Carbon-Aware Temporal Shifting

• Carbon-aware temporal shifting is defined as the deliberate scheduling of delay-tolerant tasks to align execution with periods of lower carbon intensity, thereby reducing emissions without additional energy use.
• It exploits workload flexibility—using release times, deadlines, and interruptibility—and leverages dynamic carbon signals such as average and marginal intensities to optimize scheduling.
• Empirical studies demonstrate significant emission reductions (up to 80% in some cases) while highlighting trade-offs like forecast accuracy and potential deadline misses.

Searching arXiv for recent and foundational papers on carbon-aware temporal shifting and adjacent scheduling settings. Carbon-aware temporal shifting is the practice of deferring or advancing delay-tolerant computation or data movement so that execution occurs when electricity carbon intensity is lower, thereby reducing associated emissions without reducing total energy use. In the literature, the concept appears as time-flexible cloud scheduling, day-ahead datacenter load shaping, pause-and-resume execution, slack-aware federated training, transfer scheduling, and workflow interruption, all of which use time-varying grid signals such as average carbon intensity, marginal carbon intensity, or nodal carbon intensity to align energy use with lower-carbon periods (Wiesner et al., 2021, Radovanovic et al., 2021, Arputharaj et al., 10 Sep 2025, Goldverg et al., 2024, Lee et al., 3 Jul 2026).

1. Definition, scope, and relation to adjacent strategies

Temporal shifting uses execution-time flexibility rather than geographic relocation. In the cloud setting, it exploits release times, deadlines, and interruptibility to align energy consumption with periods of lower grid carbon intensity $CI(t)$, measured in $gCO_2/kWh$ or $kgCO_2e/kWh$ depending on the study. This differs from geo-shifting, which migrates work across regions to exploit spatial carbon differences, and from demand response optimized for electricity prices or grid stability rather than emissions (Wiesner et al., 2021, Sukprasert et al., 2023).

The distinction is operationally important. Temporal shifting avoids inter-region data movement, latency penalties, compliance issues, and dependence on transcontinental availability zones, whereas geo-shifting can achieve larger reductions when low-carbon regions are available but may incur network emissions and migration overheads (Wiesner et al., 2021, Guo et al., 18 Apr 2025). In network-intensive settings, temporal shifting also applies to data transfers: the same source–destination path can have materially different emissions at different times because the sender, receiver, and intermediate network segments are supplied by grids with changing carbon intensity (Goldverg et al., 2024).

The scope of the topic has broadened well beyond batch cloud jobs. Recent work treats temporal shifting as a general mechanism for federated learning rounds with slack time, scientific workflow interruption and resumption, inter-datacenter transfer scheduling, GPU scheduling for clinical AI, and AI data center control under distribution-system carbon signals (Arputharaj et al., 10 Sep 2025, West et al., 20 Aug 2025, Rodrigues et al., 4 Jun 2025, Doshi et al., 1 Jun 2026, Lee et al., 3 Jul 2026). This suggests that the central abstraction is not a specific scheduler, but a workload with admissible temporal slack and an external carbon signal of sufficient temporal variability.

2. Carbon signals, workload flexibility, and emissions accounting

The basic signal is the time-varying carbon intensity of electricity. In one widely used cloud formulation, emissions are computed as

$E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$

or, for job $i$,

$E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$

with $x_{i,t}\in\{0,1\}$ denoting assignment decisions (Wiesner et al., 2021). Scientific workflow studies use the equivalent discrete form

$C = \sum_t CI_t P_t \Delta t,$

and distinguish average carbon intensity from marginal carbon intensity, the latter representing emissions attributable to incremental demand (West et al., 17 Mar 2025, West et al., 20 Aug 2025). In power-distribution-aware AI data center control, the signal becomes nodal carbon intensity $w_{i,t}$, derived from a carbon emission flow model coupled to distribution-system power flow rather than from a regional average alone (Lee et al., 3 Jul 2026).

