Source-linked AI summary
The Sunk Carbon Fallacy: Rethinking Carbon Footprint Metrics for Effective Carbon-Aware Scheduling
Noman Bashir, Varun Gohil, Anagha Belavadi, Mohammad Shahrad, David Irwin, Elsa Olivetti, Christina Delimitrou
TL;DR
Carbon-aware scheduling metrics can include embodied emissions that are fixed before a job is assigned, raising the question of whether they guide operational decisions effectively. The paper evaluates scheduling metrics and examines relevant factors, including through an on-premise datacenter case study. It finds that embodied-inclusive metrics can increase total carbon footprint, while operational-carbon-focused scheduling reduced it in that case study.
Problem
Method
The paper evaluates carbon-aware scheduling metrics, examines factors affecting their outcomes, and uses an on-premise datacenter case study.
Source: §1 · p. 2Results
Embodied-inclusive metrics can increase total carbon footprint, while prioritizing operational emissions led to more carbon-efficient scheduling in the real-world case study.
Source: §1 · p. 2Takeaways & Limitations
For existing hardware, scheduling should target operational emissions separately from procurement decisions affecting embodied carbon.
Source: §5 · p. 8Embodied-carbon estimates are uncertain, which can make their use in scheduling risk suboptimal decisions.
Source: §3.2 · p. 4, §5 · p. 8
Abstract
from arXiv · showhide
The rapid increase in computing demand and its corresponding energy consumption have focused attention on computing's impact on the climate and sustainability. Prior work proposes metrics that quantify computing's carbon footprint across several lifecycle phases, including its supply chain, operation, and end-of-life. Industry uses these metrics to optimize the carbon footprint of manufacturing hardware and running computing applications. Unfortunately, prior work on optimizing datacenters' carbon footprint often succumbs to the \emph{sunk cost fallacy} by considering embodied carbon emissions (a sunk cost) when making operational decisions (i.e., job scheduling and placement), which leads to operational decisions that do not always reduce the total carbon footprint. In this paper, we evaluate carbon-aware job scheduling and placement on a given set of servers for a number of carbon accounting metrics. Our analysis reveals state-of-the-art carbon accounting metrics that include embodied carbon emissions when making operational decisions can actually increase the total carbon footprint of executing a set of jobs. We study the factors that affect the added carbon cost of such suboptimal decision-making. We then use a real-world case study from a datacenter to demonstrate how the sunk carbon fallacy manifests itself in practice. Finally, we discuss the implications of our findings in better guiding effective carbon-aware scheduling in on-premise and cloud datacenters.
1 INTRODUCTION
The paper argues that including already-incurred embodied carbon in scheduling can lead to decisions that increase a datacenter’s total carbon footprint. It separates scheduling from procurement and examines how infrastructure and utilization affect carbon-aware scheduling.
1 INTRODUCTION
- Scheduling should target operational emissions, while procurement decisions address embodied emissions that cannot be changed when jobs are assigned.Scheduling occurs continuously; procurement decisions occur over longer hardware-lifecycle intervals.Source: §1 · p. 2
- SCI-like scheduling metrics can ignore the embodied carbon of idle servers, creating incentives that may increase the datacenter’s overall footprint.The resulting accounting can overlook the broader carbon impact of the server fleet, including unused hardware.Source: §1 · p. 2
- Metric outcomes depend on infrastructure: server heterogeneity can make job-level lifecycle carbon diverge from the datacenter’s total footprint.At very high or very low utilization, scheduling metric choice has little impact; at intermediate utilization, SCI can lead to inefficient scheduling.Source: §1 · p. 2
- Metrics that include embodied and operational emissions can produce suboptimal schedules and increase a datacenter’s total carbon footprint.The paper identifies utilization, operational carbon intensity, and embodied-carbon amortization as factors that can exacerbate these outcomes.Source: §1 · p. 2
- In a real-world on-premise datacenter case study, prioritizing operational emissions led to more carbon-efficient scheduling under realistic workload assumptions.The paper evaluates three metrics, including metrics that prioritize operational emissions or account for infrastructure-wide embodied emissions more appropriately than SCI.Source: §1 · p. 2
- The paper recommends choosing metrics that reflect the carbon costs relevant to operational decisions to help optimize for a lower total footprint.The paper frames avoiding the sunk carbon fallacy as a practical guideline for datacenter operators and users.Source: §1 · p. 2
2 BACKGROUND AND MOTIVATION
Prior sustainable-computing work spans carbon accounting, hardware design, and workload scheduling, while proposed metrics allocate operational and embodied emissions in different ways. This section identifies limited analysis of the incentives and outcomes of those metrics and focuses this work on workload scheduling.
Prior work on sustainable computing
- Prior work studies computing’s environmental impact, carbon-accounting challenges, hardware lifecycle tradeoffs, workload shifting, and systems for carbon-aware scheduling.The passage notes that real-world use of carbon-aware optimizations is limited, with one hyperscaler example of workload shifting.Source: §2 · p. 3
Metrics for sustainable computing
- Proposed metrics include amortized sustainability cost, CCI for component replacement and end-of-life decisions, and SCI for software carbon-footprint decisions.These metrics address different decision contexts and carbon-accounting goals.Source: §2 · p. 3
Limitations and research gaps
- Few studies analyze the incentives and outcomes of carbon-accounting metrics, and a single metric spanning the computing lifecycle may not be feasible.This work focuses on workload scheduling in public-cloud and enterprise datacenters, a lifecycle stage with scheduling and procurement decisions on different timescales.Source: §2 · p. 3
3 THE SUNK CARBON FALLACY
Section 3 frames carbon-aware placement as minimizing emissions for a fixed job set and contrasts lifecycle metrics with choices focused on operational costs. It illustrates when SCI can raise cluster emissions and generalizes why oSCI can minimize total emissions across the fleet.
