►
From YouTube: NUG Meeting 2014: Brooks
Description
No description was provided for this meeting.
If this is YOUR meeting, an easy way to fix this is to add a description to your video, wherever mtngs.io found it (probably YouTube).
A
A
So
I
got
asked
to
come
out
here:
pete
ungaro
couldn't
make
it
unfortunately,
because
we're
having
a
big
leadership
meeting
in
phoenix-
and
I
was
there
too,
but
I
you
know
they
said
you
you
go,
and
so
I
got
I've
not
been
here
for
the
40
years,
but
I
have
been
working
at
cray
for
26
or
seven,
something
like
that.
A
So
when
I
think
about
nurse
can
cray-
and
you
know
you
sort
of
think
when
did
it
all
begin
and
a
lot
of
people
might
say
1978,
because
that's
when
the
cra
one
arrived
here
and
but
of
course
everybody
knows
that
you
know
you
might
say
we
were
here
from
the
very
beginning,
because
the
cdc
6600
I
mean
everybody
knows
that's
a
seymour
cray
machine.
A
A
So
a
lot
of
times
people
talk
about
the
6600
being
the
first
supercomputer
ever
so
nurse
actually
started
with
the
first
supercomputer
that
was
ever
built.
Now
it
was
hand-me-downs
from
livermore
they'd
used
it
for
10
years
and
it
was
pretty
tired
out
by
the
time
you
got
it,
but
it
was
was
serial
one
in
a
lot
of
ways.
This
was.
I
never
worked
on
one
of
these
now
there's
one
of
these
in
the
chippewa
museum
and
apparently
paul
allen's,
trying
to
get
it
running
again,
so
that'd
be
kind
of
cool.
A
If
that
that
happens,
but
you
know
a
lot
of
times,
people
call
this
the
first
risk
computer.
You
know
it
had
a
load
store
architecture.
It
had,
I
think,
only
70,
some
operands
and
you
know
patterson
would
coin
this
term
years
and
years
later,
but
this
they
were
all
the
properties
were
in
this
machine
and
so
of
course
the
joke
at
cray
was
always.
It
stands
for
really
invented
by
seymour
cray.
A
I
wonder
whose
job
it
was
to
enact.
You
know
to
answer
those
four
acoustic
modems
you
know
manually
when,
when
users
called
in
to
use
this
thing,
of
course,
that
that
machine
was
also
the
subject
of
this
famous
memo,
which
has
now
been
50
years
from
t.j
watson.
You
know
kind
of
an
angry
memo
about
in
the
kind
of
it's
the
beginning,
the
rivalry
that
we've
had
with
ibm.
A
You
know
where
he
talks
about
the
small
lab
and
chippewa,
and
you
know
just
a
handful
of
people,
including
the
janitor,
and
how
can
we
lose
with
our
vast
development
efforts?
Now
there's
supposedly
this
legend
at
cray?
I
don't
know
if
it's
true
that
seymour,
you
know,
saw
this
memo
and
said.
The
answer
to
his
question
is
right
in
his
own
memo,
but
I
don't
know
if
that's
really
true
or
not.
A
So,
of
course
you
guys
know
what
came
after.
That
was
the
cdc
7600.
I
think
the
major
new
feature
that
that
was
important
in
this
thing
was
pipeline
functional
units.
It
was
the
first
machine
with
pipeline
functional
units.
You
would
issue
a
multiply,
takes
ten
clocks
to
go
through
the
multiply
and
on
the
6600
you'd
have
to
wait,
10
clocks
to
issue
another
multiply,
but
on
this
machine,
oh
no,
you
could
just
issue
a
multiply
right
after
as
soon
as
you
went
down
the
pipeline.
Of
course
this
became
very
important
for
vector
stuff.
A
That
would
happen
later
on
and
and
this
you
know
a
lot
of
people
at
craig
that
are
still
there.
There
are
a
few
people
left
that
actually
worked
on
this
machine
and,
of
course
this
was
freon
cool.
In
fact,
the
the
paul
allen
thing.
What
I've
heard
is
the
hardest
thing
to
get
that
old
machine
running
again
is
to
get
the
freon
refrigerant
stuff,
going
in
the
thing
of
course,
and
then
there's
the
cray
one
actually
worked
on
one
of
these
at
the
university
of
minnesota.
A
It's
the
first
cray
I
ever
got
on,
but
it
was
serial
12,
not
serial
six,
and
you
know
you
ran
the
ctss
operating
system
on
here,
which
is
probably
the
first
time
you
weren't
happy
with
what
the
vendor
gave
you
and
decided
to
do
something
else.
Instead,
so
another
one,
you
know
you've
seen
the
pictures.
You
know
these.
