Scientific applications are constantly facing a huge challenge when it comes to exploiting high-level concurrency and novel features in current and future architectures. A growing concern in the scientific community is the amount of investment needed to port existing applications to a particular architecture, and the estimate of its benefits over the long term. The application developers face a daunting task of how they should design and architect the algorithms to scale on Leadership-class systems. On the other hand, Leadership Computing Facilities are faced with challenges to invest in hardware features that meet application needs and science demands.

To address the issues about large time investments, holistic mechanisms are needed to help the application developers to port and scale their applications to future systems, reduce the barriers to entry for new applications in HPC, facilitate the design of future Leadership systems, and produce systems beneficial to the community. Performance engineering team of ALCF has been working to address the challenges with the help of SKOPE – a SKeleton framewOrk for Performance Exploration. SKOPE helps users describe, model, and explore a workload’s current and potential behavior. Given a formalized description of the workload, and using a database of hardware models, that reflect various hardware design features, SKOPE automatically analyzes, tunes, and projects the workload's performance for the target hardware. SKOPE offers readability and expressiveness in its frontend and flexibility in its backend functionality.

In this informal talk, I will go through the history of the development of SKOPE, describe the most noticeable features of the framework, and highlight the most interesting findings that SKOPE has helped to generate.