Parallel Computer Architecture

882 CHAPTER 11 Latency Tolerance

cache block of four words) and there will be no unnecessary prefetches—but the

effectiveness is not likely to be great since the prefetches are issued too late. Extend-

ing the idea of long cache blocks are one-block lookahead (OBL) schemes, in which a

reference to cache block i may trigger a prefetch of block i

+

1. Several variants for

analysis and scheduling may be used in this technique; for example, block i + 1 can

be prefetched whenever a reference to block i is detected, or only if a cache miss to i

is detected, or when i is referenced for the first time after it is prefetched. Extensions

may include prefetchin ...