Hardware performance sniffers for embedded systems profiling

摘要

Today on-chip monitoring solutions should be characterized by a reduced software and hardware overheads. So, this work deals with techniques to profile computational behavior and communication patterns of hardware/software components belonging to systems with multiple processing elements, i.e. a more general representation of on-chip embedded systems. In particular, the paper focuses on profiling techniques based on ad-hoc hardware mechanisms in order to avoid possible distortion of system behavior due to the monitoring action itself, thus satisfying unobtrusive profiling requirement. The final goal is the definition of a set of metrics to profile memory accesses and communication between processing elements, and the development of proper hardware mechanisms to support their evaluation. The profiling system architecture is described and its possible advantages over alternative software-based and hardware-based solutions are analyzed. Tests on proposed system are also done by means of a prototype implementation specifically targeting Xilinx MicroBlaze based multi-processor system.