<s>
Non-uniform	B-Operating_System
memory	I-Operating_System
access	I-Operating_System
(	O
NUMA	O
)	O
is	O
a	O
computer	B-General_Concept
memory	I-General_Concept
design	O
used	O
in	O
multiprocessing	B-Operating_System
,	O
where	O
the	O
memory	O
access	O
time	O
depends	O
on	O
the	O
memory	O
location	O
relative	O
to	O
the	O
processor	O
.	O
</s>
<s>
NUMA	O
architectures	O
logically	O
follow	O
in	O
scaling	O
from	O
symmetric	B-Operating_System
multiprocessing	I-Operating_System
(	O
SMP	O
)	O
architectures	O
.	O
</s>
<s>
Techniques	O
developed	O
by	O
these	O
companies	O
later	O
featured	O
in	O
a	O
variety	O
of	O
Unix-like	B-Operating_System
operating	I-Operating_System
systems	I-Operating_System
,	O
and	O
to	O
an	O
extent	O
in	O
Windows	B-Device
NT	I-Device
.	O
</s>
<s>
The	O
performance	O
lines	O
of	O
processors	O
and	O
memory	O
crossed	O
in	O
the	O
1960s	O
with	O
the	O
advent	O
of	O
the	O
first	O
supercomputers	B-Architecture
.	O
</s>
<s>
Many	O
supercomputer	B-Architecture
designs	O
of	O
the	O
1980s	O
and	O
1990s	O
focused	O
on	O
providing	O
high-speed	O
memory	O
access	O
as	O
opposed	O
to	O
faster	O
processors	O
,	O
allowing	O
the	O
computers	O
to	O
work	O
on	O
large	O
data	O
sets	O
at	O
speeds	O
other	O
systems	O
could	O
not	O
approach	O
.	O
</s>
<s>
For	O
commodity	O
processors	O
,	O
this	O
meant	O
installing	O
an	O
ever-increasing	O
amount	O
of	O
high-speed	O
cache	B-General_Concept
memory	I-General_Concept
and	O
using	O
increasingly	O
sophisticated	O
algorithms	O
to	O
avoid	O
cache	B-General_Concept
misses	O
.	O
</s>
<s>
But	O
the	O
dramatic	O
increase	O
in	O
size	O
of	O
the	O
operating	B-General_Concept
systems	I-General_Concept
and	O
of	O
the	O
applications	O
run	O
on	O
them	O
has	O
generally	O
overwhelmed	O
these	O
cache-processing	O
improvements	O
.	O
</s>
<s>
Multi-processor	B-Operating_System
systems	O
without	O
NUMA	O
make	O
the	O
problem	O
considerably	O
worse	O
.	O
</s>
<s>
Another	O
approach	O
to	O
addressing	O
this	O
problem	O
is	O
the	O
multi-channel	B-Architecture
memory	I-Architecture
architecture	I-Architecture
,	O
in	O
which	O
a	O
linear	O
increase	O
in	O
the	O
number	O
of	O
memory	O
channels	O
increases	O
the	O
memory	O
access	O
concurrency	O
linearly	O
.	O
</s>
<s>
AMD	O
implemented	O
NUMA	O
with	O
its	O
Opteron	B-General_Concept
processor	O
(	O
2003	O
)	O
,	O
using	O
HyperTransport	B-Device
.	O
</s>
<s>
Intel	O
announced	O
NUMA	O
compatibility	O
for	O
its	O
x86	O
and	O
Itanium	B-General_Concept
servers	O
in	O
late	O
2007	O
with	O
its	O
Nehalem	B-Device
and	O
Tukwila	B-General_Concept
CPUs	O
.	O
</s>
<s>
Both	O
Intel	O
CPU	O
families	O
share	O
a	O
common	O
chipset	B-Device
;	O
the	O
interconnection	O
is	O
called	O
Intel	B-Architecture
