Profiling HPC Applications with Low Overhead and High Accuracy

As the parallel scale of HPC applications represented by earth system models becomes larger and the computing cost becomes higher, the performance of HPC applications is increasingly critical. Profiling HPC applications accurately helps to model the applications and find the performance bottlenecks. However, due to the complexity of HPC applications, the diversity of programming languages, the differences of individual programming habits, and multiple architectures, accurate profiling becomes very tough. In this paper, we propose LPerf: a low-overhead and high-accuracy profiler for HPC applications. To reduce the profiling overhead and improve the profiling accuracy, we propose a preprocessing method which can automatically instrument with tunable granularity thus significantly reducing the run-time overhead of profiling, an aggregated caller-callee relationship which is used to locate relationship of functions efficiently, and a profiling-aware method which can precisely calculate running time of functions. The experimental results show that the error rate of profiling reaches 0.02%, and the overhead reaches 1.6%, in the earth system model named CAS-ESM. Compared with the baselines, the precision, accuracy, and overhead of LPerf have reached the state of the art.