An approach for symbolic mapping of memory references

Abstract

Understanding and tuning memory system performance is a critical issue for most scientific programs to achieve reasonable performance on current high performance systems. Users need a data-centric performance measurement infrastructure that can help them understand the precise memory references in their program that are causing poor utilization of the memory subsystem. However, a data-centric performance tool requires the mapping of memory references to the corresponding symbolic names of the data structure, which is non trivial, especially for mapping local variables and dynamically allocated data structures. In this paper we describe with examples the algorithms and extensions implemented in the SIGMA environment for symbolic mapping of memory references to data structures. © Springer-Verlag 2004.