<s>
HPC	B-Application
Challenge	I-Application
Benchmark	I-Application
combines	O
several	O
benchmarks	O
to	O
test	O
a	O
number	O
of	O
independent	O
attributes	O
of	O
the	O
performance	O
of	O
high-performance	O
computer	O
(	O
HPC	O
)	O
systems	O
.	O
</s>
<s>
The	O
HPC	B-Application
Challenge	I-Application
Benchmark	I-Application
is	O
an	O
effort	O
to	O
improve	O
visibility	O
into	O
this	O
multidimensional	O
space	O
by	O
combining	O
the	O
measurement	O
of	O
several	O
of	O
these	O
attributes	O
into	O
a	O
single	O
program	O
.	O
</s>
<s>
Although	O
the	O
performance	O
attributes	O
of	O
interest	O
are	O
not	O
specific	O
to	O
any	O
particular	O
computer	O
architecture	O
,	O
the	O
reference	O
implementation	O
of	O
the	O
HPC	B-Application
Challenge	I-Application
Benchmark	I-Application
in	O
C	B-Language
and	O
MPI	B-Application
assumes	O
that	O
the	O
system	O
under	O
test	O
is	O
a	O
cluster	B-Architecture
of	O
shared	B-Operating_System
memory	I-Operating_System
multiprocessor	I-Operating_System
systems	O
connected	O
by	O
a	O
network	B-Architecture
.	O
</s>
<s>
HPL	O
(	O
High	O
Performance	O
LINPACK	B-Application
)	O
–	O
measures	O
performance	O
of	O
a	O
solver	O
for	O
a	O
dense	O
system	O
of	O
linear	O
equations	O
(	O
global	O
)	O
.	O
</s>
<s>
STREAM	O
–	O
measures	O
sustained	O
memory	B-General_Concept
bandwidth	I-General_Concept
to/from	O
memory	O
(	O
single	O
,	O
star	O
)	O
.	O
</s>
<s>
FFT	O
–	O
performs	O
a	O
Fast	O
Fourier	O
Transform	O
on	O
a	O
large	O
one-dimensional	O
vector	O
using	O
the	O
generalized	O
Cooley	B-Algorithm
–	I-Algorithm
Tukey	I-Algorithm
algorithm	O
(	O
single	O
,	O
star	O
,	O
global	O
)	O
.	O
</s>
<s>
Communication	O
Bandwidth	O
and	O
Latency	O
–	O
MPI-centric	O
performance	O
measurements	O
based	O
on	O
the	O
b_eff	O
bandwidth/latency	O
benchmark	O
.	O
</s>
<s>
At	O
a	O
high	O
level	O
,	O
the	O
tests	O
are	O
intended	O
to	O
provide	O
coverage	O
of	O
four	O
important	O
attributes	O
of	O
performance	O
:	O
double-precision	O
floating-point	O
arithmetic	O
(	O
DGEMM	O
and	O
HPL	O
)	O
,	O
local	O
memory	B-General_Concept
bandwidth	I-General_Concept
(	O
STREAM	O
)	O
,	O
network	B-Architecture
bandwidth	O
for	O
"	O
large	O
"	O
messages	O
(	O
PTRANS	O
,	O
RandomAccess	O
,	O
FFT	O
,	O
b_eff	O
)	O
,	O
and	O
network	B-Architecture
bandwidth	O
for	O
"	O
small	O
"	O
messages	O
(	O
RandomAccess	O
,	O
b_eff	O
)	O
.	O
</s>
<s>
For	O
example	O
,	O
in	O
some	O
systems	O
HPL	O
performance	O
can	O
be	O
limited	O
by	O
network	B-Architecture
bandwidth	O
and/or	O
network	B-Architecture
latency	O
.	O
</s>
<s>
Class	O
2	O
:	O
Most	O
"	O
elegant	O
"	O
implementation	O
of	O
four	O
or	O
five	O
computational	O
kernels	O
including	O
three	O
or	O
more	O
of	O
the	O
HPC	B-Application
Challenge	I-Application
benchmarks	I-Application
.	O
</s>
