<s>
Approaches	O
to	O
supercomputer	B-Operating_System
architecture	I-Operating_System
have	O
taken	O
dramatic	O
turns	O
since	O
the	O
earliest	O
systems	O
were	O
introduced	O
in	O
the	O
1960s	O
.	O
</s>
<s>
Early	O
supercomputer	B-Operating_System
architectures	I-Operating_System
pioneered	O
by	O
Seymour	O
Cray	O
relied	O
on	O
compact	O
innovative	O
designs	O
and	O
local	O
parallelism	B-Operating_System
to	O
achieve	O
superior	O
computational	O
peak	O
performance	O
.	O
</s>
<s>
However	O
,	O
in	O
time	O
the	O
demand	O
for	O
increased	O
computational	O
power	O
ushered	O
in	O
the	O
age	O
of	O
massively	B-Operating_System
parallel	I-Operating_System
systems	O
.	O
</s>
<s>
While	O
the	O
supercomputers	B-Architecture
of	O
the	O
1970s	O
used	O
only	O
a	O
few	O
processors	B-General_Concept
,	O
in	O
the	O
1990s	O
,	O
machines	O
with	O
thousands	O
of	O
processors	B-General_Concept
began	O
to	O
appear	O
and	O
by	O
the	O
end	O
of	O
the	O
20th	O
century	O
,	O
massively	B-Operating_System
parallel	I-Operating_System
supercomputers	B-Architecture
with	O
tens	O
of	O
thousands	O
of	O
commercial	O
off-the-shelf	O
processors	B-General_Concept
were	O
the	O
norm	O
.	O
</s>
<s>
Supercomputers	B-Architecture
of	O
the	O
21st	O
century	O
can	O
use	O
over	O
100,000	O
processors	B-General_Concept
(	O
some	O
being	O
graphic	B-Architecture
units	I-Architecture
)	O
connected	O
by	O
fast	O
connections	O
.	O
</s>
<s>
Throughout	O
the	O
decades	O
,	O
the	O
management	O
of	O
heat	O
density	O
has	O
remained	O
a	O
key	O
issue	O
for	O
most	O
centralized	O
supercomputers	B-Architecture
.	O
</s>
<s>
Systems	O
with	O
a	O
massive	O
number	O
of	O
processors	B-General_Concept
generally	O
take	O
one	O
of	O
two	O
paths	O
:	O
in	O
one	O
approach	O
,	O
e.g.	O
,	O
in	O
grid	B-Architecture
computing	I-Architecture
the	O
processing	O
power	O
of	O
a	O
large	O
number	O
of	O
computers	O
in	O
distributed	O
,	O
diverse	O
administrative	O
domains	O
,	O
is	O
opportunistically	O
used	O
whenever	O
a	O
computer	O
is	O
available	O
.	O
</s>
<s>
In	O
another	O
approach	O
,	O
a	O
large	O
number	O
of	O
processors	B-General_Concept
are	O
used	O
in	O
close	O
proximity	O
to	O
each	O
other	O
,	O
e.g.	O
,	O
in	O
a	O
computer	B-Architecture
cluster	I-Architecture
.	O
</s>
<s>
In	O
such	O
a	O
centralized	O
massively	B-Operating_System
parallel	I-Operating_System
system	O
the	O
speed	O
and	O
flexibility	O
of	O
the	O
interconnect	O
becomes	O
very	O
important	O
,	O
and	O
modern	O
supercomputers	B-Architecture
have	O
used	O
various	O
approaches	O
ranging	O
from	O
enhanced	O
Infiniband	B-Architecture
systems	O
to	O
three-dimensional	O
torus	B-Operating_System
interconnects	I-Operating_System
.	O
</s>
<s>
Since	O
the	O
late	O
1960s	O
the	O
growth	O
in	O
the	O
power	O
and	O
proliferation	O
of	O
supercomputers	B-Architecture
has	O
been	O
dramatic	O
,	O
and	O
the	O
underlying	O
architectural	O
directions	O
of	O
these	O
systems	O
have	O
taken	O
significant	O
turns	O
.	O
</s>
<s>
While	O
the	O
early	O
supercomputers	B-Architecture
relied	O
on	O
a	O
small	O
number	O
of	O
closely	O
connected	O
processors	B-General_Concept
that	O
accessed	O
shared	B-Operating_System
memory	I-Operating_System
,	O
the	O
