<s>
The	O
replay	B-Device
system	I-Device
is	O
a	O
subsystem	O
within	O
the	O
Intel	B-General_Concept
Pentium	I-General_Concept
4	I-General_Concept
processor	O
.	O
</s>
<s>
Operations	O
caught	O
by	O
the	O
replay	B-Device
system	I-Device
are	O
then	O
re-executed	O
in	O
a	O
loop	O
until	O
the	O
conditions	O
necessary	O
for	O
their	O
proper	O
execution	O
have	O
been	O
fulfilled	O
.	O
</s>
<s>
The	O
replay	B-Device
system	I-Device
came	O
about	O
as	O
a	O
result	O
of	O
Intel	O
's	O
quest	O
for	O
ever-increasing	O
clock	O
speeds	O
.	O
</s>
<s>
These	O
higher	O
clock	O
speeds	O
necessitated	O
very	O
lengthy	O
pipelines	B-General_Concept
(	O
up	O
to	O
31	O
stages	O
in	O
the	O
Prescott	O
core	O
)	O
.	O
</s>
<s>
Because	O
of	O
this	O
,	O
there	O
are	O
six	O
stages	O
between	O
the	O
scheduler	O
and	O
the	O
execution	B-General_Concept
units	I-General_Concept
in	O
the	O
Prescott	O
core	O
.	O
</s>
<s>
The	O
scheduler	O
in	O
a	O
Pentium4	B-General_Concept
processor	O
is	O
so	O
aggressive	O
that	O
it	O
will	O
send	O
operations	O
for	O
execution	O
without	O
a	O
guarantee	O
that	O
they	O
can	O
be	O
successfully	O
executed	O
.	O
</s>
<s>
(	O
Among	O
other	O
things	O
,	O
the	O
scheduler	O
assumes	O
all	O
data	O
is	O
in	O
level	O
1	O
"	O
trace	B-General_Concept
cache	I-General_Concept
"	O
CPU	B-General_Concept
cache	I-General_Concept
.	O
)	O
</s>
<s>
When	O
this	O
happens	O
,	O
the	O
replay	B-Device
system	I-Device
signals	O
the	O
scheduler	O
to	O
stop	O
,	O
then	O
repeatedly	O
executes	O
the	O
failed	O
string	O
of	O
dependent	O
operations	O
until	O
they	O
have	O
completed	O
successfully	O
.	O
</s>
<s>
Not	O
surprisingly	O
,	O
in	O
some	O
cases	O
the	O
replay	B-Device
system	I-Device
can	O
have	O
a	O
very	O
bad	O
impact	O
on	O
performance	O
.	O
</s>
<s>
Under	O
normal	O
circumstances	O
,	O
the	O
execution	B-General_Concept
units	I-General_Concept
in	O
the	O
Pentium	B-General_Concept
4	I-General_Concept
are	O
in	O
use	O
roughly	O
33%	O
of	O
the	O
time	O
.	O
</s>
<s>
When	O
the	O
replay	B-Device
system	I-Device
is	O
invoked	O
,	O
it	O
will	O
occupy	O
execution	B-General_Concept
units	I-General_Concept
nearly	O
every	O
available	O
cycle	O
.	O
</s>
<s>
This	O
wastes	O
power	O
,	O
which	O
is	O
an	O
increasingly	O
important	O
architectural	O
design	O
metric	O
,	O
but	O
poses	O
no	O
performance	O
penalty	O
because	O
the	O
execution	B-General_Concept
units	I-General_Concept
would	O
be	O
sitting	O
idle	O
anyway	O
.	O
</s>
<s>
However	O
,	O
if	O
hyper-threading	B-Operating_System
is	O
in	O
use	O
,	O
the	O
replay	B-Device
system	I-Device
will	O
prevent	O
the	O
other	O
thread	O
from	O
utilizing	O
the	O
execution	B-General_Concept
units	I-General_Concept
.	O
</s>
<s>
This	O
is	O
the	O
true	O
cause	O
of	O
any	O
performance	O
degradation	O
concerning	O
hyper-threading	B-Operating_System
.	O
</s>
<s>
In	O
Prescott	O
,	O
the	O
Pentium	B-General_Concept
4	I-General_Concept
gained	O
a	O
replay	O
queue	O
,	O
which	O
reduces	O
the	O
time	O
the	O
replay	B-Device
system	I-Device
will	O
occupy	O
the	O
execution	B-General_Concept
units	I-General_Concept
.	O
</s>
<s>
In	O
other	O
cases	O
,	O
where	O
each	O
thread	O
is	O
processing	O
different	O
types	O
of	O
operations	O
,	O
the	O
replay	B-Device
system	I-Device
will	O
not	O
interfere	O
,	O
and	O
a	O
performance	O
increase	O
can	O
appear	O
.	O
</s>
<s>
This	O
explains	O
why	O
performance	O
with	O
hyper-threading	B-Operating_System
is	O
application-dependent	O
.	O
</s>
