<s>
A	O
supercomputer	B-Operating_System
operating	I-Operating_System
system	I-Operating_System
is	O
an	O
operating	B-General_Concept
system	I-General_Concept
intended	O
for	O
supercomputers	B-Architecture
.	O
</s>
<s>
Since	O
the	O
end	O
of	O
the	O
20th	O
century	O
,	O
supercomputer	B-Operating_System
operating	I-Operating_System
systems	I-Operating_System
have	O
undergone	O
major	O
transformations	O
,	O
as	O
fundamental	O
changes	O
have	O
occurred	O
in	O
supercomputer	B-Operating_System
architecture	I-Operating_System
.	O
</s>
<s>
While	O
early	O
operating	B-General_Concept
systems	I-General_Concept
were	O
custom	O
tailored	O
to	O
each	O
supercomputer	B-Architecture
to	O
gain	O
speed	O
,	O
the	O
trend	O
has	O
been	O
moving	O
away	O
from	O
in-house	O
operating	B-General_Concept
systems	I-General_Concept
and	O
toward	O
some	O
form	O
of	O
Linux	B-Application
,	O
with	O
it	O
running	O
all	O
the	O
supercomputers	B-Architecture
on	O
the	O
TOP500	B-Operating_System
list	O
in	O
November	O
2017	O
.	O
</s>
<s>
In	O
2021	O
,	O
top	O
10	O
computers	O
run	O
for	O
instance	O
Red	O
Hat	O
Enterprise	O
Linux	B-Application
(	O
RHEL	O
)	O
,	O
or	O
some	O
variant	O
of	O
it	O
or	O
other	O
Linux	B-Application
distribution	O
e.g.	O
</s>
<s>
Ubuntu	B-Operating_System
.	O
</s>
<s>
Given	O
that	O
modern	O
massively	B-Operating_System
parallel	I-Operating_System
supercomputers	B-Architecture
typically	O
separate	O
computations	O
from	O
other	O
services	O
by	O
using	O
multiple	O
types	O
of	O
nodes	B-Operating_System
,	O
they	O
usually	O
run	O
different	O
operating	B-General_Concept
systems	I-General_Concept
on	O
different	O
nodes	B-Operating_System
,	O
e.g.	O
,	O
using	O
a	O
small	O
and	O
efficient	O
lightweight	B-Device
kernel	I-Device
such	O
as	O
Compute	B-Operating_System
Node	I-Operating_System
Kernel	I-Operating_System
(	O
CNK	B-Operating_System
)	O
or	O
Compute	B-Application
Node	I-Application
Linux	I-Application
(	O
CNL	B-Application
)	O
on	O
compute	O
nodes	B-Operating_System
,	O
but	O
a	O
larger	O
system	O
such	O
as	O
a	O
Linux-derivative	O
on	O
server	O
and	O
input/output	B-General_Concept
(	O
I/O	B-General_Concept
)	O
nodes	B-Operating_System
.	O
</s>
<s>
While	O
in	O
a	O
traditional	O
multi-user	O
computer	O
system	O
job	B-General_Concept
scheduling	I-General_Concept
is	O
in	O
effect	O
a	O
tasking	O
problem	O
for	O
processing	O
and	O
peripheral	O
resources	O
,	O
in	O
a	O
massively	B-Operating_System
parallel	I-Operating_System
system	O
,	O
the	O
job	O
management	O
system	O
needs	O
to	O
manage	O
the	O
allocation	O
of	O
both	O
computational	O
and	O
communication	O
resources	O
,	O
as	O
well	O
as	O
gracefully	O
dealing	O
with	O
inevitable	O
hardware	O
failures	O
when	O
tens	O
of	O
thousands	O
of	O
processors	O
are	O
present	O
.	O
</s>
<s>
Although	O
most	O
modern	O
supercomputers	B-Architecture
use	O
the	O
Linux	B-Application
operating	I-Application
