►
From YouTube: NERSC's Ten Year Plan
Description
NERSC's Ten Year Plan, Sudip Dosanjh, NERSC Director
A
Everybody
to
guess
is
technically
the
third
day
but
day
two
of
nug
2014
start
off
today
with
nurse
director,
sudeep
dosanjh,
sudeep
and
says
that
is
rector
of
nurse.
Previously
he
had
an
extreme
scale
computing
at
Sandia
National
Laboratories
was
code
directory
of
Los
Angeles
sandy
Alliance
for
computing
extreme
awesome.
A
B
Okay,
thank
thank
you
Richard
night.
Thank
you.
All
for
coming,
I
was
heading
back
from
DC
last
night
and
my
flight
got
canceled
and
I
told
him
well,
I've
got
to
make
it
back
no
matter
what
so
I
was
happy
to
get
on.
Another
flight
I
was
scrunched
in
the
middle
seat
for
about
six
hours,
but
I
was
happy
too
happy
to
arrive
here
in
time
to
make
it
here
and
meet
all
of
you
I
guess.
You
know
this
is
a
very
special
meeting
for
us.
B
This
is
kind
of
the
kickoff
as
you've
heard
for
a
40th
year.
Our
anniversary,
and
we
do
have
some
faces
from
the
past
who
have
joined
us.
So
michael
mccoy
is
here
and
he
was
a
heavily
involved
with
with
nurse
get
Livermore
when
it
first
started,
and
you
were
the
deputy
director
for
a
while
at
the
end,
okay
and
then
where's
bill
Kramer,
it
bill
step
out
so
bill.
Kramer
is
here
and
and
and
so
he
certainly
has
a
very
long
history
with
nurse
Caswell
when
once
it
was
at
Brooke
Leigh,
so
so.
B
B
B
So
I
just
wanted
to
give
you
kind
of
an
overview
of
kind
of
where
we're
at
with
Narcy
and
where
we're
headed
over
the
next
over
the
next
the
next
decade,
and
so
as
I
mentioned
you
know,
nurse
was
established
originally
at
livermore
in
1974.
So
so
2014
will
write
it
right
at
40
years
it
had
a
long
history.
A
lot
happened
while
it
was
at
Livermore
in
in
96
it
moved
to
to
berkeley
lab
there's.
Also
a
name
change
became
the
national
energy
research
scientific
computing
center.
B
A
lot
of
our
if
you
notice
these
are
just
kind
of
highlights,
but
but
one
of
the
notable
things
that
was
going
on
while
nursed
moved
to
deburr
play
was
the
transition
from
vector
supercomputers
to
massively
parallel
computers,
and
that
was
actually
a
very,
very
challenging
time
for
a
lot
of
the
users
at
nurse.
Certainly
we
have
lots
of
even
back,
then
there
were
lots
and
lots
of
users
and
lots
of
codes
and
managing
that
transition.
That
was
actually
a
you
know.
B
Very
major
transition
for
the
for
the
community
was
was
definitely
a
challenge
and
and
in
some
ways
we're
facing
kind
of
a
similar
challenge.
Now,
as
we
go
to
multi-core,
we
have
again,
we
have
lots
of
users,
and
so
we
can't
perhaps
make
a
very
radical
shift,
but
we
recognize
that
we
need
to
start
making
a
shift
with
our
next
generation
systems.
B
And
so
so
we
do
right,
like
in
kind
of
the
you
know,
kind
of
what's
going
on
now,
with
the
move
to
multi-core
and
perhaps
a
new
programming
model
and
and
the
changes
that
are
going
to
be
required
in
the
software
to
what
we
had
successfully
managed
back
in
the
late
90s.
And
of
course
there
are
lots
of
lessons
learned
from
then
as
well
that
we
can
apply
to
now.
B
We
deployed
a
facility
wide
file
system
in
2006
and
we
started
a
collaboration
with
the
joint
genome
Institute.
So
we
provide
all
the
computing
for
jgi,
and
so
so
I'll
talk
more
about
this.
But
there's
been,
you
know,
there's
lots
of
talk
about
big
data,
but
we
deal
with
lots
of
big
data
at
nurse
as
well.
It's
more
scientific
data.
So
it's
it's
not
the
type
of
data
that
Google
deep
deals
with,
but
it's
still
very
large
quantities
of
data
that
we're
we're
dealing
with
now.
B
So
we
do
collaborate
with
computer
companies
to
deploy
advanced
HPC
and
data
resources.
I
was
at
as
Richard
mentioned.
I
was
at
La
Salle
with
Sandia
and
Los
Alamos
working
on
CL
0
it.
At
the
same
time,
nurse
was
working
on
hopper
and
those
were
sister
systems
that
were
deployed.
At
the
same
time
they
were
the
first
crepe
petascale
systems
with
a
Gemini
interconnect
right
now
we
have
we
have
accepted
and
we're
going
to
have
a
dedication
later
on
for
Edison
and
that's
the
first
Cray
petascale
system
with
Intel
processors,
aries
interconnect
and
dragonfly
topology.
B
It
was
a
serial
number
one
system
for
the
HP
CS
program
right
now,
we're
working
with
with
aces
again
they're
going
to
deploy
Trinity
in
the
2015-2016
time
frame
and
we'll
be
deploying
a
sister
sister
system,
nurse,
Cade
and
and
those
are
being
jointly
designed
on
ramps
to
exascale.
