Investigation of the Locality of Memory Accesses during Symbolic Program Execution.

摘要

This research focused on the low-level virtual address memory referencing behaviour of symbolic programs. The virtual address traces of six artificial intelligence applications and two conventional workloads executed on the Texas Instruments Explorer II were used to characterize the locality of virtual memory accesses during symbolic program execution and to compare these characteristics with the locality characteristics of conventional workloads. By using this approach, this research accomplished three overall objectives: first, it enhanced the basic understanding of symbolic program memory referencing behavior; second, it found new ways in which the architecture can be tailored for these low-level memory referencing locality characteristics; and third, it laid the foundation for developing a symbolic program workload generator model that can be used to drive memory system design tools. This research was the first systematic effort to characterize symbolic program behavior through the investigation of symbolic workloads' temporal, spatial, and structural locality. Its specific contributions are the development of new measures and methods for the analysis of program locality, an extension of Denning's model of page referencing behavior to virtual address word-level memory referencing behavior, the identification of a low-level virtual memory referencing structural locality, the identification of significant differences in the low-level virtual memory referencing behavior of symbolic and conventional workloads, and a memory system design for exploiting the structural locality characteristics of symbolic workloads. (KR)