<s>
In	O
parallel	B-Operating_System
computing	I-Operating_System
,	O
granularity	B-Operating_System
(	O
or	O
grain	O
size	O
)	O
of	O
a	O
task	B-General_Concept
is	O
a	O
measure	O
of	O
the	O
amount	O
of	O
work	O
(	O
or	O
computation	O
)	O
which	O
is	O
performed	O
by	O
that	O
task	B-General_Concept
.	O
</s>
<s>
Another	O
definition	O
of	O
granularity	B-Operating_System
takes	O
into	O
account	O
the	O
communication	O
overhead	O
between	O
multiple	O
processors	O
or	O
processing	O
elements	O
.	O
</s>
<s>
It	O
defines	O
granularity	B-Operating_System
as	O
the	O
ratio	O
of	O
computation	O
time	O
to	O
communication	O
time	O
,	O
wherein	O
computation	O
time	O
is	O
the	O
time	O
required	O
to	O
perform	O
the	O
computation	O
of	O
a	O
task	B-General_Concept
and	O
communication	O
time	O
is	O
the	O
time	O
required	O
to	O
exchange	O
data	O
between	O
processors	O
.	O
</s>
<s>
If	O
is	O
the	O
computation	O
time	O
and	O
denotes	O
the	O
communication	O
time	O
,	O
then	O
the	O
granularity	B-Operating_System
of	O
a	O
task	B-General_Concept
can	O
be	O
calculated	O
as	O
:	O
</s>
<s>
Granularity	B-Operating_System
is	O
usually	O
measured	O
in	O
terms	O
of	O
the	O
number	O
of	O
instructions	B-General_Concept
which	O
are	O
executed	B-General_Concept
in	O
a	O
particular	O
task	B-General_Concept
.	O
</s>
<s>
Alternately	O
,	O
granularity	B-Operating_System
can	O
also	O
be	O
specified	O
in	O
terms	O
of	O
the	O
execution	O
time	O
of	O
a	O
program	O
,	O
combining	O
the	O
computation	O
time	O
and	O
communication	O
time	O
.	O
</s>
<s>
Depending	O
on	O
the	O
amount	O
of	O
work	O
which	O
is	O
performed	O
by	O
a	O
parallel	O
task	B-General_Concept
,	O
parallelism	B-Operating_System
can	O
be	O
classified	O
into	O
three	O
categories	O
:	O
fine-grained	O
,	O
medium-grained	O
and	O
coarse-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
In	O
fine-grained	O
parallelism	B-Operating_System
,	O
a	O
program	O
is	O
broken	O
down	O
to	O
a	O
large	O
number	O
of	O
small	O
tasks	O
.	O
</s>
<s>
The	O
amount	O
of	O
work	O
associated	O
with	O
a	O
parallel	O
task	B-General_Concept
is	O
low	O
and	O
the	O
work	O
is	O
evenly	O
distributed	O
among	O
the	O
processors	O
.	O
</s>
<s>
Hence	O
,	O
fine-grained	O
parallelism	B-Operating_System
facilitates	O
load	B-Application
balancing	I-Application
.	O
</s>
<s>
As	O
each	O
task	B-General_Concept
processes	O
less	O
data	O
,	O
the	O
number	O
of	O
processors	O
required	O
to	O
perform	O
the	O
complete	O
processing	O
is	O
high	O
.	O
</s>
<s>
Fine-grained	O
parallelism	B-Operating_System
is	O
best	O
exploited	O
in	O
architectures	O
which	O
support	O
fast	O
communication	O
.	O
</s>
<s>
Shared	B-Operating_System
memory	I-Operating_System
architecture	I-Operating_System
which	O
has	O
a	O
low	O
communication	O
overhead	O
is	O
most	O
suitable	O
for	O
fine-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
It	O
is	O
difficult	O
for	O
programmers	O
to	O
detect	O
parallelism	B-Operating_System
in	O
a	O
program	O
,	O
therefore	O
,	O
it	O
is	O
usually	O
the	O
compilers	B-Language
 '	I-Language
responsibility	O
to	O
detect	O
fine-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
An	O
example	O
of	O
a	O
fine-grained	O
system	O
(	O
from	O
outside	O
the	O
parallel	B-Operating_System
computing	I-Operating_System
domain	O
)	O
is	O
the	O
system	O
of	O
neurons	O
in	O
our	O
brain	O
.	O
</s>
<s>
Connection	B-Device
Machine	I-Device
(	O
CM-2	O
)	O
and	O
J-Machine	B-Device
are	O
examples	O
of	O
fine-grain	O
parallel	B-Operating_System
computers	I-Operating_System
that	O
have	O
grain	O
size	O
in	O
the	O
range	O
of	O
4-5	O
μs	O
.	O
</s>
<s>
In	O
coarse-grained	O
parallelism	B-Operating_System
,	O
a	O
program	O
is	O
split	O
into	O
large	O
tasks	O
.	O
</s>
<s>
Further	O
,	O
coarse-grained	O
parallelism	B-Operating_System
fails	O
to	O
exploit	O
the	O
parallelism	B-Operating_System
in	O
the	O
program	O
as	O
most	O
of	O
the	O
computation	O
is	O
performed	O
sequentially	O
on	O
a	O
processor	O
.	O
</s>
<s>
The	O
advantage	O
of	O
this	O
type	O
of	O
parallelism	B-Operating_System
is	O
low	O
communication	O
and	O
synchronization	O
overhead	O
.	O
</s>
<s>
Message-passing	B-Architecture