supercomputers	B-Architecture
of	O
the	O
21st	O
century	O
use	O
over	O
100,000	O
processors	B-General_Concept
connected	O
by	O
fast	O
networks	O
.	O
</s>
<s>
Throughout	O
the	O
decades	O
,	O
the	O
management	O
of	O
heat	O
density	O
has	O
remained	O
a	O
key	O
issue	O
for	O
most	O
centralized	O
supercomputers	B-Architecture
.	O
</s>
<s>
Seymour	O
Cray	O
's	O
"	O
get	O
the	O
heat	O
out	O
"	O
motto	O
was	O
central	O
to	O
his	O
design	O
philosophy	O
and	O
has	O
continued	O
to	O
be	O
a	O
key	O
issue	O
in	O
supercomputer	B-Operating_System
architectures	I-Operating_System
,	O
e.g.	O
,	O
in	O
large-scale	O
experiments	O
such	O
as	O
Blue	B-Device
Waters	I-Device
.	O
</s>
<s>
There	O
have	O
been	O
diverse	O
approaches	O
to	O
heat	O
management	O
,	O
e.g.	O
,	O
the	O
Cray	B-Device
2	I-Device
pumped	O
Fluorinert	O
through	O
the	O
system	O
,	O
while	O
System	B-Device
X	I-Device
used	O
a	O
hybrid	O
liquid-air	O
cooling	O
system	O
and	O
the	O
Blue	B-Operating_System
Gene/P	I-Operating_System
is	O
air-cooled	O
with	O
normal	O
air	O
conditioning	O
temperatures	O
.	O
</s>
<s>
The	O
heat	O
from	O
the	O
Aquasar	O
supercomputer	B-Architecture
is	O
used	O
to	O
warm	O
a	O
university	O
campus	O
.	O
</s>
<s>
The	O
heat	O
density	O
generated	O
by	O
a	O
supercomputer	B-Architecture
has	O
a	O
direct	O
dependence	O
on	O
the	O
processor	O
type	O
used	O
in	O
the	O
system	O
,	O
with	O
more	O
powerful	O
processors	B-General_Concept
typically	O
generating	O
more	O
heat	O
,	O
given	O
similar	O
underlying	O
semiconductor	O
technologies	O
.	O
</s>
<s>
While	O
early	O
supercomputers	B-Architecture
used	O
a	O
few	O
fast	O
,	O
closely	O
packed	O
processors	B-General_Concept
that	O
took	O
advantage	O
of	O
local	O
parallelism	B-Operating_System
(	O
e.g.	O
,	O
pipelining	B-General_Concept
and	O
vector	B-Operating_System
processing	I-Operating_System
)	O
,	O
in	O
time	O
the	O
number	O
of	O
processors	B-General_Concept
grew	O
,	O
and	O
computing	O
nodes	O
could	O
be	O
placed	O
further	O
away	O
,	O
e.g.	O
,	O
in	O
a	O
computer	B-Architecture
cluster	I-Architecture
,	O
or	O
could	O
be	O
geographically	O
dispersed	O
in	O
grid	B-Architecture
computing	I-Architecture
.	O
</s>
<s>
As	O
the	O
number	O
of	O
processors	B-General_Concept
in	O
a	O
supercomputer	B-Architecture
grows	O
,	O
"	O
component	O
failure	O
rate	O
"	O
begins	O
to	O
become	O
a	O
serious	O
issue	O
.	O
</s>
<s>
If	O
a	O
supercomputer	B-Architecture
uses	O
thousands	O
of	O
nodes	O
,	O
each	O
of	O
which	O
may	O
fail	O
once	O
per	O
year	O
on	O
the	O
average	O
,	O
then	O
the	O
system	O
will	O
experience	O
several	O
node	O
failures	O
each	O
day	O
.	O
</s>
<s>
As	O
the	O
price/performance	O
of	O
general	B-Architecture
purpose	I-Architecture
graphic	I-Architecture
processors	I-Architecture
(	O
GPGPUs	B-Architecture
)	O
has	O
improved	O
,	O
a	O
number	O
of	O
petaflop	O
supercomputers	B-Architecture
such	O
as	O
Tianhe-I	B-Device
and	O
Nebulae	B-Device
have	O
started	O
to	O
rely	O
on	O
them	O
.	O
</s>
<s>
However	O
,	O
other	O
systems	O
such	O
as	O
the	O
K	B-Device
computer	I-Device
continue	O
to	O
use	O
conventional	O
processors	B-General_Concept
such	O
as	O
SPARC-based	O
designs	O
and	O
the	O
overall	O
applicability	O
of	O
GPGPUs	B-Architecture