QuickPath	I-Architecture
Interconnect	I-Architecture
(	O
QPI	B-Architecture
)	O
,	O
which	O
provides	O
extremely	O
high	O
bandwidth	O
to	O
enable	O
high	O
on-board	O
scalability	O
and	O
was	O
replaced	O
by	O
a	O
new	O
version	O
called	O
Intel	B-Architecture
UltraPath	I-Architecture
Interconnect	I-Architecture
with	O
the	O
release	O
of	O
Skylake	B-Architecture
(	O
2017	O
)	O
.	O
</s>
<s>
Nearly	O
all	O
CPU	O
architectures	O
use	O
a	O
small	O
amount	O
of	O
very	O
fast	O
non-shared	O
memory	O
known	O
as	O
cache	B-General_Concept
to	O
exploit	O
locality	B-General_Concept
of	I-General_Concept
reference	I-General_Concept
in	O
memory	O
accesses	O
.	O
</s>
<s>
With	O
NUMA	O
,	O
maintaining	O
cache	B-General_Concept
coherence	I-General_Concept
across	O
shared	O
memory	O
has	O
a	O
significant	O
overhead	O
.	O
</s>
<s>
Although	O
simpler	O
to	O
design	O
and	O
build	O
,	O
non-cache-coherent	O
NUMA	O
systems	O
become	O
prohibitively	O
complex	O
to	O
program	O
in	O
the	O
standard	O
von	B-Architecture
Neumann	I-Architecture
architecture	I-Architecture
programming	O
model	O
.	O
</s>
<s>
Typically	O
,	O
ccNUMA	O
uses	O
inter-processor	O
communication	O
between	O
cache	B-General_Concept
controllers	O
to	O
keep	O
a	O
consistent	O
memory	O
image	O
when	O
more	O
than	O
one	O
cache	B-General_Concept
stores	O
the	O
same	O
memory	O
location	O
.	O
</s>
<s>
Support	O
for	O
NUMA	O
in	O
operating	B-General_Concept
systems	I-General_Concept
attempts	O
to	O
reduce	O
the	O
frequency	O
of	O
this	O
kind	O
of	O
access	O
by	O
allocating	O
processors	O
and	O
memory	O
in	O
NUMA-friendly	O
ways	O
and	O
by	O
avoiding	O
scheduling	O
and	O
locking	O
algorithms	O
that	O
make	O
NUMA-unfriendly	O
accesses	O
necessary	O
.	O
</s>
<s>
Alternatively	O
,	O
cache	B-General_Concept
coherency	I-General_Concept
protocols	O
such	O
as	O
the	O
MESIF	B-General_Concept
protocol	I-General_Concept
attempt	O
to	O
reduce	O
the	O
communication	O
required	O
to	O
maintain	O
cache	B-General_Concept
coherency	I-General_Concept
.	O
</s>
<s>
Scalable	B-General_Concept
Coherent	I-General_Concept
Interface	I-General_Concept
(	O
SCI	O
)	O
is	O
an	O
IEEE	O
standard	O
defining	O
a	O
directory-based	O
cache	B-General_Concept
coherency	I-General_Concept
protocol	O
to	O
avoid	O
scalability	O
limitations	O
found	O
in	O
earlier	O
multiprocessor	B-Operating_System
systems	O
.	O
</s>
<s>
One	O
can	O
view	O
NUMA	O
as	O
a	O
tightly	O
coupled	O
form	O
of	O
cluster	B-Architecture
computing	I-Architecture
.	O
</s>
<s>
The	O
addition	O
of	O
virtual	B-Architecture
memory	I-Architecture
paging	O
to	O
a	O
cluster	O
architecture	O
can	O
allow	O
the	O
implementation	O
of	O
NUMA	O
entirely	O
in	O
software	O
.	O
</s>
<s>
Microsoft	B-Device
Windows	I-Device
