<s>
In	O
computing	O
,	O
load	B-Application
balancing	I-Application
is	O
the	O
process	O
of	O
distributing	O
a	O
set	O
of	O
tasks	B-General_Concept
over	O
a	O
set	O
of	O
resources	B-General_Concept
(	O
computing	O
units	O
)	O
,	O
with	O
the	O
aim	O
of	O
making	O
their	O
overall	O
processing	O
more	O
efficient	O
.	O
</s>
<s>
Load	B-Application
balancing	I-Application
can	O
optimize	O
the	O
response	O
time	O
and	O
avoid	O
unevenly	O
overloading	O
some	O
compute	O
nodes	O
while	O
other	O
compute	O
nodes	O
are	O
left	O
idle	O
.	O
</s>
<s>
Load	B-Application
balancing	I-Application
is	O
the	O
subject	O
of	O
research	O
in	O
the	O
field	O
of	O
parallel	B-Operating_System
computers	I-Operating_System
.	O
</s>
<s>
Among	O
other	O
things	O
,	O
the	O
nature	O
of	O
the	O
tasks	B-General_Concept
,	O
the	O
algorithmic	O
complexity	O
,	O
the	O
hardware	O
architecture	O
on	O
which	O
the	O
algorithms	O
will	O
run	O
as	O
well	O
as	O
required	O
error	B-General_Concept
tolerance	I-General_Concept
,	O
must	O
be	O
taken	O
into	O
account	O
.	O
</s>
<s>
The	O
efficiency	O
of	O
load	B-Application
balancing	I-Application
algorithms	O
critically	O
depends	O
on	O
the	O
nature	O
of	O
the	O
tasks	B-General_Concept
.	O
</s>
<s>
Therefore	O
,	O
the	O
more	O
information	O
about	O
the	O
tasks	B-General_Concept
is	O
available	O
at	O
the	O
time	O
of	O
decision	O
making	O
,	O
the	O
greater	O
the	O
potential	O
for	O
optimization	O
.	O
</s>
<s>
Perfect	O
knowledge	O
of	O
the	O
execution	B-Library
time	I-Library
of	O
each	O
of	O
the	O
tasks	B-General_Concept
allows	O
to	O
reach	O
an	O
optimal	O
load	B-Application
distribution	I-Application
(	O
see	O
algorithm	O
of	O
prefix	B-Application
sum	I-Application
)	O
.	O
</s>
<s>
Knowing	O
the	O
exact	O
execution	B-Library
time	I-Library
of	O
each	O
task	O
is	O
an	O
extremely	O
rare	O
situation	O
.	O
</s>
<s>
For	O
this	O
reason	O
,	O
there	O
are	O
several	O
techniques	O
to	O
get	O
an	O
idea	O
of	O
the	O
different	O
execution	B-Library
times	I-Library
.	O
</s>
<s>
First	O
of	O
all	O
,	O
in	O
the	O
fortunate	O
scenario	O
of	O
having	O
tasks	B-General_Concept
of	O
relatively	O
homogeneous	O
size	O
,	O
it	O
is	O
possible	O
to	O
consider	O
that	O
each	O
of	O
them	O
will	O
require	O
approximately	O
the	O
average	O
execution	B-Library
time	I-Library
.	O
</s>
<s>
If	O
,	O
on	O
the	O
other	O
hand	O
,	O
the	O
execution	B-Library
time	I-Library
is	O
very	O
irregular	O
,	O
more	O
sophisticated	O
techniques	O
must	O
be	O
used	O
.	O
</s>
<s>
Depending	O
on	O
the	O
previous	O
execution	B-Library
time	I-Library
for	O
similar	O
metadata	O
,	O
it	O
is	O
possible	O
to	O
make	O
inferences	O
for	O
a	O
future	O
task	O
based	O
on	O
statistics	O
.	O
</s>
<s>
In	O
some	O
cases	O
,	O
tasks	B-General_Concept
depend	O
on	O
each	O
other	O
.	O
</s>
<s>
Intuitively	O
,	O
some	O
tasks	B-General_Concept
cannot	O
begin	O
until	O
others	O
are	O
completed	O
.	O
</s>
<s>
Assuming	O
that	O
the	O
required	O
time	O
for	O
each	O
of	O
the	O
tasks	B-General_Concept
is	O
known	O
in	O
advance	O
,	O
an	O
optimal	O
execution	O
order	O
must	O
lead	O
to	O
the	O
minimization	O
of	O
the	O
total	O
execution	B-Library
time	I-Library
.	O
</s>
<s>
There	O
are	O
algorithms	O
,	O
like	O
job	B-General_Concept
scheduler	I-General_Concept
,	O
that	O
calculate	O
optimal	O
task	O
distributions	O
using	O
metaheuristic	B-Algorithm
methods	O
.	O
</s>
<s>
Another	O
feature	O
of	O
the	O
tasks	B-General_Concept
critical	O
for	O
the	O
design	O
of	O
a	O
load	B-Application
balancing	I-Application
algorithm	O
is	O
their	O
ability	O
to	O
be	O
broken	O
down	O
into	O
subtasks	O
during	O
execution	O
.	O
</s>
<s>
A	O
load	B-Application
balancing	I-Application
algorithm	O
is	O
"	O
static	O
"	O
when	O
it	O
does	O
not	O
take	O
into	O
account	O
the	O
state	O
of	O
the	O
system	O
for	O
the	O
distribution	O
of	O
tasks	B-General_Concept
.	O
</s>
<s>
Thereby	O
,	O
the	O
system	O
state	O
includes	O
measures	O
such	O
as	O
the	O
load	B-Operating_System
level	I-Operating_System
(	O
and	O
sometimes	O
even	O
overload	O
)	O
of	O
certain	O
processors	O
.	O
</s>
<s>
Instead	O
,	O
assumptions	O
about	O
the	O
overall	O
system	O
are	O
made	O
beforehand	O
,	O
such	O
as	O
the	O
arrival	O
times	O
and	O
resource	O
requirements	O
of	O
incoming	O
tasks	B-General_Concept
.	O
</s>
<s>
Therefore	O
,	O
static	O
load	B-Application
balancing	I-Application
aims	O
to	O
associate	O
a	O
known	O
set	O
of	O
tasks	B-General_Concept
with	O
the	O
available	O
processors	O
in	O
order	O
to	O
minimize	O
a	O
certain	O
performance	O
function	O
.	O
</s>
<s>
Static	O
load	B-Application
balancing	I-Application
techniques	O
are	O
commonly	O
centralized	O
around	O
a	O
router	O
,	O
or	O
Master	B-Operating_System
,	O
which	O
distributes	O
the	O
loads	O
and	O
optimizes	O
the	O
performance	O
function	O
.	O
</s>
<s>
This	O
minimization	O
can	O
take	O
into	O
account	O
information	O
related	O
to	O
the	O
tasks	B-General_Concept
to	O
be	O
distributed	O
,	O
and	O
derive	O
an	O
expected	O
execution	B-Library
time	I-Library
.	O
</s>
<s>
The	O
advantage	O
of	O
static	O
algorithms	O
is	O
that	O
they	O
are	O
easy	O
to	O
set	O
up	O
and	O
extremely	O
efficient	O
in	O
the	O
case	O
of	O
fairly	O
regular	O
tasks	B-General_Concept
(	O
such	O
as	O
processing	O
HTTP	B-Protocol
requests	I-Protocol
from	O
a	O
website	O
)	O
.	O
