acceptodds
Under review as a conference paper at ICLR 2027

GENO: Diversity-Preserving Green Optimization for Multi-Objective AI Service Orchestration

Abstract

The proliferation of AI inference services across edge and cloud infrastructures has elevated energy consumption and carbon emissions to critical concerns in service orchestration. On the AI-intensive five-objective formulations we study, conventional decomposition-based multi-objective optimizers exhibit what we term low-energy diversity collapse: neighborhood-restricted updates systematically under-explore the energy-efficient region of the Pareto front spanning inference energy, transport energy, carbon, latency, and reliability. To address this, we propose GENO, a green-aware service orchestration framework built around Decomposition–Dominance Coordination (DDC), a mechanism that synchronizes a dominance archive with the decomposition population via energy-priority redundancy removal and re-association. DDC is complemented by an energy-augmented state encoder, a five-objective formulation that partitions total energy into AI-inference and non-AI components, and a bounded weight-perturbation interface instantiated through a local language model. We establish two formal guarantees: DDC monotonically improves a green-weighted archive score (Proposition 1), and any preference-channel perturbation is Lipschitz-bounded in its effect on the scalarized objective (Proposition 2). Experiments on three service orchestration benchmarks show that, among the eight methods evaluated, GENO is the only one to reach a low-energy operating point (energy-per-request below 0.6 Wh and tail latency under 22 ms) on the AI-EO benchmark, improving energy-per-request over the strongest baseline by roughly a quarter to a third. Component ablation attributes −21.1% HV to DDC removal versus only −2.8% to the language interface, confirming that the gains are algorithmic.

Primary area:
optimization
Keywords:
Green Service Orchestration, Multi-Objective Optimization, AI Inference Energy, Carbon-Aware Computing

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_vnY2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper considers the problem of multi-objective service orchestration of ML and AI workloads, considering the energy cost of orchestration. The proposed GENO framework uses decomposition-dominance coordination (DDC) along with an energy-augmented state encoder to encourage finding a Pareto optimal solution that is energy optimal.

Critical strengths

The paper puts forward an interesting argument that existing methods leveraging multi-objective service orchestration result in diversity collapse, potentially resulting in a conservative exploration of the solution space.

Critical weaknesses

The novelty of the paper is unclear, both in terms of the problem considered and in terms of the methodology.

The main claim made in the abstract that the proposed method results in a monotonic improvement of the green-weighted objective is never referenced again and is not supported by evidence provided in the paper.

The use of the Tchebycheff scalarizing function to scalarize multiple objectives (5) is far from novel. Yet, the paper does not discuss the related works.

The problem description as part of the system model is not well-motivated. The system model simply states the objectives (2)-(3) without any motivation.

The description of the methodology is vague; key details of the steps in the proposed algorithm, such as decision pruning and complexity of solving the optimization, have been omitted.

The writing at several places is very verbose but at the same time the paper does not introduce the concepts or the terms discussed, making it extremely hard to follow the arguments.

The experimental evaluation does not discuss the evaluation methodology. This makes it hard to assess the relevance of the evaluation.

Questions for authors

Does the system model account for the CPU energy consumption induced by AI workloads? Eqn (1) states that only non-AI workloads lead to CPU utilization.

Why does the choice of δ=10−5\delta = 10^{-5} guarantee properly Pareto-optimal minimizers (line 198)? This claim appears to be unsupported.

Why does the proposed objective lead to a monotonic improvement? How does this differ from existing methods?

What was the evaluation methodology for the results presented in Section 5? What were the considered workloads, the load imposed, and the requirements - latency, resource utilization, etc?

Reviewer_KGfEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper treats the AI inference services across edge and cloud infrastructures as a multi-objective optimization problem with five objectives: energy, transport energy, carbon, latency, and reliability. It then uses multi-objective evolutionary algorithms to deal with it, with the assist of language model.

