<s>
Message	B-Architecture
passing	I-Architecture
is	O
an	O
inherent	O
element	O
of	O
all	O
computer	B-Architecture
clusters	I-Architecture
.	O
</s>
<s>
All	O
computer	B-Architecture
clusters	I-Architecture
,	O
ranging	O
from	O
homemade	O
Beowulfs	B-Operating_System
to	O
some	O
of	O
the	O
fastest	O
supercomputers	B-Architecture
in	O
the	O
world	O
,	O
rely	O
on	O
message	B-Architecture
passing	I-Architecture
to	O
coordinate	O
the	O
activities	O
of	O
the	O
many	O
nodes	O
they	O
encompass	O
.	O
</s>
<s>
Message	B-Operating_System
passing	I-Operating_System
in	I-Operating_System
computer	I-Operating_System
clusters	I-Operating_System
built	O
with	O
commodity	O
servers	O
and	O
switches	O
is	O
used	O
by	O
virtually	O
every	O
internet	O
service	O
.	O
</s>
<s>
Recently	O
,	O
the	O
use	O
of	O
computer	B-Architecture
clusters	I-Architecture
with	O
more	O
than	O
one	O
thousand	O
nodes	O
has	O
been	O
spreading	O
.	O
</s>
<s>
As	O
the	O
number	O
of	O
nodes	O
in	O
a	O
cluster	O
increases	O
,	O
the	O
rapid	O
growth	O
in	O
the	O
complexity	O
of	O
the	O
communication	O
subsystem	O
makes	O
message	B-Architecture
passing	I-Architecture
delays	O
over	O
the	O
interconnect	B-Architecture
a	O
serious	O
performance	O
issue	O
in	O
the	O
execution	O
of	O
parallel	B-Operating_System
programs	I-Operating_System
.	O
</s>
<s>
Specific	O
tools	O
may	O
be	O
used	O
to	O
simulate	O
,	O
visualize	O
and	O
understand	O
the	O
performance	O
of	O
message	B-Architecture
passing	I-Architecture
on	O
computer	B-Architecture
clusters	I-Architecture
.	O
</s>
<s>
Before	O
a	O
large	O
computer	B-Architecture
cluster	I-Architecture
is	O
assembled	O
,	O
a	O
trace-based	B-Operating_System
simulator	I-Operating_System
can	O
use	O
a	O
small	O
number	O
of	O
nodes	O
to	O
help	O
predict	O
the	O
performance	O
of	O
message	B-Architecture
passing	I-Architecture
on	O
larger	O
configurations	O
.	O
</s>
<s>
Following	O
test	O
runs	O
on	O
a	O
small	O
number	O
of	O
nodes	O
,	O
the	O
simulator	O
reads	O
the	O
execution	O
and	O
message	O
transfer	O
log	B-Application
files	I-Application
and	O
simulates	O
the	O
performance	O
of	O
the	O
messaging	O
subsystem	O
when	O
many	O
more	O
messages	O
are	O
exchanged	O
between	O
a	O
much	O
larger	O
number	O
of	O
nodes	O
.	O
</s>
<s>
Historically	O
,	O
the	O
two	O
typical	O
approaches	O
to	O
communication	O
between	O
cluster	O
nodes	O
have	O
been	O
PVM	B-Operating_System
,	O
the	O
Parallel	B-Operating_System
Virtual	I-Operating_System
Machine	I-Operating_System
and	O
MPI	B-Application
,	O
the	O
Message	B-Application
Passing	I-Application
Interface	I-Application
.	O
</s>
<s>
However	O
,	O
MPI	B-Application
has	O
now	O
emerged	O
as	O
the	O
de	O
facto	O
standard	O
for	O
message	B-Architecture
passing	I-Architecture
on	O
computer	B-Architecture
clusters	I-Architecture
.	O
</s>
<s>
PVM	B-Operating_System
predates	O