7	I-Device
and	O
Windows	B-Device
Server	I-Device
2008	I-Device
R2	I-Device
added	O
support	O
for	O
NUMA	O
architecture	O
over	O
64	O
logical	O
cores	O
.	O
</s>
<s>
Java	O
7	O
added	O
support	O
for	O
NUMA-aware	O
memory	O
allocator	O
and	O
garbage	B-General_Concept
collector	I-General_Concept
.	O
</s>
<s>
Linux	B-Operating_System
kernel	I-Operating_System
:	O
</s>
<s>
Version	O
2.5	O
provided	O
a	O
basic	O
NUMA	O
support	O
,	O
which	O
was	O
further	O
improved	O
in	O
subsequent	O
kernel	B-Operating_System
releases	O
.	O
</s>
<s>
Version	O
3.8	O
of	O
the	O
Linux	B-Operating_System
kernel	I-Operating_System
brought	O
a	O
new	O
NUMA	O
foundation	O
that	O
allowed	O
development	O
of	O
more	O
efficient	O
NUMA	O
policies	O
in	O
later	O
kernel	B-Operating_System
releases	O
.	O
</s>
<s>
Version	O
3.13	O
of	O
the	O
Linux	B-Operating_System
kernel	I-Operating_System
brought	O
numerous	O
policies	O
that	O
aim	O
at	O
putting	O
a	O
process	O
near	O
its	O
memory	O
,	O
together	O
with	O
the	O
handling	O
of	O
cases	O
such	O
as	O
having	O
memory	B-General_Concept
pages	I-General_Concept
shared	O
between	O
processes	O
,	O
or	O
the	O
use	O
of	O
transparent	O
huge	O
pages	O
;	O
new	O
sysctl	B-Application
settings	O
allow	O
NUMA	O
balancing	O
to	O
be	O
enabled	O
or	O
disabled	O
,	O
as	O
well	O
as	O
the	O
configuration	O
of	O
various	O
NUMA	O
memory	O
balancing	O
parameters	O
.	O
</s>
<s>
OpenSolaris	B-Operating_System
models	O
NUMA	O
architecture	O
with	O
lgroups	O
.	O
</s>
<s>
FreeBSD	B-Operating_System
added	O
support	O
for	O
NUMA	O
architecture	O
in	O
version	O
9.0	O
.	O
</s>
<s>
Silicon	O
Graphics	O
IRIX	B-Operating_System
(	O
discontinued	O
as	O
of	O
2021	O
)	O
support	O
for	O
ccNUMA	O
architecture	O
over	O
1240	O
CPU	O
with	O
Origin	O
server	B-Application
series	O
.	O
</s>
<s>
As	O
of	O
2011	O
,	O
ccNUMA	O
systems	O
are	O
multiprocessor	B-Operating_System
systems	O
based	O
on	O
the	O
AMD	B-General_Concept
Opteron	I-General_Concept
processor	O
,	O
which	O
can	O
be	O
implemented	O
without	O
external	O
logic	O
,	O
and	O
the	O
Intel	B-General_Concept
Itanium	I-General_Concept
processor	O
,	O
which	O
requires	O
the	O
chipset	B-Device
to	O
support	O
NUMA	O
.	O
</s>
<s>
Examples	O
of	O
ccNUMA-enabled	O
chipsets	B-Device
are	O
the	O
SGI	O
Shub	O
(	O
Super	O
hub	O
)	O
,	O
the	O
Intel	O
E8870	O
,	O
the	O
HP	O
sx2000	O
(	O
used	O
in	O
the	O
Integrity	O
and	O
Superdome	O
servers	O
)	O
,	O
and	O
those	O
found	O
in	O
NEC	O
Itanium-based	O
systems	O
.	O
</s>
<s>
Earlier	O
ccNUMA	O
systems	O
such	O
as	O
those	O
from	O
Silicon	O
Graphics	O
were	O
based	O
on	O
MIPS	B-Device
processors	O
and	O
the	O
DEC	O
Alpha	B-General_Concept
21364	I-General_Concept
(	O
EV7	O
)	O
processor	O
.	O
</s>
