Information Technology Reference
In-Depth Information
(a) Miss Rate.
(b) Cycles per instruction.
Fig. 2.
Normalized performance comparison of Random-LRU (25% LP) with baseline
design
replace dead or never-reused blocks early but it may also choose MRU block (or
other frequently used blocks) as victim block. So in case of these two replace-
ment policies, the advantage of one is the disadvantage of other. In our proposed
replacement policy a randomly selected victim block from
RP
is placed in
LP
.
Now if the selected victim block is a frequently used block then very soon it will
be requested again and hence moved from
LP
to
RP
. Otherwise the block will
be eventually removed from the cache. In this way our policy takes the advan-
tage of both LRU and random replacement policy. Hence it improves the miss
rate as well as the CPI.
Normally reduction in miss rate results in higher improvement of CPI. But
in our experiments some benchmarks show exceptional behavior, where the im-
provement in CPI is not as expected even when there is a large reduction in the
miss rate. This is because of the non-deterministic thread management by the
