Augmenting HPC Profilers with Analysis Capabilities

This paper focuses on framework designed and developed by integrating multiple profiler modules to profile HPC applications. There is a constant demand for application performance on High Performance Computing (HPC) systems. Multiple software tools are currently available in the performance profiling realm to help application developers to profile and identify performance bottlenecks. Application developers strive to achieve the optimum execution of applications on available hardware resources using the profiling tools of their choice. They have to comprehend all the information provided by the tools and identify the bottlenecks to arrive at possible modifications. To assimilate the bottlenecks information from the profiling tool's outputs and representations, one requires experience and expertise in the application algorithm, programming methodologies, domain knowledge, and hardware resources. Many a time it is challenging for the developers to make changes in the application program based on the profiler output. We have developed a profile and analysis framework to identify the bottlenecks in an HPC application and help the developers improve application performance. The framework is developed by augmenting the existing opensource profilers with analysing capabilities for bottleneck identification and potential performance suggestions. This paper describes the architecture of the software framework and the benefits of the adopted mechanism.