Source-linked AI summary
COAST: Congestion-Aware Start-Time Recommendations for Carbon-Aware HPC Jobs
Weibin Feng, Abhishek Dasgupta, Zeynep Duygu Tekler, Sudha Ahuja, Jin Zheng, Xun Jiang
TL;DR
Carbon-aware shifting of flexible HPC jobs can create congestion when jobs independently target the same low-carbon periods. COAST coordinates target starts as an exact potential game balancing carbon, waiting, and congestion; under idealized realization, it reduces carbon by 15.4% and peak load by 23.8% versus carbon-greedy timing while retaining 92.8% of carbon-greedy’s savings.
Problem
Independent carbon-aware timing can concentrate flexible HPC jobs in the same periods, creating congestion and undermining the practical benefit of shifting.
Method
COAST coordinates micro-batched HPC start-time recommendations as an exact potential game solved through sequential strict best-response updates.
Results
15.4% carbon reduction versus Observed, 1.5 h added waiting, 23.8% lower peak load versus carbon-greedy, and 92.8% of carbon-greedy’s savings.
Takeaways & Limitations
COAST preserves most carbon savings while avoiding the peak-load increases produced by carbon-focused baselines.
Takeaways & Limitations
The evaluation assumes accepted recommendations start at their target slots and does not model queue delay or submission-time conversion, so reported gains are an upper bound for an external advisor.
Abstract
from arXiv · showhide
High-performance computing (HPC) workloads consume substantial amounts of electricity, and their carbon emissions vary over time with the carbon intensity of grid electricity. However, uncoordinated shifting of carbon-aware HPC jobs toward low-carbon periods can concentrate recommended start times in the same time slots, creating congestion and eroding the resulting carbon benefits. This paper proposes COAST, a congestion-aware start-time recommendation mechanism that coordinates HPC jobs by balancing carbon savings against additional job delay and congestion in recommended start-time slots. COAST formulates each decision batch as an exact potential game, enabling best-response updates to converge to stable start-time recommendations. Under an idealized realization model that assumes each job can start at its recommended time, we conduct trace-driven simulations using public HPC energy data and historical grid carbon-intensity traces. The results show that COAST reduces estimated carbon emissions by 15.4\% with 1.5 hours of additional average start delay. Compared with carbon-greedy start-time recommendations, COAST reduces peak time-slot load by 23.8\% while preserving 93\% of the achievable carbon savings. These results quantify the potential benefits of congestion-aware start-time coordination for flexible, carbon-aware HPC jobs.
I. INTRODUCTION
Carbon-aware shifting can reduce HPC emissions because grid carbon intensity varies over time, but independently choosing clean periods can concentrate jobs and create congestion. COAST coordinates flexible jobs by balancing carbon, waiting time, and overlapping execution demand.
- Temporal shifting can reduce operational emissions by moving flexible HPC workloads toward periods with lower grid carbon intensity.
- Independent recommendations for the lowest-carbon periods can concentrate jobs, increasing congestion and reducing the practical benefit of carbon-aware timing.
- COAST coordinates micro-batches by balancing predicted emissions, user waiting time, and additional execution-slot congestion.
- COAST models overlapping execution demand as a finite congestion game and uses sequential strict best-response updates to reach stable recommendations.
- Trace-driven simulations quantify the carbon–delay–load trade-off against observed and carbon-greedy timing baselines.
A. System Model and Assumptions
COAST operates as an external advisory layer that batches job intents, discretizes a 48-hour future horizon into 30-minute slots, and plans target start times without replacing the scheduler. The evaluation assumes accepted recommendations start at their targets and synthesizes missing workload parameters from trace data.
- COAST collects job intents in 5-minute micro-batches and plans target starts over a 48-hour rolling horizon.
- The planning horizon uses 30-minute execution slots aligned with the grid carbon-intensity trace’s settlement resolution.
- Each job is represented by arrival, resource demand, runtime, estimated energy, deadline, and maximum allowable delay.
