Information Technology Reference
In-Depth Information
(a) Miss Rate.
(b) Cycles per instruction.
Fig. 1. Normalized performance comparison of Random-LRU (50% LP) with baseline
design
Figure 2(a) shows that random-LRU gets -4.7% to 24.8% reduction in miss
rate with an average of 9.29%. The improvements in terms of CPI is shown in
Figure 2(b). It shows that random-LRU improves performance (CPI) between
-8.0% to 46.0% with an average of 5.0%. We can see that the improvement in
25% LP is not as satisfactory as 50% LP. This is because, in case of 25% LP, 75%
ways (6 ways) from each set follows random replacement policy and only 25%
ways are follows LRU policy. From Figure 2 we can see that random-LRU (25%
LP) degrades performance of the benchmarks having higher temporal locality.
Analysis. As explained in Section 1 the applications having more number
of dead and never-reused blocks cause higher miss rate in case of LRU replace-
ment policy. On the other hand a completely random replacement policy can
Search WWH ::




Custom Search