►
From YouTube: How Ceph Performance in an ARM-Base Microserver Cluster
Description
Traditional server architecture is not a perfect fit for distributed storage because network is usually saturated with low CPU utility. And using virtual machine or multi-daemons carries the risk of multiple lost storage devices when a single server fails. Ambedded aims to address those problems with its microserver architecture designed for distributed storage and optimizing Ceph performance with this 1-to-1 architecture.
A
So
Aaron
thank
you
good
morning.
My
name
is
Aaron
I'm,
the
founder
and
the
CEO
of
a
method
technology.
Today,
I'm
going
to
share
our
experience
over
the
several
versions
in
an
arm-based
micro
server,
cluster,
okay,
okay,
first
I
were
introduced
a
little
bit
about
embedded
because
there
is
a
stop
company
and
then
I
will
each
point
out
the
issues
of
traditional
server
that
were
running
safe
with
multiple
OSD
in
a
single
server.
Okay,
and
how
will?
How
are
we
solve
this
issue
by
micro?
A
Server
based
on
up
and
also
I,
will
show
share
some
the
basic
ability
and
scale
out
performance
and
some
use
case
over
the
Hadoop
and
finally,
I
will
let
you
know
we
use
the
am
based
micro
server.
It
will
save
a
lot
of
energy
power
for
our
server
and
how
much
we
can
save
for
you
embedded.
We
have
founded
in
2013
that
we
start
with
designing
the
own
server
for
distributed
storage
until
now,
until
until
now
we
have
abroad
Phi,
Theta
or
more
of
us
origin
now
installed
in
u.s.
A
A
The
issue-
the
first
issue
is
when
the
server
is
failed.
No
matter
it
is
a
meriting
fail
or
some
components
fails.
You
will.
You
are
lost
of
the
data
immediately
inside
that
chassis
yeah.
That
is,
that
is
a
big
issue
because
you
know
safe
will
have
to
is
a
self-healing
when
it
detects
OSD
dance,
so
this
will
make
a
huge
network
zone
when
they
start
to
recover,
or
it
would
take
very,
very
long
time
to
recover.
So
this
is
a
big
issue
that
for
not
only
safe
but
of
the
distributed
storage
and.
A
Safe
is
a
network
based
storage?
It
is
not.
The
bottleneck
is
always
not
the
CPU
utility.
We
have
customer
experience
that
tell
us
no
matter.
You
have
10
gigabit
of
Ethernet
or
20
gigabits
of
Ethernet.
You
still
does
not
have
enough
bandwidth
and
the
CPU
utility
are
always
very
low,
so
CPU
is
very
idle
and
not
fully
utilized.
A
That's
why
there's
also
a
lot
of
energy
are
waste,
so
the
third
is
certainly
the
power
consumption
yeah,
normally
a
8
disk
or
a
of
the
X
server,
as
it
is
denote
you
a
condition
like
350
watts,
yeah
you
pay
the
energy
and
for
the
power
consumption.
You
also
penned
energy
for
the
cooling,
so
you
pay
double
of
your
your
expense
over
energy,
so
the
power
consumption
is
consuming
your
budget,
okay.
So
how
we
solve
the
least
kind
of
problem
by
using
a
decentralized
architecture
of
server
instead
of
traditional
one
too
many
servers.
A
First
in
this
standard,
I
would
like
to
explain
what
is
the
micro
server
architecture?
Okay
in
our
server
for
for
the
safe,
there
are
eight
independent
arm-based
our
server
each
arm
based
server
will
play
the
rule
of
an
OSD
or
it
can
be
a
monitor
or
can
be
a
radio
scale
way
so
to
give
the
best
of
bandwidth
for
each
node
or
each
OSD
from
each
micro
server.
A
We
have
a
certainly
a
data
storage
for
the
for
the
data
that
is
a
3.5
SATA
hard
drive,
or
you
can
use
2.5
SSD
for
the
OSD
data
storage
also,
it
weighs
it
has
an
independent
SSD
for
the
journal
disk,
so
the
journal
disk
is
dedicated
for
lotto
SD
and
it
is
not
shared
with
others
and
on
a
server.
