Viterbi Faculty of Electrical Engineering, Technion
Semantic locality in memory access patterns and machine learning approaches for hardware prefetching
Memory prefetching is increasingly critical to processor performance as memory latencies continue to grow. Most modern memory prefetchers rely on spatio-temporal locality to predict future memory addresses. However, many emerging applications do not manifest such regularities, and are less amenable to existing prefetching techniques. Our work aims to better understand the memory access sequences exhibited by such computer programs, their recurrence patterns, and their locality paradigms. To that end, we introduce the concept of semantic locality: a high-level abstraction of data locality that represents relations between data objects as reflected by the inherent semantics of the data structure or the traversal algorithm. We show that semantic locality transcends the existing spatio-temporal locality paradigms and generalizes them. We further present several ways of capturing semantic locality through program context correlation and dynamic code analysis. We describe multiple prefetcher designs using various machine learning models (including on-core neural networks) to implement these approaches in order to predict future memory accesses. * PhD seminar under supervision of Prof. Uri Weiser and Prof. Yoav Etsion.
Date: Sun 28 Jul 2019
Start Time: 13:30
End Time: 14:30
1061 | Electrical Eng. Building