arxiv: 2605.13489 · v1 · submitted 2026-05-13 · 💻 cs.DC

Recognition: unknown

Sustainable Graph Analytics Workload Scheduling with Evolutionary Reinforcement Learning in Edge-Cloud Systems

P. Ramicetty , H. Moore , S. Qi , A. Islam , M. Ghose , D. Milojicic , C. Bash , S. Pasricha

Authors on Pith no claims yet

Pith reviewed 2026-05-14 19:25 UTC · model grok-4.3

classification 💻 cs.DC

keywords sustainable schedulinggraph analyticsedge-cloud systemsevolutionary reinforcement learningcarbon emission reductionSLA violation minimizationworkload allocation

0 comments

The pith

A hybrid evolutionary reinforcement learning scheduler reduces SLA violations by up to 45 percent and carbon emissions by up to 12 percent for graph analytics in edge-cloud systems.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces MERSEM, a framework that combines evolutionary search with reinforcement learning to allocate and schedule graph analytics workloads across heterogeneous edge and cloud resources. Graph analytics support applications such as smart cities and IoT networks, yet their execution in mixed environments drives higher energy consumption and emissions due to varying carbon intensity and hardware differences. The evolutionary component generates diverse allocation options while the reinforcement learning component refines them locally to balance service-level agreement compliance against emission goals. A sympathetic reader would care because the reported gains suggest a concrete path to lower the environmental cost of scaling these workloads without separate hardware upgrades.

Core claim

MERSEM integrates evolutionary search with reinforcement learning to solve the problem of graph workload allocation and scheduling. The evolutionary component explores diverse global solutions, while the RL agent refines decisions through adaptive local optimization. The framework is designed to jointly minimize service-level agreement violations and carbon emissions by considering dynamic carbon intensity, resource heterogeneity, and workload characteristics.

What carries the argument

MERSEM, a multi-objective evolutionary reinforcement learning framework that pairs global exploration from evolutionary search with local policy refinement from reinforcement learning to produce workload allocations.

If this is right

Graph analytics workloads in edge-cloud environments can meet stricter service-level targets while lowering overall carbon output.
Hybrid evolutionary-RL schedulers achieve better joint optimization of delay and emission objectives than single-technique baselines.
Scheduling policies that incorporate real-time carbon intensity data produce measurable emission reductions under heterogeneous resource conditions.
The same allocation mechanism extends to dynamic workloads with varying graph sizes and arrival patterns typical of IoT and infrastructure monitoring.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same hybrid search-plus-learning structure could be tested on non-graph workloads such as model training or stream processing jobs.
If the simulation matches real conditions, the method may lower total operating costs in carbon-priced computing markets.
Physical deployment on hardware with live renewable energy signals would test whether the reported improvements survive outside controlled traces.
Coupling the framework with short-term carbon intensity forecasts could further reduce emissions by anticipating grid changes.

Load-bearing premise

The simulation environment used for experiments accurately captures real-world dynamics of carbon intensity, resource heterogeneity, and workload arrival patterns in edge-cloud systems.

What would settle it

Running MERSEM on a live edge-cloud testbed with measured carbon intensity traces and comparing actual SLA violation rates and emission totals against a baseline scheduler.

Figures

Figures reproduced from arXiv: 2605.13489 by A. Islam, C. Bash, D. Milojicic, H. Moore, M. Ghose, P. Ramicetty, S. Pasricha, S. Qi.

**Figure 1.** Figure 1: Edge-Fog-Cloud Environment. independent tasks can be executed in parallel at the same execution level. The degree of parallelism is determined by the DAG structure and the number of available CPU cores on the assigned VM. For communication modeling, we consider transmission delay for the input data of a job. 3.3 Latency Model To evaluate job execution performance in the edge-fog-cloud infrastructure, we m… view at source ↗

**Figure 2.** Figure 2: Overview of MERSEM framework. 4 Problem Formulation We consider an edge-fog-cloud computing environment for orchestrating job offloading. The system consists of a central cloud datacenter, geographically distributed fog datacenters, and a set of edge devices that generate computational jobs. Each generated job can be executed at one of the following locations: on the originating edge device, at one of the… view at source ↗

**Figure 4.** Figure 4: Comparison of (a) Carbon Emissions and (b) SLA [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 3.** Figure 3: Hyperparameter sensitivity analysis for the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