system	I-Application
,	O
each	O
manufacturer	O
has	O
made	O
its	O
own	O
specific	O
changes	O
to	O
the	O
Linux-derivative	O
they	O
use	O
,	O
and	O
no	O
industry	O
standard	O
exists	O
,	O
partly	O
because	O
the	O
differences	O
in	O
hardware	O
architectures	O
require	O
changes	O
to	O
optimize	O
the	O
operating	B-General_Concept
system	I-General_Concept
to	O
each	O
hardware	O
design	O
.	O
</s>
<s>
In	O
the	O
early	O
days	O
of	O
supercomputing	B-Architecture
,	O
the	O
basic	O
architectural	O
concepts	O
were	O
evolving	O
rapidly	O
,	O
and	O
system	B-Application
software	I-Application
had	O
to	O
follow	O
hardware	O
innovations	O
that	O
usually	O
took	O
rapid	O
turns	O
.	O
</s>
<s>
In	O
the	O
early	O
systems	O
,	O
operating	B-General_Concept
systems	I-General_Concept
were	O
custom	O
tailored	O
to	O
each	O
supercomputer	B-Architecture
to	O
gain	O
speed	O
,	O
yet	O
in	O
the	O
rush	O
to	O
develop	O
them	O
,	O
serious	O
software	O
quality	O
challenges	O
surfaced	O
and	O
in	O
many	O
cases	O
the	O
cost	O
and	O
complexity	O
of	O
system	B-Application
software	I-Application
development	O
became	O
as	O
much	O
an	O
issue	O
as	O
that	O
of	O
hardware	O
.	O
</s>
<s>
In	O
the	O
1980s	O
the	O
cost	O
for	O
software	O
development	O
at	O
Cray	O
came	O
to	O
equal	O
what	O
they	O
spent	O
on	O
hardware	O
and	O
that	O
trend	O
was	O
partly	O
responsible	O
for	O
a	O
move	O
away	O
from	O
the	O
in-house	O
operating	B-General_Concept
systems	I-General_Concept
to	O
the	O
adaptation	O
of	O
generic	O
software	O
.	O
</s>
<s>
The	O
first	O
wave	O
in	O
operating	B-General_Concept
system	I-General_Concept
changes	O
came	O
in	O
the	O
mid-1980s	O
,	O
as	O
vendor	O
specific	O
operating	B-General_Concept
systems	I-General_Concept
were	O
abandoned	O
in	O
favor	O
of	O
Unix	B-Application
.	O
</s>
<s>
By	O
the	O
early	O
1990s	O
,	O
major	O
changes	O
were	O
occurring	O
in	O
supercomputing	B-Architecture
system	B-Application
software	I-Application
.	O
</s>
<s>
By	O
this	O
time	O
,	O
the	O
growing	O
use	O
of	O
Unix	B-Application
had	O
begun	O
to	O
change	O
the	O
way	O
system	B-Application
software	I-Application
was	O
viewed	O
.	O
</s>
<s>
The	O
use	O
of	O
a	O
high	O
level	O
language	O
(	O
C	B-Language
)	O
to	O
implement	O
the	O
operating	B-General_Concept
system	I-General_Concept
,	O
and	O
the	O
reliance	O
on	O
standardized	O
interfaces	O
was	O
in	O
contrast	O
to	O
the	O
assembly	B-Language
language	I-Language
oriented	O
approaches	O
of	O
the	O
past	O
.	O
</s>
<s>
As	O
hardware	O
vendors	O
adapted	O
Unix	B-Application
to	O
their	O
systems	O
,	O
new	O
and	O
useful	O
features	O
were	O
added	O
to	O
Unix	B-Application
,	O
e.g.	O
,	O
fast	O
file	O
systems	O
and	O
tunable	O
process	O
schedulers	O
.	O
</s>
<s>
However	O
,	O
all	O
the	O
companies	O
that	O
adapted	O
Unix	B-Application
made	O
unique	O
changes	O
to	O
it	O
,	O
rather	O
than	O
collaborating	O
on	O
an	O
industry	O
standard	O
to	O
create	O
