Combining Power and Performance Modeling for Application Analysis: A Case Study Using Aspen

Scientists often rely on analytical modeling for predicting performance because it provides a flexible and fast estimate without the need for low-level details. With traditional performance modeling tools, users first develop a performance model and then repeatedly evaluate and analyze the model manually. As we observe a strong trend towards system designs with significantly more parallelism, such manual investigations prove laborious and error prone. In order to address this problem, some performance tools include a formal specification of an application's performance behavior and an abstract machine model to take advantage of composability, modularity and reusability. While these types of tools address performance at scale, they provide little insight to power and energy use at scale. Nonetheless, curtailing or controlling the power and energy use of systems at scale has been identified as a grand challenge. The goal of this paper is to demonstrate the feasibility of creating application-specific analysis tools that capture fine-grain power and energy information. The result is a prototype tool that aids in exploration of application and architecture specific power and energy tradeoffs. Results show new fine-grain, application-specific models of 3-D Fast Fourier Transform and Matrix-Matrix multiplication are accurate and may lend themselves to automation.