AAP and AAPM: Improved Prefetching Structures of the L2 Cache

Abstract

The push technique is an effective method of tolerating the Load-to-use latency. However, not every memory level can gain a sound performance improvement when the push technique is adopted. In this paper, we devise a quantitative approach to analyze workload on the bus. For balancing workload, we propose a novel architecture, which compounds the push and prefetching technique to serve different storage levels. Between the L1 and the L2 cache, we use the push technique. Between the L2 cache and memory, we utilize the prefetching technique. Focusing on the L2 cache prefetching, we propose AAP and AAPM, both of which meet the requirement of timeliness and bandwidth adaptability and can increase the IPC by 2% compared with SRP.