We
also
have
third
storage,
which
is
for
the
operating
system,
and
the
safe
software
is
rule.
Fascism,
okay,
so
the
resource
for
any
OSD
is
all
independent,
not
share
with
others.
A
Okay,
so
they
can
provide
dedicate
storage
interface
memory
CPU
in
CPU
resource
and
also
the
main
most
important
is
the
network.
Okay,
every
single
server
has
a
doer
Ethernet.
Each
one
is
2.5.
Gigabit
per
second
okay,
so
how
to
connect
them
as
a
cluster,
we
have
built
a
two
inch:
SD
switch
in
a
1u
chassis,
the
chest.
A
The
switch
provides
in
cluster
the
inch
edges
of
fabric
for
the
OSD
peering
and
also
provide
a
public
for
the
scale
out
and
can
assess
so
to
you
fully
utilize
the
eight
server
by
a
server
by
two
two
point:
five
gigabit
ethernet.
We
had
to
provide
14
Giga
bits
of
uplink,
so
this
is
a
completed
completely
eliminate
the
bottleneck
of
a
safe
cluster
yeah,
so
it
is,
leases
are
also
very
important
for
asset
performance
and
in
this
architecture,
this
to
switch
are
are
redundant
for
each
other.
Okay.
A
A
This
is
a
picture
of
the
server
you
can
go
to.
The
rain
has
boots
to
have
a
look
at
it.
Okay.
So,
as
I
mentioned,
there
are
8
micro
servers.
These
8
micro
servers
are
all
independent
and
they
are
one
MDOT
to
SSD
dedicate
for
that
server
for
its
OSD
journal.
Ok
and
the
one
server
has
its
own
hard
drive,
not
more
than
one
ok
and
all
of
the
module
dock
here
are
no
matter
is
a
hard
drive
or
the
the
server
or
the
switch
or
hot
swappable
in
certainly
include
the
power
supply
yeah.
A
A
Okay,
so
compared
to
the
this,
this
diagram
show
you
the
difference
of
one
to
many
server
when
one
server
too
many
OSD
or
versus
the
microserver
of
one
to
one
when
the
traditional
server
you
have
many
disks
in
one
server.
So
when
you
have
one
server
fails,
as
I've
mentioned,
Asli
of
the
disk
will
be
lost
yeah.
This
is
a
very,
very
big
problem,
but
for
the
microserver
architecture,
this
is
one-to-one
architecture
instead
of
one
too
many.
So
when
you
lose
a
single
server,
you
just
lost
one
OSD.
Instead
of
so
many.
A
Okay,
so
the
benefit
of
using
a
one
single
node
to
single
OSD
architecture
unsafe
the
best.
The
first
benefit
is
that
it
is
a
decentralized
architecture,
the
distribute
value
or
minimize
the
failure
domain
to
a
single
OSD
instead
of
many
instead
of
single
node,
so
that
the
big
note
is
a
micro
note
and
the
NTP
F
on
server
I
can
show
you
the
server
here.
I
have
carry
one
okay.
A
This
is
the
this
is
the
on
server.
It
is
very,
very
simple,
circuit,
very,
very
few
components.
That
means
it
has
very
very
high
mtbf,
compared
to
a
thousand
component
in
a
motherboard
with
the
chance
of
pharaoh,
is
very
low
compared
to
a
x86
server
yeah.
So,
with
the
dedicate
hardware
resource
for
their
OSD,
you
get
a
cpu
memory,
storage,
the
network
and
also
important.
A
Not
only
not
only
the
OSD
can
serve
on
this
micro
server,
we
also
use
the
server
for
monitor
and
radio
scale
way
and
even
I
sigasi
gateway.
So
you
you
need
not
no
extra
hardware
to
make
a
cluster
and
the
basic
high
availability
architecture
is
to
have
three
chases.
That
means
that
you
put
a
one
monitor
in
each
chassis
and
you
have.
If
you
use
our
replication,
you
can
use
crush
mail
to
dispute
your
data
replication
to
different
chassis
and
in
different
node
server.