These
disc
drives
always
reminded
me
of
the
laundromat,
but
we
had
dd-29s
at
minnesota.
A
These
were
19s
and
we
had
the
same
tricks
where
we
could
do
stuff
with
io
and
make
them
walk
down
the
floor
and
so
on.
It
was
kind
of
kind
of
neat
stuff,
but
the
big
innovations
in
this,
I
think,
really
were
vector-
processing,
of
course,
but
also
compilers.
We
had
a
guy
named
dick
nelson,
really
smart
guy.
He
invented
dependency
analysis
in
a
compiler.
He
wrote
the
first
compiler
and
he
wrote
it
in
assembler.
A
It
was
like
later
on,
the
compilers
were
written
in
pascal
of
all
goofy
languages,
but
back
in
those
days
the
compiler
was
written
in
assembler.
It
was
lightning
fast,
you
know,
cft
and
then
later
cft2
packaging
was
pretty
cool
with
the
with
the
freon
first
one
to
use
integrated
circuits
and,
of
course,
vector
processing.
A
Now
I
can
talk
about
transitions.
I
think
one
of
the
things
that
nurse
is
pretty
good
at
is
transitioning
users
from
one
technology
to
the
next.
So
this
is
the
first
major
transition
right.
You
went
from
a
scalar
machine
to
a
vector
machine,
but
I
claimed
this
was
the
easiest
trend
transition
and
the
reason
was
in
the
marketplace.
A
A
So
I
think
this
is
probably
a
pretty
easy
transition
from
from
the
7600
and
then,
of
course,
after
this
is
the
is
the
xmp
which
came
in
1984
shown
in
the
foreground
of
this
picture.
Of
course,
the
you
know,
steve
chen
did
this
machine
because
seymour
was
off
doing
the
crate
ii
at
this
point
in
time
and
as
a
benchmarker.
This
is
kind
of
when
I
was
starting
at
craig.
We
were
still
selling
these
things,
and
this
was
a
great
machine.
A
You
know
it
had
the
multi-port
memory
with
you
could
do
two
reads
in
a
store
it
had
this
cool
new
instruction
called
gather
scatter.
No
one
had
ever
had
that
before
we
have
a
register
full
of
operands
and
you
can
gather
out
of
it
and
and
then
it
had
this
feature
called
chaining
where
you
could
go
right
out
of
a
vector
register
right
into
a
functional
unit.
You
didn't
have
to
wait
for
the
whole
register
to
be
filled.
A
I
always
loved
this
machine
because
this
was
just
such
a
great
looking
machine.
There's
one
of
these
is
in
some
art
museum
in
in
france
somewhere.
You
know,
because
these
are
so
attractive,
crate
twos
I
mean
they.
You
know
world's
most
expensive
aquarium.
You
know
these
were
20
million
dollars
and
we
were
enamored
with
them
because
they
had
such
this
high
clock
rate,
but
there
were
a
few
innovations
in
the
in
the
in
the
area.
A
I
think
the
major
innovations
I
mean
the
floor
inert
thing
remember
the
early
crate
twos
they
had
round
round
tanks.
For
that.
The
reason
they
had
to
switch
to
this
design
was
because
the
weight
of
the
floor
nerd
so
heavy.
It
actually
warped
those
plastic
cylinders
and
they
were
starting
to
kind
of
look
like
they
might
crack.
That
stuff
was
200
a
gallon.
If
I
remember
right
so
they
went
to
this
design,
and
this
is
the
eight
processor.
It
was
the
only
one
in
the
world
that
that
was
delivered
here.
A
The
other
thing
on
this
machine
is
a
benchmarker.
You
know
it
only
had
a
single
port
memory,
so
you
you
could
read
from
this
memory
out
into
one
vector
register,
but
you
really
it's
hard
to
go
faster
than
that.
You
had
to
use
this
local
memory.
There
was
this
local
memory
and
the
compiler
was
supposed
to
figure
out
how
to
use
it
never
did.
But
if
you
wanted
to
do
something
like
matrix
multiply,
you
could
squirrel
away
your
temporaries
in
the
local
memory.
A
I
think
we'll
see
that
feature
come
into
processor
shear
again,
so
that's
that
will
come
back
and
then
we
had
this
macro
tasking
and
microtasking,
so
we're
starting
to
use
multiple
processors.
Now
I
kind
of
doubt
this
was
really
used
too
much
by
users,
because
it
didn't
really
improve
the
throughput
of
the
machine.
These
multi-processor
machines,
it
was
just
really
would
just
run
more
more
jobs
or
more
capacity
as
bill
put
it
earlier.
A
But
when
we
had
these
things
dedicated
in
the
lab,
we
would
go
and
try
to
do
this
stuff,
and
you
know
I
remember
the
thrill
of
getting
something
to
run
it
at
a
gigaflop
on
this.