</s>
<s>
However	O
,	O
there	O
is	O
still	O
some	O
statistical	O
variance	O
in	O
the	O
assignment	O
of	O
tasks	B-General_Concept
which	O
can	O
lead	O
to	O
the	O
overloading	O
of	O
some	O
computing	O
units	O
.	O
</s>
<s>
Unlike	O
static	O
load	B-Application
distribution	I-Application
algorithms	O
,	O
dynamic	O
algorithms	O
take	O
into	O
account	O
the	O
current	O
load	O
of	O
each	O
of	O
the	O
computing	O
units	O
(	O
also	O
called	O
nodes	O
)	O
in	O
the	O
system	O
.	O
</s>
<s>
In	O
this	O
approach	O
,	O
tasks	B-General_Concept
can	O
be	O
moved	O
dynamically	O
from	O
an	O
overloaded	O
node	O
to	O
an	O
underloaded	O
node	O
in	O
order	O
to	O
receive	O
faster	O
processing	O
.	O
</s>
<s>
While	O
these	O
algorithms	O
are	O
much	O
more	O
complicated	O
to	O
design	O
,	O
they	O
can	O
produce	O
excellent	O
results	O
,	O
in	O
particular	O
,	O
when	O
the	O
execution	B-Library
time	I-Library
varies	O
greatly	O
from	O
one	O
task	O
to	O
another	O
.	O
</s>
<s>
Dynamic	O
load	B-Application
balancing	I-Application
architecture	O
can	O
be	O
more	O
modular	O
since	O
it	O
is	O
not	O
mandatory	O
to	O
have	O
a	O
specific	O
node	O
dedicated	O
to	O
the	O
distribution	O
of	O
work	O
.	O
</s>
<s>
When	O
tasks	B-General_Concept
are	O
uniquely	O
assigned	O
to	O
a	O
processor	O
according	O
to	O
their	O
state	O
at	O
a	O
given	O
moment	O
,	O
it	O
is	O
a	O
unique	O
assignment	O
.	O
</s>
<s>
If	O
,	O
on	O
the	O
other	O
hand	O
,	O
the	O
tasks	B-General_Concept
can	O
be	O
permanently	O
redistributed	O
according	O
to	O
the	O
state	O
of	O
the	O
system	O
and	O
its	O
evolution	O
,	O
this	O
is	O
called	O
dynamic	O
assignment	O
.	O
</s>
<s>
Obviously	O
,	O
a	O
load	B-Application
balancing	I-Application
algorithm	O
that	O
requires	O
too	O
much	O
communication	O
in	O
order	O
to	O
reach	O
its	O
decisions	O
runs	O
the	O
risk	O
of	O
slowing	O
down	O
the	O
resolution	O
of	O
the	O
overall	O
problem	O
.	O
</s>
<s>
Parallel	B-Operating_System
computing	I-Operating_System
infrastructures	O
are	O
often	O
composed	O
of	O
units	O
of	O
different	O
computing	O
power	O
,	O
which	O
should	O
be	O
taken	O
into	O
account	O
for	O
the	O
load	B-Application
distribution	I-Application
.	O
</s>
<s>
Parallel	B-Operating_System
computers	I-Operating_System
are	O
often	O
divided	O
into	O
two	O
broad	O
categories	O
:	O
those	O
where	O
all	O
processors	O
share	O
a	O
single	O
common	O
memory	O
on	O
which	O
they	O
read	O
and	O
write	O
in	O
parallel	O
(	O
PRAM	B-Operating_System
model	O
)	O
,	O
and	O
those	O
where	O
each	O
computing	O
unit	O
has	O
its	O
own	O
memory	O
(	O
distributed	B-Operating_System
memory	I-Operating_System
model	O
)	O
,	O
and	O
where	O
information	O
is	O
exchanged	O
by	O
messages	O
.	O
</s>
<s>
For	O
shared-memory	B-Operating_System
computers	O
,	O
managing	O
write	O
conflicts	O
greatly	O
slows	O
down	O
the	O
speed	O
of	O
individual	O
execution	O
of	O
each	O
computing	O
unit	O
.	O
</s>
<s>
In	O
general	O
,	O
the	O
processors	O
each	O
have	O
an	O
internal	O
memory	O
to	O
store	O
the	O
data	O
needed	O
for	O
the	O
next	O
calculations	O
and	O
are	O
organized	O
in	O
successive	O
clusters	B-Architecture
.	O
</s>
<s>
Often	O
,	O
these	O
processing	O
elements	O
are	O
then	O
coordinated	O
through	O
distributed	B-Operating_System
memory	I-Operating_System
and	O
message	B-Architecture
passing	I-Architecture
.	O
</s>
<s>
Therefore	O
,	O
the	O
load	B-Application
balancing	I-Application
algorithm	O
should	O
be	O
uniquely	O
adapted	O
to	O
a	O
parallel	B-Operating_System
architecture	I-Operating_System
.	O
</s>
<s>
Adapting	O
to	O
the	O
hardware	O
structures	O
seen	O
above	O
,	O
there	O
are	O
two	O
main	O
categories	O
of	O
load	B-Application
balancing	I-Application
algorithms	O
.	O
</s>
<s>
On	O
the	O
one	O
hand	O
,	O
the	O
one	O
where	O
tasks	B-General_Concept
are	O
assigned	O
by	O
“	O
master	B-Operating_System
”	O
and	O
executed	O
by	O
“	O
workers	O
”	O
who	O
keep	O
the	O
master	B-Operating_System
informed	O
of	O
the	O
progress	O
of	O
their	O
work	O
,	O
and	O
the	O
master	B-Operating_System
can	O
then	O
take	O
charge	O
of	O
assigning	O
or	O
reassigning	O
the	O
workload	O
in	O
case	O
of	O
the	O
dynamic	O
algorithm	O
.	O
</s>
<s>
The	O
literature	O
refers	O
to	O
this	O
as	O
"	B-Operating_System
Master-Worker	I-Operating_System
"	I-Operating_System
architecture	O
.	O
</s>
<s>
The	O
load	B-Application
balancing	I-Application
algorithm	O
is	O
then	O
executed	O
on	O
each	O
of	O
them	O
and	O
the	O
responsibility	O
for	O
assigning	O
tasks	B-General_Concept
(	O
as	O
well	O
as	O
re-assigning	O
and	O
splitting	O
as	O
appropriate	O
)	O
is	O
shared	O
.	O
</s>
<s>
The	O
last	O
category	O
assumes	O
a	O
dynamic	O
load	B-Application
balancing	I-Application
algorithm	O
.	O
</s>
<s>
Since	O
the	O
design	O
of	O
each	O
load	B-Application
balancing	I-Application
algorithm	O
is	O
unique	O
,	O
the	O
previous	O
distinction	O
must	O
be	O
qualified	O
.	O
</s>
<s>
Thus	O
,	O
it	O
is	O
also	O
possible	O
to	O
have	O
an	O
intermediate	O
strategy	O
,	O
with	O
,	O
for	O
example	O
,	O
"	O
master	B-Operating_System
"	O
nodes	O
for	O
each	O
sub-cluster	O
,	O
which	O
are	O
themselves	O
subject	O
to	O
a	O
global	O
"	O
master	B-Operating_System
"	O
.	O
</s>
<s>
There	O
are	O
also	O
multi-level	O
organizations	O
,	O
with	O
an	O
alternation	O
between	O
master-slave	B-Operating_System
and	O
