<s>
In	O
a	O
distributed	B-Architecture
computing	I-Architecture
system	B-General_Concept
,	O
a	O
failure	B-Operating_System
detector	I-Operating_System
is	O
a	O
computer	B-Application
application	I-Application
or	O
a	O
subsystem	B-General_Concept
that	O
is	O
responsible	O
for	O
the	O
detection	O
of	O
node	B-Protocol
failures	B-General_Concept
or	O
crashes	B-General_Concept
.	O
</s>
<s>
Failure	B-Operating_System
detectors	I-Operating_System
were	O
first	O
introduced	O
in	O
1996	O
by	O
Chandra	O
and	O
Toueg	O
in	O
their	O
book	O
Unreliable	O
Failure	B-Operating_System
Detectors	I-Operating_System
for	O
Reliable	O
Distributed	B-Architecture
Systems	I-Architecture
.	O
</s>
<s>
The	O
book	O
depicts	O
the	O
failure	B-Operating_System
detector	I-Operating_System
as	O
a	O
tool	O
to	O
improve	O
consensus	B-Operating_System
(	O
the	O
achievement	O
of	O
reliability	B-Protocol
)	O
and	O
atomic	B-Operating_System
broadcast	I-Operating_System
(	O
the	O
same	O
sequence	O
of	O
messages	O
)	O
in	O
the	O
distributed	B-Architecture
system	I-Architecture
.	O
</s>
<s>
In	O
other	O
words	O
,	O
failure	B-Operating_System
detectors	I-Operating_System
seek	O
errors	O
in	O
the	O
process	B-Operating_System
,	O
and	O
the	O
system	B-General_Concept
will	O
maintain	O
a	O
level	O
of	O
reliability	B-Protocol
.	O
</s>
<s>
In	O
practice	O
,	O
after	O
failure	B-Operating_System
detectors	I-Operating_System
spot	O
crashes	B-General_Concept
,	O
the	O
system	B-General_Concept
will	O
ban	O
the	O
processes	O
that	O
are	O
making	O
mistakes	O
to	O
prevent	O
any	O
further	O
serious	O
crashes	B-General_Concept
or	O
errors	O
.	O
</s>
<s>
In	O
the	O
21st	O
century	O
,	O
failure	B-Operating_System
detectors	I-Operating_System
are	O
widely	O
used	O
in	O
distributed	B-Architecture
computing	I-Architecture
systems	O
to	O
detect	O
application	B-Error_Name
errors	I-Error_Name
,	O
such	O
as	O
a	O
software	B-Application
application	I-Application
stops	O
functioning	O
properly	O
.	O
</s>
<s>
As	O
the	O
distributed	B-Operating_System
computing	I-Operating_System
projects	I-Operating_System
(	O
see	O
List	B-Operating_System
of	I-Operating_System
distributed	I-Operating_System
computing	I-Operating_System
projects	I-Operating_System
)	O
become	O
more	O
and	O
more	O
popular	O
,	O
the	O
usage	O
of	O
the	O
failure	O
detects	O
also	O
becomes	O
important	O
and	O
critical	O
.	O
</s>
<s>
Chandra	O
and	O
Toueg	O
,	O
the	O
co-authors	O
of	O
the	O
book	O
Unreliable	O
Failure	B-Operating_System
Detectors	I-Operating_System
for	O
Reliable	O
Distributed	B-Architecture
Systems	I-Architecture
(	O
1996	O
)	O
,	O
approached	O
the	O
concept	O
of	O
detecting	O
failure	O
nodes	O
by	O
introducing	O
the	O
unreliable	O
failure	B-Operating_System
detector	I-Operating_System
.	O
</s>
<s>
They	O
describe	O
the	O
behavior	O
of	O
a	O
unreliable	O
failure	B-Operating_System
detector	I-Operating_System
in	O
a	O
distributed	B-Architecture
computing	I-Architecture
system	B-General_Concept
as	O
:	O
after	O
each	O
process	B-Operating_System
in	O
the	O
system	B-General_Concept
entered	O
a	O
local	O
failure	B-Operating_System
detector	I-Operating_System
component	B-Architecture
,	O
each	O
local	O
component	B-Architecture
will	O
examine	O
a	O
portion	O
of	O
all	O
processes	O
within	O
the	O
system	B-General_Concept
.	O
</s>
<s>
In	O
addition	O
,	O
each	O
process	B-Operating_System
must	O
also	O
contain	O
programs	B-Application