If
you
used
all
four
processors-
and
you
were
doing
something
that
involved
a
lot
of
matrix,
multiply,
it's
funny
how
things
really
haven't
changed
all
these
years.
A
So
then
came
the
c90
and
this
was
1992
and
we
put
a
ymp8
in
and
you
know,
c90
is
kind
of
a
terrible
name.
It
was
the
cray
for
the
90s,
I
guess,
but
it
was
called
the
ympc
90.
A
Originally
that
was
the
code
name,
and
then
we
were,
I
think
too
lazy
to
change
the
name,
but
this
was
the
first
one
that
didn't
go
to
a
three-letter
agency
that
you
know-
and
I
think
this
is
probably
the
fifth
one
that
went
out
that
came
to
to
nurse
and-
and
I
think
this
was
the
last
big
flagship-
vector
system
that
came
here
before
the
transition
to
massively
parallel
computing.
Now
this
machine,
I
I
looked
up
the
top
500
1993.
It
was
the
first
year
and
on
that
list
it
was
number
19.
A
and
I'm
saying
here
it
could
have
been
number
six.
So
I'll
tell
you
the
story
behind
that
you,
you
know
you.
You
run
linpack
on
this
thing,
it's
a
lot
of
matrix
multiply
and
in
the
bench
lock
mark
lab
at
cray,
we'd
figure
out
how
to
do
this.
Goofy
thing
called
strawson's
matrix
multiply
and
we
coded
up
a
version
of
linpack
on
this
machine.
It
actually
ran
at
21,
21
gigaflops.
A
A
Probably
sx4s
is
what
they
were
so
another
funny
story
about
this
machine
was
that
it
actually
had
errors
in
the
floating
point
units
and
the
first
machines.
First,
four
machines
that
went
out
had
a
bug
in
the
floating
point
unit.
You
know
how
terrifying
that
is.
A
Is
it
for
a
vendor
when
you
put
out
a
hundred
million
dollars
worth
of
stuff
and
there's
an
error
in
it,
but
lucky
luckily,
for
us
I
mean
there's
a
guy
in
my
group
named
dennis
kuba,
who
discovered
it
doing
something
with
an
fft
and
he
got
on
this
machine.
There's
a
bug
in
this
and
sure
enough
is
there
a
way
to
is
there
software
work
around?
No,
but
fortunately,
at
the
time
we
had
an
integrated
circuit
fab.
A
We
made
our
own
ships
for
this
thing,
believe
it
or
not,
and
that
is
this
building
here
and
I'll.
Tell
you
one
thing:
I
never
understood:
that's
a
railroad
track
there
why'd
you
put
a
fab
next
to
a
railroad,
I
mean
there's.
Even
today,
our
development
building
is
up
at
the
top.
A
train
goes
by
the
building
shakes.
You
know,
I
can't
can't
imagine
that
was
a
good
idea,
but
literally
we
got
that
chip
respond
and
back
in
a
machine
in
one
week.
You
know
today
this
takes
months.
A
So
meanwhile,
in
the
marketplace
this
new
style
of
computing
was
was
coming
out
and
emerging,
and
I
mean
we
kind
of
think
there
were
a
lot
of
snake
oil
salesmen
out
here
at
this
point
in
time.
You
know
it's,
they
were
kind
of.
A
Oh,
this
is
great
remember,
and
there
were
all
these
phony
claims
about
how
great
these
machines
were
and
massively
parallel
machines
and
and
so
on-
and
this
is
about
the
era
when
david
bailey
wrote
this
funny
paper
about
the
12
ways
to
fool
the
masses
when
you're
showing
results
on
massively
parallel
super
computers.
I
mean
this
was
always
a
really
funny
paper.
I
really
enjoyed
it
and
I
think
you
could
write
it
today
now,
except
now,
it's
gpu
computing
or
something
else
I
mean
it's.
A
There
was
a
lot
of
buzz
about
massively
parallel
computing
and
you
know
eugene
brooks
attack
of
the
killer
micros,
all
that
was
going
on.
We
decided
yeah.
There
was
something
to
this
and
we
better
get
on
the
ball
here.
So
steve
nelson
kicked
off
this
t3d
project
inside
cray
and
one
of
the
test
beds
came
here,
so
this
grew
out
of
this
thing
that
steve
nelson
put
together
called
the
mpp
advisory
group.
A
This
is
our
way
to
try
to
get
intimate
with
customers.
What
do
you
want?