I'll
talk
a
little
bit
about
that
and
then
we've
architected
and
deployed
data
platforms,
including
the
largest,
do
a
system
focused
on
genomics.
So
with
all
of
these
you
know
these
are
leading
edge
systems.
There's
always
lots
of
challenges
when
you're
deploying
systems
like
this.
B
B
They
allocate
their
base
and
submit
proposals
for
over
target
so
and
then
the
deputy
director
of
science
actually
prioritizes
these
over
target
regret
requests.
So
so
we
directly
support
office
of
sciences
mission.
Each
of
the
center's
also
has
a
ten
percent
directors
reserved,
but
but
ninety
percent
of
our
resources
really
get
allocated
through
the
nurse
program.
The
other
point
is
that
that
our
usage
really
shifts.
So
you
know
we
we
we
can't
pick
our
users,
you
know
we
have
to
serve.
The
people.
B
Were
selected
as
users
and
then
the
usage
really
shifts
as
deal
we
priorities
change.
So
if
you
choose,
if
you
see
changes
in,
do
a
strategic
plan
that
directly
affects
the
problems
that
are
run
at
risk,
and
so,
if
you
look
at
the
changes
from
2002
to
2012,
one
thing
that
we
noticed
was
that
there's
so
in
blue
is
2012
and
in
red
is
in
2002
and
what
you'll
see
our
increases
and
things
like
material
science,
chemistry,
climate
research,
Biosciences
earth
sciences.
B
We
really
focus
on
the
the
scientific
impact
of
our
users,
so
this
is
you
know.
This
is
a
really
bragging
about
all
you.
It's
so
it's
easier
to
do.
I
was
having
a
hard
time
bragging
about
what
we're
doing,
but
this
is
a
bragging
about
our
users
and
we
really
enable
an
amazing
breadth
of
science.
We
have
typically
about
1500
journal
publications
per
year.
More
than
10
journal
cover
stories
per
year
there
were
17
this
past
year
there
were
three
recent
Nobel
Prize
winning
projects
that
use
nursing
physics,
magazines,
2013
breakthrough
of
the
year
users.
B
Resources
to
identify
the
first
high-energy
cosmic
neutrinos,
finding
earth-like
planets
are
not
uncommon,
was
recognized
by
Wired
magazine
as
a
top
scientific
discovery
and
covered
in
the
New
York
Times
MIT
researchers
developed
a
new
approach
for
desalinating
sea
water,
and
that
was
one
of
Smithsonian
magazines.
Fifth,
surprising
scientific,
milestone
of
2012
for
of
science
magazines.
Last
insights
of
the
last
decade,
three
in
genomics
and
one
related
to
Cosmic
Microwave
Background,
those
were
all
enabled
by
nurse
resources,
help
enable
those,
and
so
so,
as
I
mentioned
there
were
17
journal
covers
in.
B
B
Just
got
a
new
Edison
t-shirt,
but
but
I
have
a
hopper
t-shirt
now
its
kind
of
falling
apart,
so
I
can't
do
it
anymore,
but
every
time
I've
gotten
on
an
airplane
with
my
hopper
t-shirt
I've
had
someone
approached
me
saying:
oh,
do
you
work
at
nurse,
and
so
so
it
has
always
been
done
so
so
I
guess.
Maybe
that
means
our
users.
Not
only
do
we
have
a
lot
of
them,
but
they
seem
to
fly
a
lot
they're
geographically
distributed.
We
have
47
states
so
in
dark,
blue
or
states
with
over
100.
B
So
we
have
many
states
that
have
over
100
users
there's
lots
with
with
over
50.
So
it
really
is
a
nurse
really
is
a
national
resource
and
there's
a
guy
could
have
also
over
late
late,
an
international
map,
and
so
we
get
lots
of
users
from
really
all
around
the
world.
We
have
lots
of
international
projects,
and
so
so
it
really
is.
It's
amazing
to
see
who's
logged
on
every
every
day.
So
so
one
of
the
consequences
of
this
is
that
we
do
have
a
very
diverse
workload.
B
So
we
have
lots
of
users,
but
we
have
over
600
different
codes
and
and
algorithms
shown
in
the
upper
right
or
is
kind
of
the
breakdown
of
the
for
algorithms,
and
so
the
slices
are
the
codes
and
overlaid
on.
There
are
the
different
algorithm.
So
you
can
see
that
that
lots
of
people
do
fusion
pick
codes,
lattice,
qcd
density,
functional
theory,
climate
molecular
dynamics,
quantum
chemistry,
so
there's
lots
of
different
different
applications
and
algorithms
that
we
need
to
support.
B
B
So
so
you
do
have
to
be
capable
of
running
at
scale,
but
then
we
also
have
to
be
able
to
support
what
we
call
high-throughput
computing,
these
these
massive
massive
numbers
of
smaller
simulations
for
statistics
or
ensembles
that
people
need
to
do
and
we
want
to
be
able
to
do
that.
You
know
seamlessly
so
that
you
don't
have
to
go
to
a
different
system
just
depending
on
the
size
of
the
job
that
you
want
to
run.
You
know
our
operational
priority
is
providing
highly
available.
B
Hpc
resources
with
really
exceptional
user
support
will
try
to
maintain
a
very
high
availability
of
users.
So
so
we
we
do
chart
the
satisfaction
rating.
You
all
have
probably
gotten
an
email
for
me
asking
you
to
complete
our
user
surveys,
but
we
do
track
those
very
seriously.
