Approaches for Scalable Performance Analysis and Optimization
Lawrence Livermore National Laboratory
The growing scale of future machines, coupled with increasing node as well as application complexity, requires new approaches to performance analysis and optimization. Tools will need to be more intuitive by mapping their data into different, application oriented domains to enable the detection of root causes for performance bottlenecks. Further, tools will need to be less monolithic and instead be adjustable to particular applications and target scenarios, since only this will allow them to exploit application level semantics for effective data filtering and attribution. While existing tools provide a wide range of raw measurement capabilities, they often short fall short with respect to these requirements.
In this talk I will present techniques that address these key challenges focusing on new visualization and data analysis techniques that provide an increased intuition behind performance results by showing correlation between performance measurements and application properties; as well as on interoperable and modular tool infrastructures that enable rapid and possibly application specific tool prototyping. This will lead us towards a new generation of tools that are truly scalable and provide application developers with the necessary mechanisms for a meaningful performance analysis and root cause detection.