Let's
try
to
build
this
thing
and
believe
it
or
not
from
the
time
that
project
started,
we
delivered
a
machine
in
18
months
later
it
is
handy
to
have
a
fab
right
next
door,
so
this
was
all
kind
of
made
out
of
ymp
like
technology
and
we
could
build
the
chips
and
and
re-spin
them
right
next
door.
A
So
this
was
an
old
150,
megahertz
ev4
and
it
had
a
ymp
front
end,
and
so
I
claimed
this
was
the
first
es
login.
You
know
because
the
ymp
had
the
mature
unicos
operating
system.
This
had
some
goofy
micro
kernel
and
I
can't
remember
if
it
was
chorus
or
mock.
I
forget
which
one
came
first,
but
we
certainly
didn't
want
to
run
the
users
compiles
on
all
that.
A
So
some
of
the
cool
stuff
in
this
machine
alice
mentioned
writing
a
book
about
this.
I
think
I
might
have
contributed
to
some
did.
I
did
I
help
you
with
that.
Okay,
there
was
this
one-sided
get
put
thing
when
this
thing
was
brand
new.
There's
no
message
passing
there
was
this
pvm
thing
from
al
geist
and
there
was
nothing
but
bob
numeric
wrote
this
get
in
a
put.
You
know
you
just
kind
of
and
and
that's
all
we
had
when
we
started,
but
it
was
really
low
latency.
A
I
think
it
was
one
or
two
microseconds.
So
if
you
think
about
it,
that's
it's
no
better.
Today
than
than
it
was
back
then
so
this
was
pretty
cool,
3d
taurus,
it
was
pretty
high
density.
You
could
get
up
to
512
processing
elements
in
one
of
these
cabinets
that
use
the
same
kind
of
cooling
as
the
ymp
there.
You
know
not
all
our
projects
work
out.
A
Well,
we
had
this
parallel
programming,
language
called
craft
and
cray
research,
adaptive,
fortran,
translator
or
something
like
that
and
it
it
was
not
great
and
go
back
to
the
12
ways
to
fool
the
masses.
I
remember
going
to
a
conference
in
edinburgh.
They
had
one
of
these.
Someone
showed
this
great
graph
scaling
all
the
way
to
512
I'm
using
craft,
but
craft.
Didn't
let
the
data
cache
work.
You
had
to
turn
the
cache
off
for
this
thing
to
work,
so
you
asked
him
it
was
a
great
scaling
curve.
But
what?
A
A
You
know,
I
think
ssp
is
a
great
idea
and
it's
a
great
thing
that
nursk
has
done,
but
it
was
the
closest
thing
of
the
time
and
there's
another
young
good-looking
guy
named
horse
simon,
who
is
at
nasa
ames
and
producing
this
thing
called
the
nas
parallel
benchmarks,
and
it
was
really
this
effort
to
kind
of
have
an
honest-to-goodness
comparison
between
different
architectures
and
machines,
running
these
algorithms
that
were
important
to
to
nasa
ames
and
we
implemented
these
things
inside
cray.
In
fact,
on
the
t3d
we
implemented
all
eight
benchmarks.
A
Using
these
two
calls
call
get
call
barrier,
and
you
can
do
them
all
with
this,
because
there
was
no
other
software
that
worked
and
actually
turned
out.
This
is
pretty
easy,
and
this
bob
newman,
as
I
mentioned,
came
up
with
this
idea.
It
went
on
to
become
the
schmem
programming
model
which
is
still
around
today,
and
he
followed
up
this
whole
idea
with
something
called
f
minus
minus,
which
eventually
became
co-ray
fortran.
A
So
I
actually
dug
up
this
paper
from
1993
that
horst
had
done.
We
must
have
drove
horse
nuts
because
we
would
send
him
new
results
like
every
week.
Every
time
we
made
something
go
a
little
bit
faster.
We
had
a
new
compiler
we'd,
go
run
them
again,
charles
grassley
and
myself,
but
you
can
the
way
he
did.
The
measurements
was
how
many
ymp
units
of
performance
you
know
sort
of,
like
that's
the
gold
standard,
a
ymp
single
processor,
and
you
could
see
on
an
8
processor
ymp.
A
We
could
get
almost
a
speed
up
of
seven
that
isn't
going
to
work
in
an
mpp
world.
Is
it
and
down
here
on
the
t3d?
You
could
see
that
with
with
128
we
could,
we
could
surpass
a
ymp8
right
and
there
weren't
a
lot
of
mpps.
That
could
do
that.
So.
A
So
there
we
are
at
64
processors,
so
there
was
128
processor
one
here.
I
think
that
was
was
here
early
and
there
was
another
fellow
I
used
to.
I
used
to
look
at
his
papers,
this
guy
named
harvey
wasserman,
and
I
tell
you
this
paper
harvey
that
you
wrote
about
the
rs
6000.
