Search for: All records

Federated Learning (FL) distributes machine learning (ML) training across edge devices to reduce data transfer overhead and protect data privacy. Since FL model training may span hundreds of devices and is thus resource- and energy-intensive, it has a significant carbon footprint. Importantly, since energy's carbon-intensity differs substantially (by up to 60×) across locations, training on the same device using the same amount of energy, but at different locations, can incur widely different carbon emissions. While prior work has focused on improving FL's resource- and energy-efficiency by optimizing time-to-accuracy, it implicitly assumes all energy has the same carbon intensity and thus does not optimize carbon efficiency, i.e., work done per unit of carbon emitted. To address the problem, we design EcoLearn, which minimizes FL's carbon footprint without significantly affecting model accuracy or training time. EcoLearn achieves a favorable tradeoff by integrating carbon awareness into multiple aspects of FL training, including i) selecting clients with high data utility and low carbon, ii) provisioning more clients during the initial training rounds, and iii) mitigating stragglers by dynamically adjusting client over-provisioning based on carbon. We implement EcoLearn and its carbon-aware FL training policies in the Flower framework and show that it reduces the carbon footprint of training (by up to 10.8×) while maintaining model accuracy and training time (within ~1%) compared to state-of-the-art approaches 

Greenhouse gas emissions from the residential sector represent a large fraction of global emissions and must be significantly curtailed to achieve ambitious climate goals. To stimulate the adoption of relevant technologies such as rooftop PV and heat pumps, governments and utilities have designedincentivesthat encourage adoption of decarbonization technologies. However, studies have shown that many of these incentives are inefficient since a substantial fraction of spending does not actually promote adoption. Further, these incentives are not equitably distributed across socioeconomic groups. In this article, we present a novel data-driven approach that adopts a holistic, emissions-based, and city-scale perspective on decarbonization. We propose an optimization model that dynamically allocates a total incentive budget to households to directly maximize the resultantcarbon emissions reduction– this is in contrast to prior work, which focuses on metrics such as the number of new installations. We leverage techniques from the multi-armed bandits problem to estimatehuman factors, such as a household’s willingness to adopt new technologies given a certain incentive. We apply our proposed dynamic incentive framework to a city in the Northeast U.S., using real household energy data, grid carbon intensity data, and future price scenarios. We compare our learning-based technique to two baselines, one “status-quo” baseline using incentives offered by a state and utility, and one simple heuristic baseline. With these baselines, we show that our learning-based technique significantly outperforms both the status-quo baseline and the heuristic baseline, achieving up to 37.88% higher carbon reductions than the status-quo baseline and up to 28.76% higher carbon reductions compared to the heuristic baseline. Additionally, our incentive allocation approach is able to achieve significant carbon reduction even in a broad set of environments, with varying values for electricity and gas prices, and for carbon intensity of the grid. Finally, we show that our framework can accommodateequity-awareconstraints to preserve an equitable allocation of incentives across socioeconomic groups while achieving 83.34% of the carbon reductions of the optimal solution on average. 

To improve data center efficiency, job schedulers often overcommit computing resources such that the sum of the maximum resource requirements across running jobs on a server exceeds its resource capacity while relying on statistical multiplexing of workloads at runtime to reduce the likelihood of saturating capacity and violating applications’ service level objectives. The challenge with overcommitting resources is that future job resource demand varies widely over time. As a result, jobs’ collective resource usage may exceed resource capacity if their periods of high demand align. Our key insight is that many jobs often exhibit some periodicity in their resource usage, which schedulers can leverage to improve resource usage predictions and job placement decisions. To leverage this insight, we show how to model jobs as simple periodic functions and then develop a period-aware placement policy that increases resource utilization while mitigating performance degradation due to jobs’ collective resource usage exceeding resource capacity. We evaluate our approach on a publicly-available job trace from a major cloud platform, and show that it yields low (<30%) error for a large fraction (∼80%) of periodic jobs. We then evaluate our period-aware placement policy on small problem instances, and show that it is much closer to the NP-hard optimal policy than current state-of-theart policies. Finally, we evaluate our approach on an industry job trace, and show that combining our periodic models and periodaware placement policy results in the best of both worlds: higher average server utilization and lower performance degradation by more than 2× compared to the existing state-of-the-art. 