We
look
at
all
the
issues
if
you
write
down
something,
someone
actually
will
go
and
look
at
it
and
we'll
we'll
try
to
figure
out.
B
What's
what's
going
on
so,
if
you're
having
issues
you
know
you
can,
you
can
call
directly
or
you
know
directly,
but
but
the
user
survey
is
actually
a
very,
very
useful
resource
for
us.
In
figuring
out
what
help
you
need,
we
also
want
to
maximize
the
productivity
of
our
users.
We've
had
some
conversations
with
a
date
good
one
hour
program
that
that
nurse
staff
has
actually
been
at
about
the
same
number
of
ftes
for
a
long
period
of
time,
and
and
during
that
time
our
number
of
users
has
grown
very
dramatically
and
so
shown
on.
B
The
left
is
the
number
of
users
and
and
and
the
the
red
line
is
the
users,
and
you
can
see
this
very
dramatic
increase
from
2000
to
where
we
were
perhaps
around
a
thousand
to
close
to
5,000
now,
and
so
we've
done
that
with
roughly
the
same
number
of
staff,
and
so
so
one
of
the
keys
here
is
is
keeping
the
number
of
tickets
per
user
down,
and
so
so.
This
is
a
very
important
metric
for
us.
B
So
we
we
have
to
have
a
plan
to
if
you,
if
you
generate
a
ticket
for
us
we
within
three
days.
We
have
to
have
some
kind
of
plan
for
trying
to
resolve
that,
and-
and
so
if
we
get
really
high
in
the
number
of
tickets
for
users-
and
we
have
more
and
more
users,
you
can
see,
we
can
really
very
easily
get
dwarfed
and
so
keeping
this
down.
So
we
threw
a
lot
of
hard
work
through
the
user
services
group.
B
B
A
that's
a
huge
change.
That's
really
been
necessary
necessary
by
this
dramatic
growth
in
the
number
of
users,
and
so
how
will
we
done
that?
So,
as
I
said,
you
know,
we
work
very
hard
to
make
the
systems
as
reliable
as
possible.
We
also
try
to
have
training
as
much
training
as
we
can
do.
We
try
to
have
very
useful
web
pages.
People
work
really
hard
on
those.
B
They
understand
that
if
you
don't
put
something
in
the
web
page
that
that
you're
going
to
get
a
phone
call
or
you
going
to
get
lots
of
phone
calls,
and
so
so
we
try
very
hard
to
do
things
that
are
scalable,
and
this
is
going
to
be
some
some.
This
has
to
be
part
of
our
strategy
also
for
moving
to
multi-core,
because
we
have
so
many
users.
We
have
to
do
things
that
are
scalable,
so
we
can't.
B
B
So
if
we
look
at
nurse
today,
this
is
the
these.
Are
the
large
systems
we've
just
just
deployed
Edison
we're
going
to
have
the
dedication
for
Edison
and
I
talked
a
little
bit
more
about
that.
We
have
hopper.
We
have
a
number
of
production
clusters
viz
and
analytics
data
transfer
nodes
all
connected
by
by
very
high
speed
networking.
We
have
a
local
scratch
on
both
of
our
large
systems,
and
then
we
have
a
global
file
system
as
well
as
well
as
archival
storage.
We
just
went
to
100
gigabit
connection
to
the
outside.
B
So
this
just
shows
the
large
systems
within
within
Office
of
Science.
Currently,
as
I
said,
we
always
have
to
at
the
same
time.
So
when
we
bring
on
nurse
aide
will
will
have
retired
hopper
by
that
time,
and
at
that
time
our
systems
will
be
north
gate
and
end
and
Edison.
But
we
really
I
was
actually
out
of
meeting
we're
talking
about
redefining
the
Linpack
benchmark,
but
we
really
don't
focus
on
the
top
line.
Peak
flops.
B
We're
really
focused
on
scientific
productivity
and
deploying
systems
that
are
that
are
useful
to
all
of
you
and
what's
happening
partially.
Is
that
the
benchmark
that
people
use
that
lint
linpack?
Even
the
developers
recognize
that
it's
becoming
more
and
more
divergent
from
the
everyday
applications
that
we
see?
B
So
we
really
focus
on
other
things,
and
one
of
the
things
we
focus
on
is
is
its
memory
having
lots
of
memory
per
node
lots
of
memory
bandwidth
for
node,
more
and
more,
the
performance
of
applications
is
limited
by
your
your
ability
to
transfer
data,
not
not
as
much
as
as
your
ability
to
compute
people
will
say
that
that
floating
point
operations
are
for
free
and
that's
because
you're
often
just
waiting
waiting
for
for
data
to
arrive.
So
you
may
you're
lucky.
B
You
know
if
you
do
a
calculation,
every
eight
or
ten
clock
cycles,
you're,
probably
actually
doing
pretty
well
so
so
so
we
work
very
hard
to
make.
The
the
in
collaboration
with
Cray
and
Intel
the
memory
bandwidth
is
per
per
node
is
very
high.
The
peak
bisection
bandwidth
of
the
new
interconnect
that
Cray
is
deployed.
B
B
So
if
you,
if
you
drove
up
or
came
up
in
the
shuttle,
you
probably
saw
our
new
you
building
and
you
probably
have
heard
something
about
it,
but
but
that
CRT
on
the
top.
