►
Description
A
You're
doing
ICT
stuff
for
research,
it's
very
different
to
doing
for
the
enterprise
and-
and
I
hope
to
highlight
that
a
little
bit
in
here
and
really
what
we
need
at
the
end
of
the
day
is
both
scale
and
performance
at
large.
And
so
we
end
up
in
the
space
where
we
have
to
create
a
fabric
for
this.
And
we
are
actually
the
workshop
for
researchers
to
go
off
and
build
things
from.
A
So
so
that
we
got
there,
so
these
are
these
spectrums,
you
know.
Sometimes
we
have
to
deal
with.
So
you
look
at
from
a
money
point
of
view.
Ninety
fifty
percent
of
the
research
income
come
then
I
said
the
research
income
come
from
only
ten
percent
of
the
researchers
that
proxies
to
rough
size
in
their
agendas,
so
I
have
to
deal
with
these
peak
guys
and
what
they
need
to
do,
and
we
have
to
deal
with
the
long
tail
which
you
can
see.
A
A
You
want
to
be
an
organization
that
doesn't
take
all
the
risk
itself.
You
interact
with
other
organizations
to
achieve
your
end
goal
and
so
I'll
talk
about
permeability
a
bit
and
this
whole
multidisciplinary
research
thing
which
research
would
be
doing
for
a
long
time
is
really
that
same
pattern
and
if
you're
in
nineteen
by
the
new
research,
we
sometimes
talk
about
these
four
paradigms.
You
know
and
that's
their
paradigms
for
discovery.
A
Chris
and
I
had
a
nice
beer
review
this
on
the
other
day
and
so
I
thought
I'd
fling
it
in
okay,
so
I'm
gonna
go
to
those
three
points
in
a
bit
of
detail
and
then
kind
of
come
back
from
there.
Okay,
so
it
send
the
pic
versus
long
tail.
We've
all
seen
these
graphs
right
and
we're
going
to
be
careful
that
we
don't
leave
dead
bodies
behind.
A
So
if
your
ICT
in
the
enterprise,
you
tend
to
worry
about
the
long
tail,
how
you
can
commoditize
something
make
something
cheapest
/,
but
for
a
number
of
stuff
you
have
and
all
your
bodies
that
end
up
lying
at
all
the
guys
doing.
Experiments
at
the
peak
end
if
you're
a
traditional
HPC
center,
I'm
going
to
try
and
look
at
into
you
at
the
time,
and
you
tend
to
be
the
opposite
way
around.
A
Where
you
only
think
of
the
peak,
the
superstar
researchers-
and
you
don't
do
anything
for
the
long
tail,
so
you're,
not
supporting
desktops
or
windows
at
all
right
and
your
bodies
are
all
at
the
long
tail
right.
You
can
see
the
bodies
on
that
a
bit
vague,
but
they're
there
right
and
there's.
Obviously
they
match
up
at
some
time.
A
We
need
to
be
able
to
create
some
fancy
supercomputer
or
some
such
some
sort
of
fancy
device
for
some
Pete
guy
on
the
fly,
just
as
we
need
to
be
able
to
provide
for
the
rigorous
ITIL
vase
services
that
researchers
will
need
to
do
at
the
end
of
they.
All
researchers
are
actually
on
the
long
tail,
because
even
peak
researchers
use
well-established
tools
that
they
coupled
together
in
some
integrated
way
to
create
their
fantastic
device
right
permeability.
A
The
example
I
got
here
is
actually
comes
from
from
oil
and
how
oils
made
right
permeability
is
the
idea
that
you've
got
some
you've
got
some
structure
which
huge
organization,
in
this
case
it's
through
a
rock
or
the
sand.
Other
think
this
is
going
to
laser
pointer.
Maybe
it
doesn't
know
that
the
yellow
beets
run.
A
And,
and
then
another
organization
might
be
the
blue,
which
is
the
fluid
in
this
case,
and
when
the
two
come
together
in
some
certain
environment,
they
can
go
off
and
do
something
create
something
new
right
in
this
case
it's.
So.
This
is
where
the
oil
of
your
carbon
and
oil
and
everything
else,
given
some
pressure
and
everything
else
and
time
and
temperature
and
everything
else
we
end
up
with
oil
and
gas
and
everything
else
I
created
something
new
together
right.
A
So
permeability
is
this
property
of
being
able
to
have
this
sort
of
a
behavior
where
you
can
let
something
else
come
in
and
together
you
make
something
better.
Multidisciplinary
research
is
is,
is
a
term
that's
often
user
research,
because
the
greatest
discoveries
always
come
at
the
edge
of
boundaries
of
disciplines
said
that
were
quite
badly
I
guess
what
I'm
really
trying
to
say
is
if
you're
working
in
a
field,
that's
well
a
star.