that	O
are	O
currently	O
suspected	O
by	O
failure	B-Operating_System
detectors	I-Operating_System
.	O
</s>
<s>
Chandra	O
and	O
Toueg	O
claimed	O
that	O
an	O
unreliable	O
failure	B-Operating_System
detector	I-Operating_System
can	O
still	O
be	O
reliable	O
in	O
detecting	O
the	O
errors	O
made	O
by	O
the	O
system	B-General_Concept
.	O
</s>
<s>
They	O
generalize	O
unreliable	O
failure	B-Operating_System
detectors	I-Operating_System
to	O
all	O
forms	O
of	O
failure	B-Operating_System
detectors	I-Operating_System
because	O
unreliable	O
failure	B-Operating_System
detectors	I-Operating_System
and	O
failure	B-Operating_System
detectors	I-Operating_System
share	O
the	O
same	O
properties	O
.	O
</s>
<s>
Furthermore	O
,	O
Chandra	O
and	O
Toueg	O
point	O
out	O
an	O
important	O
fact	O
that	O
the	O
failure	B-Operating_System
detector	I-Operating_System
does	O
not	O
prevent	O
any	O
crashes	B-General_Concept
in	O
the	O
system	B-General_Concept
,	O
even	O
if	O
the	O
crashed	O
program	O
has	O
been	O
suspected	O
previously	O
.	O
</s>
<s>
The	O
construction	O
of	O
a	O
failure	B-Operating_System
detector	I-Operating_System
is	O
an	O
essential	O
,	O
but	O
a	O
very	O
difficult	O
problem	O
that	O
occurred	O
in	O
the	O
development	O
of	O
the	O
fault-tolerant	B-General_Concept
component	B-Architecture
in	O
a	O
distributed	B-Architecture
computer	I-Architecture
system	I-Architecture
.	O
</s>
<s>
As	O
a	O
result	O
,	O
the	O
failure	B-Operating_System
detector	I-Operating_System
was	O
invented	O
because	O
of	O
the	O
need	O
for	O
detecting	O
errors	O
in	O
the	O
massive	O
information	O
transaction	O
in	O
distributed	B-Architecture
computing	I-Architecture
systems	O
.	O
</s>
<s>
The	O
classes	O
of	O
failure	B-Operating_System
detectors	I-Operating_System
are	O
distinguished	O
by	O
two	O
important	O
properties	O
:	O
completeness	O
and	O
accuracy	O
.	O
</s>
<s>
Completeness	O
means	O
that	O
the	O
failure	B-Operating_System
detectors	I-Operating_System
would	O
find	O
the	O
programs	B-Application
that	O
finally	O
crashed	O
in	O
a	O
process	B-Operating_System
,	O
whereas	O
accuracy	O
means	O
that	O
correct	O
decisions	O
that	O
the	O
failure	B-Operating_System
detectors	I-Operating_System
made	O
in	O
a	O
process	B-Operating_System
.	O
</s>
<s>
The	O
degrees	O
of	O
completeness	O
depend	O
on	O
the	O
number	O
of	O
crashed	O
process	B-Operating_System
is	O
suspected	O
by	O
a	O
failure	B-Operating_System
detector	I-Operating_System
in	O
a	O
certain	O
period	O
.	O
</s>
<s>
Strong	O
completeness	O
:	O
"	O
every	O
faulty	B-General_Concept
process	B-Operating_System
is	O
eventually	O
permanently	O
suspected	O
by	O
every	O
non-faulty	O
process.	O
"	O
</s>
<s>
Weak	O
completeness	O
:	O
"	O
every	O
faulty	B-General_Concept
process	B-Operating_System
is	O
eventually	O
permanently	O
suspected	O
by	O
some	O
non-faulty	O
process.	O
"	O
</s>
<s>
The	O
degrees	O
of	O
accuracy	O
depend	O
on	O
the	O
number	O
of	O
mistakes	O
that	O
a	O
failure	B-Operating_System
detector	I-Operating_System
made	O
in	O
a	O
certain	O
period	O
.	O
</s>
<s>
Strong	O
accuracy	O
:	O
"	O
no	O
process	B-Operating_System
is	O
suspected	O
(	O
by	O
anybody	O
)	O
before	O
it	O
crashes.	O
"	O
</s>
<s>
Weak	O
accuracy	O
:	O
"	O
some	O
non-faulty	O
process	B-Operating_System
is	O
never	O
suspected.	O
"	O
</s>
<s>
Eventual	O
strong	O
accuracy	O
:	O
"	O
no	O
non-faulty	O
process	B-Operating_System
is	O
suspected	O
after	O
some	O