- Trace evaluation takes arrival, CPU allocation, and runtime from workload data, while estimating energy with a linear per-core power model.
- Accepted recommendations are assumed to execute at their target starts, excluding queue delay and practical submission-time conversion.
- COAST uses a soft congestion penalty because the external advisory layer lacks full capacity, queue-state, and priority information.
B. Problem Formulation
The formulation assigns each job a feasible target start and evaluates carbon, waiting, and congestion costs over its execution interval. By counting overlapping occupied slots, it captures temporal load caused by full job runtimes rather than start slots alone.
- Each job selects one feasible target execution start within its horizon, arrival, delay, and deadline constraints.
- Carbon cost is computed from predicted grid carbon intensity across the job’s execution window using uniformly distributed estimated energy.
- The uniform-energy approximation can be replaced by more detailed power profiles without changing the game formulation.
- Waiting cost increases with the delay between job arrival and its target execution start.
- Execution-slot load includes background load from previously accepted recommendations and demand from jobs in the current strategy profile.
- The load model counts resource demand across every slot occupied during each runtime, capturing overlapping execution intervals.
C. Micro-Batch Congestion Game
COAST models each micro-batch as a game in which jobs choose feasible start times while accounting for carbon, waiting, and congestion effects. The formulation assigns each job a cost that combines normalized carbon emissions, waiting time, and marginal congestion.
- Each job is a player whose strategy is a feasible target start time, and its desirability depends on other jobs’ choices.
- COAST defines slot-level congestion potential over the execution slots occupied by jobs.
- A job’s marginal congestion increment is the cumulative increase in congestion potential across all slots it would occupy.
- Each job’s cost combines normalized carbon emissions, waiting time, and congestion, weighted by α, β, and γ.The normalization constants are fixed medians computed on training data and frozen thereafter.
D. COAST Algorithm and Theoretical Property
COAST initializes feasible recommendations and repeatedly applies strict best-response updates until no job can lower its cost. Its exact-potential structure guarantees finite convergence to a batch-level equilibrium, while the rolling procedure does not claim global or cross-batch optimality.
- COAST Algorithm: COAST initializes each job with a feasible start time and repeatedly selects a cost-minimizing feasible alternative.Ties retain the current start time, and updates occur only when they strictly decrease the job’s cost.
- COAST Algorithm: The algorithm outputs recommended start times after updates stop.
- Theoretical Property: COAST’s game is an exact potential game for every fixed micro-batch with fixed background load.A unilateral change in a job’s cost equals the corresponding change in the global potential.
- Theoretical Property: The marginal-contribution congestion construction yields an exact potential for any convex ψ and heterogeneous resource demands.The result does not rely on the affine special case associated with weighted congestion games.
- Theoretical Property: Sequential strict best-response updates converge in finitely many steps to a batch-level pure-strategy Nash equilibrium.
- Scope: The equilibrium is local to one micro-batch and its fixed background load, with no claim of global optimum or cross-batch equilibrium.The rolling procedure applies the algorithm batch by batch on accumulating committed load.
A. Experimental Setup
The evaluation combines a public HPC workload trace with historical carbon-intensity data in a rolling-horizon simulation. It compares COAST against observed and per-job carbon-aware baselines using carbon, delay, and load metrics on held-out windows.
- Workload Trace: The workload trace is the archive-cleaned CEA-Curie 2011 production HPC log, supplemented with modeled energy, deadlines, and maximum delays.The trace provides submission times, allocated cores, and runtimes but lacks per-job energy, deadline, and delay-tolerance records.
- Carbon Intensity Trace: The carbon trace contains 17,557 half-hourly UK Carbon Intensity API samples from 1 January 2025 to 1 January 2026.The samples range from 24 to 293 gCO2/kWh, with median 124; UK data are paired counterfactually because HPC-specific historical carbon data are unavailable.
