Making Address-Correlated Prefetching Practical

Thomas F. Wenisch, Michael Ferdman, Anastasia Ailamaki, Babak Falsafi, Andreas Moshovos

Abstract: Despite a decade of research demonstrating its efficacy, address-correlated prefetching has never been implemented in a shipping processor because it requires megabytes of metadata - too large to store practically on chip. New storage-, latency-, and bandwidth-efficient mechanisms for storing metadata off chip yield a practical design that achieves 90 percent of the performance potential of idealized on-chip metadata storage.