This
got
our
attention
in
cray
really
did
we
looked
at
this
and
this
leadership
we
had
when
you
think
back
to
the
old
craze.
You
just
compile
a
code
you're
faster
than
anyone
in
the
world.
A
All
of
a
sudden.
This
was
getting
eroded.
Now
it
was
you
just
compile
the
code,
and
if
it's
got
long
vectors,
like
pueblo
or
newt
or
whatever
those
those
codes
were,
then
you
were
good,
but
all
these
other
ones.
These
microprocessors
were
starting
to
look
pretty
good.
In
particular,
this
590
workstation
really
caught
our
attention.
It
shook
us
up
in
cray.
We
said
we
got
to
do
this
new
direction
now
part
of
it
was.
A
A
They
went
all,
I
think,
there's
j-90
still
running,
there's
probably
one
at
nurse
somewhere
still
running
dmf
back
in
a
closet,
but
basically
it
was
a
way
to
leverage
all
the
software
on
the
expensive,
vector
machines
and
kind
of
put
it
in
a
cmos
package
instead
of
an
ecl
package,
and
these
were
good
machines,
and
you
had
a
few
of
these.
You
had
jane,
I
and
I
think
part
of
this
another
thing
I
saw
as
I
looked
through
the
history.
A
You
have
a
pretty
good
track
record
of
keeping
continuous
service
to
the
user,
so
when
it
was
time
to
move
the
c90
to
berkeley
you
kind
of
put
users
on
this
for
a
while,
while
you
were
doing
the
move-
and
this
is
probably
the
last
systems
at
nurse
that
used
the
cray
floating
point
format-
remember
that
that
was
a
little
different
than
the
ieee
it
had
had
a
bigger
mantissa
or
no
bigger
exponent
and
a
smaller
mantissa.
A
A
So
this
is
probably
one
of
my
my
favorite
machines
to
ever
work
on
the
the
t3e
and
the
you
know.
You
guys
got
the
128
processor
system
in
1996.
You
know-
and
you
know,
bill's
talked
about
kind
of
making
this
a
production
supercomputer
and
I
really
couldn't
agree
more
and
it
really
wasn't
a
production
supercomputer
when
we
were
starting
out.
So
I
got
a
couple
things
I
thought
I'd
share
with
you.
We'll
start
with
the
this
bullet
called
the
r5
debacle.
A
A
So
we're
you
know
the
router
four
chip,
the
fourth
spin.
It's
got
some
problems,
we
know
exactly
what
the
problems
are.
The
logic
designers
figure
this
out.
They
say,
oh
here
it
is,
they
fix
the
verilog
or
whatever
send
it
back
to
lsi
and
two
or
three
weeks
later.
You
know
it's
not
overnight.
Now
it's
two
or
three
weeks
chip
comes
back,
they
put
it
in
blades,
they
get
it
into
the
system.
They
boot
it
all
up
kind
of
looking
at
this
is
exactly
the
way
it
was
before
and
say
this
is
impossible.
A
We
changed
the
logic
we
did
this,
they
did
that
the
other
thing
and
they
could
not
figure
it
out
until
someone
got
the
idea
of.
Let's
look
at
that
tape.
We
sent
them
and
they
and
we
realized,
we
just
respun
the
r4
chip.
So
I
think
it's
been
long
enough
where
we
can
let
let
people
know
that
we
did.
That
turned
out
not
to
be
the
long
pole
in
the
tent
of
the
schedule.
A
Then.
The
other
story
here
you
can
see
is
about
two
bills
and
four
steves.
Well,
the
bills,
as
you
might
guess,
are
bill
mccurdy
and
bill
cramer,
and
the
steves
are
all
cray
people,
and
this
had
to
do
with
a
factory
test.
Demo
in
in
chippewa
falls
in
august
of
1996.,
so
our
sales
guy
was
al
geller.
At
the
time
the
two
bills
were
coming
to
do.
Factory
lynn
was
probably
there.
I
don't
we're
coming
to
do
a
factory
trial
and
they
wanted
to
run
a
bunch
of
applications
on
128
processors.
A
Well,
we
we'd
gone
from
32
to
a
64.
everything
worked
and
then
we
went
up
to
128
and
nothing
worked.
Absolutely
nothing.
None
of
the
nurse
applications
even
started,
and
this
was
going
on
for
almost
a
week
and
a
half
two
weeks
and
no
one
could
figure
it
out.
We
didn't
have
debuggers
on
this
thing
I
mean
it
was
people
were
trying
to
figure
this
out
and
I
think
you
guys
landed
in
minneapolis.
