DynRP- Non-Intrusive Profiler for Dynamic Reconfigurability

摘要

Emerging technological areas such as machine learning, speech recognition, computer vision, autonomous robots, AI, bioinformatics involving big data, require implementation in complex heterogeneous accelerator platforms, to be able to handle data explosion with higher efficiency, lower power, and better performance. Dynamic reconfiguration in such platforms can help in run-time optimization to meet the design goals. The required optimal platform configuration can be achieved by a flexible design space exploration and appropriate task partitioning obtained through profiling computation and communication of processes in application code. This paper focuses on profiling, it being the key to the success of obtaining optimal platform configurations. It points to existing profiling techniques, their pros and cons vis-à-vis dynamic reconfigurable architectures, and the challenges in their design for obtaining optimal profiling performance. It further outlines desirable specifications for a profiler to allow dynamic real-time profiling for effective use of dynamic reconfiguration. DynRP, a non-intrusive hardware profiler for dynamic reconfiguration is proposed based on the desirable specifications, followed by its design and implementation details.