distributed	O
control	O
strategies	O
.	O
</s>
<s>
An	O
extremely	O
important	O
parameter	O
of	O
a	O
load	B-Application
balancing	I-Application
algorithm	O
is	O
therefore	O
its	O
ability	O
to	O
adapt	O
to	O
scalable	B-Architecture
hardware	O
architecture	O
.	O
</s>
<s>
This	O
is	O
called	O
the	O
scalability	B-Architecture
of	O
the	O
algorithm	O
.	O
</s>
<s>
An	O
algorithm	O
is	O
called	O
scalable	B-Architecture
for	O
an	O
input	O
parameter	O
when	O
its	O
performance	O
remains	O
relatively	O
independent	O
of	O
the	O
size	O
of	O
that	O
parameter	O
.	O
</s>
<s>
Most	O
load	B-Application
balancing	I-Application
algorithms	O
are	O
at	O
least	O
moldable	O
.	O
</s>
<s>
Especially	O
in	O
large-scale	O
computing	B-Architecture
clusters	I-Architecture
,	O
it	O
is	O
not	O
tolerable	O
to	O
execute	O
a	O
parallel	O
algorithm	O
that	O
cannot	O
withstand	O
the	O
failure	O
of	O
one	O
single	O
component	O
.	O
</s>
<s>
Therefore	O
,	O
fault	B-General_Concept
tolerant	I-General_Concept
algorithms	O
are	O
being	O
developed	O
which	O
can	O
detect	O
outages	O
of	O
processors	O
and	O
recover	O
the	O
computation	O
.	O
</s>
<s>
If	O
the	O
tasks	B-General_Concept
are	O
independent	O
of	O
each	O
other	O
,	O
and	O
if	O
their	O
respective	O
execution	B-Library
time	I-Library
and	O
the	O
tasks	B-General_Concept
can	O
be	O
subdivided	O
,	O
there	O
is	O
a	O
simple	O
and	O
optimal	O
algorithm	O
.	O
</s>
<s>
By	O
dividing	O
the	O
tasks	B-General_Concept
in	O
such	O
a	O
way	O
as	O
to	O
give	O
the	O
same	O
amount	O
of	O
computation	O
to	O
each	O
processor	O
,	O
all	O
that	O
remains	O
to	O
be	O
done	O
is	O
to	O
group	O
the	O
results	O
together	O
.	O
</s>
<s>
Using	O
a	O
prefix	B-Application
sum	I-Application
algorithm	O
,	O
this	O
division	O
can	O
be	O
calculated	O
in	O
logarithmic	O
time	O
with	O
respect	O
to	O
the	O
number	O
of	O
processors	O
.	O
</s>
<s>
If	O
,	O
however	O
,	O
the	O
tasks	B-General_Concept
cannot	O
be	O
subdivided	O
(	O
i.e.	O
,	O
they	O
are	O
atomic	B-General_Concept
)	O
,	O
although	O
optimizing	O
task	O
assignment	O
is	O
a	O
difficult	O
problem	O
,	O
it	O
is	O
still	O
possible	O
to	O
approximate	O
a	O
relatively	O
fair	O
distribution	O
of	O
tasks	B-General_Concept
,	O
provided	O
that	O
the	O
size	O
of	O
each	O
of	O
them	O
is	O
much	O
smaller	O
than	O
the	O
total	O
computation	O
performed	O
by	O
each	O
of	O
the	O
nodes	O
.	O
</s>
<s>
Most	O
of	O
the	O
time	O
,	O
the	O
execution	B-Library
time	I-Library
of	O
a	O
task	O
is	O
unknown	O
and	O
only	O
rough	O
approximations	O
are	O
available	O
.	O
</s>
<s>
Even	O
if	O
the	O
execution	B-Library
time	I-Library
is	O
not	O
known	O
in	O
advance	O
at	O
all	O
,	O
static	O
load	B-Application
distribution	I-Application
is	O
always	O
possible	O
.	O
</s>
<s>
In	O
a	O
round-robin	O
algorithm	O
,	O
the	O
first	O
request	O
is	O
sent	O
to	O
the	O
first	O
server	B-Application
,	O
then	O
the	O
next	O
to	O
the	O
second	O
,	O
and	O
so	O
on	O
down	O
to	O
the	O
last	O
.	O
</s>
<s>
Then	O
it	O
is	O
started	O
again	O
,	O
assigning	O
the	O
next	O
request	O
to	O
the	O
first	O
server	B-Application
,	O
and	O
so	O
on	O
.	O
</s>
<s>
Randomized	O
static	O
load	B-Application
balancing	I-Application
is	O
simply	O
a	O
matter	O
of	O
randomly	O
assigning	O
tasks	B-General_Concept
to	O
the	O
different	O
servers	O
.	O
</s>
<s>
If	O
,	O
on	O
the	O
other	O
hand	O
,	O
the	O
number	O
of	O
tasks	B-General_Concept
is	O
known	O
in	O
advance	O
,	O
it	O
is	O
even	O
more	O
efficient	O
to	O
calculate	O
a	O
random	O
permutation	O
in	O
advance	O
.	O
</s>
<s>
There	O
is	O
no	O
longer	O
a	O
need	O
for	O
a	O
distribution	O
master	B-Operating_System
because	O
every	O
processor	O
knows	O
what	O
task	O
is	O
assigned	O
to	O
it	O
.	O
</s>
<s>
Even	O
if	O
the	O
number	O
of	O
tasks	B-General_Concept
is	O
unknown	O
,	O
it	O
is	O
still	O
possible	O
to	O
avoid	O
communication	O
with	O
a	O
pseudo-random	O
assignment	O
generation	O
known	O
to	O
all	O
processors	O
.	O
</s>
<s>
The	O
performance	O
of	O
this	O
strategy	O
(	O
measured	O
in	O
total	O
execution	B-Library
time	I-Library
for	O
a	O
given	O
fixed	O
set	O
of	O
tasks	B-General_Concept
)	O
decreases	O
with	O
the	O
maximum	O
size	O
of	O
the	O
tasks	B-General_Concept
.	O
</s>
<s>
Less	O
work	O
:	O
Assign	O
more	O
tasks	B-General_Concept
to	O
the	O
servers	O
by	O
performing	O
less	O
(	O
the	O
method	O
can	O
also	O
be	O
weighted	O
)	O
.	O
</s>
<s>
Hash	O
:	O
allocates	O
queries	O
according	O
to	O
a	O
hash	B-Algorithm
table	I-Algorithm
.	O
</s>
<s>
Master-Worker	B-Operating_System
schemes	O
are	O
among	O
the	O
simplest	O
dynamic	O
load	B-Application
balancing	I-Application
algorithms	O
.	O
</s>
<s>
A	O
master	B-Operating_System
distributes	O
the	O
workload	O
to	O
all	O
workers	O
(	O
also	O
sometimes	O
referred	O
to	O
as	O
"	O
slaves	O
"	O
)	O
.	O
</s>
<s>
Initially	O
,	O
all	O
workers	O
are	O
idle	O
and	O
report	O
this	O
to	O
the	O
master	B-Operating_System
.	O
</s>
<s>
The	O
master	B-Operating_System
answers	O
worker	O
requests	O
and	O
distributes	O
the	O
tasks	B-General_Concept
to	O
them	O
.	O
</s>
<s>
When	O
he	O
has	O
no	O
more	O
tasks	B-General_Concept
to	O
give	O
,	O
he	O
informs	O
the	O
workers	O
so	O
that	O
they	O
stop	O
asking	O
for	O
tasks	B-General_Concept
.	O
</s>
<s>
In	O
fact	O
,	O
if	O
one	O
does	O
not	O
take	O
into	O
account	O
the	O
time	O
needed	O
