,
b
and
c
and
see
how
it
behaves
with
key
value
store.
We
want
to
see
how
it
behaves
with
the
what's
new
store.
We
might
want
to
see
how
it
babes
on
the
monitor.
It
might
be
that
the
monitor,
which
also
uses
like
kita,
adb
interface,
is
a
good
fit
there
too,
especially.
D
B
D
B
B
We
kind
of
do
that
already.
The
British
gives
you
the
omap
interface,
but
it's
sort
of
chunked
into
objects.
Oh
ok,
ok!
Oh
we
kind
of
have
that,
but
it's
it's
it's
a
little
bit
annoying
to
you.
Sometimes,
if
you
want
to,
if
you
want
to
have
a
you
know,
a
key
value
thing
that
has
like
a
billion
keys
and
as
the
user
of
Liberator,
so
you
have
to
shower
that
across
objects
which
is
kind
of
annoying
instead
of
having
the
system
do
that
for
you.
So
it's
not.
D
B
B
Make
it
part
of
liberators
yeah,
I,
guess
I
guess
it
just
depends
like
so
what
I'm
a
little
bit
leery
of
is
trying
to
solve
all
problems
with
SEF
like
there
are
other
projects
that
have
that
are
distributed,
key
value
that
focus
just
on
key
value
data
and
have
all
sort
of
semantics
around
that,
like
like
react
and
cassandra
and
all
the
stuff
in
that
category,
I'm,
depending
on
what
your
data
model
is
and
I'm
not
I,
don't
know
that
stuff
should
just
try
to
solve
all
their
problems
too.
I,
don't
know
that
is
worthwhile.
B
Ok,
but
it
does,
but
it
does
solve
some
problems
like
it.
We
do.
We
do
provide
a
key
value.
Interface
like
within
this
sort
of
object.
Data
model-
that's
already
exposed,
so
it's
useful
for
some
stuff.
It
just
doesn't
do
everything
so
I,
don't
know.
Maybe
okay.
D
Another
question:
what
do
while
we're
trying
to
walking
on
a
key
value
back
in
it?
We
we
want
to
find
a
suitable
back
in
authority,
for
the
SSD
said
have,
for
example,
current
status
is
the
of
you
two
pcie
ssds.
If
we
took
a
local,
current
storage
back
and
fast
doc,
and
hopefully
you
to
divide
the
velocity
back
end.
So
who
do
you
think
that
news
dog
can
can
solve
that
problem?
I.
B
Mean
that's,
that's
the
hope,
but
we'll
we'll
find
out
right.
Okay,
they
were
sort
of
I'm
taking
taking
everything
that
we've
learned
from
what
he's
file
stored
as
well
and
doesn't
do
well.
But
we've
learned
from
what
email
a
store
does
well
and
doesn't
do
well
and
we're
trying
to
like
take
another
stab
at
it
and
see
what
do
it
works?
Oh.
B
Yeah
we'll
see,
I
think
I
mean
I
think
the
the
goal
with
with
new
store
is
to
be
reasonably
general
purpose,
so
it
should
still
work
well
on
disks.
It'll,
basically
be
a
general
purpose,
vows
or
replacement.
So
it'll
work
well
in
disks.
It
will
work
well
on
sort
of
general
Oh
SSDs.
It
might
not
be
that
the
best
thing
for
like
really
fast,
like
TC
I
attached,
flash
or
stuff
that
has
non-volatile
memory
and
all
that
and
in
those
cases
where
you
have
like
something
that's
totally
different,
that
has
very
specific
performance
characteristics.
B
You
probably
need
to
do
something
different
than
new
store,
but
hopefully
it
will
be
a
useful
starting
point
in
that
design
or
not.
Maybe
you
just
like
start
from
scratch
and
do
I,
don't
know,
but
the
biggest
goal
I
think,
is
to
have
something:
that's
just
better
for
the
vast
majority
of
users.
Initially,
I
might
get
to
too
worried
about
the
super
fast
or
super
slow
make
sense.
B
Okay,
thanks
a
really
harmful,
yeah,
yep,
okay,
so
I
mean
just
coming
back
to
this
I
think
implementing
a
level
to
be
in
the
key
value.
Dv
interface,
lmtv
and
the
e
value
DV
interface
is:
we
should
definitely
do
it.
It's
like
super
easy
to
integrate.
We
already
have
all
the
stuff
to
switch
between
level
to
be
in
rocks.
V
will
just
add.
Lm,
DB
and
they'll
be
very
will
learn
a
lot
just
from
seeing
what
what
happens
and
then
based
on
what
we
learned
we
can
decide.
E
How
useful
is
the
tool
that
you
all
wrote
for
testing
the
monitor,
specifically
miter
saw
r
DB
with
like
different
io
types.
We
take
that
and
adapt
it
to
be
useful
for
other
testing
I.
Actually.
B
Haven't
looked
at
the
monitor
one
he
wrote
like
three
years
ago.
He
wrote
one
for
the
file
store
that
generated
an
OSD
like
workload
for
the
file
store,
but
it's
so
out
of
date.
It
looks
the
I/o
pattern.
It
generates
looks
nothing
like
what
the
OSD
actually
does,
though
it
might
be
worth
rewriting
that
or
we
could
just
run
the
OSD
I'm
kind
of
waiting
for
a.
E
E
B
Yeah,
we'll
see
I
mean
I.
Think
I
think
that
the
thing
that
worries
me
about
approaching
it
from
that
angle
is
that
we
don't
know
what
those
ratios
are
for,
what
the
OST
actually
does.
So
it's
it's
almost
easier
just
to
stick
it
under
the
OSD
and
then
throw
the
workload
at
the
OSD
that
you
expect
and
not
try
to
simulate
it.