Predicting data structures for energy efficient computing

Dynamic data structures in software applications have been shown to have a large impact on system performance. In this paper, we explore energy saving opportunities of interface-based dynamic data structures. Our results suggest that opportunities do exist in the C5 Collection, at least 16.95% and up to 97.50%. We propose an architecture for building adaptive green data structures by applying machine learning tools to build a model for predicting energy efficient data structures. Our neural network model can classify energy efficient data structures based on features such as the number of elements, frequency of operations, interface and set/bag semantics. The 10-fold cross validation results show 96.01% accuracy on the training data and 95.80% on the training validation data. Our n-gram model can accurately predict the most energy efficient data structure sequence in 19 simulated and real-world programs-on average, with more than 50% accuracy and up to 98% using a bigram predictor.

I. Introduction

The proliferation of global computing devices has pushed IT energy consumption higher and higher, raising concerns among many environmentalists. As reported in [1], information and communication technology ecosystems have approached 10% of the world's electricity generation. The need for reducing IT energy consumption and research efforts in every sector is now necessary, not just for energy saving and extending battery life, but also for the environment. There are many areas in the layers of computer systems that can be optimized for energy efficiency, from hardware [3], [4], operating system [5] to application layers [12], [13]. By making “unpolluted design” of computer systems [2], hardware engineers and software developers can also play a significant part in the fight against climate change. In this study, we focus on the application layer, primarily object-oriented software with “interface-based” implementation [6], where there are multiple choices of classes that implement the same interface. Our goal is to explore how the choice of these classes impacts the energy consumption of software applications and to find ways to intelligently switch between implementations for energy efficiency.