for	O
the	O
assignment	O
,	O
the	O
execution	B-Library
time	I-Library
would	O
be	O
comparable	O
to	O
the	O
prefix	B-Application
sum	I-Application
seen	O
above	O
.	O
</s>
<s>
This	O
lack	O
of	O
scalability	B-Architecture
makes	O
it	O
quickly	O
inoperable	O
in	O
very	O
large	O
servers	O
or	O
very	O
large	O
parallel	B-Operating_System
computers	I-Operating_System
.	O
</s>
<s>
The	O
master	B-Operating_System
acts	O
as	O
a	O
bottleneck	O
.	O
</s>
<s>
However	O
,	O
the	O
quality	O
of	O
the	O
algorithm	O
can	O
be	O
greatly	O
improved	O
by	O
replacing	O
the	O
master	B-Operating_System
with	O
a	O
task	O
list	O
that	O
can	O
be	O
used	O
by	O
different	O
processors	O
.	O
</s>
<s>
Although	O
this	O
algorithm	O
is	O
a	O
little	O
more	O
difficult	O
to	O
implement	O
,	O
it	O
promises	O
much	O
better	O
scalability	B-Architecture
,	O
although	O
still	O
insufficient	O
for	O
very	O
large	O
computing	O
centers	O
.	O
</s>
<s>
Another	O
technique	O
to	O
overcome	O
scalability	B-Architecture
problems	O
when	O
the	O
time	O
needed	O
for	O
task	O
completion	O
is	O
unknown	O
is	O
work	B-Operating_System
stealing	I-Operating_System
.	O
</s>
<s>
The	O
approach	O
consists	O
of	O
assigning	O
to	O
each	O
processor	O
a	O
certain	O
number	O
of	O
tasks	B-General_Concept
in	O
a	O
random	O
or	O
predefined	O
manner	O
,	O
then	O
allowing	O
inactive	O
processors	O
to	O
"	O
steal	O
"	O
work	O
from	O
active	O
or	O
overloaded	O
processors	O
.	O
</s>
<s>
In	O
the	O
case	O
of	O
atomic	B-General_Concept
tasks	B-General_Concept
,	O
two	O
main	O
strategies	O
can	O
be	O
distinguished	O
,	O
those	O
where	O
the	O
processors	O
with	O
low	O
load	O
offer	O
their	O
computing	O
capacity	O
to	O
those	O
with	O
the	O
highest	O
load	O
,	O
and	O
those	O
were	O
the	O
most	O
loaded	O
units	O
wish	O
to	O
lighten	O
the	O
workload	O
assigned	O
to	O
them	O
.	O
</s>
<s>
In	O
the	O
case	O
where	O
one	O
starts	O
from	O
a	O
single	O
large	O
task	O
that	O
cannot	O
be	O
divided	O
beyond	O
an	O
atomic	B-General_Concept
level	O
,	O
there	O
is	O
a	O
very	O
efficient	O
algorithm	O
"	O
Tree-Shaped	O
computation	O
"	O
,	O
where	O
the	O
parent	O
task	O
is	O
distributed	O
in	O
a	O
work	O
tree	O
.	O
</s>
<s>
This	O
induces	O
a	O
tree	B-Data_Structure
structure	I-Data_Structure
.	O
</s>
<s>
The	O
efficiency	O
of	O
such	O
an	O
algorithm	O
is	O
close	O
to	O
the	O
prefix	B-Application
sum	I-Application
when	O
the	O
job	O
cutting	O
and	O
communication	O
time	O
is	O
not	O
too	O
high	O
compared	O
to	O
the	O
work	O
to	O
be	O
done	O
.	O
</s>
<s>
To	O
avoid	O
too	O
high	O
communication	O
costs	O
,	O
it	O
is	O
possible	O
to	O
imagine	O
a	O
list	O
of	O
jobs	O
on	O
shared	B-Operating_System
memory	I-Operating_System
.	O
</s>
<s>
Therefore	O
,	O
a	O
request	O
is	O
simply	O
reading	O
from	O
a	O
certain	O
position	O
on	O
this	O
shared	B-Operating_System
memory	I-Operating_System
at	O
the	O
request	O
of	O
the	O
master	B-Operating_System
processor	O
.	O
</s>
<s>
In	O
addition	O
to	O
efficient	O
problem	O
solving	O
through	O
parallel	B-Operating_System
computations	I-Operating_System
,	O
load	B-Application
balancing	I-Application
algorithms	O
are	O
widely	O
used	O
in	O
HTTP	B-Protocol
request	O
management	O
where	O
a	O
site	O
with	O
a	O
large	O
audience	O
must	O
be	O
able	O
to	O
handle	O
a	O
large	O
number	O
of	O
requests	O
per	O
second	O
.	O
</s>
<s>
One	O
of	O
the	O
most	O
commonly	O
used	O
applications	O
of	O
load	B-Application
balancing	I-Application
is	O
to	O
provide	O
single	O
Internet	O
service	O
from	O
multiple	O
servers	O
,	O
sometimes	O
known	O
as	O
a	O
server	B-Operating_System
farm	I-Operating_System
.	O
</s>
<s>
Commonly	O
load-balanced	B-Application
systems	O
include	O
popular	O
web	O
sites	O
,	O
large	O
Internet	B-Protocol
Relay	I-Protocol
Chat	I-Protocol
networks	I-Protocol
,	O
high-bandwidth	O
File	B-Protocol
Transfer	I-Protocol
Protocol	I-Protocol
(	O
FTP	O
)	O
sites	O
,	O
Network	B-Protocol
News	I-Protocol
Transfer	I-Protocol
Protocol	I-Protocol
(	O
NNTP	B-Protocol
)	O
servers	O
,	O
Domain	B-Protocol
Name	I-Protocol
System	I-Protocol
(	O
DNS	B-Protocol
)	O
servers	O
,	O
and	O
databases	O
.	O
</s>
<s>
Round-robin	B-Protocol
DNS	I-Protocol
is	O
an	O
alternate	O
method	O
of	O
load	B-Application
balancing	I-Application
that	O
does	O
not	O
require	O
a	O
dedicated	O
software	O
or	O
hardware	O
node	O
.	O
</s>
<s>
In	O
this	O
technique	O
,	O
multiple	O
IP	B-Protocol
addresses	I-Protocol
are	O
associated	O
with	O
a	O
single	O
domain	O
name	O
;	O
clients	O
are	O
given	O
IP	O
in	O
a	O
round-robin	O
fashion	O
.	O
</s>
<s>
Another	O
more	O
effective	O
technique	O
for	O
load-balancing	O
using	O
DNS	B-Protocol
is	O
to	O
delegate	O
as	O
a	O
sub-domain	O
whose	O
zone	O
is	O
served	O
by	O
each	O
of	O
the	O
same	O
servers	O
that	O
are	O
serving	O
the	O
website	O
.	O
</s>
<s>
However	O
,	O
the	O
zone	O
file	O
for	O
on	O
each	O
server	B-Application
is	O
different	O
such	O
that	O
each	O
server	B-Application
resolves	O
its	O
own	O
IP	B-Protocol
Address	I-Protocol
as	O
the	O
A-record	O
.	O
</s>
<s>
On	O
server	B-Application
one	O
the	O
zone	O
file	O
for	O
reports	O
:	O
</s>
<s>
On	O
server	B-Application
two	O
the	O
same	O
zone	O
file	O
contains	O
:	O
</s>
<s>
This	O
way	O
,	O
when	O
a	O
server	B-Application
is	O
down	O
,	O
its	O
DNS	B-Protocol
will	O
not	O
respond	O
and	O
the	O
web	O
