►
From YouTube: #vBrownBag 2014 OpenStack Paris - Boyan Krosnov – Block Storage: When Ceph is not enough
Description
#vBrownBag 2014 OpenStack Paris - Boyan Krosnov – Block Storage: When Ceph is not enough
A
I've
come
from
I
worked
in
IT
infrastructure
system,
since
99
I
never
been
doing
this
to
the
storage
for
about
three
and
a
half
years
now
in
a
company
called
stop
pool,
I've
lived
all
over
the
world,
seen
all
kinds
of
crazy
infrastructure
things
and
distributed
storage
distributed.
Software-Defined
storage
is
fairly
new
and
it
is
basically
storage
that
works
on
standard
servers,
standard
hard
drives
standard
network
and
all
the
storage
magic
happens
in
software
that
runs
on
these
terminal
servers
and
the
reason
now
is
the
good
time
for
this
for
distributed
storage.
A
There
are
a
couple
of
main
reasons:
they're
a
lot
more,
but
there
are
a
couple
main
reasons:
money's
low
costs,
high
performance
networks
like
this
switch,
for
example,
is
at
around
sixty
four
five
dollars
per
port,
the
whole
box
around
four
thousand
dollars
and
the
other
reason
is
multi-core
processors.
So
the
standard
server
today
he
has
about,
say,
20
cars,
turning
it
around
three
gigahertz
and
he
has
lots
of
RAM
so
for
a
software-defined
system
bit.
You
know
software-defined
Network,
software-defined
storage.
A
What
we
do
is
we
take
a
small
fraction
of
that,
a
small
fraction
of
that
server
say
one
core,
maybe
two
cars
and
we
try
to
deliver
high
performance
in
this
case
storage
system.
Out
of
this
one,
or
maybe
two
cpu
cores-
and
this
is
what
we
do
so
stop
with
a
distributed.
Locklear
high-performance
storage
system
running
on
standard
x86
hardware
and
we
deliver
it
as
software
and
services.
So
we
help
you
design.
A
system
will
install
it.
A
We
have
latest
etc,
and
we
do
this
for
hosting
a
virtual
machines
in
containers
for
service
providers
and
private
clouds
and
in
a
lot
of
cases
where
people
choose
torpor.
It's
because
of
performance
and
efficiency.
It's
not
because
of
some
other
say
features
so
the
the
problem
with
block
storage
in
the
linux
ecosystem
and
in
the
OpenStack
ecosystem
is
that
clouds
need
a
high
performance
block
storage
system
for
a
lot
of
workload.
A
A
So,
to
give
you
an
example
of
you
know
the
type
of
performance
difference
you
can
get
between
one
system
India,
very,
even
though
they're
both
distributed
block
storage
systems
right.
So
we
run
these
tests
on
the
same
Hardware
same
hard
drives
a
message,
the
same
servers
we
gave
each
system,
you
know
12
drives
and
three
SSDs
or
just
12
hard
drives
and
the
performance
difference.
A
You
see
something
like
this,
so
this
is
a
I
ops
on
the
x-axis
and
CP
usage
on
the
y-axis,
so
you
get
we
safe
when
you
really
load
the
system,
you
get
something
like
forty
percent
CPU
usage
on
all
of
the
storage
nodes
and
because
these
are
hyper
threaded
systems.
This.
This
really
means
that
they're
kind
of
fully
occupied
their
kind
of
saturation
CPU.
A
They
can't
really
do
a
lot
more
and
we
store
where
you
know
down
there
than
under
five
percent
for
even
though,
at
the
same
time
you
know,
store
positive
earnings
like
15,000
tie-ups
and
it's
using
four
percent
CPU.
So
this
is
this
is
the
kind
of
performance
difference
you
would
see
so
between
one
system
in
another.
So,
like
eight
times
difference
in
performance
on
the
same
hardware,
eight
times
more
I
ops,
while
at
the
same
time
having
six
times
less
or
10
times
less
CPU
usage.
Okay.
A
A
So
if
your
storage
system
is
very
severe
hungry,
if
it
needs
a
lot
of
CPU,
then
you
can't
combine
it
with
other
workloads
on
the
same
server
in
the
storage
system
only
is
fixed
to
using
one
or
two
cpu
cores.
Then
you
can
use
the
other
cpu
cores
for
for
your
computer
workloads,
and
this
called
the
converged
architecture
in
a
converged
architecture.
You
make
it
to
the
tripe
types
of
games.
You
can
get
RS
a
half.