"	O
Unix	B-Application
for	O
supercomputers	B-Architecture
"	O
.	O
</s>
<s>
This	O
was	O
partly	O
because	O
differences	O
in	O
their	O
architectures	O
required	O
these	O
changes	O
to	O
optimize	O
Unix	B-Application
to	O
each	O
architecture	O
.	O
</s>
<s>
Thus	O
as	O
general	O
purpose	O
operating	B-General_Concept
systems	I-General_Concept
became	O
stable	O
,	O
supercomputers	B-Architecture
began	O
to	O
borrow	O
and	O
adapt	O
the	O
critical	O
system	O
code	O
from	O
them	O
and	O
relied	O
on	O
the	O
rich	O
set	O
of	O
secondary	O
functions	O
that	O
came	O
with	O
them	O
,	O
not	O
having	O
to	O
reinvent	O
the	O
wheel	O
.	O
</s>
<s>
However	O
,	O
at	O
the	O
same	O
time	O
the	O
size	O
of	O
the	O
code	O
for	O
general	O
purpose	O
operating	B-General_Concept
systems	I-General_Concept
was	O
growing	O
rapidly	O
.	O
</s>
<s>
By	O
the	O
time	O
Unix-based	O
code	O
had	O
reached	O
500,000	O
lines	O
long	O
,	O
its	O
maintenance	O
and	O
use	O
was	O
a	O
challenge	O
.	O
</s>
<s>
This	O
resulted	O
in	O
the	O
move	O
to	O
use	O
microkernels	B-Operating_System
which	O
used	O
a	O
minimal	O
set	O
of	O
the	O
operating	B-General_Concept
system	I-General_Concept
functions	O
.	O
</s>
<s>
Systems	O
such	O
as	O
Mach	B-Operating_System
at	O
Carnegie	O
Mellon	O
University	O
and	O
ChorusOS	B-Operating_System
at	O
INRIA	O
were	O
examples	O
of	O
early	O
microkernels	B-Operating_System
.	O
</s>
<s>
The	O
separation	O
of	O
the	O
operating	B-General_Concept
system	I-General_Concept
into	O
separate	O
components	O
became	O
necessary	O
as	O
supercomputers	B-Architecture
developed	O
different	O
types	O
of	O
nodes	B-Operating_System
,	O
e.g.	O
,	O
compute	O
nodes	B-Operating_System
versus	O
I/O	B-General_Concept
nodes	B-Operating_System
.	O
</s>
<s>
Thus	O
modern	O
supercomputers	B-Architecture
usually	O
run	O
different	O
operating	B-General_Concept
systems	I-General_Concept
on	O
different	O
nodes	B-Operating_System
,	O
e.g.	O
,	O
using	O
a	O
small	O
and	O
efficient	O
lightweight	B-Device
kernel	I-Device
such	O
as	O
CNK	B-Operating_System
or	O
CNL	B-Application
on	O
compute	O
nodes	B-Operating_System
,	O
but	O
a	O
larger	O
system	O
such	O
as	O
a	O
Linux-derivative	O
on	O
server	O
and	O
I/O	B-General_Concept
nodes	B-Operating_System
.	O
</s>
<s>
The	O
CDC	B-Device
6600	I-Device
,	O
generally	O
considered	O
the	O
first	O
supercomputer	B-Architecture
in	O
the	O
world	O
,	O
ran	O
the	O
Chippewa	B-Operating_System
Operating	I-Operating_System
System	I-Operating_System
,	O
which	O
was	O
then	O
deployed	O
on	O
various	O
other	O
CDC	B-Device
6000	I-Device
series	I-Device
computers	O
.	O
</s>
<s>
The	O
Chippewa	O
was	O
a	O
rather	O
simple	O
job	B-General_Concept
control	I-General_Concept
oriented	O
system	O
derived	O
from	O
the	O
earlier	O
CDC	B-Device
3000	I-Device
,	O
but	O
it	O
influenced	O
the	O
later	O
KRONOS	B-Operating_System
and	O
SCOPE	B-Operating_System
systems	O
.	O
</s>
<s>