The declining cost of solar photovoltaics (PV) combined with strong federal and state-level incentives have resulted in a high number of residential solar PV installations in the US. However, these installations are concentrated in particular regions, such as California, and demographics, such as high-income Asian neighborhoods. This inequitable distribution creates an illusion that further increasing residential solar installations will become increasingly challenging if it is not already prohibitive. Furthermore, while the inequity in solar installations has received attention, no prior comprehensive work has been done on understanding whether our current trajectory of residential solar adoption is energy- and carbon-efficient. In this paper, we reveal the hidden energy and carbon cost of the inequitable distribution of existing installations. Using US-based data on carbon offset potential—the amount of avoided carbon emissions from using rooftop PV instead of electric grid energy—and the number of existing solar installations, we surprisingly observe that locations and demographics with a higher carbon offset potential have fewer existing installations. For instance, neighborhoods with relatively higher black population have 7.4% higher carbon offset potential than average but 36.7% fewer installations; lower-income neighborhoods have 14.7% higher potential and 47% fewer installations; Republican-leaning states have 23.8% higher potential and 60.8% fewer installations. We propose several equity- and carbon-aware solar siting strategies that prioritize developing solar in certain areas based on their characteristics – these strategies may inform, for example, the development of targeted incentives. In evaluating these strategies, we develop SunSight, a toolkit that combines simulation/visualization tools and our relevant datasets, which we are releasing publicly. Our projections show that a multi-objective siting strategy can address two problems at once – namely, it can improve societal outcomes in terms of distributional equity and simultaneously improve the carbon-efficiency (i.e., climate impact) of current installation trends by up to 39.8%. 

We introduce and study spatiotemporal online allocation with deadline constraints (SOAD), a new online problem motivated by emerging challenges in sustainability and energy. In SOAD, an online player completes a workload by allocating and scheduling it on the points of a metric space (X,d) while subject to a deadlineT. At each time step, a service cost function is revealed that represents the cost of servicing the workload at each point, and the player must irrevocably decide the current allocation of work to points. Whenever the player moves this allocation, they incur a movement cost defined by the distance metricd(⋅, ⋅) that captures, e.g., an overhead cost. SOAD formalizes the open problem of combining general metrics and deadline constraints in the online algorithms literature, unifying problems such as metrical task systems and online search. We propose a competitive algorithm for SOAD along with a matching lower bound establishing its optimality. Our main algorithm, ST-CLIP, is a learning-augmented algorithm that takes advantage of predictions (e.g., forecasts of relevant costs) and achieves an optimal consistency-robustness trade-off. We evaluate our proposed algorithms in a simulated case study of carbon-aware spatiotemporal workload management, an application in sustainable computing that schedules a delay-tolerant batch compute job on a distributed network of data centers. In these experiments, we show that ST-CLIP substantially improves on heuristic baseline methods. 

As computing demand continues to grow, minimizing its environmental impact has become crucial. This paper presents a study on carbon-aware scheduling algorithms, focusing on reducing carbon emissions of delay-tolerant batch workloads. Inspired by the Follow the Leader strategy, we introduce a simple yet efficient meta-algorithm, called FTL, that dynamically selects the most efficient scheduling algorithm based on real-time data and historical performance. Without fine-tuning and parameter optimization, FTL adapts to variability in job lengths, carbon intensity forecasts, and regional energy characteristics, consistently outperforming traditional carbon-aware scheduling algorithms. Through extensive experiments using real-world data traces, FTL achieves 8.2% and 14% improvement in average carbon footprint reduction over the closest runner-up algorithm and the carbon-agnostic algorithm, respectively, demonstrating its efficacy in minimizing carbon emissions across multiple geographical regions.¹