MPI	B-Application
and	O
was	O
developed	O
at	O
the	O
Oak	O
Ridge	O
National	O
Laboratory	O
around	O
1989	O
.	O
</s>
<s>
It	O
provides	O
a	O
set	O
of	O
software	O
libraries	O
that	O
allow	O
a	O
computing	O
node	O
to	O
act	O
as	O
a	O
"	O
parallel	B-Operating_System
virtual	I-Operating_System
machine	I-Operating_System
"	O
.	O
</s>
<s>
It	O
provides	O
run-time	O
environment	O
for	O
message-passing	B-Architecture
,	O
task	O
and	O
resource	O
management	O
,	O
and	O
fault	O
notification	O
and	O
must	O
be	O
directly	O
installed	O
on	O
every	O
cluster	O
node	O
.	O
</s>
<s>
PVM	B-Operating_System
can	O
be	O
used	O
by	O
user	O
programs	O
written	O
in	O
C	B-Language
,	O
C++	O
,	O
or	O
Fortran	B-Application
,	O
etc	O
.	O
</s>
<s>
Unlike	O
PVM	B-Operating_System
,	O
which	O
has	O
a	O
concrete	O
implementation	O
,	O
MPI	B-Application
is	O
a	O
specification	O
rather	O
than	O
a	O
specific	O
set	O
of	O
libraries	O
.	O
</s>
<s>
The	O
design	O
of	O
MPI	B-Application
drew	O
on	O
various	O
features	O
available	O
in	O
commercial	O
systems	O
of	O
the	O
time	O
.	O
</s>
<s>
The	O
MPI	B-Application
specifications	O
then	O
gave	O
rise	O
to	O
specific	O
implementations	O
.	O
</s>
<s>
MPI	B-Application
implementations	O
typically	O
use	O
TCP/IP	B-Protocol
and	O
socket	O
connections	O
.	O
</s>
<s>
MPI	B-Application
is	O
now	O
a	O
widely	O
available	O
communications	O
model	O
that	O
enables	O
parallel	B-Operating_System
programs	I-Operating_System
to	O
be	O
written	O
in	O
languages	O
such	O
as	O
C	B-Language
,	O
Fortran	B-Application
,	O
Python	B-Language
,	O
etc	O
.	O
</s>
<s>
The	O
MPI	B-Application
specification	O
has	O
been	O
implemented	O
in	O
systems	O
such	O
as	O
MPICH	B-Library
and	O
Open	B-Library
MPI	I-Library
.	O
</s>
<s>
Computer	B-Architecture
clusters	I-Architecture
use	O
a	O
number	O
of	O
strategies	O
for	O
dealing	O
with	O
the	O
distribution	O
of	O
processing	O
over	O
multiple	O
nodes	O
and	O
the	O
resulting	O
communication	O
overhead	O
.	O
</s>
<s>
Some	O
computer	B-Architecture
clusters	I-Architecture
such	O
as	O
Tianhe-I	B-Device
use	O
different	O
processors	O
for	O
message	B-Architecture
passing	I-Architecture
than	O
those	O
used	O
for	O
performing	O
computations	O
.	O
</s>
<s>
Tiahnhe-I	O
uses	O
over	O
two	O
thousand	O
FeiTeng-1000	B-Device
processors	O
to	O
enhance	O
the	O
operation	O
of	O
its	O
proprietary	O
message	B-Architecture
passing	I-Architecture
system	O
,	O
while	O
computations	O
are	O
performed	O
by	O
Xeon	B-Device
and	O
Nvidia	B-Device
Tesla	I-Device
processors	O
.	O
</s>
<s>
One	O
approach	O
to	O
reducing	O
communication	O
overhead	O
is	O
the	O
use	O
of	O
local	O
neighborhoods	O
(	O
also	O
called	O
locales	B-Operating_System
)	O
for	O
specific	O
tasks	O
.	O
</s>
<s>
However	O
,	O
given	O
that	O
in	O
many	O
cases	O
the	O
actual	O
topology	B-Architecture
of	O
the	O
computer	B-Architecture