**Figure 5.** Figure 5: Comparison of (a) Carbon Emissions and (b) SLA [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 7.** Figure 7: Comparison of (a) Carbon Emissions and (b) SLA [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗

read the original abstract

Graph analytics powers modern intelligent systems such as smart cities, cyber-physical infrastructure, IoT security, and large-scale social networks. As these workloads scale in complexity, their execution in heterogeneous edge-cloud environments results in higher energy use and carbon emission footprint. To address this challenge, we propose MERSEM, a multi-objective evolutionary reinforcement learning framework for sustainable edge-cloud system management. MERSEM integrates evolutionary search with reinforcement learning (RL) to solve the problem of graph workload allocation and scheduling. The evolutionary component explores diverse global solutions, while the RL agent refines decisions through adaptive local optimization. The framework is designed to jointly minimize service-level agreement (SLA) violations and carbon emissions by considering dynamic carbon intensity, resource heterogeneity, and workload characteristics. Experimental results demonstrate that MERSEM outperforms the state-of-the-art with up to 45% SLA violation reductions and up to 12% carbon emission reductions.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MERSEM's evolutionary-RL hybrid for sustainable graph scheduling is a reasonable idea but its 45% SLA and 12% carbon claims rest on an unvalidated custom simulator.

read the letter

The main thing here is that the authors built MERSEM to combine evolutionary search with reinforcement learning for scheduling graph analytics workloads across edge-cloud systems. The evolutionary part explores broad options while the RL agent tunes local decisions, all aimed at cutting both SLA violations and carbon emissions under varying carbon intensity and resource mixes. That specific hybrid for this multi-objective sustainable scheduling task is what they present as new. The paper does a solid job framing the practical problem: graph workloads in smart cities, IoT, and similar settings are growing and their energy footprint matters, so treating carbon as a first-class objective alongside performance is timely. The high-level description of how the two techniques complement each other is clear enough to follow. The soft spots are in the evaluation. All the headline numbers—up to 45% fewer SLA violations and 12% lower carbon emissions—come from a custom simulator, yet the abstract supplies no details on workload traces, carbon intensity data sources, baselines, statistical tests, or any cross-check against real production patterns. Without sensitivity analysis on simulator parameters or external validation, those margins could easily be artifacts of how the environment was modeled rather than evidence of the algorithm itself. This leaves the central claims difficult to assess. The work is aimed at people doing scheduling and green computing in distributed systems. A reader interested in multi-objective RL ideas for heterogeneous setups might pick up useful framing, but anyone needing reproducible results would want more on the experimental controls. It deserves peer review because the problem area is relevant and the approach is coherent on its own terms; referees can push for the missing validation and baseline details that would make the results usable.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes MERSEM, a multi-objective evolutionary reinforcement learning framework for scheduling graph analytics workloads in heterogeneous edge-cloud systems. The method integrates evolutionary search for global solution exploration with an RL agent for adaptive local optimization, with the goal of jointly minimizing SLA violations and carbon emissions while accounting for dynamic carbon intensity, resource heterogeneity, and workload characteristics. The central claim, stated in the abstract, is that experimental results demonstrate MERSEM outperforming the state-of-the-art with up to 45% reductions in SLA violations and up to 12% reductions in carbon emissions.

Significance. If the experimental results hold after proper validation, the work would represent a meaningful advance in sustainable distributed computing. Graph analytics workloads are increasingly deployed in energy-sensitive edge-cloud environments supporting smart cities, IoT, and cyber-physical systems; a hybrid evolutionary-RL scheduler that demonstrably reduces both SLA violations and carbon footprint could influence practical system design and stimulate further research on multi-objective optimization under real-time carbon constraints.

major comments (2)

[Experimental evaluation] Experimental evaluation section: The headline quantitative claims (up to 45% SLA violation reductions and 12% carbon emission reductions) rest entirely on results from a custom simulator, yet no cross-validation of the simulator's carbon-intensity traces, resource-heterogeneity model, or graph-workload arrival process against external real-world measurements is reported. This omission directly undermines confidence in whether the reported margins reflect algorithmic improvement or simulator artifacts.
[Methodology] Methodology section: The integration of the evolutionary component with the RL agent is presented only at the level of high-level pseudocode without equations, sensitivity analysis, or ablation studies on simulator parameters (e.g., carbon-spike frequency or bursty-arrival rates). Without these, it is impossible to bound the risk that the observed gains are tied to untested assumptions in the simulation environment.

