Managing Static (Leakage) Power - Computer Architecture Techniques for Power-Efficiency - page 146

access. These are called decay-induced misses and cost not only in energy (reducing the energy
benefit of decay) but also in performance.
The more aggressive the choice of decay interval, the most mistakes are experienced.
The larger the energy cost of a mistake (L2 access), the less the overall benefit from decay. To
reflect this dependence, the right graph of Figure 5.7, plots several curves, each corresponding
to a different relative cost for an L2 access. The ratio that specifies the relative cost is the
L2access:leakage ratio, defined in [ 127 ]. As decay moves toward smaller decay intervals (and
with more costly L2 accesses), it starts to lose its benefit and may even go into negative territory
in the extremes. However, with a decay interval of 8000 cycles or more (for the simulated
systems in [ 127 ]), decay-induced misses are so few that the relative cost of an L2 access
becomes irrelevant. It is around these decay intervals where decay provides its maximal benefit.
Further results can be found in Hanson's work with Hrishikesh, Agarwal, Keckler, and
Burger [ 93 ]. Hanson's work is one of the most detailed and extensive studies on cache decay and
provides comparisons with two other leakage-saving techniques for caches. A detailed technical
report by Hanson et al. greatly expands on the initial results reported for decay [ 94 ].

☞ directions to improve cache decay : Cache decay is based on the generational behavior of

cache lines, and as such sits on a robust foundation. This allows it to work well in a wide
range of conditions. There is room for improvement, though, over the initial proposal, on
a number of aspects:

Decay-induced misses . Cache decay has become synonymous with non-state-
preserving techniques because of its use of the gated- V dd mechanism to turn off
cache lines. By mistakenly turning off cache lines in their live time, caches incur
decay-induced misses that hurt both the energy savings as well as performance. In
retrospect, decay can be thought of as the technique to detect dead lines in the
cache. What to do with this information is a whole different matter: dead lines can
be turned off (gated- V dd ), put into a drowsy mode, replaced, compressed, dupli-
cated for reliability, etc. In a more recent work, the decay policy of detecting dead
lines is used in conjunction with the drowsy mechanism that puts these lines into a
low-leakage state.

Measuring time in cycles . The decay interval is measured in cycles, which is a quantity
that depends on architectural features. This makes it difficult to reason about decay
intervals, especially across different programs, or in the same program but across
different platforms. A better choice would be a more independent “time” metric
such as the number of intervening accesses between two consecutive accesses to the
same cache line. This is entirely a property of the application and does not depend