in	O
general	O
purpose	O
high	B-Architecture
performance	I-Architecture
computing	I-Architecture
applications	O
has	O
been	O
the	O
subject	O
of	O
debate	O
,	O
in	O
that	O
while	O
a	O
GPGPU	B-Architecture
may	O
be	O
tuned	O
to	O
score	O
well	O
on	O
specific	O
benchmarks	O
its	O
overall	O
applicability	O
to	O
everyday	O
algorithms	O
may	O
be	O
limited	O
unless	O
significant	O
effort	O
is	O
spent	O
to	O
tune	O
the	O
application	O
towards	O
it	O
.	O
</s>
<s>
However	O
,	O
GPUs	O
are	O
gaining	O
ground	O
and	O
in	O
2012	O
the	O
Jaguar	B-Device
supercomputer	I-Device
was	O
transformed	O
into	O
Titan	B-Device
by	O
replacing	O
CPUs	O
with	O
GPUs	O
.	O
</s>
<s>
As	O
the	O
number	O
of	O
independent	O
processors	B-General_Concept
in	O
a	O
supercomputer	B-Architecture
increases	O
,	O
the	O
way	O
they	O
access	O
data	O
in	O
the	O
file	B-Application
system	I-Application
and	O
how	O
they	O
share	O
and	O
access	O
secondary	O
storage	O
resources	O
becomes	O
prominent	O
.	O
</s>
<s>
Over	O
the	O
years	O
a	O
number	O
of	O
systems	O
for	O
distributed	B-Protocol
file	I-Protocol
management	I-Protocol
were	O
developed	O
,	O
e.g.	O
,	O
the	O
IBM	B-Application
General	I-Application
Parallel	I-Application
File	I-Application
System	I-Application
,	O
BeeGFS	B-Application
,	O
the	O
Parallel	B-Application
Virtual	I-Application
File	I-Application
System	I-Application
,	O
Hadoop	B-Application
,	O
etc	O
.	O
</s>
<s>
A	O
number	O
of	O
supercomputers	B-Architecture
on	O
the	O
TOP100	B-Operating_System
list	O
such	O
as	O
the	O
Tianhe-I	B-Device
use	O
Linux	B-Application
's	O
Lustre	B-Application
file	I-Application
system	I-Application
.	O
</s>
<s>
The	O
CDC	B-Device
6600	I-Device
series	O
of	O
computers	O
were	O
very	O
early	O
attempts	O
at	O
supercomputing	B-Architecture
and	O
gained	O
their	O
advantage	O
over	O
the	O
existing	O
systems	O
by	O
relegating	O
work	O
to	O
peripheral	B-Device
devices	I-Device
,	O
freeing	O
the	O
CPU	O
(	O
Central	B-General_Concept
Processing	I-General_Concept
Unit	I-General_Concept
)	O
to	O
process	O
actual	O
data	O
.	O
</s>
<s>
With	O
the	O
Minnesota	O
FORTRAN	B-Application
compiler	O
the	O
6600	O
could	O
sustain	O
500kiloflops	O
on	O
standard	O
mathematical	O
operations	O
.	O
</s>
<s>
Other	O
early	O
supercomputers	B-Architecture
such	O
as	O
the	O
Cray	B-Device
1	I-Device
and	O
Cray	B-Device
2	I-Device
that	O
appeared	O
afterwards	O
used	O
a	O
small	O
number	O
of	O
fast	O
processors	B-General_Concept
that	O
worked	O
in	O
harmony	O
and	O
were	O
uniformly	O
connected	O
to	O
the	O
largest	O
amount	O
of	O
shared	B-Operating_System
memory	I-Operating_System
that	O
could	O
be	O
managed	O
at	O
the	O
time	O
.	O
</s>
<s>
These	O
early	O
architectures	O
introduced	O
parallel	B-Operating_System
processing	I-Operating_System
at	O
the	O
processor	O
level	O
,	O
with	O
innovations	O
such	O
as	O
vector	B-Operating_System
processing	I-Operating_System
,	O
in	O
which	O
the	O
processor	O
can	O
perform	O
several	O
operations	O
during	O
one	O
clock	O
cycle	O
,	O
rather	O
than	O
having	O
to	O
wait	O
for	O
successive	O
cycles	O
.	O
</s>
<s>
In	O
time	O
,	O
as	O
the	O
number	O
of	O
processors	B-General_Concept
increased	O
,	O
different	O
architectural	O
