To summarize, we have seen that in most cases the selective prefetching scheme performs noticeably better than the indiscriminate scheme. The advantage comes primarily from a reduction in prefetching overhead while still maintaining a comparable savings in memory stall time.
Now that we demonstrated the success of the algorithm, the next question is whether these speedups can only be achieved by carefully tuning the compile-time parameters that describe the memory hierarchy (several of them being rather heuristic in nature), or whether the algorithm is fairly robust at achieving good performance. We address this question in the following section.