esprof: A Generic Profiling Infrastructure for Multi/Many-Core Embedded Systems

摘要

Embedded systems are moving to multi-core and many-core platforms to satisfy the ever-growing demand for computing power. Complex task like autonomous driving will include both real-time applications and high-performance applications. Research has shown that traditional RTOS can be optimized to provide good scalability for parallel computing. The lack of performance analysis tools, however, is one of the major obstacles for identifying bottlenecks. In this paper, we first discuss the characteristics and challenges of existing tools used in high-performance computing. To meet the requirements of embedded systems, a generic infrastructure for creating customized premiers flexibly and effortlessly is then presented. As an example, a scalable and optimized call graph profiler has been developed. At last, we evaluate our profiler and existing profiler by measuring benchmark application on a 36-core platform. The results show that our profiler can provide much higher accuracy with very low overhead.