issues	O
emerged	O
.	O
</s>
<s>
Two	O
issues	O
that	O
need	O
to	O
be	O
addressed	O
as	O
the	O
number	O
of	O
processors	B-General_Concept
increases	O
are	O
the	O
distribution	O
of	O
memory	O
and	O
processing	O
.	O
</s>
<s>
In	O
the	O
distributed	B-Operating_System
memory	I-Operating_System
approach	O
,	O
each	O
processor	O
is	O
physically	O
packaged	O
close	O
with	O
some	O
local	O
memory	O
.	O
</s>
<s>
The	O
memory	O
associated	O
with	O
other	O
processors	B-General_Concept
is	O
then	O
"	O
further	O
away	O
"	O
based	O
on	O
bandwidth	O
and	O
latency	O
parameters	O
in	O
non-uniform	B-Operating_System
memory	I-Operating_System
access	I-Operating_System
.	O
</s>
<s>
In	O
the	O
1960s	O
pipelining	B-General_Concept
was	O
viewed	O
as	O
an	O
innovation	O
,	O
and	O
by	O
the	O
1970s	O
the	O
use	O
of	O
vector	B-Operating_System
processors	I-Operating_System
had	O
been	O
well	O
established	O
.	O
</s>
<s>
By	O
the	O
1980s	O
,	O
many	O
supercomputers	B-Architecture
used	O
parallel	O
vector	B-Operating_System
processors	I-Operating_System
.	O
</s>
<s>
The	O
relatively	O
small	O
number	O
of	O
processors	B-General_Concept
in	O
early	O
systems	O
,	O
allowed	O
them	O
to	O
easily	O
use	O
a	O
shared	B-Operating_System
memory	I-Operating_System
architecture	I-Operating_System
,	O
which	O
allows	O
processors	B-General_Concept
to	O
access	O
a	O
common	O
pool	O
of	O
memory	O
.	O
</s>
<s>
In	O
the	O
early	O
days	O
a	O
common	O
approach	O
was	O
the	O
use	O
of	O
uniform	B-Operating_System
memory	I-Operating_System
access	I-Operating_System
(	O
UMA	O
)	O
,	O
in	O
which	O
access	O
time	O
to	O
a	O
memory	B-General_Concept
location	I-General_Concept
was	O
similar	O
between	O
processors	B-General_Concept
.	O
</s>
<s>
The	O
use	O
of	O
non-uniform	B-Operating_System
memory	I-Operating_System
access	I-Operating_System
(	O
NUMA	O
)	O
allowed	O
a	O
processor	O
to	O
access	O
its	O
own	O
local	O
memory	O
faster	O
than	O
other	O
memory	B-General_Concept
locations	I-General_Concept
,	O
while	O
cache-only	B-Operating_System
memory	I-Operating_System
architectures	I-Operating_System
(	O
COMA	O
)	O
allowed	O
for	O
the	O
local	O
memory	O
of	O
each	O
processor	O
to	O
be	O
used	O
as	O
cache	O
,	O
thus	O
requiring	O
coordination	O
as	O
memory	O
values	O
changed	O
.	O
</s>
<s>
As	O
the	O
number	O
of	O
processors	B-General_Concept
increases	O
,	O
efficient	O
interprocessor	B-Operating_System
communication	I-Operating_System
and	O
synchronization	O
on	O
a	O
supercomputer	B-Architecture
becomes	O
a	O
challenge	O
.	O
</s>
<s>
For	O
instance	O
,	O
in	O
the	O
early	O
1980s	O
,	O
in	O
the	O
Cray	B-Device
X-MP	I-Device
system	O
,	O
shared	B-Operating_System
registers	I-Operating_System
were	O
used	O
.	O
</s>
<s>
In	O
this	O
approach	O
,	O
all	O
processors	B-General_Concept
had	O
access	O
to	O
shared	B-Operating_System
registers	I-Operating_System
that	O
did	O
not	O
move	O
data	O
back	O
and	O
forth	O
but	O
were	O
only	O
used	O
for	O
interprocessor	B-Operating_System
communication	I-Operating_System
and	O
synchronization	O
.	O
</s>
<s>
However	O
,	O
inherent	O
challenges	O
in	O
managing	O
a	O
large	O
amount	O
of	O
shared	B-Operating_System
memory	I-Operating_System
among	O
many	O