The	O
first	O
Cray-1	B-Device
was	O
delivered	O
to	O
the	O
Los	O
Alamos	O
Lab	O
with	O
no	O
operating	B-General_Concept
system	I-General_Concept
,	O
or	O
any	O
other	O
software	O
.	O
</s>
<s>
Los	O
Alamos	O
developed	O
the	O
application	O
software	O
for	O
it	O
,	O
and	O
the	O
operating	B-General_Concept
system	I-General_Concept
.	O
</s>
<s>
The	O
main	O
timesharing	O
system	O
for	O
the	O
Cray	B-Device
1	I-Device
,	O
the	O
Cray	B-Operating_System
Time	I-Operating_System
Sharing	I-Operating_System
System	I-Operating_System
(	O
CTSS	O
)	O
,	O
was	O
then	O
developed	O
at	O
the	O
Livermore	O
Labs	O
as	O
a	O
direct	O
descendant	O
of	O
the	O
Livermore	B-Operating_System
Time	I-Operating_System
Sharing	I-Operating_System
System	I-Operating_System
(	O
LTSS	O
)	O
for	O
the	O
CDC	B-Device
6600	I-Device
operating	B-General_Concept
system	I-General_Concept
from	O
twenty	O
years	O
earlier	O
.	O
</s>
<s>
In	O
developing	O
supercomputers	B-Architecture
,	O
rising	O
software	O
costs	O
soon	O
became	O
dominant	O
,	O
as	O
evidenced	O
by	O
the	O
1980s	O
cost	O
for	O
software	O
development	O
at	O
Cray	O
growing	O
to	O
equal	O
their	O
cost	O
for	O
hardware	O
.	O
</s>
<s>
That	O
trend	O
was	O
partly	O
responsible	O
for	O
a	O
move	O
away	O
from	O
the	O
in-house	O
Cray	B-Operating_System
Operating	I-Operating_System
System	I-Operating_System
to	O
UNICOS	B-Operating_System
system	O
based	O
on	O
Unix	B-Application
.	O
</s>
<s>
In	O
1985	O
,	O
the	O
Cray-2	B-Device
was	O
the	O
first	O
system	O
to	O
ship	O
with	O
the	O
UNICOS	B-Operating_System
operating	B-General_Concept
system	I-General_Concept
.	O
</s>
<s>
Around	O
the	O
same	O
time	O
,	O
the	O
EOS	B-Operating_System
operating	B-General_Concept
system	I-General_Concept
was	O
developed	O
by	O
ETA	O
Systems	O
for	O
use	O
in	O
their	O
ETA10	B-Operating_System
supercomputers	B-Architecture
.	O
</s>
<s>
Written	O
in	O
Cybil	B-Language
,	O
a	O
Pascal-like	O
language	O
from	O
Control	O
Data	O
Corporation	O
,	O
EOS	B-Operating_System
highlighted	O
the	O
stability	O
problems	O
in	O
developing	O
stable	O
operating	B-General_Concept
systems	I-General_Concept
for	O
supercomputers	B-Architecture
and	O
eventually	O
a	O
Unix-like	B-Operating_System
system	I-Operating_System
was	O
offered	O
on	O
the	O
same	O
machine	O
.	O
</s>
<s>
The	O
lessons	O
learned	O
from	O
developing	O
ETA	O
system	B-Application
software	I-Application
included	O
the	O
high	O
level	O
of	O
risk	O
associated	O
with	O
developing	O
a	O
new	O
supercomputer	B-Operating_System
operating	I-Operating_System
system	I-Operating_System
,	O
and	O
the	O
advantages	O
of	O
using	O
Unix	B-Application
with	O
its	O
large	O
extant	O
base	O
of	O
system	B-Application
software	I-Application
libraries	O
.	O
</s>
<s>
By	O
the	O
middle	O
1990s	O
,	O
despite	O
the	O
extant	O
investment	O
in	O
older	O
operating	B-General_Concept
systems	I-General_Concept
,	O
the	O
trend	O
was	O
