<s>
NAS	B-Device
Parallel	I-Device
Benchmarks	I-Device
(	O
NPB	O
)	O
are	O
a	O
set	O
of	O
benchmarks	O
targeting	O
performance	O
evaluation	O
of	O
highly	O
parallel	B-Operating_System
supercomputers	B-Architecture
.	O
</s>
<s>
They	O
are	O
developed	O
and	O
maintained	O
by	O
the	O
NASA	O
Advanced	O
Supercomputing	B-Architecture
(	O
NAS	O
)	O
Division	O
(	O
formerly	O
the	O
NASA	O
Numerical	O
Aerodynamic	O
Simulation	O
Program	O
)	O
based	O
at	O
the	O
NASA	O
Ames	O
Research	O
Center	O
.	O
</s>
<s>
Traditional	O
benchmarks	O
that	O
existed	O
before	O
NPB	O
,	O
such	O
as	O
the	O
Livermore	B-Operating_System
loops	I-Operating_System
,	O
the	O
LINPACK	B-Application
Benchmark	I-Application
and	O
the	O
,	O
were	O
usually	O
specialized	O
for	O
vector	O
computers	O
.	O
</s>
<s>
They	O
generally	O
suffered	O
from	O
inadequacies	O
including	O
parallelism-impeding	O
tuning	O
restrictions	O
and	O
insufficient	O
problem	O
sizes	O
,	O
which	O
rendered	O
them	O
inappropriate	O
for	O
highly	O
parallel	B-Operating_System
systems	O
.	O
</s>
<s>
Equally	O
unsuitable	O
were	O
full-scale	O
application	O
benchmarks	O
due	O
to	O
high	O
porting	O
cost	O
and	O
unavailability	O
of	O
automatic	O
software	O
parallelization	B-Operating_System
tools	O
.	O
</s>
<s>
As	O
a	O
result	O
,	O
NPB	O
were	O
developed	O
in	O
1991	O
and	O
released	O
in	O
1992	O
to	O
address	O
the	O
ensuing	O
lack	O
of	O
benchmarks	O
applicable	O
to	O
highly	O
parallel	B-Operating_System
machines	I-Operating_System
.	O
</s>
<s>
new	O
parallel-aware	O
algorithmic	O
and	O
software	O
methods	O
,	O
</s>
<s>
Secondly	O
,	O
problem	O
sizes	O
of	O
NPB	O
1	O
lagged	O
behind	O
the	O
development	O
of	O
supercomputers	B-Architecture
as	O
the	O
latter	O
continued	O
to	O
evolve	O
.	O
</s>
<s>
NPB	O
2.3	O
of	O
1997	O
was	O
the	O
first	O
complete	O
implementation	O
in	O
MPI	B-Application
.	O
</s>
<s>
It	O
shipped	O
with	O
serial	O
versions	O
of	O
the	O
benchmarks	O
consistent	O
with	O
the	O
parallel	B-Operating_System
versions	O
and	O
defined	O
a	O
problem	O
size	O
Class	O
W	O
for	O
small-memory	O
systems	O
.	O
</s>
<s>
NPB	O
2.4	O
of	O
2002	O
offered	O
a	O
new	O
MPI	B-Application
implementation	O
and	O
introduced	O
another	O
still	O
larger	O
problem	O
size	O
Class	O
D	O
.	O
It	O
also	O
augmented	O
one	O
benchmark	O
with	O
I/O	B-General_Concept
-intensive	O
subtypes	O
.	O
</s>
<s>
NPB	O
3	O
retained	O
the	O
MPI	B-Application
implementation	O
from	O
NPB	O
2	O
and	O
came	O
in	O
more	O
flavors	O
,	O
namely	O
OpenMP	B-Application
,	O
Java	B-Language
and	O
High	B-Language
Performance	I-Language
Fortran	I-Language
.	O
</s>
<s>
These	O
new	O
parallel	B-Operating_System
implementations	O
were	O
derived	O
from	O
the	O
serial	O
codes	O
in	O
NPB	O
2.3	O
with	O
additional	O
optimizations	O
.	O
</s>
<s>
Based	O
on	O
the	O
single-zone	O
NPB	O
3	O
,	O
a	O
set	O
of	O
multi-zone	O
benchmarks	O
taking	O
advantage	O
of	O
the	O
MPI/OpenMP	O
hybrid	O
programming	O
model	O
were	O
released	O
under	O
the	O
name	O
NPB-Multi-Zone	O
(	O
NPB-MZ	O
)	O
for	O
"	O
testing	O
the	O
effectiveness	O
of	O
multi-level	O
and	O
hybrid	O
parallelization	B-Operating_System
paradigms	O
and	O
tools	O
"	O
.	O
</s>
<s>
All	O
three	O
benchmarks	O
have	O
multi-zone	O
versions	O
SP	O
Scalar	O
Pentadiagonal	B-Algorithm
LU	O
Lower-Upper	O
symmetric	B-Algorithm
Gauss-Seidel	B-Algorithm
UA	O
Unstructured	O
Adaptive	O
NPB	O
3.1	O
Solve	O
Heat	O
equation	O
with	O
convection	O
and	O
diffusion	O
from	O
moving	O
ball	O
.	O
</s>