service	O
does	O
not	O
receive	O
any	O
traffic	O
.	O
</s>
<s>
If	O
the	O
line	O
to	O
one	O
server	B-Application
is	O
congested	O
,	O
the	O
unreliability	O
of	O
DNS	B-Protocol
ensures	O
less	O
HTTP	B-Protocol
traffic	O
reaches	O
that	O
server	B-Application
.	O
</s>
<s>
Furthermore	O
,	O
the	O
quickest	O
DNS	B-Protocol
response	O
to	O
the	O
resolver	O
is	O
nearly	O
always	O
the	O
one	O
from	O
the	O
network	O
's	O
closest	O
server	B-Application
,	O
ensuring	O
geo-sensitive	O
load-balancing	O
.	O
</s>
<s>
A	O
short	O
TTL	B-Protocol
on	O
the	O
A-record	O
helps	O
to	O
ensure	O
traffic	O
is	O
quickly	O
diverted	O
when	O
a	O
server	B-Application
goes	O
down	O
.	O
</s>
<s>
Another	O
approach	O
to	O
load	B-Application
balancing	I-Application
is	O
to	O
deliver	O
a	O
list	O
of	O
server	B-Application
IPs	O
to	O
the	O
client	O
,	O
and	O
then	O
to	O
have	O
the	O
client	O
randomly	O
select	O
the	O
IP	O
from	O
the	O
list	O
on	O
each	O
connection	O
.	O
</s>
<s>
This	O
essentially	O
relies	O
on	O
all	O
clients	O
generating	O
similar	O
loads	O
,	O
and	O
the	O
Law	O
of	O
Large	O
Numbers	O
to	O
achieve	O
a	O
reasonably	O
flat	O
load	B-Application
distribution	I-Application
across	O
servers	O
.	O
</s>
<s>
It	O
has	O
been	O
claimed	O
that	O
client-side	O
random	O
load	B-Application
balancing	I-Application
tends	O
to	O
provide	O
better	O
load	B-Application
distribution	I-Application
than	O
round-robin	B-Protocol
DNS	I-Protocol
;	O
this	O
has	O
been	O
attributed	O
to	O
caching	O
issues	O
with	O
round-robin	B-Protocol
DNS	I-Protocol
,	O
that	O
in	O
the	O
case	O
of	O
large	O
DNS	B-Protocol
caching	O
servers	O
,	O
tend	O
to	O
skew	O
the	O
distribution	O
for	O
round-robin	B-Protocol
DNS	I-Protocol
,	O
while	O
client-side	O
random	O
selection	O
remains	O
unaffected	O
regardless	O
of	O
DNS	B-Protocol
caching	O
.	O
</s>
<s>
With	O
this	O
approach	O
,	O
the	O
method	O
of	O
delivery	O
of	O
a	O
list	O
of	O
IPs	O
to	O
the	O
client	O
can	O
vary	O
and	O
may	O
be	O
implemented	O
as	O
a	O
DNS	B-Protocol
list	O
(	O
delivered	O
to	O
all	O
the	O
clients	O
without	O
any	O
round-robin	O
)	O
,	O
or	O
via	O
hardcoding	O
it	O
to	O
the	O
list	O
.	O
</s>
<s>
If	O
a	O
"	O
smart	O
client	O
"	O
is	O
used	O
,	O
detecting	O
that	O
a	O
randomly	O
selected	O
server	B-Application
is	O
down	O
and	O
connecting	O
randomly	O
again	O
,	O
it	O
also	O
provides	O
fault	B-General_Concept
tolerance	I-General_Concept
.	O
</s>
<s>
For	O
Internet	O
services	O
,	O
a	O
server-side	O
load	B-Application
balancer	I-Application
is	O
usually	O
a	O
software	O
program	O
that	O
is	O
listening	O
on	O
the	O
port	B-Protocol
where	O
external	O
clients	O
connect	O
to	O
access	O
services	O
.	O
</s>
<s>
The	O
load	B-Application
balancer	I-Application
forwards	O
requests	O
to	O
one	O
of	O
the	O
"	O
backend	O
"	O
servers	O
,	O
which	O
usually	O
replies	O
to	O
the	O
load	B-Application
balancer	I-Application
.	O
</s>
<s>
This	O
allows	O
the	O
load	B-Application
balancer	I-Application
to	O
reply	O
to	O
the	O
client	O
without	O
the	O
client	O
ever	O
knowing	O
about	O
the	O
internal	O
separation	O
of	O
functions	O
.	O
</s>
<s>
Some	O
load	B-Application
balancers	I-Application
provide	O
a	O
mechanism	O
for	O
doing	O
something	O
special	O
in	O
the	O
event	O
that	O
all	O
backend	O
servers	O
are	O
unavailable	O
.	O
</s>
<s>
This	O
might	O
include	O
forwarding	O
to	O
a	O
backup	O
load	B-Application
balancer	I-Application
or	O
displaying	O
a	O
message	O
regarding	O
the	O
outage	O
.	O
</s>
<s>
It	O
is	O
also	O
important	O
that	O
the	O
load	B-Application
balancer	I-Application
itself	O
does	O
not	O
become	O
a	O
single	B-Architecture
point	I-Architecture
of	I-Architecture
failure	I-Architecture
.	O
</s>
<s>
Usually	O
,	O
load	B-Application
balancers	I-Application
are	O
implemented	O
in	O
high-availability	O
pairs	O
which	O
may	O
also	O
replicate	O
session	O
persistence	O
data	O
if	O
required	O
by	O
the	O
specific	O
application	O
.	O
</s>
<s>
Certain	O
applications	O
are	O
programmed	O
with	O
immunity	O
to	O
this	O
problem	O
,	O
by	O
offsetting	O
the	O
load	B-Application
balancing	I-Application
point	O
over	O
differential	O
sharing	O
platforms	O
beyond	O
the	O
defined	O
network	O
.	O
</s>
<s>
Numerous	O
scheduling	O
algorithms	O
,	O
also	O
called	O
load-balancing	O
methods	O
,	O
are	O
used	O
by	O
load	B-Application
balancers	I-Application
to	O
determine	O
which	O
back-end	O
server	B-Application
to	O
send	O
a	O
request	O
to	O
.	O
</s>
<s>
Simple	O
algorithms	O
include	O
random	O
choice	O
,	O
round	B-Algorithm
robin	I-Algorithm
,	O
or	O
least	O
connections	O
.	O
</s>
<s>
More	O
sophisticated	O
load	B-Application
balancers	I-Application
may	O
take	O
additional	O
factors	O
into	O
account	O
,	O
such	O
as	O
a	O
server	B-Application
's	O
reported	O
load	O
,	O
least	O
response	O
times	O
,	O
up/down	O
status	O
(	O
determined	O
by	O
a	O
monitoring	O
poll	O
of	O
some	O
kind	O
)	O
,	O
a	O
number	O
of	O
active	O
connections	O
,	O
geographic	O
location	O
,	O
capabilities	O
,	O
or	O
how	O
much	O
traffic	O
it	O
has	O
recently	O
been	O
assigned	O
.	O
</s>
<s>
An	O
important	O
issue	O
when	O
operating	O
a	O
load-balanced	B-Application
service	O
is	O
how	O
to	O
handle	O
information	O
that	O
must	O
be	O
kept	O
across	O
the	O
multiple	O
requests	O
in	O
a	O
user	O
's	O
session	O
.	O
</s>
<s>
If	O
this	O
