Memory Characterization of Embedded Applications

One method commonly used by computer architects to study application memory behavior is trace driven simulation of synthetic address traces. Leveraging synthetic traces reduces the time needed to gather the diverse set of application traces needed for effective testing, and provides programmers with knowledge to tune their programs for better runtime performance. In this paper, we present a technique for accurately modeling memory performance of applications executing on an embedded platform. In order to more accurately capture application memory behavior, we present novel techniques for modeling the bursty nature of compulsory cache misses and sequences of loads and stores. We show that our model can be used to generate synthetic address traces that achieve cache performance within 3% on average of the target application.