<s>
A	O
lightweight	O
kernel	O
(	O
LWK	O
)	O
operating	B-General_Concept
system	I-General_Concept
is	O
one	O
used	O
in	O
a	O
large	O
computer	O
with	O
many	O
processor	B-General_Concept
cores	O
,	O
termed	O
a	O
parallel	B-Operating_System
computer	I-Operating_System
.	O
</s>
<s>
A	O
massively	B-Operating_System
parallel	I-Operating_System
high-performance	B-Architecture
computing	I-Architecture
(	O
HPC	O
)	O
system	O
is	O
particularly	O
sensitive	O
to	O
operating	B-General_Concept
system	I-General_Concept
overhead	O
.	O
</s>
<s>
Traditional	O
multi-purpose	O
operating	B-General_Concept
systems	I-General_Concept
are	O
designed	O
to	O
support	O
a	O
wide	O
range	O
of	O
usage	O
models	O
and	O
requirements	O
.	O
</s>
<s>
The	O
computing	O
overhead	O
of	O
these	O
processes	O
leads	O
to	O
an	O
unpredictable	O
amount	O
of	O
processor	B-General_Concept
time	O
available	O
to	O
a	O
parallel	O
application	O
.	O
</s>
<s>
A	O
very	O
common	O
parallel	B-Application
programming	I-Application
model	I-Application
is	O
referred	O
to	O
as	O
the	O
bulk	B-Application
synchronous	I-Application
parallel	I-Application
model	O
which	O
often	O
employs	O
Message	B-Application
Passing	I-Application
Interface	I-Application
(	O
MPI	O
)	O
for	O
communication	O
.	O
</s>
<s>
If	O
one	O
processor	B-General_Concept
takes	O
longer	O
to	O
reach	O
that	O
point	O
than	O
all	O
the	O
other	O
processors	O
,	O
everyone	O
must	O
wait	O
.	O
</s>
<s>
Unpredictable	O
operating	B-General_Concept
system	I-General_Concept
overhead	O
is	O
one	O
significant	O
reason	O
a	O
processor	B-General_Concept
might	O
take	O
longer	O
to	O
reach	O
the	O
synchronization	O
point	O
than	O
the	O
others	O
.	O
</s>
<s>
Custom	O
lightweight	B-Device
kernel	I-Device
operating	I-Device
systems	I-Device
,	O
used	O
on	O
some	O
of	O
the	O
fastest	O
computers	O
in	O
the	O
world	O
,	O
help	O
alleviate	O
this	O
problem	O
.	O
</s>
<s>
The	O
IBM	B-Operating_System
Blue	I-Operating_System
Gene	I-Operating_System
line	O
of	O
supercomputers	B-Architecture
runs	O
various	O
versions	O
of	O
CNK	B-Operating_System
operating	I-Operating_System
system	I-Operating_System
.	O
</s>
<s>
The	O
Cray	B-Device
XT4	I-Device
and	O
Cray	B-Device
XT5	I-Device
supercomputers	B-Architecture
run	O
Compute	B-Application
Node	I-Application
Linux	I-Application
while	O
the	O
earlier	O
XT3	O
ran	O
the	O
lightweight	O
kernel	O
Catamount	B-Operating_System
which	O
was	O
based	O
on	O
SUNMOS	O
.	O
</s>
<s>
Sandia	O
and	O
University	O
of	O
New	O
Mexico	O
researchers	O
began	O
work	O
on	O
SUNMOS	O
for	O
the	O
Intel	B-Device
Paragon	I-Device
in	O
the	O
early	O
1990s	O
.	O
</s>
<s>
This	O
operating	B-General_Concept
system	I-General_Concept
evolved	O
into	O
the	O
Puma	O
,	O
Cougar	O
-	O
which	O
achieved	O
the	O
first	O
teraflop	O
on	O
ASCI	B-Device
Red	I-Device
-	O
and	O
Catamount	B-Operating_System
on	O
Red	B-Operating_System
Storm	I-Operating_System
.	O
</s>
<s>
Targeted	O
at	O
massively	B-Operating_System
parallel	I-Operating_System
environments	O
composed	O
of	O
thousands	O
of	O
processors	O
with	O
distributed	O
memory	O
and	O
a	O
tightly	O
coupled	O
network	O
.	O
</s>
<s>
LWK	O
implementations	O
vary	O
,	O
but	O
all	O
strive	O
to	O
provide	O
applications	O
with	O
predictable	O
and	O
maximum	O
access	O
to	O
the	O
central	B-General_Concept
processing	I-General_Concept
unit	I-General_Concept
(	O
CPU	O
)	O
and	O
other	O
system	O
resources	O
.	O
</s>
<s>
One	O
such	O
example	O
is	O
the	O
Portals	B-Operating_System
network	I-Operating_System
programming	I-Operating_System
application	I-Operating_System
programming	I-Operating_System
interface	I-Operating_System
(	O
API	O
)	O
.	O
</s>
<s>
Lightweight	B-Device
kernel	I-Device
operating	I-Device
systems	I-Device
assume	O
access	O
to	O
a	O
small	O
set	O
of	O
nodes	O
that	O
are	O
running	O
full-service	O
operating	B-General_Concept
systems	I-General_Concept
to	O
offload	O
some	O
of	O
the	O
necessary	O
services	O
:	O
login	O
access	O
,	O
compiling	O
environments	O
,	O
batch	O
job	O
submission	O
,	O
and	O
file	O
I/O	O
.	O
</s>
<s>
By	O
restricting	O
services	O
to	O
only	O
those	O
that	O
are	O
absolutely	O
necessary	O
and	O
by	O
streamlining	O
those	O
that	O
are	O
provided	O
,	O
the	O
overhead	O
(	O
sometimes	O
called	O
noise	O
)	O
of	O
the	O
lightweight	O
operating	B-General_Concept
system	I-General_Concept
is	O
minimized	O
.	O
</s>
<s>
This	O
allows	O
a	O
significant	O
and	O
predictable	O
amount	O
of	O
the	O
processor	B-General_Concept
cycles	O
to	O
be	O
given	O
to	O
the	O
parallel	O
application	O
.	O
</s>
<s>
Since	O
the	O
application	O
can	O
make	O
consistent	O
progress	O
on	O
each	O
processor	B-General_Concept
,	O
they	O
will	O
reach	O
their	O
synchronization	O
points	O
faster	O
,	O
ideally	O
at	O
the	O
same	O
time	O
.	O
</s>
<s>
The	O
last	O
supercomputers	B-Architecture
running	O
lightweight	O
kernels	O
are	O
the	O
remaining	O
IBM	O
Bluegene	B-Operating_System
systems	O
running	O
CNK	B-Operating_System
.	O
</s>