processors	B-General_Concept
resulted	O
in	O
a	O
move	O
to	O
more	O
distributed	O
architectures	O
.	O
</s>
<s>
During	O
the	O
1980s	O
,	O
as	O
the	O
demand	O
for	O
computing	O
power	O
increased	O
,	O
the	O
trend	O
to	O
a	O
much	O
larger	O
number	O
of	O
processors	B-General_Concept
began	O
,	O
ushering	O
in	O
the	O
age	O
of	O
massively	B-Operating_System
parallel	I-Operating_System
systems	O
,	O
with	O
distributed	B-Operating_System
memory	I-Operating_System
and	O
distributed	B-Protocol
file	I-Protocol
systems	I-Protocol
,	O
given	O
that	O
shared	B-Operating_System
memory	I-Operating_System
architectures	I-Operating_System
could	O
not	O
scale	O
to	O
a	O
large	O
number	O
of	O
processors	B-General_Concept
.	O
</s>
<s>
Hybrid	O
approaches	O
such	O
as	O
distributed	B-Operating_System
shared	I-Operating_System
memory	I-Operating_System
also	O
appeared	O
after	O
the	O
early	O
systems	O
.	O
</s>
<s>
The	O
computer	B-Architecture
clustering	I-Architecture
approach	O
connects	O
a	O
number	O
of	O
readily	O
available	O
computing	O
nodes	O
(	O
e.g.	O
</s>
<s>
personal	O
computers	O
used	O
as	O
servers	O
)	O
via	O
a	O
fast	O
,	O
private	O
local	B-General_Concept
area	I-General_Concept
network	I-General_Concept
.	O
</s>
<s>
The	O
activities	O
of	O
the	O
computing	O
nodes	O
are	O
orchestrated	O
by	O
"	O
clustering	O
middleware	B-General_Concept
"	O
,	O
a	O
software	O
layer	O
that	O
sits	O
atop	O
the	O
nodes	O
and	O
allows	O
the	O
users	O
to	O
treat	O
the	O
cluster	O
as	O
by	O
and	O
large	O
one	O
cohesive	O
computing	O
unit	O
,	O
e.g.	O
</s>
<s>
via	O
a	O
single	B-Operating_System
system	I-Operating_System
image	I-Operating_System
concept	O
.	O
</s>
<s>
Computer	B-Architecture
clustering	I-Architecture
relies	O
on	O
a	O
centralized	O
management	O
approach	O
which	O
makes	O
the	O
nodes	O
available	O
as	O
orchestrated	O
shared	B-Application
servers	I-Application
.	O
</s>
<s>
It	O
is	O
distinct	O
from	O
other	O
approaches	O
such	O
as	O
peer-to-peer	O
or	O
grid	B-Architecture
computing	I-Architecture
which	O
also	O
use	O
many	O
nodes	O
,	O
but	O
with	O
a	O
far	O
more	O
distributed	B-Architecture
nature	I-Architecture
.	O
</s>
<s>
By	O
the	O
21st	O
century	O
,	O
the	O
TOP500	B-Operating_System
organization	O
's	O
semiannual	O
list	O
of	O
the	O
500	O
fastest	B-Operating_System
supercomputers	I-Operating_System
often	O
includes	O
many	O
clusters	O
,	O
e.g.	O
</s>
<s>
the	O
world	O
's	O
fastest	O
in	O
2011	O
,	O
the	O
K	B-Device
computer	I-Device
with	O
a	O
distributed	B-Operating_System
memory	I-Operating_System
,	O
cluster	B-Architecture
architecture	I-Architecture
.	O
</s>
<s>
in	O
a	O
cluster	B-Architecture
architecture	I-Architecture
)	O
the	O
speed	O
and	O
flexibility	O
of	O
the	O
interconnect	O
becomes	O
very	O
important	O
.	O
</s>
<s>
Modern	O
supercomputers	B-Architecture
have	O
taken	O
different	O
approaches	O
to	O
address	O
this	O
issue	O
,	O
e.g.	O
</s>
<s>
Tianhe-1	B-Device
uses	O
a	O
proprietary	O
high-speed	O
network	O
based	O
on	O
the	O
Infiniband	B-Architecture
QDR	O
,	O
enhanced	O
with	O
FeiTeng-1000	B-Device
CPUs	O
.	O
</s>
<s>
On	O
the	O
other	O
hand	O
,	O
the	O
Blue	B-Operating_System
Gene/L	I-Operating_System
system	O
uses	O
a	O
