Characterizing Phase Behavior for Dynamically Reconfigurable Architectures

Previous researches have shown most programs have phase behavior. We would like to take advantage of the phase behavior of applications to dynamically reconfigure an embedded platform in order to achieve more energy efficiency and performance. In this study, we first outline some aspects of the phase behavior with emphasis also on the influence of the Operating System (OS). We developed a tool chain to characterize the phase behavior of workloads (benchmarks) for embedded reconfigurable systems. Eight typical workloads are analyzed by using our tool chain. The results show that i) phase behavior can be relevant in workloads for embedded reconfigurable systems; ii) phase behavior (including the number of phases and the time spent in each phase) is reflected in several types of system metrics; iii) there are more phases when OS activities are involved for the same program; iv) the time spent by each phase of a program varies in a wide range. We believe the outcomes of this work can be used to guide the dynamic reconfiguration of components in embedded reconfigurable systems.