With the rapid growth of Artificial Intelligence (AI) applications supported by deep learning (DL), the energy efficiency of these applications has an increasingly large impact on sustainability. We introduce Smaragdine, a new energy accounting system for tensor-based DL programs implemented with TensorFlow. At the heart of Smaragdine is a novel white-box methodology of energy accounting: Smaragdine is aware of the internal structure of the DL program, which we call tensor-aware energy accounting. With Smaragdine, the energy consumption of a DL program can be broken down into units aligned with its logical hierarchical decomposition structure. We apply Smaragdine for understanding the energy behavior of BERT, one of the most widely used language models. Layer-by-layer and tensor-by-tensor, Smaragdine is capable of identifying the highest energy/power-consuming components of BERT. Furthermore, we conduct two case studies on how Smaragdine supports downstream toolchain building, one on the comparative energy impact of hyperparameter tuning of BERT, the other on the energy behavior evolution when BERT evolves to its next generation, ALBERT.

Power modeling is an essential building block for computer systems in support of energy optimization, energy profiling, and energy-aware application development. We introduce Vesta , a novel approach to modeling the power consumption of applications with one key insight: language runtime events are often correlated with a sustained level of power consumption. When compared with the established approach of power modeling based on hardware performance counters (HPCs), Vesta has the benefit of solely requiring application-scoped information and enabling a higher level of explainability, while achieving comparable or even higher precision. Through experiments performed on 37 real-world applications on the Java Virtual Machine (JVM), we find the power model built by Vesta is capable of predicting energy consumption with a mean absolute percentage error of 1.56 % , while the monitoring of language runtime events incurs small performance and energy overhead.

Developing energy-aware applications is a well known approach to software-based energy optimization. This promising approach is however faced with a significant hurdle when deployed to the environments shared among multiple applications, where the energy consumption effected by one application may erroneously be observed by another application. We introduce Eflect, a novel software framework for disentangling the energy consumption of co-running applications. Our key idea, called energy virtualization, enables each energy-aware application to be only aware of the energy consumption effected by its execution. Eflect is unique in its lightweight design: it is a purely application-level solution that requires no modification to the underlying hardware or system software. Experiments show Eflect incurs low overhead with high precision. Furthermore, it can seamlessly port existing application-level energy frameworks --- one for energy-adaptive approximation and the other for energy profiling --- to shared environments while retaining their intended effectiveness.

Energy accounting is a fundamental problem in energy management, defined as attributing global energy consumption to individual components of interest. In this paper, we take on this problem at the application level, where the components for accounting are application logical units, such as methods, classes, and packages. Given a Java application, our novel runtime system Chappie produces an energy footprint, i.e., the relative energy consumption of all programming abstraction units within the application.