Towards Datacenter Environmental Sustainability Using Carbon Depreciation Models

TL;DR

This work addresses data-center sustainability by incorporating non-linear carbon depreciation to allocate embodied carbon ($C_{em}$) across hardware lifetimes and by accounting for secondary carbon from idle periods. It combines vendor-based embodied-carbon estimates (via SCARIF) with operational carbon costs ($C_{op}$) driven by grid carbon intensity ($CI_a$, $CI_i$) and per-job workloads, using depreciation rules to influence provisioning and scheduling. The findings show that non-linear depreciation (DB/DDB/SYD) can promote longer hardware lifetimes and higher utilization, reducing total carbon by about 28–57% depending on lifetimes, while including secondary carbon further incentivizes efficient use and discourages over-provisioning. These insights enable more sustainable datacenter provisioning policies and scheduler designs that balance QoS with environmental impact.

Abstract

Recently, the growing need for increasingly capable computing resources to be available on-demand has led to the prosperity of data centers. These data centers have led to several challenges and opportunities to address the environmental impacts from this computing resource. Conventional thinking has been concerned with minimizing energy usage of data centers to address sustainability. However, due to energy efficiency trends and renewable energy integration, recent evidence has demonstrated that embodied carbon is increasingly important and calls for improvements in data center provisioning strategies. In this paper we propose to adopt carbon depreciation models to better encourage the longer lifetime of hardware in the data center. Carbon depreciation models apply a higher proportion of embodied carbon to newly provisioned servers. This promotes provisioning fewer new servers to service jobs only with strict quality-of-service (QoS) constraints and extending lifetime of existing servers whose embodied carbon has already been mostly recovered. Along with carbon depreciation, we make the case that both embodied and operational carbon from server idle time must also be recovered during active jobs. This promotes provisioning strategies that maintain high rates of utilization. We show that prior carbon accounting strategies are counterproductive for sustainability with a greedy job scheduler that attempts to minimize carbon under QoS constraints as they price jobs as 25% cheaper on new versus old hardware. Our approach uses a greedy scheduler that prefers older hardware due to non-linear carbon depreciation promoting sustainable provisioning. Our approach reduces carbon by between 28--57% depending on assumptions for server lifetimes.

Figures (11)

• Figure 1: Greenhouse gas emissions from Meta Hyperscalars from 2014--2021. The initial concern of operational GHG emissions rise (black line) is projected to be mitigated by renewable integration (green line). Embodied GHG started to be reported in ICT only after 2019 (red line). The reverse projection (dotted red line) shows embodied carbon from ICT may dominate hyperscalar GHG emissions since 2014 Gupta-Tutorial22.
• Figure 2: Impact of different depreciation functions for embodied carbon $C_{em}$.
• Figure 3: The embodied and operational carbon cost per job under different throughputs. The carbon intensities of active and idle states are set to 0.188 and 0.019 kg/kWh.
• Figure 4: Carbon cost comparison in 2020 between year 4 V100 server and year 1 A10G server under 4 different depreciation models. The required throughput is set to 200 infer/s.
• Figure 5: Carbon cost comparison between estimations without and with considering the secondary carbon for year 4 V100 server. When the secondary carbon recovery is not considered, the embodied and operational carbon costs in different utilizations stay the same (blue bars). There is no incentive for users to prefer a high utilization of the server in this case. However, secondary carbon accounting promotes high server utilization (red bars).