That's
the
artist's
rendition,
and
this
is
a
little
bit.
It's
there's
been
every
every
few
days.
You
see
significant
progress,
but
but
it's
looking
more
and
more
like
a
building,
it's
pretty
imposing
as
you
look
up.
B
The
big
thing
for
us
is
that
it's
going
to
give
us
a
lot
more
space
for
the
systems
and
a
lot
more
power,
and
the
other
thing
is
that
it
will
bring
nurse
back
to
currently
we're
in.
For
those
of
you
visited
us
we're
in
downtown
Oakland,
so
it
will
bring
us
back
onto
the
hill
and
so
will
be
co-located
with
the
research
division
and
with
es
net
and
and
we'll
have
certainly
be
much
closer
to
UC
Berkeley.
B
So
so
there'll
be
a
lot
of
collaborations
that
this
this
will
foster,
but
it
will
also
give
us
lots
of
floor
space
for
44
nurse
gate
and
nurse
9.
The
building
is
being
designed
for
exceptional
energy
efficiency,
so
it's
natural
air
and
water
pooling
there
are
no
no
chillers
we're
trying
to
do
heat
recovery.
The
of
course
you
don't
get
I.
Guess
we
in
California
we
get
cold
when
it's
50
degrees,
but
but
we
can't.
B
So
we're
going
to
try
to
make
it
as
as
painless
as
possible
for
all
of
you,
you
can't
make
a
move
of
the
size
without
some
disruption.
We
work
very
closely
with
some
of
the
folks
from
Livermore.
Actually
who've
done.
Some
similar
moves
to
understand
some
of
the
issues
that
that
we
might
be
facing
as
we
as
we
make
this
change.
B
We're
going
to
move
the
systems
and
phases,
hopper
will
stay
actually
in
in
oakland
and
will
be
retired
there,
but
Edison
will
move
and
will
deploy
nurse
gate
in
the
in
the
new
building.
So
Edison.
You
know
we're
trying
to
keep
it
down
for
as
little
time
as
possible,
and
so
so
for
the
smaller
clusters.
You
know
we
can.
We
can
probably
do
the
moves
in
about
two
weeks.
B
So
we've
been
working
very
hard
on
on
nurse
eight,
you
know
Edison,
we
made
a
decision
not
to
deploy
GPUs
within
Edison
would
have
given
them.
A
much
higher
peak
would
have
done
much
better
on
linpack,
but
are
you
know
for
a
majority
of
our
users?
It
didn't
look
like
all
of
our
users
were
quite
ready
for
that
jump.
We
do
think
that
in
the
2015-2016
time
frame,
there's
been
lots
of
work
done
and
I
think
people
are
better
poised
to
make
the
leap
we've
identified.
B
People
who
we
need
to
work
with
and
I
think
it
really
is
necessary
to
you
know,
show
some
some
data
on
that
later,
but
I
think
it
really
is
necessary
to
make
this
change
to
multi-core
and
and
more
energy-efficient
architectures.
So
the
mission
need
for
for
nurse
gate
is
actually
the
users
need
is
much
higher,
but
we
establish
the
mission.
Need
is
at
least
10
*
hopper
on
on
a
representative
set
of
do
ii
benchmarks,
but
but
we
would
like
to
get
that
on
real
deal.
B
We
applications
we
do
need
to
provide
very
high
bandwidth
access
to
existing
data.
There's
lots
of
data
already
stored
here,
and
so
so
so
northgate
needs
to
integrate
within
that
environment
very
seamlessly
and
I
said
it.
As
I
said,
we
need
to
have
a
plan
to
begin
transitioning
our
user
base,
so
this
should
actually
not
say
not
yet
known.
It's
just
not
public.
B
So
we're
hoping
we'll
be
able
to
announce
this.
We
have
an
independent
project
review
later
this
month,
and
so
our
hope
is.
This
was
a
joint
project
with
Los
Alamos
and
Sandia
on
Trinity.
So
it
doesn't
really
make
sense
for
one
of
us
to
announce
it,
because
everyone
will
know
what
the
other
system
is,
and
so
so
we're
going
to
try
to
do
a
joint
announcement,
hopefully
in
early
March-
and
it
looks
like
we're
on
track
for
that.
B
So
so
at
one
of
the
one
of
the
telecoms
in
March
or
April
will
certainly
give
you
all
a
lot
more
detail
about
what
what
we
can
say
about
nurse
gate.
It's
still
going
to
be
an
early
announcement
of
the
technology,
and
so
we
need
to
work
with
the
technology
providers
well,
to
figure
out
exactly
what
we,
what
we
can
say,
but
I
think
it
will
be
evident.
I
mean
I.
Think
people
understand
what
what
changes
that
that
need
to
be
made
to
begin
preparing
for
nurse
gate
and
really
no
matter
what
the
technology
is.
B
There's
a
lot
of
similarity
in
all
the
systems
and
really
multiple
levels
of
code
modification
may
be
necessary.
You'll
need
to
expose
more
on
node
parallelism,
increased
application
vectorization
if
there
was
a
coprocessor
architecture
than
you
have
worried
about
locality
directives,
if
there's
some
local
scratch
or
scratch
path
than
then
something
you'd
have
to
look
at.
As
far
as
how
do
you
most
effectively
use
that?
Can
the
compiler
just
make
use
of
that?
Or
do
we
need
to
add
code
directives
to
be
able
to
effectively
use
that?