A
A
A
A
For
for
here
we
have
five
Intel
Xeon
server
for
the
kind
as
a
loader
of
the
test,
so
that
ten
gigabit
switch
and
we
use
three
three
units
of
microserver
on-base
microserver,
with
21
SSD
and
SSD
journal
and
three
monitors
and
totally
I
will
create
a
forty
RBD.
This
IBD
are
for
all
of
the
kind.
Well
one
clients
half
is
on
a
BD
and
make
the
FIO
test
over
are
over
it
and
aggregate
the
total
bandwidth
from
all
the
server's
client
servers.
A
Okay,
so,
first
of
all,
a
list
diagram
shows
that
the
scale
out
performance
we
start
from
one
chassis,
which
is
seven
OSDs
and
add
another
chassis.
To
have
total
14,
then
add
the
third
to
have
total
21.
Oh
s,
DS
you
can
see
the
I
io
PS,
the
IRS
of
read
and
write
is
very,
very
linear,
scaled
from,
for
example,
this
is
nine
thousand.
This
is
almost
18.
This
is
almost
27
case
for
the
right
yeah
so
can
achieve
this.
Is
that
proof?
A
Channels
on
the
network
to
compare
if
we
have
all
of
the
caster
every
chassis
as
only
210,
Giga
or
utilize
order
for
10
Giga
Network,
how?
What
is
the
difference?
Okay,
so
here
we
do
some
4k
random
right
test
which
increases
our
rights
not
number
of
times.
Then
we
compare
its
aggregate
IOPS
of
20,
Giga,
uplink
and
40
giggling.
It
shows
that
average
you
get
of
50
percent
of
total
bandwidth
increased.
A
A
The
data
loss
in
that
disk
we
heaved
to
all
of
the
other
15
OSDs
yeah,
so
compared
to
a
traditional
disk
array,
you
will
take
a
very
very
long
time
more
than
41
hours.
It
is
on
two
and
half
day
to
just
recover.
Oh
one,
disk,
yeah,
so
safe
is
very,
very
powerful
because
all
of
the
OSD
will
take
over
the
reveal.
If
you
have
more
OSD
the
time
of
recover
will
getting
less
and
less
linearly.
We
can
show
you
that
we
can
it
scale
our
linearly.
A
Okay,
next,
our
so
I
would
like
to
share
a
use
case
in
a
telecom
company
in
Taiwan
that
is
a
biggest
telecom
company
in
Taiwan
they
have,
they
have
a
cluster
which
is
use
Hadoop
file
system.
For
the
big
data
analysis,
the
scenery
is
that
they
collect
different
data
source
from
everywhere
and
connect
to
a
staging
server
and
convert
it
to
a
free
copy
of
HDFS
and
each
H
HDFS
has
its
data
node.
Every
data
node
is
a
server
and
is
a
computing
node
with
storage.
So
it's
like
a
hyper-converged
infrastructure.
A
Computing
and
storage
are
combined
together,
but
when
they
grow
their
cluster
in
the
storage
getting
bigger
and
bigger
up
to
now,
they
have
more
than
two
petabytes
of
data
stored
in
that
will
still
grow
very
fast.
The
issue
is
that
the
HDFS,
when
you
want
to
scale
out
your
storage
capacity,
you
have
to
scale
out
your
computing
as
well.
That
is
a
very,
very
expensive.
They
don't
the
computing
power
they
ready
very
enough
because
they
have
hundreds
of
server
yeah.
So
they
want
to
separate
the
storage
and
the
compute
and
the
computing
node.
A
They
are
not
expect
to
have
a
very
high
performance
external
internal
storage,
but
they
can
tearing
their
data
to
have
hard
data
as
they
use
HDFS,
but
the
history
and
the
codata
were
sold
in
some
other
whales.
So
we
tried
to
work
with
them
with
many
different
ways
from
safe
to
lady
works
is
a
Hadoop.
A
If
you
find
the
papers
and
you
get
internet
to
search,
you
will
find
several
ways.