A
So
it
took
about
2
300
years
ago,
where
a
guy
managed
to
work
out
a
technique
to
machine
lenses,
really
well,
he
put
in
the
tube
and
he
gave
it
to
some
people.
First,
people
style
booking
up
the
sky,
and
eventually
people
started
looking
down
low
and
the
biology
started,
because
we
were
able
to
take
pretty
pictures
or
see
for
the
first
time
the
structure
of
leaves
and
the
bugs
and
everything
else
right.
A
So
microscopes
are
really
important
and
telescopes
are
really
important
because
they,
let
us
see
something
that
we
have
not
been
able
to
see
before.
If
you're,
a
peak
researcher
you're
building
your
own,
you
microscope
right.
I'll
come
back
to
the
microscope
in
the
tick.
But
what
was
really
important
was
a
piece
of
technology,
machining,
brass,
but
also
machining.
A
The
lens
was
the
really
important
part
that
created
the
paradigm
shift
that
allowed
for
many
observations
to
be
made
from
these
new
microscopes
and
they
obviously
see
okay
might
be
a
bit
small,
but
I'm
just
going
to
highlight
to
a
few
lines
which
are
really
really
important.
This
is
a
graph
from
1875
to
now
of
innovations.
If
you
follow
on
Tripoli,
you
might
have
seen
the
publication
around
this
early
this
year
or
last
year.
That
sort
of
got
me
thinking.
A
I
want
to
present
this
in
a
slightly
different
way
and
and
what
it
talks
about,
and
if
you
understand,
interest,
work
or
compounded,
growth
works.
It's
the
compare,
the
growth
rate
of
various
innovations,
so
in
here
we've
got
things
like
the
speed
of
traveling
over
oceans
and
all
those
sorts
of
things
here,
I've
this
user
ones
to
do
I,
guess
electronic
technology,
this
blue
line.
Can
anyone
guess
what
that
blue
line
is?
If
you
can't
read
it,
it's
Moore's
law
right!
A
So
from
roughly
this
time,
we've
been
sitting
on
moore's
law
and
I'm
going
to
come
back
and
afterwards
and
say:
that's
pretty
much
dominated
what
this
third
paradigm
and
why
it
exists
right.
This
green
line.
Can
anyone
who
can't
read
it
see
what
that
is
tell
what
that
is.
It's
the
number
of
devices
on
the
internet,
if
you
like
the
internet
of
things
right,
and
this
thing
has
been
going
for
50
years-
moore's
law
50
years
running.
A
If
you
took
the
number
one
apply
this
graph
way
to
50
years,
you'd
be
at
150
billion
as
a
number,
your
app
okay,
it's
a
massive
growth.
Nothing
in
mankind's
been
like
it.
Internet
of
Things
is
presently
like
it
and
is
expected
to
keep
on
going
like
that
for
F,
probably
so
for
50
years,
the
very
bottom
Nick
down
the
bottom
down
there
is
is
a
a
same
line
for
imaging
sensors,
have
bowel
do
sensors
now
that
one's
subjective,
whether
it's
going
to
follow
but
the
rent?
A
The
point
is
that
these
two
combined
and
they
guess
with
hard
disk
storage
I,
should
probably
add
that
in
law
and
in
sort
of
an
SSD
storage
storage,
flash
storage.
But
this
should
be
alone'
spectrum
right.
We
could
probably
see
it
will
have
a
similar
growth
and
these
guys
are
leading
to
what's
probably
the
the
technology
tool
that
leads
to
the
fourth
paradigm,
which
is
all
these
data
centric
things
so
I'm
just
going
to
go
to
the
next
slide.
B
A
Don't
try
so
being
a
mathematician
computer
scientist
I
refuse
to
not
put
equations
I'm
in
any
of
my
talks.
I've
got
equations,
so
so
the
fourth
the
first
program,
which
is
just
your
microscope.
It's
all
about
observations.
What
I
get
is
my
answer.
I
can
count
things
right.
The
2nd
Para
die
is
where
we
started:
creating
physical
laws
like
Newton's
and
the
laws
behind
normals,
like
statistics
and
everything
else
right,
but
mankind
can
computer
more
right
to
end
up
with
equations.
A
The
third
paradigm
really
talks
about
we're
computing,
enabled
us
to
compute,
really
big
and
complicated
models.
Now
the
eff
has
gotten
really
really
large.
It's
complicated,
it's
a
system
of
equations,
it's
whatever
else,
and
we
need
a
computer,
the
computer
in
a
reasonable
time.