three-dimensional	O
torus	B-Operating_System
interconnect	I-Operating_System
with	O
auxiliary	O
networks	O
for	O
global	O
communications	O
.	O
</s>
<s>
A	O
similar	O
torus	O
was	O
used	O
by	O
the	O
Cray	B-Device
T3E	I-Device
.	O
</s>
<s>
Massive	O
centralized	O
systems	O
at	O
times	O
use	O
special-purpose	O
processors	B-General_Concept
designed	O
for	O
a	O
specific	O
application	O
,	O
and	O
may	O
use	O
field-programmable	B-Architecture
gate	I-Architecture
arrays	I-Architecture
(	O
FPGA	B-Architecture
)	O
chips	O
to	O
gain	O
performance	O
by	O
sacrificing	O
generality	O
.	O
</s>
<s>
molecular	O
dynamics	O
and	O
Deep	B-Algorithm
Crack	I-Algorithm
,	O
for	O
breaking	O
the	O
DES	B-Algorithm
cipher	B-Application
.	O
</s>
<s>
Grid	B-Architecture
computing	I-Architecture
uses	O
a	O
large	O
number	O
of	O
computers	O
in	O
distributed	O
,	O
diverse	O
administrative	O
domains	O
.	O
</s>
<s>
An	O
example	O
is	O
BOINC	B-Operating_System
a	O
volunteer-based	B-Operating_System
,	O
opportunistic	O
grid	O
system	O
.	O
</s>
<s>
Some	O
BOINC	B-Operating_System
applications	O
have	O
reached	O
multi-petaflop	O
levels	O
by	O
using	O
close	O
to	O
half	O
a	O
million	O
computers	O
connected	O
on	O
the	O
internet	O
,	O
whenever	O
volunteer	O
resources	O
become	O
available	O
.	O
</s>
<s>
However	O
,	O
these	O
types	O
of	O
results	O
often	O
do	O
not	O
appear	O
in	O
the	O
TOP500	B-Operating_System
ratings	O
because	O
they	O
do	O
not	O
run	O
the	O
general	O
purpose	O
Linpack	B-Application
benchmark	O
.	O
</s>
<s>
Although	O
grid	B-Architecture
computing	I-Architecture
has	O
had	O
success	O
in	O
parallel	O
task	O
execution	O
,	O
demanding	O
supercomputer	B-Architecture
applications	O
such	O
as	O
weather	B-General_Concept
simulations	I-General_Concept
or	O
computational	O
fluid	O
dynamics	O
have	O
remained	O
out	O
of	O
reach	O
,	O
partly	O
due	O
to	O
the	O
barriers	O
in	O
reliable	O
sub-assignment	O
of	O
a	O
large	O
number	O
of	O
tasks	O
as	O
well	O
as	O
the	O
reliable	O
availability	O
of	O
resources	O
at	O
a	O
given	O
time	O
.	O
</s>
<s>
In	O
quasi-opportunistic	B-Operating_System
supercomputing	I-Operating_System
a	O
large	O
number	O
of	O
geographically	O
disperse	B-Architecture
computers	I-Architecture
are	O
orchestrated	O
with	O
built-in	B-General_Concept
safeguards	I-General_Concept
.	O
</s>
<s>
The	O
quasi-opportunistic	O
approach	O
goes	O
beyond	O
volunteer	B-Operating_System
computing	I-Operating_System
on	O
a	O
highly	O
distributed	B-Architecture
systems	I-Architecture
such	O
as	O
BOINC	B-Operating_System
,	O
or	O
general	O
grid	B-Architecture
computing	I-Architecture
on	O
a	O
system	O
such	O
as	O
Globus	O
by	O
allowing	O
the	O
middleware	B-General_Concept
to	O
provide	O
almost	O
seamless	O
access	O
to	O
many	O
computing	B-Architecture
clusters	I-Architecture
so	O
that	O
existing	O
programs	O
in	O
languages	O
such	O
as	O
Fortran	B-Application
or	O
C	B-Language
can	O
be	O
distributed	O
among	O
multiple	O
computing	O
resources	O
.	O
</s>
<s>
Quasi-opportunistic	B-Operating_System
supercomputing	I-Operating_System
aims	O
to	O
provide	O
a	O
higher	O
quality	O
of	O
service	O
than	O
opportunistic	B-Architecture
resource	I-Architecture
sharing	I-Architecture
.	O
</s>