- Simulation and Baselines: The rolling-horizon simulation solves 5-minute micro-batches over a fresh 48-hour horizon and honors accepted recommendations at execution.Baselines are Observed, Carbon-greedy, and Carbon-plus-waiting; only COAST couples jobs.
- Metrics and Split: The evaluation reports carbon, average added waiting, peak slot load, and load variance across 224 randomly split windows.There are 67 training and 157 held-out test windows, with weights and normalizers calibrated only on training data.
B. Evaluation Results
Across 157 held-out windows, COAST preserves most carbon savings while reducing congestion, with results stable across windows and workload–carbon alignments.
- Carbon–waiting–load trade-off: COAST reduces carbon emissions by 15.4% versus Observed while retaining 92.8% of carbon-greedy’s saving at 1.5 hours of added waiting.The 95% confidence interval for the reduction versus Observed is [13.7, 17.2].
- Carbon–waiting–load trade-off: COAST lowers peak execution-slot load by 20.2% versus Observed and 23.8% versus carbon-greedy, while carbon-focused baselines raise peak load.COAST also achieves the lowest load variance, about 35% below Observed.
- Effect of congestion awareness: Increasing γ yields greater peak reduction with only a small carbon loss and modest waiting increase; the selected γ = 5×10^-4 lies near the trade-off curve’s knee.The operating point was frozen after training-window selection.
- Reshaping load over time: COAST spreads jobs across nearby feasible slots rather than forming carbon-greedy spikes, explaining lower aggregate peak load and load variance.Figure 1 reports the lowest peak load and load variance for COAST among the compared methods.
- Stability and sensitivity: The reported improvements remain stable across the 157 held-out windows and are not sensitive to particular workload–carbon pairings.Crossed averages over 20 workload windows and 20 carbon alignments remain close to the main results.
- Algorithm overhead: Best-response updates converge rapidly: 98% of 25,116 non-empty micro-batches reach stable assignments in one sweep, with a mean of 1.02 sweeps.Each micro-batch solves in a median of 0.07 ms; these are indicative solver times, not an end-to-end runtime benchmark.
V. DISCUSSION AND LIMITATIONS
COAST is evaluated as an advisory layer under deliberately idealized assumptions, so its reported gains represent an upper bound. The evaluation also uses actual execution-window carbon intensity and estimates missing per-job energy for the CEA-Curie trace.
- COAST assumes accepted recommendations are realized at their target slots, without modeling queue delay or translating target start times into submission times.This makes the reported gains an upper bound for an external advisor.
- Candidates are scored using the actual carbon intensity of their execution window rather than forecasts.
- Because the CEA-Curie trace lacks per-job energy, the evaluation estimates it with a linear model.
VI. CONCLUSION
COAST coordinates flexible HPC start-time recommendations as an exact potential game, aiming to retain carbon savings while avoiding load concentration. Across 157 held-out workload days, it reduced carbon and peak load relative to carbon-agnostic and carbon-greedy timing.
- Coordination: COAST coordinates each micro-batch as an exact potential game, with best-response updates converging to stable start-time recommendations.The mechanism internalizes each job’s marginal contribution to congestion while balancing carbon savings and waiting.
- Results: 15.4% lower carbon than no shifting was achieved while reducing peak load by 20.2% versus no shifting and 23.8% versus carbon-greedy timing.The reported evaluation covers a year of grid carbon-intensity data and 157 held-out workload days.
- Evaluation: Table III reports per-window means over 157 held-out test windows, with 95% confidence-interval half-widths and weights frozen at α = 4, β = 1, γ = 5×10−4.The table caption defines Δ as method minus Observed and marks lower-is-better metrics with ↓.
- Evaluation: Figure 2 evaluates policies in the signed carbon–peak plane and examines how γ changes carbon reduction, peak change, and added waiting.The frozen γ = 5×10−4 point was selected on training data.
- Robustness: Table IV compares main 157-window results with crossed carbon-alignment means over 400 workload–carbon pairs.The crossed analysis provides a sensitivity comparison across carbon alignments.