A
You
were
probably
in
the
car
on
the
way
over
and
steve
piatz
realized,
there's
this
constant
that
times
out
and
as
we
double
the
size
of
the
machine
with
the
the
time
to
launch
the
job
just
exceeded
the
constant
he
figured
it
out
with
print
statements.
All
he
did
was
change
the
constant
and,
and
everything
just
worked.
So
you
guys
showed
up-
and
you
know
you
didn't
know
that
we
were.
We
were
terrified.
We
were
about
to
be
raked
over
the
coals.
Everything
went
perfectly
and
you
actually
thought
that
we
knew
what
we
were
doing.
A
From
a
benchmarker's
perspective,
I
think
some
of
the
neat
stuff
in
here
were
e-registers,
which
was
a
steve
scott
idea,
stream
buffers,
which
was
an
idea
from
rick
kessler.
Instead
of
a
third
level
cache,
we
had
these
smart
buffers
that
would
read
ahead
and
there
were
six
of
them.
So
if
you
had
six
right-hand
sides,
this
thing
worked
really
well
the
whole
os
the
whole
unicos
mk
architecture
unicos
ran
on
the
on
the
login
nodes,
and
then
this
micro
kernel
ran
everywhere
else.
A
You
know
that's
sort
of
the
same
idea
we're
doing
today
on
on
edison
colray,
fortran,
checkpoint,
restart,
pvm,
eventually
mpi,
and
we
got
quite
a
bit
of
help
from
the
edinboro
folks
to
make
the
the
mpi
good
use
the
message
queues
and
do
all
that.
A
But
this
is
the
big
transition,
of
course,
to
go
from
one
memory
to
lots
of
memories,
and
I
think
you
did
a
good
job-
and
I
remember
this
guy
too,
and
I'm
glad
that
he
showed
up,
because
this
was
our
proudest
achievement
on
this
machine
was
to
run
the
first
code
at
a
teraflop,
and
I
know
there
were
guys
in
here
from
pittsburgh
and
other
places.
But
you
look
exactly
the
same.
I
mean
it's.
A
So
this
was
all
going
swimmingly.
Well
at
cray
we
were
pretty
excited.
We
thought
we
were
going
to
make
this
transition,
of
course,
and
then
and
then
this
happened
and
we
had
these
two
machines
now
sgi
bought
us,
and
that
was
okay,
but
we
had
these
two
machines.
You
know
this
one
was
the
distributed
memory
this
one
they
wanted
to
push
the
shared
memory.
A
The
origin
2000,
as
you
can
imagine
you
know,
whoever
gets
bought,
is
the
one
that
has
to
kind
of
follow
the
orders
and
over
time
it
was
kind
of
sad
that
one
kind
of
dropped
out
of
the
picture
and
all
the
focus
went
into
this
line,
and-
and
I
thought
that
was
kind
of
a
shame,
because
I
thought
making
the
leap
to
distributed
computing.
This
is
almost
like
going
backwards
and
and
but
that
was
what
what
we
ended
up
doing.
A
So
there
was
a
t3f
development
project
that
was
cancelled
and,
and
that
was
a
bit
upsetting
and
we
did
our
best
with
this
to
try
to
sell
as
many
as
possible.
This
is
what
ended
up
going
into
nasa
ames
big
systems
built
out
of
this,
but
I
don't
think
it
ever
quite
worked
out
as
well
and
of
course
you
know
you
didn't
buy
a
cray
the
next
time.
This
wasn't
a
cray,
but
you
you
turned
out,
you
bought
it
from
a
cray
guy.
A
You
just
he
just
didn't
know
he
was
a
cray
guy
at
the
time.
So,
honestly,
when
I
talked
to
pete,
he
says
he
says
boy.
You
know
I
can
remember
going
there
when
he
was
trying
to
sell,
probably
sp1,
and
he
said
it's
like
a
cray
museum
there,
you
just
just
the
beautiful
machines
you
know
he's
coming
in
with
these
black
boxes
so
kind
of
in
this
transition.
A
I
kind
of
left
the
benchmarking
group
and
I
went
into
product
management,
and
so
I
started
to
help
try
to
sell
these
machines
and-
and
I
got
a
little
bit
excited
with
the
red
storm
thing,
and
I
can
actually
took
a
brief
time
away
from
craig
and
came
back
and
I
was
pretty
excited
when
we
got
the
call
from
is
either
from
lynn
rippy
or
from
from
bill
that
hey.
A
We
want,
you
know,
come
in
and
talk
to
us,
and
there
was
some
room
here
in
this
building
where
we
had
to
spend
weeks,
and
it
reminded
me
of
being
inside
a
submarine.
It
was
this
gray
room
and-
and
there
were
pipes
coming
through
and
they
didn't
let
us
out
until
you
know.