information	O
is	O
stored	O
locally	O
on	O
one	O
backend	O
server	B-Application
,	O
then	O
subsequent	O
requests	O
going	O
to	O
different	O
backend	O
servers	O
would	O
not	O
be	O
able	O
to	O
find	O
it	O
.	O
</s>
<s>
This	O
might	O
be	O
cached	O
information	O
that	O
can	O
be	O
recomputed	O
,	O
in	O
which	O
case	O
load-balancing	O
a	O
request	O
to	O
a	O
different	O
backend	O
server	B-Application
just	O
introduces	O
a	O
performance	O
issue	O
.	O
</s>
<s>
Ideally	O
,	O
the	O
cluster	O
of	O
servers	O
behind	O
the	O
load	B-Application
balancer	I-Application
should	O
not	O
be	O
session-aware	O
,	O
so	O
that	O
if	O
a	O
client	O
connects	O
to	O
any	O
backend	O
server	B-Application
at	O
any	O
time	O
the	O
user	O
experience	O
is	O
unaffected	O
.	O
</s>
<s>
This	O
is	O
usually	O
achieved	O
with	O
a	O
shared	O
database	O
or	O
an	O
in-memory	O
session	O
database	O
like	O
Memcached	B-Application
.	O
</s>
<s>
One	O
basic	O
solution	O
to	O
the	O
session	O
data	O
issue	O
is	O
to	O
send	O
all	O
requests	O
in	O
a	O
user	O
session	O
consistently	O
to	O
the	O
same	O
backend	O
server	B-Application
.	O
</s>
<s>
A	O
significant	O
downside	O
to	O
this	O
technique	O
is	O
its	O
lack	O
of	O
automatic	O
failover	B-General_Concept
:	O
if	O
a	O
backend	O
server	B-Application
goes	O
down	O
,	O
its	O
per-session	O
information	O
becomes	O
inaccessible	O
,	O
and	O
any	O
sessions	O
depending	O
on	O
it	O
are	O
lost	O
.	O
</s>
<s>
Assignment	O
to	O
a	O
particular	O
server	B-Application
might	O
be	O
based	O
on	O
a	O
username	O
,	O
client	O
IP	B-Protocol
address	I-Protocol
,	O
or	O
random	O
.	O
</s>
<s>
Because	O
of	O
changes	O
in	O
the	O
client	O
's	O
perceived	O
address	O
resulting	O
from	O
DHCP	B-Protocol
,	O
network	B-Protocol
address	I-Protocol
translation	I-Protocol
,	O
and	O
web	O
proxies	O
this	O
method	O
may	O
be	O
unreliable	O
.	O
</s>
<s>
Random	O
assignments	O
must	O
be	O
remembered	O
by	O
the	O
load	B-Application
balancer	I-Application
,	O
which	O
creates	O
a	O
burden	O
on	O
storage	O
.	O
</s>
<s>
If	O
the	O
load	B-Application
balancer	I-Application
is	O
replaced	O
or	O
fails	O
,	O
this	O
information	O
may	O
be	O
lost	O
,	O
and	O
assignments	O
may	O
need	O
to	O
be	O
deleted	O
after	O
a	O
timeout	O
period	O
or	O
during	O
periods	O
of	O
high	O
load	O
to	O
avoid	O
exceeding	O
the	O
space	O
available	O
for	O
the	O
assignment	O
table	O
.	O
</s>
<s>
The	O
random	O
assignment	O
method	O
also	O
requires	O
that	O
clients	O
maintain	O
some	O
state	O
,	O
which	O
can	O
be	O
a	O
problem	O
,	O
for	O
example	O
when	O
a	O
web	O
browser	O
has	O
disabled	O
the	O
storage	O
of	O
cookies	B-Application
.	O
</s>
<s>
Sophisticated	O
load	B-Application
balancers	I-Application
use	O
multiple	O
persistence	O
techniques	O
to	O
avoid	O
some	O
of	O
the	O
shortcomings	O
of	O
any	O
one	O
method	O
.	O
</s>
<s>
To	O
prevent	O
a	O
database	O
from	O
becoming	O
a	O
single	B-Architecture
point	I-Architecture
of	I-Architecture
failure	I-Architecture
,	O
and	O
to	O
improve	O
scalability	B-Architecture
,	O
the	O
database	O
is	O
often	O
replicated	O
across	O
multiple	O
machines	O
,	O
and	O
load	B-Application
balancing	I-Application
is	O
used	O
to	O
spread	O
the	O
query	O
load	O
across	O
those	O
replicas	O
.	O
</s>
<s>
Microsoft	O
's	O
ASP.net	B-Application
State	O
Server	B-Application
technology	O
is	O
an	O
example	O
of	O
a	O
session	O
database	O
.	O
</s>
<s>
All	O
servers	O
in	O
a	O
web	B-Operating_System
farm	I-Operating_System
store	O
their	O
session	O
data	O
on	O
State	O
Server	B-Application
and	O
any	O
server	B-Application
in	O
the	O
farm	O
can	O
retrieve	O
the	O
data	O
.	O
</s>
<s>
One	O
way	O
to	O
achieve	O
this	O
is	O
to	O
use	O
a	O
browser	B-Application
cookie	I-Application
,	O
suitably	O
time-stamped	O
and	O
encrypted	O
.	O
</s>
<s>
Another	O
is	O
URL	B-Protocol
rewriting	I-Protocol
.	O
</s>
<s>
Storing	O
session	O
data	O
on	O
the	O
client	O
is	O
generally	O
the	O
preferred	O
solution	O
:	O
then	O
the	O
load	B-Application
balancer	I-Application
is	O
free	O
to	O
pick	O
any	O
backend	O
server	B-Application
to	O
handle	O
a	O
request	O
.	O
</s>
<s>
However	O
,	O
this	O
method	O
of	O
state-data	O
handling	O
is	O
poorly	O
suited	O
to	O
some	O
complex	O
business	O
logic	O
scenarios	O
,	O
where	O
session	O
state	O
payload	O
is	O
big	O
and	O
recomputing	O
it	O
with	O
every	O
request	O
on	O
a	O
server	B-Application
is	O
not	O
feasible	O
.	O
</s>
<s>
URL	B-Protocol
rewriting	I-Protocol
has	O
major	O
security	O
issues	O
because	O
the	O
end-user	O
can	O
easily	O
alter	O
the	O
submitted	O
URL	O
and	O
thus	O
change	O
session	O
streams	O
.	O
</s>
<s>
Yet	O
another	O
solution	O
to	O
storing	O
persistent	O
data	O
is	O
to	O
associate	O
a	O
name	O
with	O
each	O
block	O
of	O
data	O
,	O
and	O
use	O
a	O
distributed	B-Operating_System
hash	I-Operating_System
table	I-Operating_System
to	O
pseudo-randomly	O
assign	O
that	O
name	O
to	O
one	O
of	O
the	O
available	O
servers	O
,	O
and	O
then	O
store	O
that	O
block	O
of	O
data	O
in	O
the	O
assigned	O
server	B-Application
.	O
</s>
<s>
Hardware	O
and	O
software	O
load	B-Application
balancers	I-Application
may	O
have	O
a	O
variety	O
of	O
special	O
features	O
.	O
</s>
<s>
The	O
fundamental	O
feature	O
of	O
a	O
load	B-Application
balancer	I-Application
is	O
to	O
be	O
able	O
to	O
distribute	O
incoming	O
requests	O
over	O
a	O
