Information Technology Reference
In-Depth Information
TABLE 4.9:
Parallel Switching Activity Optimization in Set-Associative Caches
What it Does
Examples
Technique
Phased access
Finds a tag match first and then
accesses the data way
Hitachi SH3 [
95
]
Alpha 21264 [
87
]
Sequential access
Sequentially accesses ways starting
from the way most likely to hit
(MRU)
Kessler et al. [
133
]
Way prediction
Predicts and accesses the cache
way that contains the data in
parallel to tag match
MRU prediction,
Inhue et al.
[
109
],
Selective Direct
Mapping,
Powel et al. [
183
],
Multi-MRU
, Zhang et al.
[
242
], Zhu et al. [
249
]
Way Selection
Deterministically Selects the cache
way that contains the data prior
to tag match
Location Cache
, Min et al.
[
168
],
Way Halting,
Zhang et al.
[
241
],
Decaying Bloom filters
,
Keramidasetal.[
131
]
in in-order issue processors. However, if the increased cache latency can be tolerated this is a
straightforward and efficient technique to use. Case in point, the Alpha's L2 cache were the
additional L2 latency can be easily tolerated by the out-of-order core.
4.9.2 Sequentially Accessed Set-Associative Cache
A sequentially accessed set-associative cache was first proposed by Kessler, Jooss, Lebeck, and
Hill [
133
] as an inexpensive method to implement associativity. In a set-associative organization
the sequential access scheme is depicted in Figure 4.25. Initially, only the most likely cache
way to produce a hit is probed. In Figure 4.25, the first probe is chosen to be the most recently
used (MRU) way. This feature was first proposed for the IBM 370 cache [
48
]. Additional
information is needed to point to the MRU way in each set. However, if the cache features a
true-LRU replacement algorithm the MRU information can be readily extracted from the LRU
list—incurring of course additional latency. Failing to find the requested data in the MRU way
starts a cyclic
sequential
search of the remaining ways.
Search WWH ::
Custom Search