First,
is
the
Java
parking
for
for
Hadoop
and
we
make
it
works,
but
the
performance
is
very
poor
because
you
have
a
safe
is
based
on
C
and
and
Hadoop
is
based
on
Java.
So
the
province
is
very
cool
and
we
also
tried
the
way
of
Amazon
like
s3
a
we
make
it
work
with
s38,
but
its
provenance
is
not
good
as
well.
So,
finally,.
A
Finally,
we
use
a
safe
file
system.
Is
this
local
file
system
and
then
it
is
not
100%
fulfill
their
needs,
but
it
solve
a
lot
of
their
problem.
Yeah
to
solve
the
computing
storage
can
separate
with
the
storage
and
reduce
their
cost
and
when
the
server,
because
are
there
Hadoop
for
assisting
the
disk,
is
the
inside
every
computing
note
when
any
disk
fails
is
making
that
trouble.
You
have
to
immediately
to
replace
the
hard
drive
in
the
server.
So
that's
a
very
high
load
of
operation
in
the
maintenance.
A
Before
we
making
production,
we
have
to
make
a
proof
of
concept
in
the
lab,
so
beside
the
safest
system.
As
mentioned,
we
have
worked
on
other
solution
like
s3
a
or
the
Java
project.
No,
no
others.
Work
is
same
for
the
say,
file
system.
We
use
it
as
a
local
file
system.
We
make
a
small
benchmark
so
in
the
benchmark
we
compare
to
how
to
private
file
system.
A
A
A
Yeah
most
of
the
most
of
the
test
set,
shows
each
performance,
but
for
the
Reid
we
got
a
slower
performance
on
safer,
safe,
a
system
yeah.
But
this
is
not.
This
is
not
a
issue
for
them,
because
this
is
they
want
to
have
a
trade-off
between
very
high,
expensive
and
very
high
power
consumption
of
hdf
hdfs
server
or
versus
our
low
power
server
yeah.
So
we
got
better
performance
over
the
right,
but
a
little
bit
lower
performance,
fuller
for
the
reed.
A
Okay,
so
our
product
is
a
combination
of
three
key
components:
one
is
the
arm
server
and
in
the
center
in
the
core
is
safe
country.
We
are
using
duo,
it
is
our
safe
and
the
search
on
top
of
safe.
We
have
unified
a
virtual
storage
manager,
which
is
a
web
use
interface
for
the
user.
We
banked
lease
is
because,
when
we
promote
our
product
to
our
customer
many
of
the
customer,
they
have
a
issue
that
they
say
yes,
self
is
very
powerful.
A
A
We
made
this
problem,
so
we
come
back
to
think
and
decide
to
make
a
safe
you
to
interface
for
operands
that
make
it
a
much
easier
and
lower
the
barrier.
The
skill
of
the
user?
Yes,
so
make
it
very
simple.
Yeah
later
I
can
show
you
video
about
Mini's
about
our
user
interface
on
the
video
yeah,
and
we
are
demoing
in
ready.
Booze
welcome
to
come
to
the
rabbit
rabbits.
We
can
talk
about
if
you
had
any
questions
and
our
product
was
a
winner
of
interrupts
2016.
Yes,.
A
So
now
we
have
our
own
server.
What
is
the
next,
because,
two
years
ago,
when
we
designed
this
server,
we
have
only
32
bits
arm
server,
chip
to
use
and
most
of
the
software,
including
safe
on,
is
not
supported
by
it
does
not
support
the
safe,
and
so
that
does
not
support
arm.
So
we
have
to
do
any
most
of
things
from
0
until
now,
everything
by
ourselves,
yeah
and
now
the
64
bits
arm
is
become
reality.
A
So
we
are
on
going
to
design
a
new
pear
phone
with
same
form
factor,
but
it
is
64
bits
observer
that
can
give
us
more
memory
to
use,
because
with
now
we
have
only
2
gigabytes
of
memory.
We
have
to
very
carefully
to
use
all
of
this
limit
limited
memory,
but
is
still
even
it
get
a
very
good
performance,
but
we
think
we
can
make
it
much
better
if
we
have
more
memories
and
also
for
the
SSDI
OSD.