Table 3 illustrates the dynamic behavior of three systems:
an uncached system (simulating a disabled instruction cache), a cached
system with the bit-encoding approach, and a conventional cached
system. Column 3 indicates the number of instructions executed. The
hit ratio (percentage of cache hits of all instruction references) is
shown for the bit-encoded system in column 4 and for conventional
caches in column 5. Column 6 shows the percentage of executed
instruction references which were classified as conflicts on a cached
system. Column 7 indicates the estimated execution time in cycles for
an uncached system. The percentage of cycles required for a
bit-encoded system (column 8) and for a conventional cached system
(column 9) are compared to an uncached system. The execution time is
calculated based on the equations 1, 2,
and 3 for n=9.
The bit-encoding approach results in lower hit ratios (72-98%) than on a conventional cached system (92-99%). Yet, caches are often disabled for critical real-time tasks to provide the predictability required by scheduling analysis. Thus, the bit-encoding approach should be compared to an uncached system. The bit-encoding method requires only 13-39% of the cycles used by the uncached systems. This provides a speedup of programs by a factor of 3-8 without sacrificing the predictability of a program's execution time. The result resembles the improvement over critical real-time tasks which require caches to be disabled. The results improve considerably as the cache size increases and entire programs fit into cache. The execution time required for a conventional cached system is only about 14% of an uncached system, but the predictability also decreases to the point where it becomes insufficient for scheduling analysis of critical tasks. This can be explained as follows:
Conflicts correspond to the instructions whose cache behavior could not be predicted prior to execution in a conventional cached system. The dynamic percentage of conflict references is higher than the static percentage given in Table 2 since conflicts typically occur in loops. Since 5-25% of the instructions executed were conflicts, the execution time of programs cannot be predicted as tightly in conventional cached systems with traditional timing tools, especially for smaller caches. However, more recent work by the authors shows that the instruction categorization of the static simulator may be used by a more sophisticated timing tool to provide tight worst-case execution time predictions with a 4-9 times speedup over uncached system using a conventional instruction cache [].
