Managing Static (Leakage) Power - Computer Architecture Techniques for Power-Efficiency - page 165

instead of using gated- V dd to turn-off cache lines, these are put in drowsy mode when they are
“decayed.” 15 The result is that would-be decay-induced misses turn into simple wake-ups from
the drowsy mode. Consequently, the decay interval for the drowsy mode does not need to be
as conservative as in the gated- V dd mode.
The comparison, thus, turns into finding a break-even point: how much should the (non-
state-preserving) decay penalty cost in order for decay (which saves more leakage but incurs
decay-induced misses) to outperform the drowsy mode. A slower L2 makes the decay penalty
more costly, wiping out the advantage of saving more. But the break even point also changes
with temperature: the benefit of L1 decay is much higher at higher temperatures where it can
outperform a drowsy L1 even when backed by a relatively slow L2.
Decay drowsy hybrids on different cache levels : In contrast, to their conclusions regarding

the L1, Li, Kadayif, Tsai, Vijaykrishnan, Kandemir, Irwin, and Sivasubramaniam examine
in detail state-preserving versus non-state-preserving strategies in the L2 and conclude that
non-state-preserving decay policies in the L2 do not perform well compared to drowsy policies
[ 154 ]. Simply put, the penalty for an L2 decay-induced miss—which has to go to memory—is
just too high. Obviously, one has to be very conservative in the L2 not to incur decay-induced
misses. This means using long decay intervals. This, in turn, implies that an overly conservative
decay will probably not outpace the drowsy mode in leakage savings enough to turn the balance
in its favor. Nevertheless, Li et al. conclude that combining cache decay in the L1 with drowsy
policies in the L2 yields the best results overall if one considers the whole cache hierarchy [ 154 ].

+

☞ hybrid policies for the L1 and L2 : Mirroring the Speculative-II and Speculative-IV

decay policies discussed in Section 5.2.4, the Speculative-I and Speculative-III policies also
deactivate L2 lines when they are transferred to the L1 but—this time—using the state-
preserving drowsy mode instead of the gated- V dd [ 154 ]. These two drowsy L2 policies are
shown in Figure 5.15 (see also Figure 5.11 for a comparison with the gated- V dd version).
From the discussion so far, it is not surprising that Speculative-I turns out to be the best-
performing policy in the L2: it saves considerable leakage while hurting performance very
little.

On top of these two L2 policies, Li et al. examine more policies for the combined
two-level cache hierarchy. Overall, the best policy in terms of EDP for the whole hierarchy

15 As a matter of terminology, “ decay cache ” denotes the non-state-preserving version of decay while “ drowsy cache ”
means the state-preserving (drowsy mode) form. The change in drowsy caches from the Simple policy to the
generational policy is also followed in subsequent work, attempting to level the field between the two approaches
by utilizing the same policy and emphasizing only the state-preserving (gated- V dd ) versus non-state-preserving
(DVS) aspect.