B
So
this
is
something
that
will
be
will
be,
as
I
said,
we'll
be
talking
to
you
more
about
this
just
shows
notionally
what
we're
looking
at
we're.
Looking
at
a
system
that,
in
terms
of
system
peak,
would
be
20
to
40
pedo
flops.
The
system
memory
would
be
around
a
petabyte.
A
lot
of
this
depends
on.
There
are
also
lots
of
budget
lots
of
options
in
the
contract,
so
so
part
of
this
is
to
be
a
little
bit
fuzzy
on
what
the
system
is,
but
part
of
it
is
also
that
there
are.
B
There
are
options
to
make
make
changes,
because
again
this
is
still
early,
and
so
so
we
don't
exactly
know
what
Congress
is
going
to
do
over
the
next.
The
next
the
next
couple
of
years,
the
note
performance
will
be
up
by
you-
know
something
like
a
factor
of
five
we're
hoping
to
see
some
increase
in
node
memory,
bandwidth,
you're,
you're,
probably
going
to
you
know
the
higher
number
would
be
is
if
there's
some
some
scratch
pad
or
some
some
some
memory.
B
That's
on
package
we're
going
to
see
a
lot
more
concurrency
in
terms
of
system
size.
We
will
see,
you
know
some
increase,
it's
not
going
to
be
a
huge
increase
in
the
number
of
nodes.
The
main
thing
you're
going
to
see
is
a
lot
more
node
concurrency
increasing
the
node
interconnect
connect
bandwidth
in
this
time.
Time
frame
is
really
a
chat.
Is
it
is
a
challenge,
so
it
may
actually
be
nurse
night
or
you
see
really
a
big
big
boost
in
this,
but
but
we
are
working
to
see
if
we
can
improve
that.
B
So
so,
we'll
need
to
use
a
number
of
different
approaches
to
begin
preparing.
The
user
community
we're
having
with
the
vendor
and
the
technology
provider,
will
be
working
with
North
kinases
to
we'll
be
doing
developer
workshops
early
test
beds.
An
early
test
bed
for
users
would
be
engaging
with
the
the
application
teams
want
a
partner
and
within
leverage.
Existing
efforts
are
certainly
a
lot
going
on
with
the
community,
and
so
it's
not
like
nurses
going
to
do
this
all
by
ourselves.
B
So
we
need
to
coordinate
with
with
oakridge
and
argon
and
the
other
centers,
and
we
really
need
to
do
widespread
training.
As
I
said,
you
know,
we
have
lots
of
users
and
so
to
manage
is
transition
for
everyone.
We
really
need
to
do.
Things
are
scalable,
so
we
want
to
post
workshops
online
training
create
easy-to-follow
online
documentation
to
help
users
as
much
as
possible.
B
Ultimately,
all
of
those
are
necessary
to
get
good
performance,
but
what
you
are
seeing
in
all
cases
is
that
the
blue
line,
as
you
do
this
work
becomes
goes
down
as
well,
and
so
so
one
benefit
of
getting
ready
for
these
next-generation
architectures
is
that
actually
your
code
will
run
better
on
the
Xeon
on
xeon
clusters
and
xeon
work
stations
as
well?
So
so
so
that's
that's
a
that's
I
think
that's
a
very
important
part
of
this
as
well.
B
So
we
spend
a
lot
of
time
talking
with
you
all,
partly
through
requirements,
reviews
partly
through
meetings
like
this,
to
figure
out
what
we
need
to
be
deploying
in
the
future.
What
direction
we
need
to
be
heading,
so
we
do
requirements
with
six
program
offices,
so
within
the
Office
of
Science,
we
try
to
do
the
reviews
every
33
years,
so
we're
almost
through
our
second
set
of
second
set
of
reviews.
B
So
we
pick
a
target
so,
like
would
say
2017,
so
we
go
and
have
a
discussion
with
the
program
managers
and
the
scientists
we
try
to
get
a
representative
set
of
the
user.
So
typically
we
want
people
at
the
meeting
who
represent
more
than
fifty
percent
of
the
usage
within
that
within
that
program
office.
We
try
to
really
have
a
discussion
about
science
goals
and
representative
use
cases.
Oftentimes.
B
The
scientists
aren't
really
that
interested
in
flops
and
bites
and
bandwidth
and
all
these
things
they're
really
there's
certain
science
problems
that
they
need
to
solve
in
a
certain
time
frame,
and
then
we
work
with
them
to
go
backwards
based
on
those
so
science
that
requirements
did
to
estimate
what
the
what
the
computational
requirements
are.
And
then
we
do
rescale
the
estimates
to
account
for
users,
not
at
the
meeting.
B
We
aggregate
results
across
the
six
program
offices
and
then
we
try
to
validate
through
in-depth
collaborations
things
like
the
this
meeting
here
and
the
the
user
surveys.
This
does
tend
to
underestimate
the
need,
because
you're
you
know,
sometimes
you
can
anticipate,
who
might
be
a
future
user,
but
oftentimes,
you
don't
know
who's
going
to
be
really
a
new
user
in
five
years.
B
So
so
this
does
try
to
underestimate
things.
But
this
just
shows
in
the
black
line,
is
kind
of
the
overall
trend
and
in
red
are
the
actual
nurse
hours
delivered.