If
we
didn't,
we
only
made
a
discovery
right,
so
the
microscope
has.
This
is
this?
Is
this
sort
of
a
computer
can
process
this
to
give
us
the
Y?
The
fourth
paradigm
is
where
the
big
is
actually
on
the.
Why
it's
actually
on
the
the
observations
at
the
end,
it's
so
larger.
A
The
humor
can't
get
knowledge
out
of
it
by
themselves
and
need
something
else
to
help
us
and
computers
and
storage
is
helping
us
do
that,
and
the
fourth
paradigm
really
comes
this
way
in
these
forms,
and
these
are
terms
now
which,
which
you
might
be
starting
to
feel.
The
pressures
from
your
own
research
groups
are
your
own
businesses.
A
The
one
that
it
really
is
attributed
to
is
is
data
mining
and
in
data
mining
we
actually
flip
the
equation
around
from
the
big
day
there.
Why
we're
actually
trying
to
work
out
what
F
is
data-
tells
us
what
the
function
the
model
is
from
the
data,
and
we
use
algorithms
to
do
that.
So
we
need
fast
computers
to
do
that,
and
we
need
fast
storage
to
do
that,
because
the
two
working
in
unison
in
my
own
work
and
disciplines.
A
This
is
actually
they're
the
most
interesting
story
and
this
actually
tries
to
join
both
big
large
amounts
of
data.
We've
large
large
models,
the
people
who'll
be
doing
this
style
of
stuff
for
a
long
time
are
your
atmospheric
guys.
The
guys
are
predicting
weather
right.
They
have
complicated
physical
bait
physics
based
models,
they're
running
needs
to
be
curious
to
run,
and
they
have
tons
of
observational
data,
which
is
now
the
other
problem,
and
they
got
to
mash
the
two
together.
A
Now
the
stress
in
your
system
is
even
worse
right:
it's
not
just
data
mining,
so
it's
not
big
stupid
computers.
It's
the
two
come
together
right,
and
this
is
really
interesting-
is
both
a
big
right
and
lo
and
behold.
Both
these
are
only
useful
where
we
know
what
we're
looking
for.
Some
human
intuition
need
to
exist.
A
First,
to
know
that
we're
trying
to
do
these
physics
or
trying
to
do
this
style
of
data
mining
or
whatever
else
there
are
certain
things
which
the
human
mind
is
still
better
at
and
will
always
be
better
at
and
that
you
just
need
the
intuition.
Our
brain
is
wired.
A
certain
way
to
do
things
right
and
visualization
is
still
really
really
important,
and
so
to
that
end,
where
I
was
hoping
to
take,
you
guys
was
a
what
we
call
the
K
facility.
This
thing
is
beautiful.
A
Imagine
having
really
really
large
LCD
TVs!
You
got
at
home
wrapped
around
you,
it's
really
really
bright.
There's
something
like
80
million
megapixels
worth
of
a
real
estate
and
join
up
here
is
am
some
histology
images?
You
know
if
you
go
the
pathology
and
they
take
a
slice,
can
and
they've
stained
it.
A
You
can
see
things,
it's
that
good,
that
actually
one
whole
image
and
pathology
is
now
only
reduced
in
in
sort
of
in
size
by
10,
so
that
if
this
thing
was
ten
times
higher
resolution,
we'd
be
showing
the
native
resolution
to
show
you
higher
resolution.
Histology
images
are
that's
how
high
they
are
most
researchers
or
clinicians
using
it
have
to
use
like
google
map.
Oh
google
earth
type
applications
where
they
scrolling
scrolling
scrolling
and
zoom
zoom
zoom
zoom
zoom
zoom,
and
they
do
a
lot
of
that
to
get
through
on
these
sorts
of
facilities.
A
A
Sorry,
hello,
yeah!
Well,
there
are
there's
some
light
up
lighting
and
mood
lighting
and
stuff
that
does
exist
on
it,
so
we
were
hoping
to
go
here
and
there
some
great
entertaining
things.
If
we,
if
you
ever
come
to
marsh
again
or
river
hostess
again,
I
will
be
happy
to
do
it.
Unfortunately,
we
weren't
able
to
go
there
today.
Do
some
Dean's
bumped
us
off
for
their
own
uses,
so
the
21st
so
we
talked
about
this
is
why
we
talked
about
the
21st
century
microscope,
and
this
is
what
the
microscope
looks
like
today.
A
A
Then
we
have
to
have
then
the
filters
and
how
you
try
and
make
the
zoom
and
everything
else
is
actually
the
software.
We
run
on
big
computers
to
process
that
right.