Workload flexibility is characterized by three recurring attributes. The first is the time window: jobs may have a release time $r_i$, a deadline $gCO_2/kWh$0, or a slack horizon $gCO_2/kWh$1. The second is execution structure: some jobs are non-interruptible and must occupy one contiguous interval, whereas others can be checkpointed, chunked, or resumed across multiple low-carbon slots. The third is issuance pattern: ad hoc jobs can usually only be deferred into the future, while scheduled jobs can often be moved earlier or later within a window (Wiesner et al., 2021, Arputharaj et al., 10 Sep 2025).

The same abstractions recur across domains. In federated learning, each client must complete $gCO_2/kWh$2 local rounds within an extended horizon $gCO_2/kWh$3, and permissible slots are chosen from the $gCO_2/kWh$4 lowest-carbon positions in that horizon (Arputharaj et al., 10 Sep 2025). In scientific workflows, tasks are atomic, but workflow-level pause and resume is feasible because intermediate data are persisted; interrupted scheduling therefore acts between task windows rather than within individual tasks (West et al., 20 Aug 2025). In data movement, the flexible object is not CPU time but bytes transferred over a deadline-constrained interval, optionally split across several low-carbon windows (Goldverg et al., 2024, Rodrigues et al., 4 Jun 2025).

3. Scheduling formulations and algorithmic patterns

A canonical continuous objective minimizes time-integrated emissions: $gCO_2/kWh$5 while a discrete cloud formulation minimizes

$gCO_2/kWh$6

subject to capacity, time-window, and continuity constraints (Wiesner et al., 2021). From this common core, the literature develops several scheduler families.

One family uses direct slot selection within a deadline window. Examples include greedy lowest-$gCO_2/kWh$7-first placement for nightly jobs, contiguous-interval heuristics for non-interruptible jobs, and chunk-based heuristics for interruptible jobs (Wiesner et al., 2021). A related online formulation is the pause-and-resume problem, where a workload must accept exactly $gCO_2/kWh$8 slots by deadline $gCO_2/kWh$9, carbon intensity is revealed sequentially, and switching between run and pause incurs cost $kgCO_2e/kWh$0. The proposed double-threshold algorithms are optimal among deterministic online algorithms for both the minimization and maximization variants under bounded inputs (Lechowicz et al., 2023).

A second family performs aggregate capacity shaping rather than per-job scheduling. Google’s Carbon-Intelligent Compute Management computes day-ahead Virtual Capacity Curves (VCCs), hourly caps on flexible reservations that preserve overall daily capacity while suppressing flexible execution in forecasted high-carbon hours (Radovanovic et al., 2021). In this model, higher-tier interactive services remain unaffected, and flexible throughput is protected through risk-inflated daily targets and chance constraints on power capping.

A third family integrates temporal shifting with workload-specific structure. In federated learning, a scheduler jointly chooses client participation and slot assignment under a carbon budget using an $kgCO_2e/kWh$1-fair objective, then corrects selection bias with U-FedAvg and mitigates temporal imbalance through a final full-participation fine-tuning phase (Arputharaj et al., 10 Sep 2025). In precedence-constrained data processing, $kgCO_2e/kWh$2 uses a threshold $kgCO_2e/kWh$3 that combines carbon intensity with precedence-driven task importance, ensuring that bottleneck tasks continue to run even in high-carbon periods while low-importance tasks are deferred (Lechowicz et al., 13 Feb 2025). In inter-datacenter transfers, LinTS formulates slot-level throughput allocation as a linear program and then maps throughput decisions to thread counts using a throughput–threads model (Rodrigues et al., 4 Jun 2025).

These formulations differ in decision granularity—start times, slot assignments, throughput rates, aggregate capacity caps, or pause/resume states—but they all operationalize the same principle: emissions are reduced when flexible demand is moved from higher-carbon to lower-carbon intervals, subject to workload and infrastructure constraints.