So
you
see
a
slight
dip,
which
is
when
Franklin
was
turned
off
and
now
that
we're
deploying
Edison
you're,
seeing
that
that
green
triangle-
that's
a
that's
hopper,
plus
Edison,
and
you
can
see
that
that,
since
we
went
with
a
Z
on
architecture,
we
didn't
deploy
GPUs,
we
were
actually
falling
significantly
below
the
trend
line.
The
black
arrows
are.
B
These
aggregate
needs
from
these
user
requirements,
workshops
and
those
are
actually
much
higher
than
the
trend
line.
Even,
and
so
you
can
see
that
not
only
were
you
know,
we're
way
short
of
the
user
requirements,
but
we're
also
falling
below
the
trendline
and
that's
why
we
need
to
transition
to
a
more
energy-efficient
architecture,
and
so
so
with
nurse
8.
You
see
a
range,
that's
mainly
driven
by
budget,
but
but
we
need
it
will
help
us
begin
to
get
closer
to
that
that
trend
line
as
well.
B
The
other
thing
is
a
you:
can
you
can
kind
of
project
out
from
these
requirements?
Reviews
and-
and
you
can
see
that
that
they're
increasing
very
rapidly
and
in
the
2018
timeframe,
the
aggregate
needs
of
our
users
will
actually
be
in
the
exascale
regime.
So
so
so
you
hear
lots
of
talk
about
exascale,
but
but
just
looking
at,
what's
going
on
and
trying
to
project
with
our
users,
it's
pretty
clear
that
to
do
the
science
if
they
they
need
to
do
in
2018.
2019
really
requires
exascale
computing,
so
this
just
shows
I
asked.
B
We
always
show
these
things
on
log
plots,
but
if
you
show
it
on
linear
plot,
you
actually
see
it
really
is
a
big
gap
between
between
kind
of
what
we're
looking
at
and
the
user
requirements
or
or
needs
are
much
much
higher,
and
we
also
ask
people
about
about
storage,
networking
other
things,
and
so
so
the
the
needs
that
we're
requiring
that
we're
getting
from
these
requirements.
Reviews
also
pointed
big
shortfalls
and
things
like
archival
storage.
As
far
as
what
the
what
the
users
need.
B
B
This
just
shows
there's
lots
of
different
different
plots
like
this
could
show.
But
this
is
just
one
weekend
last
year
and
you
see
data
traffic.
They
were
doing
a
particular
study
down
at
slack
and
you
could
just
see
the
the
data
traffic
very
high
speed
data
traffic
to
to
nurse
for
the
entire
weekend,
and
so
this
often
happens
is
they're
running
some.
Some
big
experiment
at
some
some
facility
and
they're,
really
in
undated
with
data
and
they've,
got
to
take
that
data
somewhere
to
analyze
and
try
to
get
scientific
understanding.
B
And
so
we
see
these
huge
inflows
of
data
traffic
depending
on.
What's
going
on
at
the
difference,
do
II
experimental
facilities,
and
so
so
I
guess
I
can't
won't
be
able
to
say
this
a
lot
longer.
But
but
you
know
this
is
one
of
the
things
that
surprised
me.
So
I
find
something
else.
That
surprises
me,
but
one
of
the
things
that
surprised
me
is
that
that
nurse
users
actually
import
more
data
than
they
export,
which
is
you
know
you
think
about
yourself
as
a
supercomputing
center.
B
What
you
think
of
is
that
people
will
do
simulations
and
take
data
away,
and
so
they
are
taking
lots
of
data
way
where
it's
exporting
and
this
these
numbers
are
actually
much
higher.
Now
we
need
to
update
them
because
they're
they're,
typically
above
a
petabyte
a
month,
especially
since
we've
gone
100
gigabit
networking,
but
we
were
in
blue,
is
how
much
data
we
were
exporting
and
in
red
is
how
much
we're
importing,
and
so
you
can
see
a
lot
of
data
is
coming
in.
B
So
you
could
say
well
where's
that
coming
from
to
be
honest,
some
of
that
does
come
from
the
other
centers
because
they
have
inside
out
allocations
that
end
and
suddenly,
all
that
data
ends
up
coming
to
nurse.
But
a
lot
of
that
data
is
also
experimental
data
coming
from
from
cosmology
or
high
energy
physics
or
from
for
from
daya
Bay,
or
get
lots
of
traffic
from
jgi.
So
lots
of
different
facilities
are
always
transferring
data
from
from.
E
B
The
advanced
light
source
here
from
slack
and
so
so
you'd
expect
that
well
with
all
this
data
coming
here,
that
people
must
be
doing
something
with
it
right
and
we
have
lots
of
examples
of
a
scientific
discovery.
That's
kind
of
the
traditional
uses
of
hpc,
but
we're
also
seeing
that
data
analysis
is
playing
a
key
role
in
scientific
discovery.
So
a
lot
of
the
highlights
that
we're
having
here
at
nurse
we're
seeing
more
and
more
examples
where
its
extreme
data
science
or
related
to
extreme
data
science.
So
there's
the
palomar
transient
factory.
B
B
There's
an
example
of
the
one
of
the
the
very
early
discovery
of
type
1a
supernovae
in
the
last
40
years,
so
those
discovered
overnight
and
then
instantly
the
the
scientists
called
collaborators
around
the
world
and
really
telescopes
from
around
the
world
were
refocused
on
that
on
the
on
the
same
supernova.
So
they
could,
they
could
follow
its
birth.