So
it
was
the
first
negotiation
for
this,
but
there
were
a
number
of
things
that
we
had
to
do
to
the
machine
to
make
it
usable
at
nurse.
A
You
know
one
of
them
had
to
do
with
connecting
into
this
nurse
global
file
system.
So
you
guys
kind
of
got
us
developing
this
dvs
technology
so
that
we
could.
We
could
mount
gpfs
on
this
system
and
this
was
a
fragile
system.
There
were
issues
with
it.
You
know
you'd
like
to
design
a
system
without
a
single
point
of
failure,
and
this
thing
had
thousands
of
single
points
of
failure,
because
if
you
lost
a
single
lane
on
the
sea
star,
the
thing
would
come
down,
so
it's
very
sensitive
and
so
on.
A
A
So
we
did
an
upgrade
in
place
here
on
this
machine
from
a
dual
core
to
quad
core
and
dan
unger
came
up
with
this
method
of
how
we
could
do
it
and
again,
back
to
this
we
tried
to
keep
service
open
the
whole
time
and,
and
so
we
would
take
down
every
other
column
or
something
like
that
and
and
keep
the
remaining
stuff
away.
You
know
you
guys
had
orchestrated
a
pretty
good
plan
to
do
this
and
I
think
that
was
a
worthwhile
upgrade.
A
The
other
thing
I
want
to
mention
is
bill
made
the
decision
not
to
go
to
catamount
on
this
machine.
Catamount
was
the
what
we
had
the
mature
operating
system.
We
were
transitioning
to
linux,
but
we
weren't
ready
to
do
it
on
102
cabinets,
but
you
guys
pushed
us
to
get
it
going
and-
and
I
think
that
piece
of
it
worked
out-
okay
in
the
end.
A
Yes,
so
it's
and
of
course
it
really
wasn't
102
cabinets.
Everybody
knows
that
it
was
really
108
because
of
that
darn
pole
in
the
middle
of
the
computer
room,
and
we
had
to
put
in
that
column
of
empty
cabinets
and
put
them
up
on
those
earthquake
deals
around
and
so
on.
So
this
ran
now
franklin
lives
on.
We
took
these
cabinets
back
and
they've
been
remanufactured
and
they've
gone
out
as
xc6
nanos
they've
gone
out
as
eureka
systems,
so
franklin
was
an.
It
was
an
organ
donor
at
the
end
of
its
life.
A
Did
I
go
into
a
landfill
now?
I
know
you're
thinking.
How
do
you
come
up
with
these
great
names?
You
know
now
that
you're
in
product
management
jeff,
you
know
what
we're
going
to
call
that
thing.
When
we
productize
redstorm,
we
were
going
to
call
it
cray,
rs,
believe
it
or
not,
and
then
we
found
out
from
the
australians
that
rs
in
australia
means
the
same
thing
as
bs
does
here,
except
with
a
rat
instead
of
a
bull,
so
we
thought.
Maybe
this
is
not
such
a
good
name.
A
So
next
door
to
me
was
william
white,
who
was
the
product
manager
for
the
t3e,
and
he
had
these
plastic
injection
mold
letters
on
his
door
when
we
were
trying
to
do
this
alpha
cluster
and
it
said
cray
t3x.
I
said
those
are
nice
letters,
let's
just
check
there,
you
go
that'll
do,
and
that
was
how
that
came
about.
A
So
then,
you
know
things
started
to
get
better
right.
We
did
the
gemini
interconnect
and
gemini
was
just
way
way
better
than
sea
star,
and
I
was
par
among
the
people
that
made
the
decision
to
put
the
gemini
chip
into
the
taurus
infrastructure
instead
of
the
hypercube
infrastructure,
which
is
where
it
was
supposed
to
go.
So
that
has
the
good
and
the
bad.
A
I
think
if
it
was
hypercube,
we
probably
couldn't
have
built
blue
waters,
and
I
don't
think
we
could
have
done
the
upgrades
in
place
that
we
did
at
at
oak
ridge,
but
the
the
the
downside
is
when
you
start
to
make
big
tauruses,
you
have
to
start
to
care
about
placement
and
everything
as
you've
sort
of
discovered,
but
now
all
of
a
sudden
boy
we're
getting
24
core
nodes.
A
We
get
this
worm
swap
thing
working,
we
get
better
mpi
latency
and
this
is
just
way
less
fragile
than
the
than
the
xt4
was
and
we
have
cool
murals.
Now
that
was
kind
of
neat.
You
know
where
that
came
from
was
there
was
a
there
was
a
customer
down
in
urduc.
They
wanted
us
to
spray
camouflage
on
the
machine,
so
we
you
know
chippewa
falls,
there's
guys
that
spray
paint
camo
all
the
time.