number	O
of	O
backend	O
servers	O
in	O
the	O
cluster	O
according	O
to	O
a	O
scheduling	O
algorithm	O
.	O
</s>
<s>
TLS	B-Protocol
(	O
or	O
its	O
predecessor	O
SSL	O
)	O
acceleration	O
is	O
a	O
technique	O
of	O
offloading	O
cryptographic	O
protocol	O
calculations	O
onto	O
specialized	O
hardware	O
.	O
</s>
<s>
Depending	O
on	O
the	O
workload	O
,	O
processing	O
the	O
encryption	O
and	O
authentication	O
requirements	O
of	O
a	O
TLS	B-Protocol
request	O
can	O
become	O
a	O
major	O
part	O
of	O
the	O
demand	O
on	O
the	O
Web	O
Server	B-Application
's	O
CPU	O
;	O
as	O
the	O
demand	O
increases	O
,	O
users	O
will	O
see	O
slower	O
response	O
times	O
,	O
as	O
the	O
TLS	B-Protocol
overhead	O
is	O
distributed	O
among	O
Web	O
servers	O
.	O
</s>
<s>
To	O
remove	O
this	O
demand	O
on	O
Web	O
servers	O
,	O
a	O
balancer	O
can	O
terminate	O
TLS	B-Protocol
connections	O
,	O
passing	O
HTTPS	B-Protocol
requests	O
as	O
HTTP	B-Protocol
requests	I-Protocol
to	O
the	O
Web	O
servers	O
.	O
</s>
<s>
The	O
downside	O
of	O
this	O
approach	O
is	O
that	O
all	O
of	O
the	O
TLS	B-Protocol
processing	O
is	O
concentrated	O
on	O
a	O
single	O
device	O
(	O
the	O
balancer	O
)	O
which	O
can	O
become	O
a	O
new	O
bottleneck	O
.	O
</s>
<s>
Some	O
load	B-Application
balancer	I-Application
appliances	O
include	O
specialized	O
hardware	O
to	O
process	O
TLS	B-Protocol
.	O
</s>
<s>
Instead	O
of	O
upgrading	O
the	O
load	B-Application
balancer	I-Application
,	O
which	O
is	O
quite	O
expensive	O
dedicated	O
hardware	O
,	O
it	O
may	O
be	O
cheaper	O
to	O
forgo	O
TLS	B-Protocol
offload	O
and	O
add	O
a	O
few	O
web	O
servers	O
.	O
</s>
<s>
Also	O
,	O
some	O
server	B-Application
vendors	O
such	O
as	O
Oracle/Sun	O
now	O
incorporate	O
cryptographic	O
acceleration	O
hardware	O
into	O
their	O
CPUs	O
such	O
as	O
the	O
T2000	O
.	O
</s>
<s>
F5	O
Networks	O
incorporates	O
a	O
dedicated	O
TLS	B-Protocol
acceleration	I-Protocol
hardware	O
card	O
in	O
their	O
local	O
traffic	O
manager	O
(	O
LTM	O
)	O
which	O
is	O
used	O
for	O
encrypting	O
and	O
decrypting	O
TLS	B-Protocol
traffic	O
.	O
</s>
<s>
One	O
clear	O
benefit	O
to	O
TLS	B-Protocol
offloading	O
in	O
the	O
balancer	O
is	O
that	O
it	O
enables	O
it	O
to	O
do	O
balancing	O
or	O
content	O
switching	O
based	O
on	O
data	O
in	O
the	O
HTTPS	B-Protocol
request	O
.	O
</s>
<s>
Load	B-Application
balancers	I-Application
can	O
provide	O
features	O
such	O
as	O
SYN	B-Protocol
cookies	I-Protocol
and	O
delayed-binding	O
(	O
the	O
back-end	O
servers	O
do	O
n't	O
see	O
the	O
client	O
until	O
it	O
finishes	O
its	O
TCP	O
handshake	O
)	O
to	O
mitigate	O
SYN	O
flood	O
attacks	O
and	O
generally	O
offload	O
work	O
from	O
the	O
servers	O
to	O
a	O
more	O
efficient	O
platform	O
.	O
</s>
<s>
HTTP	B-Protocol
compression	I-Protocol
reduces	O
the	O
amount	O
of	O
data	O
to	O
be	O
transferred	O
for	O
HTTP	B-Protocol
objects	O
by	O
utilising	O
gzip	O
compression	O
available	O
in	O
all	O
modern	O
web	O
browsers	O
.	O
</s>
<s>
The	O
trade-off	O
is	O
that	O
this	O
feature	O
puts	O
additional	O
CPU	O
demand	O
on	O
the	O
load	B-Application
balancer	I-Application
and	O
could	O
be	O
done	O
by	O
web	O
servers	O
instead	O
.	O
</s>
<s>
Different	O
vendors	O
use	O
different	O
terms	O
for	O
this	O
,	O
but	O
the	O
idea	O
is	O
that	O
normally	O
each	O
HTTP	B-Protocol
request	O
from	O
each	O
client	O
is	O
a	O
different	O
TCP	O
connection	O
.	O
</s>
<s>
This	O
feature	O
utilises	O
HTTP/1.1	B-Protocol
to	O
consolidate	O
multiple	O
HTTP	B-Protocol
requests	I-Protocol
from	O
multiple	O
clients	O
into	O
a	O
single	O
TCP	O
socket	O
to	O
the	O
back-end	O
servers	O
.	O
</s>
<s>
The	O
load	B-Application
balancer	I-Application
can	O
buffer	O
responses	O
from	O
the	O
server	B-Application
and	O
spoon-feed	O
the	O
data	O
out	O
to	O
slow	O
clients	O
,	O
allowing	O
the	O
webserver	O
to	O
free	O
a	O
thread	O
for	O
other	O
tasks	B-General_Concept
faster	O
than	O
it	O
would	O
if	O
it	O
had	O
to	O
send	O
the	O
entire	O
request	O
to	O
the	O
client	O
directly	O
.	O
</s>
<s>
An	O
option	O
for	O
asymmetrical	O
load	B-Application
distribution	I-Application
,	O
where	O
request	O
and	O
reply	O
have	O
different	O
network	O
paths	O
.	O
</s>
<s>
Some	O
balancers	O
can	O
hide	O
HTTP	B-Protocol
error	O
pages	O
,	O
remove	O
server	B-Application
identification	O
headers	O
from	O
HTTP	B-Protocol
responses	O
,	O
and	O
encrypt	O
cookies	B-Application
so	O
that	O
end	O
users	O
cannot	O
manipulate	O
them	O
.	O
</s>
<s>
Most	O
load	B-Application
balancers	I-Application
can	O
send	O
requests	O
to	O
different	O
servers	O
based	O
on	O
the	O
URL	O
being	O
requested	O
,	O
assuming	O
the	O
request	O
is	O
not	O
encrypted	O
(	O
HTTP	B-Protocol
)	O
or	O
if	O
it	O
is	O
encrypted	O
(	O
via	O
HTTPS	B-Protocol
)	O
that	O
the	O
HTTPS	B-Protocol
request	O
is	O
terminated	O
(	O
decrypted	O
)	O
at	O
the	O
load	B-Application
balancer	I-Application
.	O
</s>
<s>
Firewalls	B-Application
can	O
prevent	O
direct	O
connections	O
to	O
backend	O
servers	O
,	O
for	O
network	O
security	O
reasons	O
.	O
</s>
<s>
Intrusion	O
prevention	O
systems	O
offer	O
application	O
layer	O
security	O
in	O
addition	O
to	O
the	O
network/transport	O
layer	O
offered	O
by	O
firewall	B-Application
security	O
.	O
</s>
<s>
Load	B-Application
balancing	I-Application
can	O
be	O
useful	O
in	O
applications	O
with	O
redundant	O
communications	O