4. Empirical behavior across application domains

Reported benefits depend strongly on regional carbon-intensity variability, slack length, interruptibility, and forecast quality. In a simulation over Germany, Great Britain, France, and California in 2020, nightly non-interruptible 30-minute jobs achieved about $kgCO_2e/kWh$4 savings in Germany and about $kgCO_2e/kWh$5 savings in California with $kgCO_2e/kWh$6-hour windows and $kgCO_2e/kWh$7 forecast error, while a StyleGAN2-ADA-inspired machine learning workload achieved about $kgCO_2e/kWh$8 to $kgCO_2e/kWh$9 reductions under a semi-weekly interrupting strategy (Wiesner et al., 2021).

Domain	Setting	Reported outcome
Cloud workloads	Nightly jobs and a machine learning project	$E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$0-hour nightly shifting yielded about $E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$1 savings in California; semi-weekly interrupting scheduling yielded about $E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$2–$E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$3 reductions (Wiesner et al., 2021)
Federated learning	Slack-aware client and time-slot scheduling	With $E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$4 hours, clients with high CI variability saw up to $E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$5 emission reduction; $E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$6 of clients reduced at least $E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$7 (Arputharaj et al., 10 Sep 2025)
Inter-datacenter transfers	LinTS scheduling over 72 hours	Up to $E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$8 lower emissions versus the worst case and up to $E = \sum_t \mathrm{Energy}(t)\cdot CI(t),$9 lower emissions versus FCFS while meeting all deadlines (Rodrigues et al., 4 Jun 2025)
Scientific workflows	Seven Nextflow workflows	Temporal shifting was capable of decreasing emissions by over $i$0, and resource scaling capable of decreasing emissions by $i$1 (West et al., 20 Aug 2025)
Clinical AI GPU workloads	Night-window batching versus CUCA$i$2	The overnight rule closed about $i$3 of the average $i$4 gap from urgency-only to CUCA$i$5; CarbonShift let about $i$6 of the most urgent jobs miss their deadline (Doshi et al., 1 Jun 2026)

Operational deployments and system-level studies show more moderate but still material effects. In Google’s datacenter system, some clusters saw roughly $i$7 reductions in flexible load during peak-carbon hours, corresponding to about $i$8 power reductions in those hours, while a controlled campus experiment found about $i$9–$E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$0 average power reduction during the highest-carbon hours on shaped days (Radovanovic et al., 2021). For network-intensive data movement, hourly path-average carbon intensity on a UC→TACC path ranged from $E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$1 to $E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$2 over 51 hours, indicating nearly $E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$3 variation solely due to time-of-day and day-to-day grid conditions (Goldverg et al., 2024).

The literature therefore does not support a single magnitude claim. Instead, it documents a spectrum ranging from a few percent in low-variability grids or tight windows to much larger reductions when slack is long, interruption is feasible, or the underlying carbon signal has strong diurnal or weekly structure.

5. Uncertainty, practical limits, and recurring trade-offs

Temporal shifting is fundamentally limited by forecast quality, workload structure, and the carbon variability of the underlying grid. In the cloud experiments over 2020, the measured mean absolute error for National Grid ESO’s 48-hour carbon-intensity forecasts was about $E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$4, roughly $E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$5 of the annual mean, and simulated forecast noise reduced gains more strongly for interrupting strategies than for non-interrupting contiguous scheduling (Wiesner et al., 2021). More explicitly uncertainty-aware work shows that point forecasts can mis-rank cleaner and dirtier days; conformal prediction intervals achieved target coverages close to $E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$6, $E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$7, and $E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$8 across several regions and prevented a $E_i = \sum_t x_{i,t}\, e_i(t)\, CI(t),$9 increase in emissions in a temporal-shifting case study and a $x_{i,t}\in\{0,1\}$0 increase in a spatial case study for a 20 MW datacenter cluster (Li et al., 2024).

