Matryoshka: A Coalesced Delta Sequence PrefetcherMatryoshka: A Coalesced Delta Sequence Prefetcher

To learn complex memory access patterns effectively, many spatial data prefetchers have been proposed that characterize the patterns as fixed-length delta sequences. However, because complex patterns are variable in workloads, it is difficult for fixed-length delta sequences to recognize them with both high competitive coverage and accuracy. That is, longer delta sequences increase accuracy at a lower probability of pattern matching, while shorter delta sequences increase coverage at a higher probability of false predictions. A classical strategy is to introduce the multiple matching mechanism associated with variable-length delta sequences, but sequences have to be redundantly stored in multiple tables. We observe that shorter delta sequences can coalesce into longer sequences of a fixed length during learning processes. At the same time, the shorter ones can be extracted from the longer ones during matching processes. By leveraging this property, we can improve the storage efficiency of the multiple matching mechanism. In this paper, we propose a novel low-overhead prefetcher, named Matryoshka, that supports the multiple matching mechanism with high efficiency. Instead of maintaining variable-length delta sequences in multiple tables, Matryoshka coalesces variable-length delta sequences into fixed-length delta sequences, which can be maintained with a single pattern table. Concerning the evaluation of a simulated single-core system using memory-intensive workloads of SPEC 2017, Matryoshka outperforms the state-of-the-art prefetcher IPCP by 6.5% and surpasses SPP+PPF by 2.9% with 26x lower storage overhead.