So
that's
resulted
in
lots
of
refereed,
publications
and
nature
articles
and
and-
and
it
really
makes
heavy
use
of
the
science
gateway
nodes
to
share
the
data
among
the
collaboration
as
well.
B
B
Solving
the
puzzle
of
the
neutrino,
with
with
the
daya
Bay,
there
are
lots
of
their
detectors
that
they
transfer
data
tuner
scanned,
analyzed
it
and
they
were
able
to
measure
the
theta
13
neutrino
parameter,
which
is
the
last
and
most
elusive
piece
of
a
long-standing
puzzle.
Why
neutrinos
appear
to
vanish
as
they
travel
and
so
so
that
really
used
a
high-performance
computing
for
simulation
and
analysis.
B
The
the
data
came
in
to
archival
storage
and
use
the
data
transfer
capabilities,
including
es
net
and
I
use
the
the
the
nurse
global
file
system
and
then
science
gateways
for
distributing
the
results,
and
this
was
one
of
science
magazines,
top
10
2012
breakthroughs
the
clock
mission.
There's
a
lot
of
right
up.
It
was
one
of
physics
world's
top
ten
breakthroughs
of
2013,
but
a
european
space
agency
satellite
mission
to
measure
the
temperature
and
polarization
of
the
Cosmic
Microwave
Background,
realizing
the
full
scientific
potential
of
this
required
really
large
computational
resources.
So.
D
B
The
data
nurse
was
the
primary
computing
site
for
for
Punk
and
all
that
data
was
transferred
here
and
analyzed
and-
and
we
have
a
materials
project
where
this
was
recently
cover
a
cover
on
scientific
american.
But
but
we
have
lots
of
users,
lots
of
companies
that
are
using
it,
but
but
we're
we're
providing
as
the
infrastructure
so
that
that
people
can
screen
materials
using
computational
ins,
that's
much
cheaper
than
making
them
in
lab.
We've
had
more
than
35,000
inorganic
materials
calculated
in
two
years
and
those
are
coupled
with
online
design
and
search
search
tools.
B
So
I
could
give
you
a
very
long
talk
on
this.
So
if
you're
interested
sometime,
we
can
talk
about
it.
You
know,
there's
lots
of
discussion
about
exascale
I
was,
as
Richard
mentioned.
I
was
on
the
exascale
initiative,
steering
committee
for
do
II
and
we
developed
a
roadmap
for
do
we
and
identified
some
of
the
key
challenges.
You
know.
There's
lots
of
discussion
of
big
data.
You
know
granted
for
nurse
gets
its
its
scientific
data,
but
there's
lots
of
discussion
of
big
data,
and
sometimes
these
things
are
viewed
as
being
orthogonal
but
they're.
B
Really
they
really
faced
the
same
computing
challenges
so
in
terms
of
things
like
like
energy
efficiency,
in
terms
of
a
lot
more
concurrency,
your
inability
to
move
data
at
the
same
rate
as
you
can,
you
can
perform
the
computations.
All
those
are
really
common
challenges
between
big
data
and
concurrency
and,
like
I,
said
big.
B
B
This
is
kind
of
leveled
off
some,
but
the
cost
per
genome
had
been
going
down
pretty
dramatically
and
when
we
look
at
things
like
the
advanced
light
source
and
and
an
stand
at
slac,
the
data
rate
has
been
growing
and
growing,
and
you
know
we
can
see
that
in
the
future.
You
know
we're
going
to
be
at
terabit
per
second
type
data
rates
and
and
we're
seeing
that
data
sets
are
going
to
be
going
to
hundreds
of
petabytes,
and
so
so.
This
is
something
that
we
do
expect
to
play.
B
B
You
know
we
have
to
do
great
in
operations
because
really
anything
we
do
has
to
be
built
on
that
base,
but
but
on
top
of
our
operational
base,
we
see
that
we
need
to
deploy
exascale,
usable,
exascale,
computing
and
storage
systems
for
our
users.
Just
as
you
look
at
the
demand
for
computing
within
the
Office
of
Science,
it's
really.
It's
really
amazing
to
see
those
those
curves.
We
need
to
start
transitioning
the
codes
and
for
nurse.
B
This
is
a
big
challenge,
because
we
can't
spend
three
ftes
per
code,
helping
them
transition
to
many
core
architectures,
and
we
do
think
we
have
a
critical
role
in
terms
of
influencing
the
computer
industry
to
ensure
that
future
systems
really
meet
the
mission
needs
of
the
office
of
science,
but
really
meet
the
needs
of
our
broad
base
of
users
and
that's
something
of
very
much
of
interest
to
the
computer
companies
as
well.
They
don't
want
to
deploy
systems
just
for
one
application,
it's
hard
enough
to
get
them
to
pay
attention
to
high
performance
computing.
B
E
B
So
so
I
think
with
all
of
these
you
know
we
made
the
decision
to
wait.
Our
preference
would
be
to
have
a
self-hosted
architecture,
so
we
think
that
if
you
avoid
right
now,
people
are
trying
to
connect
accelerators
through
a
PCI
connection
and
that
really
you're
really
being
limited
in
many
cases
by
by
the
very
poor
bandwidth
that
you
have,
and
so
our
hope
with
northgate
would
be
for
it
to
be
a
self-hosted
architecture
and
that
that
should
that
should
help
quite
a
bit
a
bit.
B
But
to
be
frank,
there
is
a
significant
challenge,
so
so
getting
good
performance
on
these
next
generation
systems
isn't
always
going
to
be
easy.