toward	O
the	O
use	O
of	O
Unix-based	O
systems	O
,	O
which	O
also	O
facilitated	O
the	O
use	O
of	O
interactive	O
graphical	B-Application
user	I-Application
interfaces	I-Application
(	O
GUIs	B-Application
)	O
for	O
scientific	O
computing	O
across	O
multiple	O
platforms	O
.	O
</s>
<s>
The	O
move	O
toward	O
a	O
commodity	O
OS	O
had	O
opponents	O
,	O
who	O
cited	O
the	O
fast	O
pace	O
and	O
focus	O
of	O
Linux	B-Application
development	O
as	O
a	O
major	O
obstacle	O
against	O
adoption	O
.	O
</s>
<s>
As	O
one	O
author	O
wrote	O
"	O
Linux	B-Application
will	O
likely	O
catch	O
up	O
,	O
but	O
we	O
have	O
large-scale	O
systems	O
now	O
"	O
.	O
</s>
<s>
Nevertheless	O
,	O
that	O
trend	O
continued	O
to	O
gain	O
momentum	O
and	O
by	O
2005	O
,	O
virtually	O
all	O
supercomputers	B-Architecture
used	O
some	O
Unix-like	B-Operating_System
OS	I-Operating_System
.	O
</s>
<s>
These	O
variants	O
of	O
Unix	B-Application
included	O
IBM	B-Application
AIX	I-Application
,	O
the	O
open	B-Application
source	I-Application
Linux	B-Application
system	O
,	O
and	O
other	O
adaptations	O
such	O
as	O
UNICOS	B-Operating_System
from	O
Cray	O
.	O
</s>
<s>
By	O
the	O
end	O
of	O
the	O
20th	O
century	O
,	O
Linux	B-Application
was	O
estimated	O
to	O
command	O
the	O
highest	O
share	O
of	O
the	O
supercomputing	B-Architecture
pie	O
.	O
</s>
<s>
The	O
IBM	B-Operating_System
Blue	I-Operating_System
Gene	I-Operating_System
supercomputer	B-Architecture
uses	O
the	O
CNK	B-Operating_System
operating	I-Operating_System
system	I-Operating_System
on	O
the	O
compute	O
nodes	B-Operating_System
,	O
but	O
uses	O
a	O
modified	O
Linux-based	O
kernel	O
called	O
I/O	B-Application
Node	I-Application
Kernel	I-Application
(	O
INK	B-Application
)	O
on	O
the	O
I/O	B-General_Concept
nodes	B-Operating_System
.	O
</s>
<s>
CNK	B-Operating_System
is	O
a	O
lightweight	B-Device
kernel	I-Device
that	O
runs	O
on	O
each	O
node	O
and	O
supports	O
a	O
single	O
application	O
running	O
for	O
a	O
single	O
user	O
on	O
that	O
node	O
.	O
</s>
<s>
For	O
the	O
sake	O
of	O
efficient	O
operation	O
,	O
the	O
design	O
of	O
CNK	B-Operating_System
was	O
kept	O
simple	O
and	O
minimal	O
,	O
with	O
physical	O
memory	O
being	O
statically	O
mapped	O
and	O
the	O
CNK	B-Operating_System
neither	O
needing	O
nor	O
providing	O
scheduling	O
or	O
context	O
switching	O
.	O
</s>
<s>
CNK	B-Operating_System
does	O
not	O
even	O
implement	O
file	B-General_Concept
I/O	I-General_Concept
on	O
the	O
compute	O
node	O
,	O
but	O
delegates	O
that	O
to	O
dedicated	O
I/O	B-General_Concept
nodes	B-Operating_System
.	O
</s>
<s>
However	O
,	O
given	O
that	O
on	O
the	O
Blue	B-Operating_System
Gene	I-Operating_System
multiple	O
compute	O
nodes	B-Operating_System
share	O
a	O
single	O
I/O	B-General_Concept
node	O
,	O
the	O
I/O	B-General_Concept
node	O
operating	B-General_Concept
system	I-General_Concept
does	O
require	O
multi-tasking	O
,	O
hence	O
the	O
selection	O
of	O