time	O
since	O
the	O
end	O
of	O
the	O
initial	O
period	O
of	O
chaos	O
as	O
the	O
time	O
at	O
which	O
the	O
last	O
crash	B-General_Concept
occurs.	O
"	O
</s>
<s>
Eventual	O
weak	O
accuracy	O
:	O
"	O
after	O
some	O
initial	O
period	O
of	O
confusion	O
,	O
some	O
non-faulty	O
process	B-Operating_System
is	O
never	O
suspected.	O
"	O
</s>
<s>
Failure	B-Operating_System
detectors	I-Operating_System
can	O
be	O
categorized	O
in	O
the	O
following	O
eight	O
types	O
:	O
</s>
<s>
The	O
properties	O
of	O
these	O
failure	B-Operating_System
detectors	I-Operating_System
are	O
described	O
below	O
:	O
</s>
<s>
In	O
a	O
nutshell	O
,	O
the	O
properties	O
of	O
failure	B-Operating_System
detectors	I-Operating_System
depend	O
on	O
how	O
fast	O
the	O
failure	B-Operating_System
detector	I-Operating_System
detects	O
actual	O
failures	B-General_Concept
and	O
how	O
well	O
it	O
avoids	O
false	O
detection	O
.	O
</s>
<s>
A	O
perfect	O
failure	B-Operating_System
detector	I-Operating_System
will	O
find	O
all	O
errors	O
without	O
any	O
mistakes	O
,	O
whereas	O
a	O
weak	O
failure	B-Operating_System
detector	I-Operating_System
will	O
not	O
find	O
any	O
errors	O
and	O
make	O
numerous	O
mistakes	O
.	O
</s>
<s>
Different	O
types	O
of	O
failure	B-Operating_System
detectors	I-Operating_System
can	O
be	O
obtained	O
by	O
changing	O
the	O
properties	O
of	O
failure	B-Operating_System
detectors	I-Operating_System
.	O
</s>
<s>
The	O
first	O
examples	O
show	O
that	O
how	O
to	O
increase	O
completeness	O
of	O
a	O
failure	B-Operating_System
detector	I-Operating_System
,	O
and	O
the	O
second	O
example	O
shows	O
that	O
how	O
to	O
change	O
one	O
type	O
of	O
the	O
failure	B-Operating_System
detector	I-Operating_System
to	O
another	O
.	O
</s>
<s>
It	O
functions	O
by	O
boosting	O
the	O
completeness	O
of	O
a	O
failure	B-Operating_System
detector	I-Operating_System
.	O
</s>
<s>
From	O
the	O
example	O
above	O
,	O
if	O
p	O
crashes	B-General_Concept
,	O
then	O
the	O
weak-detector	O
will	O
eventually	O
suspect	O
it	O
.	O
</s>
<s>
All	O
failure	B-Operating_System
detectors	I-Operating_System
in	O
the	O
system	B-General_Concept
will	O
eventually	O
suspect	O
the	O
p	O
because	O
of	O
the	O
infinite	O
loop	O
created	O
by	O
failure	B-Operating_System
detectors	I-Operating_System
.	O
</s>
<s>
This	O
example	O
also	O
shows	O
that	O
a	O
weak	O
completeness	O
failure	B-Operating_System
detector	I-Operating_System
can	O
also	O
suspect	O
all	O
crashes	B-General_Concept
eventually	O
.	O
</s>
<s>
The	O
inspection	O
of	O
crashed	O
programs	B-Application
does	O
not	O
depend	O
on	O
completeness	O
.	O
</s>
<s>
The	O
following	O
are	O
correctness	O
arguments	O
to	O
satisfy	O
the	O
algorithm	O
of	O
changing	O
a	O
failure	B-Operating_System
detector	I-Operating_System
W	O
to	O
a	O
failure	B-Operating_System
detector	I-Operating_System
S	O
.	O
The	O
failure	B-Operating_System
detector	I-Operating_System
W	O
is	O
weak	O
in	O
completeness	O
,	O
and	O
the	O
failure	B-Operating_System
detector	I-Operating_System
S	O
is	O
strong	O
in	O
completeness	O
.	O
</s>
<s>
If	O
all	O
arguments	O
above	O
are	O
satisfied	O
,	O
the	O
reduction	O
of	O
a	O
weak	O
failure	B-Operating_System
detector	I-Operating_System
W	O
to	O
a	O
strong	O
failure	B-Operating_System
detector	I-Operating_System
S	O
will	O
agree	O
with	O
the	O
algorithm	O
within	O
the	O
distributed	B-Architecture
computing	I-Architecture
system	B-General_Concept
.	O
</s>