Some
people
are
already
working
on
it,
so
we've
actually
done
some
some
detailed
analysis
of
which
codes
are
more
ready,
which
ones
need
more
help
and
we're
trying
to
target
the
ones
that
we
need
help,
and
we
also
want
to
coordinate
with
oakridge
and
argon.
B
We
don't
want
to
be
working
on
the
same
set
of
codes,
but
then
again
we
also
need
to
make
sure
that
that
that
the
transition
that
we're
doing
is
is
somehow
related
to
the
transition
that
they're
doing
so.
There
is
some
need
for
coordination
among
the
centers
as
well
so
bill.
You
had
a
good
razor
in.
Are
you
okay,.
B
C
B
So
so,
so
that's
why
we
really
have
these
two
strategic
objectives,
these
two,
what
we've
been
calling
initiatives
they're
very
interrelated,
but
you
know
I
think
that
the
top
one
is
really
aimed
at
our
more
of
our
traditional
work.
Workflow,
which
a
lot
of
it
is
is
solving
these
scientific
problems.
These
pdes
or
or
or
other
things,
and
the
bottom
is-
is
at
related
data.
So
we
think
that
both
are
both
are
important
for
our
future.
Whether
you
can
deploy
a
single
system
that
meets
both
sets
of
needs.
That's
a
that's
something
that
we're
exploring.
B
So
I
would
say
that
that
two
thirds
or
three
quarters
of
the
the
issues
as
to
why
people
are
deploying
different
systems
here
have
to
do
with
more
software
as
opposed
to
architectural
differences.
And
so
so
we
are
looking
at
on
our
large
systems.
Can
we
provide
the
same
set
of
software
services
that
people
need?
Thank.
A
B
B
We
are
working
with
jaye
GI
to
see
if
parts
of
their
workflow
can
can
run
on
Edison
and
hopper
they'd
love
to
be
able
to
use
some
of
their
their
cap
allocation
for
some
other
their
computing
right
now.
Gene
pool
is
very
much
tailored
to
their
workflow
right
and
so
so.
Making
that
transition
is
not
not.
B
Not
particularly
easy,
but
it's
something
that
we're
going
to
be
exploring
with
with
JD
I.
We
will
be
moving
gene
pool
into
CRT,
so
I
would
expect
us
to
have
a
you
know:
the
collaboration
with
with
Jean
with
JD
I,
to
continue
for
a
long
period
of
time.
The
collaboration
with
PDS
south
and
the
high
energy
physics
community
is
even
longer,
and
so
I
would
so
we're
certainly
expecting
to
move
so
we'll
certainly
be
moving.
Mendel
will
certainly
be
you
know.
B
Our
expectation
would
be
that
PDS
f
is,
you
know,
will
continue
in
in
a
you
know.
That
would
be
a
longer
you
no
longer
term,
but
we
are
looking
at.
Can
we
can
we
move
some
of
that
workflow
to
northgate,
and
so
some
of
the
things
that
we're
looking
at
is
potentially
having
a
data
partition
on
northgate
and
that's
something
we
can
talk
with.
B
You
will
certainly
be
able
to
talk
with
you
more
about
in
a
few
months,
and
it
would
be
certainly
useful
to
understand
how
useful
that
would
be
for
some
of
these
for
some
of
these
problems,
but
the
idea
would
be
to
have
you
know
you
could
have
a
more
x86
park-based
partition
that
had
more
memory
that
also
had
access
to
the
first
buffer
that
potentially
had
a
different
software
stack.
And
so
could
you
meet
some
of
these
needs?
B
D
So
there's
currently
demand
for
infrastructure
to
host
a
human
related
data,
particularly
with
the
brain
initiative
in
many
people
here
in
Albiol
actually
participate
in
those
initiatives
that
really
to
human
data
is
there?
What's
the
likelihood,
the
nurse
will
come
to
you
host
human
related
data
and
be
HIPAA
compliant
and
so
on.
B
Yes,
so
it
depends
on
the
the
the
the
types
of
data.
That's
that's
something
we're
looking
at
if
it's
a
if
it's
HIPAA
compliant
or
you
know,
if
it
requires,
there's
some
data
that
isn't
associated
with
a
a
particular
person
right
and
so
that
that
doesn't
have
the
same
security
issues
as
if
it's
tied
to
a
particular
person
right,
if
it's
so
so
so
we
are
looking
at
different.
Could
we
provide
different
levels
of
security
for
different
types
of
data?
B
B
B
But
that's
always
an
issue
when
you
start
moving
into
data.
You
know
if
it's
data
from
slack
they're,
probably
not
that
that
that
consider
the
experimental
data
has
value
for
long
period
of
time.
You
could
argue
what
the
value
of
computational
data
is
over
time,
but
experimental
data
typically
becomes
more
valuable
with
time,
because
you
can't
go
take
a
snapshot
of
the
of
the
night
sky
or
you
can't
go
rerun
a
accelerators
experiment.
B
So
once
you
have
that
day
that
you
really
want
to
hold
on
to
it
for
a
long
period
of
time,
so
they
have
lots
of
issues
with
resilience
and
longevity.
If
you
add
privacy
to
it,
you
know
so
a
lot
of
the
data
we
deal
with
doesn't
have
this
privacy
issue,
but
but
but
that's
something
we
have
to.
We
have
to
look
at.