the	O
Linux-based	O
operating	B-General_Concept
system	I-General_Concept
.	O
</s>
<s>
While	O
in	O
traditional	O
multi-user	O
computer	O
systems	O
and	O
early	O
supercomputers	B-Architecture
,	O
job	B-General_Concept
scheduling	I-General_Concept
was	O
in	O
effect	O
a	O
task	O
scheduling	O
problem	O
for	O
processing	O
and	O
peripheral	O
resources	O
,	O
in	O
a	O
massively	B-Operating_System
parallel	I-Operating_System
system	O
,	O
the	O
job	O
management	O
system	O
needs	O
to	O
manage	O
the	O
allocation	O
of	O
both	O
computational	O
and	O
communication	O
resources	O
.	O
</s>
<s>
It	O
is	O
essential	O
to	O
tune	O
task	O
scheduling	O
,	O
and	O
the	O
operating	B-General_Concept
system	I-General_Concept
,	O
in	O
different	O
configurations	O
of	O
a	O
supercomputer	B-Architecture
.	O
</s>
<s>
A	O
typical	O
parallel	O
job	B-General_Concept
scheduler	I-General_Concept
has	O
a	O
master	B-Operating_System
scheduler	I-Operating_System
which	O
instructs	O
some	O
number	O
of	O
slave	O
schedulers	O
to	O
launch	O
,	O
monitor	O
,	O
and	O
control	O
parallel	B-Operating_System
jobs	I-Operating_System
,	O
and	O
periodically	O
receives	O
reports	O
from	O
them	O
about	O
the	O
status	O
of	O
job	O
progress	O
.	O
</s>
<s>
Some	O
,	O
but	O
not	O
all	O
supercomputer	B-Architecture
schedulers	O
attempt	O
to	O
maintain	O
locality	O
of	O
job	O
execution	O
.	O
</s>
<s>
The	O
PBS	B-Application
Pro	I-Application
scheduler	I-Application
used	O
on	O
the	O
Cray	B-Device
XT3	I-Device
and	O
Cray	B-Device
XT4	I-Device
systems	O
does	O
not	O
attempt	O
to	O
optimize	O
locality	O
on	O
its	O
three-dimensional	O
torus	B-Operating_System
interconnect	I-Operating_System
,	O
but	O
simply	O
uses	O
the	O
first	O
available	O
processor	O
.	O
</s>
<s>
On	O
the	O
other	O
hand	O
,	O
IBM	O
's	O
scheduler	O
on	O
the	O
Blue	B-Operating_System
Gene	I-Operating_System
supercomputers	B-Architecture
aims	O
to	O
exploit	O
locality	O
and	O
minimize	O
network	O
contention	O
by	O
assigning	O
tasks	O
from	O
the	O
same	O
application	O
to	O
one	O
or	O
more	O
midplanes	O
of	O
an	O
8x8x8	O
node	O
group	O
.	O
</s>
<s>
The	O
Slurm	B-Operating_System
Workload	I-Operating_System
Manager	I-Operating_System
scheduler	O
uses	O
a	O
best	O
fit	O
algorithm	O
,	O
and	O
performs	O
Hilbert	B-General_Concept
curve	I-General_Concept
scheduling	I-General_Concept
to	O
optimize	O
locality	O
of	O
task	O
assignments	O
.	O
</s>
<s>
Several	O
modern	O
supercomputers	B-Architecture
such	O
as	O
the	O
Tianhe-2	B-Device
use	O
Slurm	B-Operating_System
,	O
which	O
arbitrates	O
contention	O
for	O
resources	O
across	O
the	O
system	O
.	O
</s>
<s>
Slurm	B-Operating_System
is	O
open	B-Application
source	I-Application
,	O
Linux-based	O
,	O
very	O
scalable	O
,	O
and	O
can	O
manage	O
thousands	O
of	O
nodes	B-Operating_System
in	O
a	O
computer	O
cluster	O
with	O
a	O
sustained	O
throughput	O
of	O
over	O
100,000	O
jobs	O
per	O
hour	O
.	O
</s>