So
rack
bond
is
a
cloud
infrastructure
which
is
I,
guess
essentially
my
baby
and
and
massive,
which
is
the
other
thing,
is
our
GPU
sort
of
based
or
imaging-based
supercomputer,
which
is
effectively
now
built
on
top
of
our
cloud
infrastructure
and
and
then
there's
how
people
interact
with
it.
A
So
there's
the
K
facility
and
it's
bringing
those
desktops
and
that
rich
environment
all
the
way
to
to
to
the
individuals,
own
laptop
sort
of
arouse
and
people
orchestrate
to
create
integrated
tools
along
this
way
and
theta
has
to
move
in
and
out
through
this.
So
our
data
providing
infrastructure
has
to
help
store
all
this
stuff
and
make
it
fast
into
here
and
faster
to
hear
and
and
shareable
impermeable,
so
others
from
other
researchers
institutions
can
get
through
it,
the
general
public
or
or
constrained.
A
A
So
back
to
the
extremes
or
spectrums,
so
we
need
self
service
because
the
peak
guys
are
going
to
do
it
themselves.
We
don't
do
what
they
do
it
right.
They've
got
the
researchers
and
the
applied
guys
to
do
it
front
ends.
You
know.
Sips
is
sort
of
dead
right
in
certain
ways
like
it's,
it's
it's
the
front
ends
Alan
merge
right.
A
We
need
the
infrastructure
to
be
accessible.
There's
no
reason
why
only
my
nice
guy
should
be
able
to
do
this
sort
of
thing,
any
researcher
in
Australia
better,
do
it
and
we
then
need
aggregation
of
cost.
We
need
institutions
to
buy
rights
and
then
then
give
their
rights
out
to
their
own
researchers.
So
so
the
Monash
is
only
a
tenant
and
all
these
things
and
then
the
paradigms.
Well,
we
need
scale.
A
Clearly
we
need
low,
latency
and
bandwidth
and
all
the
things
that
suited
was
talking
about
so
that
we
can
make
this
fantastic
first
to
market
first
discovery
type
of
things,
so
Seph
is
that
fabric
fabric
for
us
so
much
OpenStack,
and
so
it's
Neutron
right.
We
all
we
sort
of
need
to
be
in
this
sort
of
software-defined
world.
We
need
to
be
so
that
we're
not
asking
our
IT
guys
to
go
off
through
a
job
ticket
to
go
off
and
do
something.
A
It's
got
to
be
on
a
dashboard
or
through
a
Python
script,
where
they
orchestrate
the
hardware
they
need
for
the
purpose
that
they
need
it
right,
and
so
it's
more
like
a
fabric,
it's
more
self-service
and
it's
those
researchers,
not
us
who
will
actually
create
the
verticals
right.
That's
really
different
to
how
most
enterprises
think
now
Merc
is
our.
Is
the
research
center
itself?
A
You
know
which
is
just
the
cloud
technology
of
high-performance
computing
and
all
the
bits
and
pieces
all
the
people,
and
it's
ask
people,
including
all
the
IT
guys
involved
Richard
the
workshop
people
come
to
ours
because
it's
a
bit.
You
know
we
have
better
tools
and
better
capability
at
trying
to
help
make
a
little
alarm.
A
A
Okay,
so
so
what
we
do,
is
we
conceptualize?
Some,
you
know
infrastructure
products
is
we've
come
up
with
these
terms
that
a
national
scale
or
a
state
scale
at
least
to
get?
They
got
some
properties,
but
essentially,
you
know
stuff
directly
connected
to
the
cloud
stiff
storage
sort
of
comes
off
here.
You
know
and
those
sorts
of
things,
but
essentially
we
have
one
or
two
great,
once
f
cluster
or
clusters
that
provide
into
this
space
and
Blair
will
go
through
some
of
that
detail.
A
The
interesting
ones,
probably
are
one
that
we've
developed
in-house
called
my
TARDIS,
which
is
very
purposely
around
data
management
for
instrument
data,
it's
agnostic
to
the
instrument
and
that
there's
something
a
bit
special
that
allows
the
researcher
to
control
accessibility
in
the
life
cycle.
Of
that
of
you
know,
of
a
piece
of
data
comes
off
an
instrument.
It
also
like
allows
the
facility
to
do
the
same.
A
A
I
know
that
there's
nothing
compared
to
100
kilobytes
of
some
other
people,
but
it's
still
quite
significant
I
guess.
The
other
point
I
want
to
raise
is
we're
pretty
new
at
SEF
actually
we're.
Thirdly,
I
was
going
to
ask
you
how
long
we've
been
doing
it
for
two
years
Rivera
new.
