Managing Static (Leakage) Power - Computer Architecture Techniques for Power-Efficiency - page 154

TABLE 5.3: Non-inclusion Policies. Adapted from [ 154 ].

Policy Sub-block is Decayed Sub-block is Reactivated

Conservative When L1 block becomes DIRTY On writeback

Speculative-II When block is transferred to L1 When block is accessed in the L2

(even for writeback)

Speculative-IV When block is transferred to L1 When L1 block is evicted

severe performance loss. Because this decay adaptation takes into account how the inter-access
times vary with the number of accesses, it also outperforms both the adaptive per-cache-line
decay and the AMC technique.
A matter of inclusion : Li, Kadayif, Tsai, Vijaykrishnan, Kandemir, Irwin, and Sivasubra-
maniam examine L2 decay from a different perspective [ 154 ]. They make the observation that
cache lines do not have to be powered-on in the L2 while they are live in the L1. In essence,
they make the cache hierarchy non-inclusive for live lines. Their proposed policies work at the
granularity of L1 lines. In case the L2 lines are larger, they are divided into sub-blocks equal to
L1 lines; but to simplify the discussion here, we assume that the L1 and the L2 line sizes are
the same. Gated- V dd , as usual, turns off individual L2 lines.
Li et al. propose several policies and also compare the use of decay (non-state-preserving)
and drowsy (state-preserving) techniques. The latter are discussed in Section 5.3. The proposed
policies differ on when an L2 line is decayed and when it is reactivated. Table 5.3 lists the
three non-state-preserving policies using the terminology of the Li et al. paper [ 154 ], while
Figure 5.11 shows graphically their behavior.
The conservative policy only decays L2 lines that are no longer needed because they
are modified in the L1 ( Figure 5.11). On a writeback, the L1 line is put back into the L2.
The other two policies (Speculative-II and Speculative-IV) decay an L2 line as soon as it is
transferred to the L1. In Speculative-II, an evicted L1 line does not go back to the L2 unless
it is dirty. If it is dirty, the L2 line is reactivated to receive the writeback. Otherwise, the L2
line is powered-up on demand upon the next access which exposes the full penalty of an L2
decay-induced miss ( Figure 5.11). In contrast, in Speculative-I V the L2 line is reactivated
when the L1 line is evicted (even when it is clean). When the L2 is accessed again, it is likely
that the line is ready and waiting to be transferred back to the L1, hiding the reactivation
cost.