Managing Static (Leakage) Power - Computer Architecture Techniques for Power-Efficiency - page 145

32KB decay cache standard caches ratio=100 ratio=20 ratio=10 ratio=5

0.12 1.0

4KB standard

1.8

0.10

0.8

8KB standard

0.08

0.6

16KB standard

0.06

32KB standard

0.4

0.04

0.2

0.02

0.0

0.00

orig 512K 64K 8K 1K

0 8 16 24 32

decay interval(cycles)

active size(KB)

FIGURE 5.7: Left : Active size versus miss rate for decay caches. Right : Normalized leakage energy as
a function of the decay interval for various L2Access:leak ratios. Reproduced from [ 127 ]. Copyright 2001
IEEE.

situation the global signal is distributed serially from one cache line to another, giving each
line the chance to complete its writeback before proceeding to the next. Forcing an early
writeback of dirty data is not necessarily bad for performance. In fact, prior to cache decay,
Lee, Tyson, and Farrens proposed an “ eager ” writeback technique that yielded performance
benefits by not bothering the cache with writebacks when it is servicing performance-critical
misses [ 151 ].

Results : Cache decay has been extensively tested using the SPEC2000 in many cache
configurations (e.g., instruction, data, L1, L2, direct mapped, set associative, and for many
cache sizes). Overall, decay is very successful in switching off a significant part of the cache, on
the order of 70% for Level-1 caches, impacting a minimal performance penalty of a few percent
(less than 4%) [ 127 ].
Figure 5.7 shows a comparison of a single 32KB decay cache with standard caches of
various sizes (4K, 8K, 16K and 32KB). What changes in the decay cache is the decay interval
going from infinite (far right) to 1024 cycles (1Kc) at the far left. As the decay interval changes,
the “ active size ” of the decay cache—the average part that remains powered on—and its miss rate
are plotted. The decay curve is consistently below the curve for the standard caches. This graph
shows that a decaying cache is always better than a standard cache: for the same size, the decay
cache has a lower miss rate or, alternatively, for the same miss rate it has a smaller active footprint.
This is a result of selectively keeping active the most important items in the decay cache.
However, decay can also result in energy and performance penalties. This is reflected
in the graph on the right of Figure 5.7 which shows the overall benefit of decay in the form
of normalized leakage . Normalized leakage is the ratio of the new leakage divided by the old
leakage. The old leakage is the leakage before decay; the new leakage is the leakage with
decay but augmented with the additional dynamic power consumed by the extra switching to
implement decay and the extra power due to decay “mistakes.” Because of the destructive nature
of the gated- V dd mechanism, mistakenly switching off a cache line results in an additional L2