minor comments (1)

[Abstract] Abstract: The performance numbers are stated without any reference to the specific baselines, statistical tests, or workload traces employed, which would allow readers to assess the claims at a glance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript to provide additional validation, equations, and analyses as outlined.

read point-by-point responses

Referee: [Experimental evaluation] Experimental evaluation section: The headline quantitative claims (up to 45% SLA violation reductions and 12% carbon emission reductions) rest entirely on results from a custom simulator, yet no cross-validation of the simulator's carbon-intensity traces, resource-heterogeneity model, or graph-workload arrival process against external real-world measurements is reported. This omission directly undermines confidence in whether the reported margins reflect algorithmic improvement or simulator artifacts.

Authors: We agree that explicit cross-validation would increase confidence in the results. The simulator parameters are grounded in established models from the carbon-aware computing and edge-cloud literature. In the revision we will add a new subsection on simulator fidelity that includes direct comparisons against publicly available real-world carbon intensity traces and workload datasets, plus a limitations discussion on any gaps that remain. revision: yes
Referee: [Methodology] Methodology section: The integration of the evolutionary component with the RL agent is presented only at the level of high-level pseudocode without equations, sensitivity analysis, or ablation studies on simulator parameters (e.g., carbon-spike frequency or bursty-arrival rates). Without these, it is impossible to bound the risk that the observed gains are tied to untested assumptions in the simulation environment.

Authors: We accept that the current presentation is insufficiently detailed. The revised manuscript will replace the high-level pseudocode with explicit mathematical formulations of the evolutionary-RL integration, and will include new ablation studies together with sensitivity analyses on parameters such as carbon-spike frequency and bursty arrival rates to quantify robustness. revision: yes

Circularity Check

0 steps flagged

No circularity: performance claims rest on simulation experiments, not self-referential derivations

full rationale

The paper proposes the MERSEM framework combining evolutionary search with RL for graph workload scheduling to jointly minimize SLA violations and carbon emissions. All quantitative results (up to 45% SLA reductions, 12% carbon reductions) are obtained via comparative experiments inside a custom simulator; no equations, fitted parameters, or self-citations are presented that reduce these outcomes to inputs by construction. The derivation chain consists of algorithmic description followed by external validation against SOTA baselines, remaining self-contained without self-definitional loops or load-bearing self-citations.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities; the framework is described at the level of high-level integration of existing techniques.

pith-pipeline@v0.9.0 · 5480 in / 1026 out tokens · 31095 ms · 2026-05-14T19:25:30.193358+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages

[1]

Alsharif et al

M.H. Alsharif et al. 2025. Survey of energy-efficient fog computing: Techniques and recent advances.Energy Reports13 (2025), 1739–1763. doi:10.1016/j.egyr. 2025.01.039

work page doi:10.1016/j.egyr 2025
[2]

Chancerel et al

M. Chancerel et al . 2026. Life-cycle carbon assessment of intermittent Fog services fully powered by renewable energy sources. InProceedings of the 18th IEEE/ACM International Conference on Utility and Cloud Computing (UCC ’25). Association for Computing Machinery, New York, NY, USA, Article 24, 10 pages. doi:10.1145/3773274.3774279

work page doi:10.1145/3773274.3774279 2026
[3]

Deb et al

K. Deb et al. 2002. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation6, 2 (2002), 182–197. doi:10.1109/ 4235.996017

work page arXiv 2002
[4]

Montes de Oca, Thomas Stützle, and Marco Dorigo

K. Deb et al. 2014. An Evolutionary Many-Objective Optimization Algorithm Using Reference-Point-Based Nondominated Sorting Approach, Part I: Solving Problems With Box Constraints.IEEE Transactions on Evolutionary Computation 18, 4 (2014), 577–601. doi:10.1109/TEVC.2013.2281535

work page doi:10.1109/tevc.2013.2281535 2014
[5]

Etemadi et al

M. Etemadi et al. 2021. A cost-efficient auto-scaling mechanism for IoT applica- tions in fog computing environment: a deep learning-based approach.Cluster Computing24, 4 (Dec. 2021), 3277–3292. doi:10.1007/s10586-021-03307-2

work page doi:10.1007/s10586-021-03307-2 2021
[6]

