Scalable parallel performance measurement and analysis tools - state-of-the-art and future challenges

Abstract

Current large-scale HPC systems consist of complex configurations with a huge number of potentially heterogeneous components. As the systems get larger, their behavior becomes more and more dynamic and unpredictable because of hard- and software re-configurations due to fault recovery and power usage optimizations. Deep software hierarchies of large, complex system software and middleware components are required to operate such systems. Therefore, porting, adapting and tuning applications to today's complex systems is a complicated and time-consuming task. Sophisticated integrated performance measurement, analysis, and optimization capabilities are required to efficiently utilize such systems. This article will summarize the state-of-the-art of scalable and portable parallel performance tools and the challenges these tools are facing on future extreme-scale and big data systems.