links	O
.	O
</s>
<s>
A	O
failover	B-General_Concept
arrangement	O
would	O
mean	O
that	O
one	O
link	O
is	O
designated	O
for	O
normal	O
use	O
,	O
while	O
the	O
second	O
link	O
is	O
used	O
only	O
if	O
the	O
primary	O
link	O
fails	O
.	O
</s>
<s>
Using	O
load	B-Application
balancing	I-Application
,	O
both	O
links	O
can	O
be	O
in	O
use	O
all	O
the	O
time	O
.	O
</s>
<s>
TRILL	B-Protocol
(	O
TRansparent	B-Protocol
Interconnection	I-Protocol
of	I-Protocol
Lots	I-Protocol
of	I-Protocol
Links	I-Protocol
)	O
facilitates	O
an	O
Ethernet	O
to	O
have	O
an	O
arbitrary	O
topology	O
,	O
and	O
enables	O
per	O
flow	O
pair-wise	O
load	O
splitting	O
by	O
way	O
of	O
Dijkstra	B-Algorithm
's	I-Algorithm
algorithm	I-Algorithm
,	O
without	O
configuration	O
and	O
user	O
intervention	O
.	O
</s>
<s>
The	O
catalyst	O
for	O
TRILL	B-Protocol
was	O
an	O
event	O
at	O
Beth	O
Israel	O
Deaconess	O
Medical	O
Center	O
which	O
began	O
on	O
13	O
November	O
2002	O
.	O
</s>
<s>
The	O
concept	O
of	O
Rbridges	O
 [ sic ] 	O
was	O
first	O
proposed	O
to	O
the	O
Institute	O
of	O
Electrical	O
and	O
Electronics	O
Engineers	O
in	O
the	O
year	O
2004	O
,	O
whom	O
in	O
2005	O
rejected	O
what	O
came	O
to	O
be	O
known	O
as	O
TRILL	B-Protocol
,	O
and	O
in	O
the	O
years	O
2006	O
through	O
2012	O
devised	O
an	O
incompatible	O
variation	O
known	O
as	O
Shortest	B-Protocol
Path	I-Protocol
Bridging	I-Protocol
.	O
</s>
<s>
The	O
IEEE	O
approved	O
the	O
IEEE	B-Protocol
802.1aq	I-Protocol
standard	O
in	O
May	O
2012	O
,	O
also	O
known	O
as	O
Shortest	B-Protocol
Path	I-Protocol
Bridging	I-Protocol
(	O
SPB	O
)	O
.	O
</s>
<s>
SPB	O
allows	O
all	O
links	O
to	O
be	O
active	O
through	O
multiple	O
equal-cost	O
paths	O
,	O
provides	O
faster	O
convergence	O
times	O
to	O
reduce	O
downtime	O
,	O
and	O
simplifies	O
the	O
use	O
of	O
load	B-Application
balancing	I-Application
in	O
mesh	O
network	O
topologies	O
(	O
partially	O
connected	O
and/or	O
fully	O
connected	O
)	O
by	O
allowing	O
traffic	O
to	O
load	O
share	O
across	O
all	O
paths	O
of	O
a	O
network	O
.	O
</s>
<s>
They	O
use	O
sophisticated	O
load	B-Application
balancing	I-Application
to	O
shift	O
traffic	O
from	O
one	O
path	O
to	O
another	O
to	O
avoid	O
network	B-Protocol
congestion	I-Protocol
on	O
any	O
particular	O
link	O
,	O
and	O
sometimes	O
to	O
minimize	O
the	O
cost	O
of	O
transit	O
across	O
external	O
networks	O
or	O
improve	O
network	B-Protocol
reliability	I-Protocol
.	O
</s>
<s>
Another	O
way	O
of	O
using	O
load	B-Application
balancing	I-Application
is	O
in	O
network	O
monitoring	O
activities	O
.	O
</s>
<s>
Load	B-Application
balancers	I-Application
can	O
be	O
used	O
to	O
split	O
huge	O
data	O
flows	O
into	O
several	O
sub-flows	O
and	O
use	O
several	O
network	O
analyzers	O
,	O
each	O
reading	O
a	O
part	O
of	O
the	O
original	O
data	O
.	O
</s>
<s>
Load	B-Application
balancing	I-Application
is	O
widely	O
used	O
in	O
data	B-Operating_System
center	I-Operating_System
networks	O
to	O
distribute	O
traffic	O
across	O
many	O
existing	O
paths	O
between	O
any	O
two	O
servers	O
.	O
</s>
<s>
In	O
general	O
,	O
load	B-Application
balancing	I-Application
in	O
datacenter	B-Operating_System
networks	O
can	O
be	O
classified	O
as	O
either	O
static	O
or	O
dynamic	O
.	O
</s>
<s>
Static	O
load	B-Application
balancing	I-Application
distributes	O
traffic	O
by	O
computing	O
a	O
hash	O
of	O
the	O
source	O
and	O
destination	O
addresses	O
and	O
port	B-Protocol
numbers	I-Protocol
of	O
traffic	O
flows	O
and	O
using	O
it	O
to	O
determine	O
how	O
flows	O
are	O
assigned	O
to	O
one	O
of	O
the	O
existing	O
paths	O
.	O
</s>
<s>
Dynamic	O
load	B-Application
balancing	I-Application
assigns	O
traffic	O
flows	O
to	O
paths	O
by	O
monitoring	O
bandwidth	O
use	O
on	O
different	O
paths	O
.	O
</s>
<s>
A	O
comprehensive	O
overview	O
of	O
load	B-Application
balancing	I-Application
in	O
datacenter	B-Operating_System
networks	O
has	O
been	O
made	O
available	O
.	O
</s>
<s>
Load	B-Application
balancing	I-Application
is	O
often	O
used	O
to	O
implement	O
failover	B-General_Concept
—	O
the	O
continuation	O
of	O
service	O
after	O
the	O
failure	O
of	O
one	O
or	O
more	O
of	O
its	O
components	O
.	O
</s>
<s>
The	O
components	O
are	O
monitored	O
continually	O
(	O
e.g.	O
,	O
web	O
servers	O
may	O
be	O
monitored	O
by	O
fetching	O
known	O
pages	O
)	O
,	O
and	O
when	O
one	O
becomes	O
unresponsive	O
,	O
the	O
load	B-Application
balancer	I-Application
is	O
informed	O
and	O
no	O
longer	O
sends	O
traffic	O
to	O
it	O
.	O
</s>
<s>
When	O
a	O
component	O
comes	O
back	O
online	O
,	O
the	O
load	B-Application
balancer	I-Application
starts	O
rerouting	O
traffic	O
to	O
it	O
.	O
</s>
<s>
For	O
this	O
to	O
work	O
,	O
there	O
must	O
be	O
at	O
least	O
one	O
component	O
in	O
excess	O
of	O
the	O
service	O
's	O
capacity	O
(	O
N+1	B-Application
redundancy	I-Application
)	O
.	O
</s>
<s>
This	O
can	O
be	O
much	O
less	O
expensive	O
and	O
more	O
flexible	O
than	O
failover	B-General_Concept
approaches	O
where	O
every	O
single	O
live	O
component	O
is	O
paired	O
with	O
a	O
single	O
backup	O
component	O
that	O
takes	O
over	O
in	O
the	O
event	O
of	O
a	O
failure	O
(	O
dual	B-Error_Name
modular	I-Error_Name
redundancy	I-Error_Name
)	O
.	O
</s>
<s>
Some	O
RAID	B-Architecture
systems	O
can	O
also	O
utilize	O
hot	B-General_Concept
spare	I-General_Concept
for	O
a	O
similar	O
effect	O
.	O
</s>