Gupta et al

A. Gupta et al. 2024. Unmanned aerial vehicles in collaboration with fog comput- ing network for improving quality of service.International Journal of Communi- cation Systems37, 9 (2024), e5759. doi:10.1002/dac.5759

work page doi:10.1002/dac.5759 2024
[7]

Hogade, S

N. Hogade, S. Pasricha, and H. J. Siegel. 2022. Energy and Network Aware Work- load Management for Geographically Distributed Data Centers.IEEE Transactions on Sustainable Computing7, 2 (2022), 400–413. doi:10.1109/TSUSC.2021.3086087

work page doi:10.1109/tsusc.2021.3086087 2022
[8]

Huang et al

C. Huang et al. 2022. Resource Scheduling and Energy Consumption Optimization Based on Lyapunov Optimization in Fog Computing.Sensors22, 9 (2022). doi:10. 3390/s22093527

work page 2022
[9]

Islam, M

A. Islam, M. Ghose, and S. Pasricha. 2025. An RL-Based Framework for Task Offloading and Resource Allocation in Energy Harvesting-Based Multi-Access Edge Computing.IEEE Transactions on Network and Service Management22, 6 (2025), 5360–5374. doi:10.1109/TNSM.2025.3600109

work page doi:10.1109/tnsm.2025.3600109 2025
[10]

Moore, S

H. Moore, S. Qi, N. Hogade, D. Milojicic, C. Bash, and S. Pasricha. 2025. Sustainable Carbon-Aware and Water-Efficient LLM Scheduling in Geo-Distributed Cloud Datacenters. InACM Great Lakes Symposium on VLSI (GLSVLSI) (GLSVLSI ’25). Association for Computing Machinery, New York, NY, USA, 929–934. doi:10. 1145/3716368.3735301

work page arXiv 2025
[11]

Pournazari et al

J. Pournazari et al. 2025. A Carbon-Aware Task Offloading Framework for Sustain- able Fog Computing. In2025 IEEE International Conference on High Performance Computing and Communications (HPCC). IEEE Computer Society, Los Alamitos, CA, USA, 1077–1084. doi:10.1109/HPCC67675.2025.00156

work page doi:10.1109/hpcc67675.2025.00156 2025
[12]

S. Qi, D. Milojicic, C. Bash, and S. Pasricha. 2024. SHIELD: Sustainable Hybrid Evolutionary Learning Framework for Carbon, Wastewater, and Energy-Aware Data Center Management. InProceedings of the 14th International Green and Sustainable Computing Conference(Toronto, ON, Canada)(IGSC ’23). Association for Computing Machinery, New York, NY, USA, 56–62. do...

work page doi:10.1145/3634769 2024
[13]

S. Qi, D. Milojicic, C. Bash, and S. Pasricha. 2026. SHIELD-EB: Sustainable Hybrid Evolutionary-Boosting Framework for Carbon, Wastewater, and Cost- Aware Datacenter Management.IEEE Access14 (2026), 40878–40898. https: //doi.org/10.1109/access.2026.3674066

work page doi:10.1109/access.2026.3674066 2026
[14]

S. A. Ali Rezaee. 2020.Jobs (DAG workflow) and tasks dataset with near 50k job instances and 1.3 Millions of tasks.doi:10.5281/zenodo.4667690

work page doi:10.5281/zenodo.4667690 2020
[15]

S. Sinha. 2025. Global Cellular IoT Connectivity Tracker and Forecast

work page 2025
[16]

Sonmez et al

C. Sonmez et al. 2017. EdgeCloudSim: An environment for performance evalua- tion of Edge Computing systems. In2017 Second International Conference on Fog and Mobile Edge Computing (FMEC). 39–44. doi:10.1109/FMEC.2017.7946405

work page doi:10.1109/fmec.2017.7946405 2017
[17]

Wang et al

Y. Wang et al. 2024. A hybrid PSO and GA algorithm with rescheduling for task offloading in device–edge–cloud collaborative computing.Cluster Computing28, 2 (Nov. 2024), 18 pages. doi:10.1007/s10586-024-04851-3

work page doi:10.1007/s10586-024-04851-3 2024
[18]

Wu et al

Q. Wu et al. 2024. Delay-Sensitive Task Offloading in Vehicular Fog Computing- Assisted Platoons.IEEE Transactions on Network and Service Management21, 2 (2024), 2012–2026. doi:10.1109/TNSM.2023.3322881

work page doi:10.1109/tnsm.2023.3322881 2024