A
We
took
the
doors
down,
sprayed
them
and
said,
and
then
a
guy
manufacturing
said
we
could
do
inserts
and
then
they
could
put
whatever
they
want
on
there,
and
we
always
like
this
idea,
because
it's
sort
of
like
saying
this
is
one
machine:
it's
not
a
bunch
of
individual
nodes.
This
is
not
a
cluster.
This
is
one
mpp
supercomputer
and
that's
kind
of
what
we're
you
know.
A
I'm
glad
you
guys
do
a
great
job
with
with
the
pictures
you
put
on
here,
one
more
naming
thing:
how
do
you
know
they
come
up
with
these
goofy
names?
Ecoflex
eco,
because
of
course
you
got
to
be
green.
You
know,
even
though
it's
got
a
bunch
of
refrigerant
and
stuff,
probably
didn't
so
good
phase
liquid
exchange,
it's
pretty
good.
We
just
made
that
up
and
this
fella
in
the
storage
group
tom
sherman.
He
says
that
is
the
dumbest
name.
A
A
It
was
actually
a
loud
machine,
you
know-
and
I
see
in
the
health
and
safety
thing
in
the
bathroom
that
you
know
where
earplugs
are
on
this.
It's
not
as
bad.
I
tell
you
with
a
crate
worth
100th
anniversary
of
mustard
gas.
You
know
and
the
cray
2
you
take
a
burn
in
there.
It
makes
mustard
gas,
so
that's
a
little
more
serious
and
then
finally
edison.
A
That's
what
we're
excited
to
have
this
machine
here
and
it
is
the
first
xc30
that's
gone
out
to
a
customer
site
and
again
there's
an
interim
plan
right.
We
came
out
with
a
small
machine.
I
think
you
guys
did
the
right
stuff.
You
got
the
storage
in
you
got
everything
tested
and
then
we
grew
the
through
the
compute
partition.
A
So
it
was
a
pleasure
to
work
with
the
procurement
team
on
this
one,
even
though
it
was
different
from
the
the
you
guys
have
changed
every
time.
There's
new
people
in
these
in
these
procurement
things,
except
for
lynn,
she's,
always
there.
A
So
I
mean
this
was
an
x86
machine
again,
but
it
was
different.
Gem
and
I
went
to
aries
amd
went
to
intel.
Copper
went
to
optics
vertical,
went
to
horizontal,
so
we
used
to
go
vertical
cooling
now
we're
doing
horizontal,
cooling
and
loud
went
to
quiet,
thankfully,
and
we
did
our
best
at
cray
to
make
this
a
smooth
transition
for
the
users
and
administrators,
and
we
tried
to
do
that
by
using
the
same
os
by
using
the
same
compilers.
I
know
we
didn't
have
pgi
on
there.
We
do
have
it
now.
A
If
you
do
need
it,
but
we
did
our.
We
did
our
best
to
make
it
a
smooth
transition,
and
I
think
this
has
been
a
good
good
transition
from
what
I
see
you
know
there's
the
machine
when,
when
it
was
it's
a
used
machine,
did
you
know
that
it
was
the
darpa
demo
machine,
and
this
was
september
of
of
12,
and
this
demo
went
off
without
a
hitch.
A
I
mean
everything
with
the
ares
was
good
and
we
we
took
the
skins
off
and
changed
the
doors,
and
then
that
became
your
first
four
cabinets
and
then
we
delivered
that
and
announced
it
at
super
computing
and
we
have
a
bunch
of
orders.
I
don't
know
how
many
petaflop
scale
systems
we
have,
but
we
just
had
a
record
year
with
a
25
revenue
growth.
It's
all
this
stuff
and
in
super
computing.
A
So
I
think
that's
pretty
neat
stuff.
So
it's
been
a
great
40
years.
Our
slogan,
you
know
used
to
be
world's
fastest
computers.
We
made
it
a
little
more,
I
don't
know
complicated.
We
build
computational
tools
that
help
change
the
world
and
you
notice
help
is
really
you
know.
At
the
end
of
the
day,
we
just
make
a
tool.
It's
just
a
thing.
You
know,
and
it
doesn't
really
do
anything
until
it
gets
in
the
hand
of
a
center
there's
user
services.
A
Folks,
there's
a
bunch
of
users
doing
really
smart
stuff
and-
and
we
know
that
it's
our
customers,
who
do
the
real
heavy
lifting
and
actually
do
that
change
the
world
stuff.
But
it's
just
a
pleasure
to
have
been
part
of
it
for
for
40
years.
So
so
thanks
for
a
great
40
years
I
and
on
behalf
of
cray,
I
hope
we
have
a
good
relationship
for
the
for
the
next
four.
So
thanks.