Locality characteristics of symbolic programs

By analyzing the virtual address traces of artificial intelligence applications, characteristics have been found in the temporal, spatial, and structural locality of the virtual memory references during symbolic program execution. These locality characteristics differ significantly from those of conventional workloads. This analysis is made using the author's two-state semi-Markov model of memory referencing behavior and not only reveals aspects of temporal and spatial locality that are much stronger in symbolic workloads, but also uncovers a high degree of structural locality in both types of workloads. Based on these findings, a unique memory system design that exploits these special reference locality characteristics of symbolic workloads is proposed.<
>