We
took
a
big
punt
right
to
go
down
this
path
and
we
made
it
work
for
us.
A
C
A
D
How
does
this
thing
with
hey
cool
all
right?
So
sorry,
if
I
hope,
I
missed
anybody
in
a
list
here,
I
was
trying
to
do
this
it
to
that
topic
last
night,
so
going
better
emails,
so
there's
quite
a
long
list
of
people
that
error,
obviously
myself
Jericho,
who
I
don't
think
is
here
today,
reference
way
or
in
the
audience
and
Craig.
D
So
at
monash,
one
of
the
interesting
things
that
we've
actually
I
guess
challenges
that
we've
had
with
SEF
and
and
some
of
the
other
technology
around
here
in
the
cloud
space
is
that
we
have
a
corporate
IT
group
and
also
the
research
center
that
have
for
a
while,
been
really
I
mean
the
part
of
organizationally
in
the
same
space,
but
operate
very
independently,
or
at
least
very
sinister,
very
siloed
fashion,
and
so
we've
been
trying
to
sort
of
break
down
some
of
those
walls.
And
so
that's
why
we've
got
a
long
list
of
people
hit.
D
B
D
A
D
Stole
it
off
me,
the
first
time
around
so
yeah.
Of
course,
it
all
began
with
the
cloud
so
we're
one
of
the
nodes
of
the
nectar
research
cloud,
which
is
a
a
national
project
in
Australia
that
was
funded
through
subside
scheme
from
the
road
government
back
just
post-gfc,
and
that
set
up
this
national
cloud.
D
So
in
early
2013,
we
had
the
first
node
of
rap
Mon
deployed
through
nectar,
so
we
had
our
own
local
club
and
everything
was
awesome,
except
nobody
had
any
persistent
storage,
unfortunately,
so
nectar
funded
all
the
compute
and
ephemeral
storage
enough
for
all
your
servers
to
run,
but
nobody
had
any
volumes.
There
was
object
storage,
but
none
of
our
uses
knew
what
to
do
with
it.
So
we
had.
We
fortunately
had
a
bit
of
an
object,
storage
hardware
and
the
existing
national.
D
The
existing
object,
storage
at
University
of
Melbourne
was
not
under
capacity
pressure,
and
we
had
no
way
to
federal
rate
was
Swift
at
that
point.
So
we
had
this
sphere
the
sphere
storage
hardware
and
decided
to
try
and
step
on
it,
and
so
nectar
were
gracious
enough
to.
Let
us
use
that
because
it
was
technically
their
hardware,
and
so
that's
when
we
started
on
cuddlefish,
so
I
was
actually
just
trying
to
rember
this
early
today,
because
the
bobtail
cuttlefish,
dumpling,
were
all
kind
of
released
in
2013.
D
I
was
a
fairly
prolific
EFSF,
and
that
was
really,
I
think,
when
Seth
probably
started
to
be
taken
seriously.
It
was
we
actually
looked
at
it
when
we
were
planning
our
cloud
node
and
it
was
just
when
ink
tank
had
come
out
of
dreamhost
and
the
website
was
this
horrible
dinky
little
thing
and
it
was
just
looked
all
a
bit
too
fragile,
really
and
then
then,
like
three
months
later,
looks
it
again,
sir.
Oh
no
I
can
take
this
seriously
now.
D
Yeah,
so
don't
we
just
go
through
a
bit
of
show-and-tell
that
stuff
we've
got
on
the
ground,
so
we've
got
three
sip
clusters
now
running
a
mash,
so
the
first
one
was
this
one
that
we
built
out
of
the
Swift
hardware:
88
del
r,
7
30
XD
nodes,
you'll
notice,
there's
no
SSD
in
them.
Of
course,
we
didn't
bother
buying
that
for
swift.
It
wasn't
a
requirement
at
the
time
and
they've.
D
Actually
they've
done
pretty
well
I
mean
that
gave
us
the
confidence
to
keep
moving
ahead
with
with
SEF,
even
despite
the
the
trade-offs
with
the
hardware
there.
So
it
basically
met
the
user
expectation
tests.
You
could
have
a
sale,
a
Windows
Server
on
the
cloud
or
something
doing
a
Windows
Explorer
file
copy
at
over
300
mega.
Second,
you
know
so
that
was
that
was
good
tick
and
that
that
is
currently
got
about
60
terabytes
used
there's.
Actually
a
hundred
and
thirty
five
terabytes
of
storage
committed
the
error.
D
If
we
look
at
at
our
OpenStack
cinder,
so
we're
quite
over
provisioned,
but
we
have
some,
you
know
is
coming
I'll
talk
about
in
a
little
bit
of
a
moment.
