An Efficient Architecture for Dynamic Profiling of Multicore Systems

摘要

Application profiling is an important step in the design and optimization of embedded systems. Accurately identifying and analyzing the execution of frequently executed computational kernels is needed to effectively optimize the system implementation, both at design time and runtime. In a traditional design process, it suffices to perform the profiling and optimization steps offline, during design time. The offline profiling guides the design space exploration, hardware software codesign, or power and performance optimizations. When the system implementation can be finalized at design time, this approach works well. However, dynamic optimization techniques, which adapt and reconfigure the system at runtime, require dynamic profiling with minimum runtime overheads. Existing profiling methods are usually software based and incur significant overheads that may be prohibitive or impractical for profiling embedded systems at runtime. In addition, these profiling methods typically focus on profiling the execution of specific tasks executing on a single processor core, but do not consider accurate and holistic profiling across multiple processor cores. Directly utilizing existing profiling approaches and naively combining isolated profiles from multiple processor cores can lead to significant profile inaccuracies of up to 35%. To address these challenges, a hardware-based dynamic application profiler for non-intrusively and accurately profiling software applications in multicore embedded systems is presented. The profiler provides a detailed execution profile for computational kernels and maintains profile accuracy across multiple processor cores. The hardware-based profiler achieves an average error of less than 0.5% for the percentage execution time of profiled applications while being area efficient.