architecture	O
takes	O
a	O
long	O
time	O
to	O
communicate	O
data	O
among	O
processes	O
which	O
makes	O
it	O
suitable	O
for	O
coarse-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
Cray	B-Device
Y-MP	I-Device
is	O
an	O
example	O
of	O
coarse-grained	O
parallel	B-Operating_System
computer	I-Operating_System
which	O
has	O
a	O
grain	O
size	O
of	O
about	O
20s	O
.	O
</s>
<s>
Medium-grained	O
parallelism	B-Operating_System
is	O
used	O
relatively	O
to	O
fine-grained	O
and	O
coarse-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
Medium-grained	O
parallelism	B-Operating_System
is	O
a	O
compromise	O
between	O
fine-grained	O
and	O
coarse-grained	O
parallelism	B-Operating_System
,	O
where	O
we	O
have	O
task	B-General_Concept
size	O
and	O
communication	O
time	O
greater	O
than	O
fine-grained	O
parallelism	B-Operating_System
and	O
lower	O
than	O
coarse-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
Most	O
general-purpose	O
parallel	B-Operating_System
computers	I-Operating_System
fall	O
in	O
this	O
category	O
.	O
</s>
<s>
Intel	B-Device
iPSC	I-Device
is	O
an	O
example	O
of	O
medium-grained	O
parallel	B-Operating_System
computer	I-Operating_System
which	O
has	O
a	O
grain	O
size	O
of	O
about	O
10ms	O
.	O
</s>
<s>
Fine-grained	O
parallelism	B-Operating_System
:	O
Assume	O
there	O
are	O
100	O
processors	O
that	O
are	O
responsible	O
for	O
processing	O
the	O
10*10	O
image	O
.	O
</s>
<s>
This	O
is	O
an	O
example	O
of	O
fine-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
Medium-grained	O
parallelism	B-Operating_System
:	O
Consider	O
that	O
there	O
are	O
25	O
processors	O
processing	O
the	O
10*10	O
image	O
.	O
</s>
<s>
This	O
is	O
an	O
example	O
of	O
medium-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
Coarse-grained	O
parallelism	B-Operating_System
:	O
Further	O
,	O
if	O
we	O
reduce	O
the	O
processors	O
to	O
2	O
,	O
then	O
the	O
processing	O
will	O
take	O
50	O
clock	O
cycles	O
.	O
</s>
<s>
This	O
case	O
illustrates	O
coarse-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
Granularity	B-Operating_System
is	O
closely	O
tied	O
to	O
the	O
level	O
of	O
processing	O
.	O
</s>
<s>
The	O
highest	O
amount	O
of	O
parallelism	B-Operating_System
is	O
achieved	O
at	O
instruction	O
level	O
,	O
followed	O
by	O
loop-level	O
parallelism	B-Operating_System
.	O
</s>
<s>
At	O
instruction	O
and	O
loop	O
level	O
,	O
fine-grained	O
parallelism	B-Operating_System
is	O
achieved	O
.	O
</s>
<s>
Typical	O
grain	O
size	O
at	O
instruction-level	O
is	O
20	O
instructions	B-General_Concept
,	O
while	O
the	O
grain-size	O
at	O
loop-level	O
is	O
500	O
instructions	B-General_Concept
.	O
</s>
<s>
At	O
the	O
sub-routine	O
(	O
or	O
procedure	O
)	O
level	O
the	O
grain	O
size	O
is	O
typically	O
a	O
few	O
thousand	O
instructions	B-General_Concept
.	O
</s>
<s>
Medium-grained	O
parallelism	B-Operating_System
is	O
achieved	O
at	O
sub-routine	O
level	O
.	O
</s>
<s>
Granularity	B-Operating_System
can	O
be	O
in	O
the	O
range	O
of	O
tens	O
of	O
thousands	O
of	O
instructions	B-General_Concept
.	O
</s>
<s>
Coarse-grained	O
parallelism	B-Operating_System
is	O
used	O
at	O
this	O
level	O
.	O
</s>
<s>
Granularity	B-Operating_System
affects	O
the	O
performance	O
of	O
parallel	B-Operating_System
computers	I-Operating_System
.	O
</s>
<s>
Using	O
fine	O
grains	O
or	O
small	O
tasks	O
results	O
in	O
more	O
parallelism	B-Operating_System
and	O
hence	O
increases	O
the	O
speedup	B-Operating_System
.	O
</s>
<s>
However	O
,	O
synchronization	O
overhead	O
,	O
scheduling	B-Application
strategies	O
etc	O
.	O
</s>
<s>
Increasing	O
parallelism	B-Operating_System
alone	O
cannot	O
give	O
the	O
best	O
performance	O
.	O
</s>
<s>
In	O
order	O
to	O
reduce	O
the	O
communication	O
overhead	O
,	O
granularity	B-Operating_System
can	O
be	O
increased	O
.	O
</s>
<s>
Hence	O
optimal	O
performance	O
is	O
achieved	O
between	O
the	O
two	O
extremes	O
of	O
fine-grained	O
and	O
coarse-grained	O
parallelism	B-Operating_System
.	O
</s>
<s>
Various	O
studies	O
have	O
proposed	O
their	O
solution	O
to	O
help	O
determine	O
the	O
best	O
granularity	B-Operating_System
to	O
aid	O
parallel	B-Operating_System
processing	I-Operating_System
.	O
</s>