We
run
that
at
two
replicas
and
that's
actually
never
been
a
problem
touch
wood.
Of
course
it's
as
kind
of
a
smallish
cluster,
so
the
failure
rates
a
little
bit,
I
guess
in
our
favor
there.
D
D
Rds
I
was
the
sister
or
the
brother
program
to
nectar,
which
was
all
about
research
data
storage,
and
so
we
created
this
computational
storage
product,
which
was
the
safe
storage
to
go
with
Monash
too,
and
so
we
changed
a
few
things
from
we
did
it
on
purpose
this
time.
So
we
because
we
added
SSDs
for
journals.
D
Although
we
still
have
internal
Don
you're
waiting
for
it
to
go
really
to
vga,
because
this
cluster
is
in
production.
So
this
one
is
now
currently
it's
about.
So
it's
about
300,
terabytes,
usable
we've
got
a
few.
We've
got
a
few
OSD
spear
they're
actually
got
a
bunch
of
nodes
of
the
same
configuration
ready
to
add
in
later
once
once
capacity
starts
pumping
up
of
it.
D
So
then,
then
we
did
the
big
one.
So
this
one
is
the
the
RDS
cluster
that
we're
doing
for
public,
facing
object,
storage
and
virtual
nerves,
and
so
on
this.
This
guy
of
course
got
dedicated
bonds.
It
has
a
cached
here
for
the
road
ice
gateway.
Cached
is
just
spinning
disk,
but
just
faster
disk.
It
has
a
mix
of
56,
gig
and
10
gig
networking.
So
the
storage
servers
themselves,
the
main
OST
nodes
is
33
of
those
again
the
720
XDS.
D
Each
of
those
also
has
an
MD
1200
j-bot
attached
to
it,
so
it's
144
terabytes
of
raw
storage
per
node
and
that's
all
hanging
off
only
20
de
which
Sudha
won't
be
very
happy
that,
but
you
know
we
had
one
of
the
really
nice
things
about
set.
Is
you
can
let
you
play
with
you
know
we
had
a
very
clear
budget
for
what
we
needed
to
meet
to
get
our
price
per
terabyte
and
we
were
able
to
tweak
off
that
around
and
choose.
We
were
making
those
trade-offs.
D
D
We
also
what
we
do.
I
went
back
to
jewel,
socket,
of
course,
with
these
big
nodes
bit
more
ram.
I
think
if
we
were
doing
this
all
again,
we
would
probably
go
denser.
That
was
one
thing
that
was
a
real
concern
at
the
time
and
there
was
a
lot
of
conflicting
information
about
making
those
hardware
choices.
You
know
you
need
this
much
ram
for
so
many
eos
DS.
D
D
If
we
had
have
gone
with,
you
know,
sixty
drive
Chace's
or
something
like
that.
That
probably
would
have
gone
into
a
maybe
a
four
and
three
and
a
bit
rack
footprint,
and
the
other
thing
that
was
different
here
is
that
we
moved
over
to
a
rail,
and
this
was
actually.
This
was
the
first
rail
seven
deployed
like
your
first
big
relative
and
deployment
in
Monash
still
really
is
like,
and
that
that,
of
course
cost
us
a
few
issues
along
the
way,
as
well
with
just
getting
to
learn
things
like
network
manager
and
so
forth.
D
D
You've
got
the
cash
tier
nodes
over
here,
and
this
is
sort
of
roughly
the
data
pools
that
you've
got.
You
know
there's
a
few
of
them
left
out
here,
but
I
think
I've
captured
the
main
ones
so
coming
out
of
those
external
drives.
We've
got
our
RVD
storage
at
three
replicas.
We've
also
got
a
set
of
test
data
pool
on
those
discs
with
replicas,
so
they're
using
the
same
crusher
all
set
those
pools.
D
You've
got
the
83
erasure
code
coming
out
of
the
internal
drive
in
this
chassis,
and
on
top
of
that
is
the
case
year
for
the
radar
scope
way
buckets
as
well
coming
off
the
internal
hard
drives
on
the
couch
denotes
the
cat
nodes
and
those
cat
nodes
also
have
this.
If
s
metadata
on
them
and
so
then
going
a
layer
up,
we've
got
a
few
hypervisors
which
offer
the
presentation
services
for
action.
Big
user
interaction
here
so
you've
got.
We've
got
virtualizer
a
das
gateway.
D
I'll
show
you
our
radars
gateway,
hae
architecture
on
the
next
slide
as
well,
because
that
was
that
was
actually
one
of
the
challenges
that
we
had
was
kind
of.