cluster	I-Architecture
nodes	O
and	O
their	O
interconnections	O
may	O
not	O
be	O
known	O
to	O
application	O
developers	O
,	O
attempting	O
to	O
fine	O
tune	O
performance	O
at	O
the	O
application	O
program	O
level	O
is	O
quite	O
difficult	O
.	O
</s>
<s>
Given	O
that	O
MPI	B-Application
has	O
now	O
emerged	O
as	O
the	O
de	O
facto	O
standard	O
on	O
computer	B-Architecture
clusters	I-Architecture
,	O
the	O
increase	O
in	O
the	O
number	O
of	O
cluster	O
nodes	O
has	O
resulted	O
in	O
continued	O
research	O
to	O
improve	O
the	O
efficiency	O
and	O
scalability	O
of	O
MPI	B-Application
libraries	O
.	O
</s>
<s>
These	O
efforts	O
have	O
included	O
research	O
to	O
reduce	O
the	O
memory	O
footprint	O
of	O
MPI	B-Application
libraries	O
.	O
</s>
<s>
From	O
the	O
earliest	O
days	O
MPI	B-Application
provided	O
facilities	O
for	O
performance	O
profiling	O
via	O
the	O
PMPI	O
"	O
profiling	O
system	O
"	O
.	O
</s>
<s>
The	O
need	O
for	O
more	O
information	O
resulted	O
in	O
the	O
development	O
of	O
the	O
MPI-Peruse	O
system	O
.	O
</s>
<s>
Peruse	O
provides	O
a	O
more	O
detailed	O
profile	O
by	O
enabling	O
applications	O
to	O
gain	O
access	O
to	O
state-changes	O
within	O
the	O
MPI-library	O
.	O
</s>
<s>
Peruse	O
can	O
work	O
with	O
the	O
PARAVER	B-Operating_System
visualization	O
system	O
.	O
</s>
<s>
PARAVER	B-Operating_System
has	O
two	O
components	O
,	O
a	O
trace	O
component	O
and	O
a	O
visual	O
component	O
for	O
analyze	O
the	O
traces	O
,	O
the	O
statistics	O
related	O
to	O
specific	O
events	O
,	O
etc	O
.	O
</s>
<s>
PARAVER	B-Operating_System
may	O
use	O
trace	O
formats	O
from	O
other	O
systems	O
,	O
or	O
perform	O
its	O
own	O
tracing	O
.	O
</s>
<s>
Thus	O
PARAVER	B-Operating_System
summarizes	O
them	O
to	O
allow	O
users	O
to	O
visualize	O
and	O
analyze	O
them	O
.	O
</s>
<s>
When	O
a	O
large	O
scale	O
,	O
often	O
supercomputer	B-Architecture
level	O
,	O
parallel	O
system	O
is	O
being	O
developed	O
,	O
it	O
is	O
essential	O
to	O
be	O
able	O
to	O
experiment	O
with	O
multiple	O
configurations	O
and	O
simulate	O
performance	O
.	O
</s>
<s>
There	O
are	O
a	O
number	O
of	O
approaches	O
to	O
modeling	O
message	B-Architecture
passing	I-Architecture
efficiency	O
in	O
this	O
scenario	O
,	O
ranging	O
from	O
analytical	O
models	O
to	O
trace-based	B-Operating_System
simulation	I-Operating_System
and	O
some	O
approaches	O
rely	O
on	O
the	O
use	O
of	O
test	O
environments	O
based	O
on	O
"	O
artificial	O
communications	O
"	O
to	O
perform	O
synthetic	O
tests	O
of	O
message	B-Architecture
passing	I-Architecture
performance	O
.	O
</s>
<s>
Systems	O
such	O
as	O
BIGSIM	B-Operating_System
provide	O
these	O
facilities	O
by	O
allowing	O
the	O
simulation	O
of	O
performance	O
on	O
various	O
node	B-Architecture
topologies	I-Architecture
,	O
message	B-Architecture
passing	I-Architecture
and	O
scheduling	O
strategies	O
.	O
</s>
