Next: Summary Up: Performance of Prefetching Previous: Loop Splitting

Software Pipelining

The effectiveness of the software pipelining algorithm is reflected by the pf-miss figures in Table 4. A large number means that the prefetches are either not issued early enough, in which case the line does not return to the primary cache by the time it is referenced, or not issued late enough, in which case the line has already been replaced in the cache before it is referenced. The results indicate that the scheduling algorithm is generally effective. The exceptions are CHOLSKY and TOMCATV, where over a third of the prefetched references are not found in the cache. The problem in these cases is that cache conflicts remove prefetched data from the primary cache before it can be referenced. However, there is still a performance advantage since the data tends to remain in the secondary cache, and therefore the primary-miss penalty is reduced, as shown earlier in Table 2.



Next: Summary Up: Performance of Prefetching Previous: Loop Splitting


Robert French