Managing Static (Leakage) Power - Computer Architecture Techniques for Power-Efficiency - page 166

L1 generation L1 generation

live

dead time

L1 timeline

eviction

L2 timeline

{

Periods L2 line
is in drowsy

Speculative-III

mode Speculative-I (non-dirty line)

Drowsy wake-up penalty

FIGURE 5.15: Time diagrams for the non-inclusion, state-preserving, drowsy policies. Speculative-I
puts the L2 line in drowsy mode as soon as it is transferred to the L1. The L2 line remains in drowsy
mode until the next L2 access (provided that there was no writeback of a dirty L1 line). The difference in
Speculative-III is that the L2 line is awaken as soon as the L1 line is evicted (even if it is clean). Adapted
from [ 154 ].

is a state-destroying decay in the L1 coupled with either state-preserving decay in the L2 or
with the Speculative-I. State-destroying decay at both cache levels, although saving the most
leakage, does not fare as well in EDP.
Decay drowsy hybrid in the same cache : Zhang et al. were the first to use a hybrid state-

destroying and state-preserving scheme for the instruction cache [ 246 ]. The control of the
leakage mode is left to the compiler which chooses the state-preserving mode for cache lines
containing instructions that are expected to appear again in the future and the state-destroying
mode for instructions that are not coming back.
In the data cache, Meng, Sherwood, and Kastner take the hybrid idea one step further
and combine decay and drowsy techniques, not on different cache levels but, in the same cache.
Their goal is to examine the limits of leakage reduction assuming oracle knowledge—full
knowledge of the access stream. They demonstrate optimal results for state-destroying (gated-
V dd ) decay and the drowsy mode in the instruction and the data L1. Interestingly, their study
shows that a hybrid scheme combining the decay and drowsy approaches yields the best results
in a single cache [ 164 ].
The hybrid scheme is quite simple: it starts by putting a cache line in drowsy mode after a
very short period of inactivity and proceeds to turn off the line completely if this period of inactiv-
ity persists beyond a larger “decay interval.” This is as close to the optimal as a practical approach
can get since it attacks the problems of decay and drowsy mode in a complementary fashion.
The drowsy mode problem of not saving enough leakage is solved by eventually completely
switching off the cache line. Whereas, the two main decay problems are solved as follows. First,
the problem of decay-induced misses is solved by waiting for a long time to make sure that a

+