<s>
The	O
quasi-opportunistic	O
approach	O
enables	O
the	O
execution	O
of	O
demanding	O
applications	O
within	O
computer	O
grids	O
by	O
establishing	O
grid-wise	O
resource	O
allocation	O
agreements	O
;	O
and	O
fault	B-General_Concept
tolerant	I-General_Concept
message	O
passing	O
to	O
abstractly	O
shield	O
against	O
the	O
failures	O
of	O
the	O
underlying	O
resources	O
,	O
thus	O
maintaining	O
some	O
opportunism	O
,	O
while	O
allowing	O
a	O
higher	O
level	O
of	O
control	O
.	O
</s>
<s>
The	O
air-cooled	O
IBM	B-Operating_System
Blue	I-Operating_System
Gene	I-Operating_System
supercomputer	B-Operating_System
architecture	I-Operating_System
trades	O
processor	O
speed	O
for	O
low	O
power	O
consumption	O
so	O
that	O
a	O
larger	O
number	O
of	O
processors	B-General_Concept
can	O
be	O
used	O
at	O
room	O
temperature	O
,	O
by	O
using	O
normal	O
air-conditioning	O
.	O
</s>
<s>
The	O
second-generation	O
Blue	B-Operating_System
Gene/P	I-Operating_System
system	O
has	O
processors	B-General_Concept
with	O
integrated	O
node-to-node	O
communication	O
logic	O
.	O
</s>
<s>
The	O
K	B-Device
computer	I-Device
is	O
a	O
water-cooled	B-General_Concept
,	O
homogeneous	O
processor	O
,	O
distributed	B-Operating_System
memory	I-Operating_System
system	O
with	O
a	O
cluster	B-Architecture
architecture	I-Architecture
.	O
</s>
<s>
It	O
uses	O
more	O
than	O
80,000	O
SPARC64	O
VIIIfx	O
processors	B-General_Concept
,	O
each	O
with	O
eight	O
cores	B-Architecture
,	O
for	O
a	O
total	O
of	O
over	O
700,000	O
coresalmost	O
twice	O
as	O
many	O
as	O
any	O
other	O
system	O
.	O
</s>
<s>
Although	O
it	O
is	O
more	O
powerful	O
than	O
the	O
next	O
five	O
systems	O
on	O
the	O
TOP500	B-Operating_System
list	O
combined	O
,	O
at	O
824.56MFLOPS/W	O
it	O
has	O
the	O
lowest	O
power	O
to	O
performance	O
ratio	O
of	O
any	O
current	O
major	O
supercomputer	B-Architecture
system	O
.	O
</s>
<s>
The	O
follow	O
up	O
system	O
for	O
the	O
K	B-Device
computer	I-Device
,	O
called	O
the	O
PRIMEHPC	O
FX10	O
uses	O
the	O
same	O
six-dimensional	O
torus	B-Operating_System
interconnect	I-Operating_System
,	O
but	O
still	O
only	O
one	O
processor	O
per	O
node	O
.	O
</s>
<s>
Unlike	O
the	O
K	B-Device
computer	I-Device
,	O
the	O
Tianhe-1A	B-Device
system	O
uses	O
a	O
hybrid	O
architecture	O
and	O
integrates	O
CPUs	O
and	O
GPUs	O
.	O
</s>
<s>
It	O
uses	O
more	O
than	O
14,000	O
Xeon	B-Device
general-purpose	O
processors	B-General_Concept
and	O
more	O
than	O
7,000	O
Nvidia	B-Device
Tesla	I-Device
general-purpose	O
graphics	O
processing	O
units	O
(	O
GPGPUs	B-Architecture
)	O
on	O
about	O
3,500	O
blades	B-Architecture
.	O
</s>
<s>
It	O
has	O
112	O
computer	O
cabinets	O
and	O
262	O
terabytes	O
of	O
distributed	B-Operating_System
memory	I-Operating_System
;	O
2	O
petabytes	O
of	O
disk	O
storage	O
is	O
implemented	O
via	O
Lustre	B-Application
clustered	O
files	O
.	O
</s>
<s>
Tianhe-1	B-Device
uses	O
a	O
proprietary	O
high-speed	O
communication	O
network	O
to	O
connect	O
the	O
processors	B-General_Concept
.	O
</s>
<s>
The	O
proprietary	O
interconnect	O
network	O
was	O
based	O
on	O
the	O
Infiniband	B-Architecture
QDR	O
,	O
enhanced	O
with	O
Chinese	O
made	O
FeiTeng-1000	B-Device
CPUs	O
.	O
</s>
<s>
In	O
the	O
case	O