<s>
A	O
well	O
known	O
model	O
is	O
Hockney	O
's	O
model	O
which	O
simply	O
relies	O
on	O
point	B-Architecture
to	I-Architecture
point	I-Architecture
communication	I-Architecture
,	O
using	O
T	O
=	O
L	O
+	O
(	O
M	O
/	O
R	O
)	O
where	O
M	O
is	O
the	O
message	O
size	O
,	O
L	O
is	O
the	O
startup	O
latency	O
and	O
R	O
is	O
the	O
asymptotic	O
bandwidth	O
in	O
MB/s	O
.	O
</s>
<s>
Gunawan	O
and	O
Cai	O
then	O
generalized	O
this	O
further	O
by	O
introducing	O
cache	B-General_Concept
size	I-General_Concept
,	O
and	O
separated	O
the	O
messages	O
based	O
on	O
their	O
sizes	O
,	O
obtaining	O
two	O
separate	O
models	O
,	O
one	O
for	O
messages	O
below	O
cache	B-General_Concept
size	I-General_Concept
,	O
and	O
one	O
for	O
those	O
above	O
.	O
</s>
<s>
Specific	O
tools	O
may	O
be	O
used	O
to	O
simulate	O
and	O
understand	O
the	O
performance	O
of	O
message	B-Architecture
passing	I-Architecture
on	O
computer	B-Architecture
clusters	I-Architecture
.	O
</s>
<s>
For	O
instance	O
,	O
CLUSTERSIM	O
uses	O
a	O
Java-based	O
visual	O
environment	O
for	O
discrete-event	B-General_Concept
simulation	I-General_Concept
.	O
</s>
<s>
In	O
this	O
approach	O
computed	O
nodes	O
and	O
network	B-Architecture
topology	I-Architecture
is	O
visually	O
modeled	O
.	O
</s>
<s>
Jobs	O
and	O
their	O
duration	O
and	O
complexity	O
are	O
represented	O
with	O
specific	O
probability	O
distributions	O
allowing	O
various	O
parallel	O
job	B-General_Concept
scheduling	I-General_Concept
algorithms	O
to	O
be	O
proposed	O
and	O
experimented	O
with	O
.	O
</s>
<s>
The	O
communication	O
overhead	O
for	O
MPI	B-Application
message	B-Architecture
passing	I-Architecture
can	O
thus	O
be	O
simulated	O
and	O
better	O
understood	O
in	O
the	O
context	O
of	O
large-scale	O
parallel	O
job	O
execution	O
.	O
</s>
<s>
Other	O
simulation	O
tools	O
include	O
MPI-sim	O
and	O
BIGSIM	B-Operating_System
.	O
</s>
<s>
MPI-Sim	O
is	O
an	O
execution-driven	O
simulator	O
that	O
requires	O
C	B-Language
or	O
C++	O
programs	O
to	O
operate	O
.	O
</s>
<s>
Unlike	O
MPI-Sim	O
,	O
BIGSIM	B-Operating_System
is	O
a	O
trace-driven	O
system	O
that	O
simulates	O
based	O
on	O
the	O
logs	O
of	O
executions	O
saved	O
in	O
files	O
by	O
a	O
separate	O
emulator	O
program	O
.	O
</s>
<s>
BIGSIM	B-Operating_System
includes	O
an	O
emulator	O
,	O
and	O
a	O
simulator	O
.	O
</s>
<s>
The	O
emulator	O
stores	O
information	O
of	O
sequential	O
execution	O
blocks	O
(	O
SEBs	O
)	O
for	O
multiple	O
processors	O
in	O
log	B-Application
files	I-Application
,	O
with	O
each	O
SEB	O
recording	O
the	O
messages	O
sent	O
,	O
their	O
sources	O
and	O
destinations	O
,	O
dependencies	O
,	O
timings	O
,	O
etc	O
.	O
</s>
<s>
The	O
simulator	O
reads	O
the	O
log	B-Application
files	I-Application
and	O
simulates	O
them	O
,	O
and	O
may	O
star	O
additional	O
messages	O
which	O
are	O
then	O
also	O
stored	O
as	O
SEBs	O
.	O
</s>
