<s>
rCUDA	B-General_Concept
,	O
which	O
stands	O
for	O
Remote	O
CUDA	B-Architecture
,	O
is	O
a	O
type	O
of	O
middleware	B-General_Concept
software	O
framework	O
for	O
remote	O
GPU	B-Architecture
virtualization	O
.	O
</s>
<s>
Fully	O
compatible	O
with	O
the	O
CUDA	B-Architecture
application	O
programming	O
interface	O
(	O
API	B-General_Concept
)	O
,	O
it	O
allows	O
the	O
allocation	O
of	O
one	O
or	O
more	O
CUDA-enabled	O
GPUs	B-Architecture
to	O
a	O
single	O
application	O
.	O
</s>
<s>
Each	O
GPU	B-Architecture
can	O
be	O
part	O
of	O
a	O
cluster	B-Architecture
or	O
running	O
inside	O
of	O
a	O
virtual	B-Architecture
machine	I-Architecture
.	O
</s>
<s>
The	O
approach	O
is	O
aimed	O
at	O
improving	O
performance	O
in	O
GPU	B-Architecture
clusters	O
that	O
are	O
lacking	O
full	O
utilization	O
.	O
</s>
<s>
GPU	B-Architecture
virtualization	O
reduces	O
the	O
number	O
of	O
GPUs	B-Architecture
needed	O
in	O
a	O
cluster	B-Architecture
,	O
and	O
in	O
turn	O
,	O
leads	O
to	O
a	O
lower	O
cost	O
configuration	O
–	O
less	O
energy	O
,	O
acquisition	O
,	O
and	O
maintenance	O
.	O
</s>
<s>
The	O
recommended	O
distributed	O
acceleration	O
architecture	O
is	O
a	O
high	B-Architecture
performance	I-Architecture
computing	I-Architecture
cluster	B-Architecture
with	O
GPUs	B-Architecture
attached	O
to	O
only	O
a	O
few	O
of	O
the	O
cluster	B-Architecture
nodes	O
.	O
</s>
<s>
When	O
a	O
node	O
without	O
a	O
local	O
GPU	B-Architecture
executes	O
an	O
application	O
needing	O
GPU	B-Architecture
resources	O
,	O
remote	O
execution	O
of	O
the	O
kernel	O
is	O
supported	O
by	O
data	O
and	O
code	O
transfers	O
between	O
local	O
system	O
memory	O
and	O
remote	O
GPU	B-Architecture
memory	O
.	O
</s>
<s>
rCUDA	B-General_Concept
is	O
designed	O
to	O
accommodate	O
this	O
client-server	B-Operating_System
architecture	I-Operating_System
.	O
</s>
<s>
On	O
one	O
end	O
,	O
clients	O
employ	O
a	O
library	O
of	O
wrappers	O
to	O
the	O
high-level	O
CUDA	B-Architecture
Runtime	O
API	B-General_Concept
,	O
and	O
on	O
the	O
other	O
end	O
,	O
there	O
is	O
a	O
network	O
listening	O
service	O
that	O
receives	O
requests	O
on	O
a	O
TCP	B-Protocol
port	I-Protocol
.	O
</s>
<s>
Several	O
nodes	O
running	O
different	O
GPU-accelerated	O
applications	O
can	O
concurrently	O
make	O
use	O
of	O
the	O
whole	O
set	O
of	O
accelerators	O
installed	O
in	O
the	O
cluster	B-Architecture
.	O
</s>
<s>
The	O
client	O
forwards	O
the	O
request	O
to	O
one	O
of	O
the	O
servers	O
,	O
which	O
accesses	O
the	O
GPU	B-Architecture
installed	O
in	O
that	O
computer	O
and	O
executes	O
the	O
request	O
in	O
it	O
.	O
</s>
<s>
Time-multiplexing	B-Protocol
the	O
GPU	B-Architecture
,	O
or	O
in	O
other	O
words	O
sharing	O
it	O
,	O
is	O
accomplished	O
by	O
spawning	O
different	O
server	O
processes	O
for	O
each	O
remote	O
GPU	B-Architecture
execution	O
request	O
.	O
</s>
<s>
The	O
rCUDA	B-General_Concept
middleware	B-General_Concept
enables	O
the	O
concurrent	O
usage	O
of	O
CUDA-compatible	O
devices	O
remotely	O
.	O
</s>
<s>
rCUDA	B-General_Concept
employs	O
either	O
the	O
InfiniBand	O
network	O
or	O
the	O
socket	O
API	B-General_Concept
for	O
the	O
communication	O
between	O
clients	O
and	O
servers	O
.	O
</s>
<s>
rCUDA	B-General_Concept
can	O
be	O
useful	O
in	O
three	O
different	O
environments	O
:	O
</s>
<s>
To	O
reduce	O
the	O
number	O
of	O
GPUs	B-Architecture
installed	O
in	O
High	O
Performance	O
Clusters	O
.	O
</s>
<s>
In	O
commodity	O
networks	O
,	O
to	O
offer	O
access	O
to	O
a	O
few	O
high	O
performance	O
GPUs	B-Architecture
concurrently	O
to	O
many	O
students	O
.	O
</s>
<s>
Virtual	B-Architecture
Machines	I-Architecture
.	O
</s>
<s>
To	O
enable	O
the	O
access	O
to	O
the	O
CUDA	B-Architecture
facilities	O
on	O
the	O
physical	O
machine	O
.	O
</s>
<s>
The	O
current	O
version	O
of	O
rCUDA	B-General_Concept
(	O
v20.07	O
)	O
supports	O
CUDA	B-Architecture
version	O
9.0	O
,	O
excluding	O
graphics	O
interoperability	O
.	O
</s>
<s>
rCUDA	B-General_Concept
v20.07	O
targets	O
the	O
Linux	B-Application
OS	I-Application
(	O
for	O
64-bit	O
architectures	O
)	O
on	O
both	O
client	O
and	O
server	O
sides	O
.	O
</s>
<s>
CUDA	B-Architecture
applications	O
do	O
not	O
need	O
any	O
change	O
in	O
their	O
source	O
code	O
in	O
order	O
to	O
be	O
executed	O
with	O
rCUDA	B-General_Concept
.	O
</s>