Upper-bound analyses indicate that ideal and practical outcomes differ sharply. Across 123 regions, temporal-only scheduling with perfect foresight and one-year slack achieved a global average reduction of about $x_{i,t}\in\{0,1\}$1, or about $x_{i,t}\in\{0,1\}$2 of the global mean CI, but with practical 24-hour slack the average temporal reduction fell to about $x_{i,t}\in\{0,1\}$3–$x_{i,t}\in\{0,1\}$4, or about $x_{i,t}\in\{0,1\}$5 of the global average. The same study argues that simple policies often yield most of these reductions and that the relative benefit of carbon-aware scheduling decreases as the energy supply becomes greener (Sukprasert et al., 2023).

A second recurring trade-off is that carbon minimization may conflict with other objectives. In clinical AI, carbon-first stress-test policies such as CarbonGreedy and CarbonShift produced low average $x_{i,t}\in\{0,1\}$6 but disrupted urgent scheduling, with CarbonShift allowing about $x_{i,t}\in\{0,1\}$7 of critical jobs to miss their deadlines in the simulator (Doshi et al., 1 Jun 2026). In WAN-aware geo-shifting, ignoring network emissions can reverse an apparently favorable migration decision: for network-heavy jobs, WAN carbon can exceed compute by $x_{i,t}\in\{0,1\}$8, and traceroute-only methods can underestimate path carbon by $x_{i,t}\in\{0,1\}$9–$C = \sum_t CI_t P_t \Delta t,$0 (Guo et al., 18 Apr 2025). More generally, temporal shifting improves carbon outcomes less than spatial shifting in some cloud studies, and when both are combined the incremental benefit of temporal adjustment is often positive but smaller than the spatial effect (Attenni et al., 9 Dec 2025).

These results caution against treating temporal shifting as a universally dominant control lever. Its effectiveness is contingent rather than absolute.

6. Deployment patterns and research directions

Operational guidance in the literature is comparatively consistent. Suitable workloads include short-running batch jobs, CI/CD, backups, ETL, periodic analytics, simulations, ML training and sweeps, reporting pipelines, and other jobs with explicit deadlines, known durations, and checkpointability (Wiesner et al., 2021). Production integration typically starts by attaching metadata such as release time, deadline, duration, and interruptibility to jobs, then extending schedulers or controllers to consume carbon forecasts. Concrete examples include Slurm and HTCondor extensions, Kubernetes CronJobs with flexible execution windows, carbon-aware queue delays for serverless systems, and scheduler-agnostic admission control through hourly capacity caps (Wiesner et al., 2021, Radovanovic et al., 2021).

Current research directions broaden the objective beyond operational carbon. One strand couples temporal shifting with 24/7 carbon-free energy procurement, modeling daily conservation windows and inter-site virtual links; in that setting, the costs of 24/7 CFE are reduced by $C = \sum_t CI_t P_t \Delta t,$1 EUR/MWh for every additional percentage of flexible load (Riepin et al., 2024). Another strand co-optimizes carbon with water and land-use footprints, finding that temporal shifting also decreases those footprints, though to a lesser extent than spatial shifting (Attenni et al., 9 Dec 2025). Reliability-aware formulations add embodied carbon and server lifetime degradation, achieving up to $C = \sum_t CI_t P_t \Delta t,$2 total carbon reduction in a heterogeneous data-center model that jointly schedules batch deferral, interactive migration, and backup allocation (Zhang et al., 1 Apr 2025).

Future work in the cited literature repeatedly returns to the same themes: multi-objective optimization over cost, peak power, and emissions; integration of temporal and spatial shifting; improved average and marginal carbon forecasting; explicit modeling of correlated forecast errors; production-grade middleware for declarative time flexibility and checkpointing; and fairness controls to prevent systematic deferral of the same jobs or tenants (Wiesner et al., 2021, Arputharaj et al., 10 Sep 2025, Li et al., 2024, Rodrigues et al., 4 Jun 2025). A plausible implication is that the mature form of carbon-aware temporal shifting will be neither a standalone heuristic nor a single optimization routine, but a scheduler interface that exposes temporal flexibility, forecast uncertainty, and system priorities as first-class control variables.