Okay,
we
wanted.
We
wanted
a
high
availability
service,
but
there's
actually
no.
As
far
as
I
know,
there
isn't
a
recipe
up
there
to
go
and
to
build
one
of
those
things.
D
D
So
how
do
we
get
Dennis?
Round-Robin
gives
us
the
scale
out.
We've
got
a
checkbox,
incapable
I've
d
for
failover
of
AJ
properties.
I
said
that
they're
not
a
single
point
of
failure
and
then,
of
course,
RadarScope
gateway,
just
scales
out
as
you
like
as
well,
and
so
that's
that's
gone
really
well
and
Jericho
did
a
really
awesome
job.
Putting
all
that,
together
from
my
diagram,
sir.
D
So
we've
got
as
I
said
before:
we've
got
a
little
bit
of
new
capacity
ready
to
go
in
the
monash
to
zone.
We've
got
another
10
nodes
ready
to
be
added
when
we
need
them
so
that
cluster
will
go
to
27
nodes,
another
another
nine.
So
in
that
in
the
big
RDS
cluster
of
the
moment,
we
only
deployed
24
of
those
noes
to
begin
with,
because
three
petabytes
of
storage
is
still
going
to
take
quite
a
long
time
to
fill
up
and
then
we're
also
refreshing.
D
The
Monash
one
costs
are
the
one
that
didn't
have
any
SSDs
in
it
or
anything
before
so
del
actually
have
a
really
nice
box
in
the
current
range
us
and
30
XD,
which
gives
you
16
three
and
a
half
inch
drives
in
a
two-hour
chassis
that
can
be
0
s.
Ds
two
discs
in
the
back
for
the
OS
and
we've
checked
in
via
me
in
for
the
journals,
so
that'll
be
interesting.
D
D
So
some
of
their
pain,
points
and
nits
would
surf
surf
itself
has
been
really
quite
solid,
I
think
no,
no
major
problems,
no,
no
data
loss.
That's
that's
a
big
thing,
that's
important
for
us,
but
it
can
be
quite
opaque
when
things
do
go
wrong.
You
know
when
the
clusters,
the
cost
is
either
kind
of
it's
either.
Okay,
everything's
hunky-dory
or
its
warning
and
warning
can
be
anywhere
from
you
know:
I!
Okay,
I'm
almost
fine.
D
Don't
worry
about
me
to
know
you
really
need
to
look
what's
going
on
here,
because
I'm
not
healthy
at
all,
so
you
this
is
this.
This
example
here
is
a
little
Seth
des
snapshot.
I
took
when
we
had
a
network
partition
in
in
the
big
yes
cluster
and
you
know
so
you
can't
it
there's
really
not
like
I'm,
not
a
lot
of
clues
to
go
on
here.
Oh
it's
kind
of
it's
a
lot
like
it's
a
lot
of
information
to
take
in
all
at
once.
D
Obviously
we
had
a
Mon
down,
so
that
was
kind
of
that
was
the
first
place.
We
went
to
look
and
try
to
figure
out
what's
going
wrong
with
the
moms,
and
then
it
was
clear
I'll
that
guy
can't
talk
to
the
other
guy
at
these
two
don't
know
where
he
is:
okay,
there's
something
screwing
with
the
network,
but
yet
that's
I,
think
I
think
we
ended
up.
D
So
what
one
of
the
things
that
we're
doing
with
that
that
picture
I
showed
you
before
is
we've
got
virtual
nares
boxes
using
our
BD.
Is
their
storage
back
end?
How
do
you
optimize
that
picture?
So
we've
we've
done
a
bit
of
we've
done
a
bit
of
work
in
the
space
reason.
Looking
at
okay,
if
we
want
to
run
say
a
ZFS
box
on
top
of
SEF.
D
It
was
the
case
that
pge
repair
was
just
like
a
roulette.
You
know
you
would
just
if,
if
the,
if
the
primary
OSD
was
the
one
that
was
bad,
then
you
copy
the
bad
data
to
the
others.
I
think
that
may
have
been
fixed
already
now,
so
you
know
that
that's
good,
but
I
mean
our
current
policy
is
basically
just
even
for
a
single
uncorrectable
read
aris
a
on
the
drive
is
just
to
kill
the
whole
drive
and
I
guess.
D
That's
that's
also
an
icing
standardize
on
because
it
that
you
can
just
do
that
for
any
any
disk
problem,
but
be
interested
to
hear
how
other
people
handle
that
and
I
guess
part
of
the
reason
we
do.
That
is
also
because
we're
using
raid
controllers
so
we're
using
virtual
raid
0
drives
for
hours
days
and
that
introduces
that
works.