3.1 Analysis Setup
- The analysis schedules profiled jobs onto available, non-power-proportional servers with instantaneous placement, but its carbon estimates account only for processors.The setup draws inventory information from academic clusters and assumes known job performance and energy characteristics.Source: §3.1 · p. 3
3.2 Carbon-Aware Scheduling Metrics
- SCI combines a job’s operational and embodied emissions per functional unit; tSCI allocates infrastructure-level embodied carbon, while oSCI bases placement on operational emissions.The paper’s simplified oSCI calculation accounts only for energy used by the server running the job.Source: §3.2 · p. 4
- SCI requires embodied-carbon estimates that can be uncertain, while tSCI needs detailed, time-varying datacenter data and costly online attribution that end users may lack.The paper includes tSCI for completeness and says oSCI can achieve the same scheduling outcomes.Source: §3.2 · pp. 4–5
3.3 An Illustrative Example
- 24.10% higher cluster carbon results when SCI selects the energy-inefficient server, while tSCI and oSCI align on the minimum-footprint placement.The comparison reports total cluster emissions as all servers’ embodied carbon plus operational carbon from active servers.Source: Table 2 · p. 5, §3.3 · p. 5
- SCI can favor a server with lower embodied carbon despite worse energy efficiency, and its cluster-level penalty is greatest at intermediate utilization.In the example, the peak discrepancy occurs at 27.3% utilization; its magnitude and location vary with the server set, base power, and job granularity.Source: §3.3 · pp. 5–6
- The SCI penalty also depends on the embodied-to-operational carbon mix, server lifespan, and benchmark: the affected server pairs can change across performance benchmarks.At 0% embodied share, using SA raises system carbon by 48%; at 100%, operational carbon is zero and server choice does not matter. Longer expected lifespans lower amortized embodied carbon, magnifying the system-level impact of lower embodied values.Source: §3.3 · p. 6
3.4 Generalization of Outcomes
- oSCI minimizes operational carbon, the only cost reducible post-purchase, and ensures minimum total carbon across embodied and operational phases across the datacenter.When k is zero or equals N, SCI and oSCI select the same servers.Source: §3.4 · p. 6
4 CASE STUDY: AN ACADEMIC DATACENTER
The academic-datacenter case study finds that SCI-based scheduling can increase total carbon, with the increase depending on embodied-carbon amortization and cluster utilization. It evaluates this effect on a 740-server inventory using a one-time placement of jobs from a year-long trace.
4 CASE STUDY: AN ACADEMIC DATACENTER
- The inventory spans 740 servers and 15 processor types, averages 9.5 years old, and contains 17,633 KgCO2 of processor embodied carbon.The servers range from 3 to 17 years old.Source: §4 · p. 7
- With a five-year expected lifespan, the study either sets embodied carbon to zero for servers older than five years or amortizes it over years operational.Both approaches are used to examine how accounting choices affect server rankings and added carbon.Source: §4 · p. 7
- Placement uses a 2016 year-long trace of 14M jobs, filters jobs requiring over 32 cores, and assigns jobs once by submission time without trace replay.Multiple jobs can share a server, avoiding stranded resources; realistic trace replay is outside the study’s scope.Source: §4 · pp. 7–8
- Both embodied-carbon amortization approaches produce nearly 30% higher datacenter carbon under SCI; across the trace’s 40–80% utilization, SCI raises footprint by >15%.The figures compare added carbon under the two amortization approaches.Source: Figure 7 · p. 8, Figure 8 · p. 8, §4 · p. 8
- The first approach adds >5% at 27–78% utilization, while the second typically exceeds 10% across 13–80%.The first approach amortizes embodied carbon across the lifespan; the second does so during the first five years.Source: Figure 7 · p. 8, Figure 8 · p. 8, §4 · p. 8
5 IMPLICATIONS AND CONCLUSION
The paper recommends separating scheduling from procurement: use operational carbon to schedule already-procured servers, and use SCI when comparing hardware for replacement. It notes that accounting choices and uncertain embodied-carbon estimates can complicate lifecycle metrics.
5 IMPLICATIONS AND CONCLUSION
- SCI’s embodied-to-operational carbon ratio can shift with the chosen expected lifespan and embodied-carbon accounting approach.The paper notes these choices can perturb comparisons across servers.Source: §5 · p. 8
- SCI can favor older, lower-efficiency servers because their embodied carbon per score may be lower, potentially assigning them base demand and newer efficient servers to peaks.The paper describes this as a worst-case ordering.Source: §5 · p. 8
- Scheduling and procurement should be decoupled: use SCI to compare replacement options, but schedule procured servers to reduce operational costs.New hardware for capabilities unavailable on existing servers is described as SCI-agnostic.Source: §5 · p. 8
- tSCI’s uncertain embodied-carbon estimates risk incorrect resource allocation, and accurately tracking it at public-cloud scale incurs significant overhead.The paper says that overhead prohibits tSCI’s use in that setting.Source: §5 · p. 8
- For scheduling procured hardware, oSCI targets operational emissions, the component directly reducible after purchase, and avoids the sunk carbon fallacy.The paper says this applies to on-premise and cloud infrastructure.Source: §5 · p. 8