• Energy Research

With the growing availability of large-scale datasets, and the popularization of affordable storage and computational capabilities, the energy consumed by AI is becoming a growing concern. To address this issue, in recent years, studies have focused on demonstrating how AI energy efficiency can be improved by tuning the model training strategy. Nevertheless, how modifications applied to datasets can impact the energy consumption of AI is still an open question. To fill this gap, in this exploratory study, we evaluate if data-centric approaches can be utilized to improve AI energy efficiency. To achieve our goal, we conduct an empirical experiment, executed by considering 6 different AI algorithms, a dataset comprising 5,574 data points, and two dataset modifications (number of data points and number of features). Our results show evidence that, by exclusively conducting modifications on datasets, energy consumption can be drastically reduced (up to 92.16%), often at the cost of a negligible or even absent accuracy decline. As additional introductory results, we demonstrate how, by exclusively changing the algorithm used, energy savings up to two orders of magnitude can be achieved. In conclusion, this exploratory investigation empirically demonstrates the importance of applying data-centric techniques to improve AI energy efficiency. Our results call for a research agenda that focuses on data-centric techniques, to further enable and democratize Green AI. 11 pages, 3 figures, 2 tables. Accepted at the 8th ICT for Sustainability Conference (ICT4S) 2022

Green ICT is a young and pioneering field. Therefore, as often pointed out in the literature, studies evaluating the main research activities and the general direction of this new and continuously evolving research field are scarce and often incomplete. This study presents a quantitative analysis, through a systematic literature review, of the main activities, trends and issues that can be found in the Green ICT literature. The research reports the analysis of various characteristics of the studies gathered for this review, such as addressed type of effect and year of publication. It also led to the identification of the most recurrent issues of the research and development of Green ICT strategies. Finally, this study proposes a new category of effect (people awareness) that, even if often addressed by the field, is not included in current Green ICT frameworks.

Background:Digital infrastructures, i.e., ICT systems, or system-of-systems, providing digital capabilities, such as storage and computational services, are experiencing an ever-growing demand for data consumption, which is only expected to increase in the future. This trend leads to a question we need to answer: How can we evolve digital infrastructures to keep up with the increasing data demand in a sustainable way?Objective:The goal of this study is to understand what is the future of sustainable digital infrastructures, in terms of: which solutions are, or will be, available to sustainably evolve digital infrastructures, and which are the related adoption factors, impediments, and open problems.Method:We carried out a 3-phase mixed-method qualitative empirical study, comprising semi-structured interviews, followed by focus groups, and a plenary session with parallel working groups. In total, we conducted 13 sessions involving 48 digital infrastructure practitioners and researchers.Results:From our investigation emerges a landscape for sustainable digital infrastructures, composed of 30 solutions, 5 adoption factors, 4 impediments, and 13 open problems. We further synthesized our results in 4 incremental scenarios, which outline the future evolution of sustainable digital infrastructures.Conclusions:From an initial shift from on-premise to the cloud, as time progresses, digital infrastructures are expected to become increasingly distributed, till it will be possible to dynamically allocate resources by following time, space, and energy. Numerous solutions will support this change, but digital infrastructures are envisaged to be able to evolve sustainably only by (i) gaining a wider awareness of digital sustainability, (ii) holding every party accountable for their sustainability throughout value chains, and (iii) establishing cross-domain collaborations.

Architectural Technical Debt (ATD) regards sub-optimal design decisions that bring short-term benefits to the cost of long-term gradual deterioration of the quality of the architecture of a software system. The identification of ATD strongly influences the technical and economic sustainability of software systems and is attracting growing interest in the scientific community. During the years several approaches for ATD identification have been conceived, each of them addressing ATD from different perspectives and with heterogeneous characteristics. In this paper we apply the systematic mapping study methodology for identifying, classifying, and evaluating the state of the art on ATD identification from the following three perspectives: publication trends, characteristics, and potential for industrial adoption. Specifically, starting from a set of 509 potentially relevant studies, we systematically selected 47 primary studies and analyzed them according to a rigorously-defined classification framework. The analysis of the obtained results supports both researchers and practitioners by providing (i) an assessment of current research trends and gaps in ATD identification, (ii) a solid foundation for understanding existing (and future) research on ATD identification, and (iii) a rigorous evaluation of its potential for industrial adoption.

Software and IT usage are continuously growing to keep our society active and manage our individual lives. But as they grow, their energy demand is exploding. By 2030, data centers alone will already consume some 10% of the global electricity.1 Including the Internet, telecommunications, and embedded devices, the energy consumption will be one-third of the global demand. Understanding that end users only consume what we offer, it is the community of software developers who must become active in ecologic behaviors. Green IT is the call of today. Each single line of code that we develop today may still be running years from now on zillions of processors, eating energy and contributing to global climate change.

A promise of cloud computing is the reduction of energy footprint enabled by economies of scale. Unfortunately, little research is available on how cloud consumers can reduce their energy footprint when running software in the public cloud. Moreover, cloud consumers do not have full access to information regarding their cloud infrastructure usage, which is required to understand the impact of design decisions on energy usage. The purpose of our study is to support cloud consumers in developing energy-efficient workloads in the public cloud. To achieve our goal, we collaborated with a large cloud solution provider to discover an initial set of reusable architectural tactics for software energy efficiency. Starting from interviews with 17 practitioners, we reviewed and selected available tactics to improve the energy efficiency of individual workloads in the public cloud, and synthetized the identified tactics in a reusable model. In addition, we conducted a case study to assess the impact of utilizing a tactic, which was selected following a prioritization provided by the practitioners. Our results demonstrate the possibility to architect cloud workloads for energy efficiency through reasoning and estimation of resource optimization. However, the process is not (yet) straightforward due to the current lack of transparency of cloud providers.

Software energy efficiency has gained increasing attention of the research community. How to improve it, however, still lacks evidence. Specifically, the impact of code smell refactoring on energy efficiency has been scarcely investigated. In the pilot study here reported, we investigate the impact on performance and energy consumption of refactoring well-known code smells on Java software applications. In order to understand if software metrics can be used as indicators of the energy impact of refactoring, we also measured the variation caused by refactoring on a set of well-established software metrics. We conducted a controlled experiment using state-of-the-art power measurement equipment. Statistical hypothesis testing and effect size estimation were performed on the experimental results. Results show that in one out of three applications, refactoring each smell significantly impacted power- and energy consumption. Specifically, refactoring Feature Envy and Long Method smells led to a 49% energy efficiency improvement. No software metrics significantly correlated with execution time, power or energy consumption. In conclusion, refactoring code smells can significantly improve software energy efficiency. The magnitude of the impact may depend on application properties, e.g. size or age. Further research is needed to understand the relationship between software metrics and energy efficiency.