Just
fine
and
you
get
extra
benefit
of
the
right
cash.
But
it
does.
You
have
to
be
aware
that
that
introduces
another
just
another
layer
of
complexity
in
operations
between
SEF
and
the
hard
way.
D
So
when
it
comes
to
doing
debugging
and
so
forth,
then
you
have
to
mess
around
with
the
dell
tools
to
reset
virtual
drives
and
so
forth.
The
other
thing
we're
actually
we're
going
through
a
problem
at
the
moment
that
we're
debugging
a
performance
issue
on
on
that
big
cluster
and
it
sort
of
seems
like
there's.
D
Actually,
we
don't
yet
have
a
really
standard
way
to
look
at
these
problems
and
when
I
was
trying
to
looking
for
things,
I
actually
came
across
a
nice
wiki
page
where
somebody
started
and
proposed
a
framework
to
do
this,
but
it
hasn't
gone
anywhere
yet
so
I
think
that
would
be.
That
would
be
a
really
useful
resource.
D
D
D
Well,
we
do.
We
do
really
like
the
idea
how
to
get
transparent
compression,
because
a
lot
of
the
data
sets
we
deal
with
if
they're,
just
for
the
cases
where
it's
an
export
to
researchers
desktop,
they
are
all
generally
quite
compressible
where
we
have
where
we
control
the
application.
Front-End
like
with
my
TARDIS,
then
we
can
do
some
of
that
at
the
application
level,
but
for
the
further
complexity,
I
would
I
would
be
happy
to
throw
all
that
stuff
away
and
just
go
to
surface
only,
even
without
having
compression
and
so
on.
A
B
D
D
We
are
looking
at
so
for,
for
the
virtual
nares
devices
would
use
TSM
to
back
up
to
the
file
backups
there
we're
not
sure
really
what
we
still
need
to
implement
something
for
a
toast
gateway,
but
that,
on
the
face
of
it,
seems
like
a
relatively
easy
problem,
because
there's
plenty
of
open
source
tools
out
there
that
can
slip
stuff
out
of
an
s3
endpoint
and
put
it
somewhere
else,
whether
it's
file
or
whatever,
and
so
we
have.
We
have
large
tape,
libraries
and
so
forth.
D
So
we
can
go
to
you
know
we
can
dump
it
on
to
say
an
NFS
export.
That's
then
HS
end
or
something
like
that.
No,
not
so!
No
we're
not
doing
a
synchronous,
replication
or
anything
like
that
or
or
RadarScope
way.
Regions.
At
this
point,
yep,
but
Colin,
who
is
now
gone,
has
been
asking
a
lot
of
questions
of
vendors
about
that
some
stuff
there,
because
we
are
with
definitely
interested
in
having
a
backup
copy
in
it.
Another
data
center.
D
Why
you
know
Venkat
talked
about
the
what
d
converged,
architecture
I
think!
That's!
You
start
to
you
start
to
see
the
need
for
that
once
you
once
you've
got
to
that
scale.
I
think
it
at
the
sort
of
at
the
scale
were
you
just
supporting
some
block
storage
for
medium
sized
cloud
deployment?
It's
not
doing
doing
a
complete
forklift
upgrade
or
refresh
of
your
storage
nodes
is
not
you
know
not
that
big
a
deal,
but
for
something
like
that
we
talking
to
me
it
does.
D
B
D
D
Ordering
of
moms,
iced,
teas
and
so
on,
and
then
we
always
picked
like
say
we
always
pick
the
kid
like
a
canary
node
and
did
the
one
first.
As
long
as
four
minor
versions,
that's
fine,
and
so
maybe
leave
that
running
for
24
hours,
make
sure,
there's
no
hiccups
and
then
do
the
rest,
but
with
never
what
never
been
and
always
actually
I
guess,
like
I
watched,
Seth
uses
pretty
closely
and
I
had
already
at
least
learned
before
we
even
wanted
to
do.
D
D
D
D
B
A
All
those
sorts
of
things
we
had
a
lot
of
cases
where
researchers
will
come
and
say,
look
I
need
to
do
this
sort
of
de
novo
genomics
mapping
which
kills
disk
and
so
we're
getting
a
lot
of
requests
for
fibre
channel
links
essentially
to
be
created,
for
you
know,
x
y&z,
so
we
kind
of
knew
there.
There
are
the
sort
of
problems
that
we
needed
to
know
that
we
could
solve
with
this
infrastructure,
and
it
became
pretty
clear
that
some
of
those
alternatives
wouldn't
actually
really
have
for
short
round
of
reeds
and
things
like
that.