of	O
the	O
interconnect	O
the	O
system	O
is	O
twice	O
as	O
fast	O
as	O
the	O
Infiniband	B-Architecture
,	O
but	O
slower	O
than	O
some	O
interconnects	O
on	O
other	O
supercomputers	B-Architecture
.	O
</s>
<s>
The	O
limits	O
of	O
specific	O
approaches	O
continue	O
to	O
be	O
tested	O
,	O
as	O
boundaries	O
are	O
reached	O
through	O
large	O
scale	O
experiments	O
,	O
e.g.	O
,	O
in	O
2011	O
IBM	O
ended	O
its	O
participation	O
in	O
the	O
Blue	B-Device
Waters	I-Device
petaflops	O
project	O
at	O
the	O
University	O
of	O
Illinois	O
.	O
</s>
<s>
The	O
Blue	B-Device
Waters	I-Device
architecture	O
was	O
based	O
on	O
the	O
IBM	O
POWER7	B-Device
processor	O
and	O
intended	O
to	O
have	O
200,000	O
cores	B-Architecture
with	O
a	O
petabyte	O
of	O
"	O
globally	B-General_Concept
addressable	I-General_Concept
memory	O
"	O
and	O
10	O
petabytes	O
of	O
disk	O
space	O
.	O
</s>
<s>
The	O
goal	O
of	O
a	O
sustained	O
petaflop	O
led	O
to	O
design	O
choices	O
that	O
optimized	O
single-core	O
performance	O
,	O
and	O
hence	O
a	O
lower	O
number	O
of	O
cores	B-Architecture
.	O
</s>
<s>
The	O
lower	O
number	O
of	O
cores	B-Architecture
was	O
then	O
expected	O
to	O
help	O
performance	O
on	O
programs	O
that	O
did	O
not	O
scale	O
well	O
to	O
a	O
large	O
number	O
of	O
processors	B-General_Concept
.	O
</s>
<s>
The	O
large	O
globally	B-General_Concept
addressable	I-General_Concept
memory	O
architecture	O
aimed	O
to	O
solve	O
memory	B-General_Concept
address	I-General_Concept
problems	O
in	O
an	O
efficient	O
manner	O
,	O
for	O
the	O
same	O
type	O
of	O
programs	O
.	O
</s>
<s>
Blue	B-Device
Waters	I-Device
had	O
been	O
expected	O
to	O
run	O
at	O
sustained	O
speeds	O
of	O
at	O
least	O
one	O
petaflop	O
,	O
and	O
relied	O
on	O
the	O
specific	O
water-cooling	O
approach	O
to	O
manage	O
heat	O
.	O
</s>
<s>
IBM	O
released	O
the	O
Power	O
775	O
computing	O
node	O
derived	O
from	O
that	O
project	O
's	O
technology	O
soon	O
thereafter	O
,	O
but	O
effectively	O
abandoned	O
the	O
Blue	B-Device
Waters	I-Device
approach	O
.	O
</s>
<s>
the	O
Cyclops64	B-Device
system	O
uses	O
a	O
"	O
supercomputer	B-Architecture
on	O
a	O
chip	O
"	O
approach	O
,	O
in	O
a	O
direction	O
away	O
from	O
the	O
use	O
of	O
massive	O
distributed	O
processors	B-General_Concept
.	O
</s>
<s>
Each	O
64-bit	O
Cyclops64	B-Device
chip	O
contains	O
80	O
processors	B-General_Concept
,	O
and	O
the	O
entire	O
system	O
uses	O
a	O
globally	B-General_Concept
addressable	I-General_Concept
memory	O
architecture	O
.	O
</s>
<s>
The	O
processors	B-General_Concept
are	O
connected	O
with	O
non-internally	O
blocking	O
crossbar	O
switch	O
and	O
communicate	O
with	O
each	O
other	O
via	O
global	O
interleaved	O
memory	O
.	O
</s>
<s>
There	O
is	O
no	O
data	B-General_Concept
cache	I-General_Concept
in	O
the	O
architecture	O
,	O
but	O
half	O
of	O
each	O
SRAM	B-Architecture
bank	O
can	O
be	O
used	O
as	O
a	O
scratchpad	O
memory	O
.	O
</s>
<s>
Although	O
this	O
type	O
of	O
architecture	O
allows	O
unstructured	O
parallelism	B-Operating_System
in	O
a	O
dynamically	O
non-contiguous	O
memory	O
system	O
,	O
it	O
also	O
produces	O
challenges	O
in	O
the	O
efficient	O
mapping	O
of	O
parallel	O
algorithms	O
to	O
a	O
many-core	B-Architecture
system	O
.	O
</s>