Critical strengths
  1. An emerging optimization problem widely used in various applications.

  2. It is good to see the authors to consider five objectives, making a challenge, but relevant optimization scenario.

Critical weaknesses
  1. One critical issue lies in the novelty of the proposed method. One major contribution is the proposed dominance and decomposition during the evolutionary search process. However, the idea of combing dominance and decomposition in the design of MOEAs has been widely used and seen since the beginning of 2010s, such as in NSGA-III, MOEA/DD, bi-criteria evolution.

  2. One novelty claimed in the paper is that the proposed method can locate energy-efficient region of the Pareto front, compared to standard MOEAs. However, this is not new, just preference-based evolutonary multi-objective optimization, where the decision-maker prefers solutions with good quality in terms of energy.

  3. The authors explain the motivation of why needs language models by stating that 'Operators may hold high-level preferences that are hard to encode as fixed coefficients—for instance, temporarily favoring low-carbon nodes during high renewable availability.' However, it seems to me such preferencs can be easily encoded into the search process of MOEAs. Actually, when applying to a real-world optimization scenario, MOEAs are expected to incorporate such problem-specific knowledge, which has been widely implemented in various applications.

  4. I don't understand why you need formal guarantees, e.g., proposition 2. How are they useful and relevant for your algorithm?

Questions for authors

I would like the authors to respond to the four weaknesses aboce, along with one additional question below.

How did you set the reference point for the HV calculation?

Reviewer_UDGkOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GENO, a green-aware multi-objective optimization framework for AI service orchestration that jointly optimizes five objectives: AI inference energy, non-AI/transport energy, carbon emissions, latency, and reliability. The core technical contribution is Decomposition–Dominance Coordination (DDC), which couples a decomposition-based evolutionary population with a non-dominated archive using energy-prioritized redundancy removal and energy-priority re-association to prevent what the authors call low-energy diversity collapse. The method is complemented by an energy-augmented state encoder, a clean partitioning of total energy into inference and non-inference components, and an optional, bounded preference-perturbation interface (instantiated by a local LLM) with a Lipschitz bound on its impact; experiments on three benchmarks suggest consistent hypervolume improvements and substantial energy reductions versus seven baselines.

Critical strengths

Novelty

  • The DDC mechanism is a well-motivated and concrete algorithmic that synchronizes decomposition-based subproblems with a dominance archive and uses energy-prioritized rules to preserve diversity in the green region of the Pareto front.
  • The five-objective formulation that explicitly partitions total energy into inference and non-inference components is clean and avoids double-counting, which is often a pitfall in systems-oriented energy accounting.
  • The energy-augmented state encoder that includes hardware energy efficiency and grid carbon intensity is a sensible addition that aligns representation with the optimization objectives.
  • The bounded weight-perturbation interface with a provable Lipschitz effect provides a principled safety property for policy-guidance channels, separating core optimization from preference nudges.

Experimental rigor and validation

  • Comparative evaluation across three benchmarks with seven baselines under matched function-evaluation budgets and statistical testing (Wilcoxon with Holm–Bonferroni) is thorough.
  • The component ablations are detailed and attribute most gains to DDC and energy-aware modeling, with the language interface contributing minimally, which strengthens the causal interpretation.
  • Sensitivity and scaling analyses (carbon noise, network size, workload growth) are included, and convergence behavior is reported.

Clarity of presentation

  • The problem formulation, objective decomposition, and the DDC mechanism are clearly described, with algorithmic pseudo-code and explicit constraints.
  • The role of the LLM advisor is clearly scoped as optional and bounded, avoiding over-claiming.

Significance of contributions

  • The paper tackles a timely and important problem—energy and carbon-aware orchestration of AI services—and shows non-trivial energy reductions while preserving strong Pareto quality.
  • The identification of a failure mode (low-energy diversity collapse) in standard decomposition-based MOEAs for AI-intensive workloads is a useful observation for the community.
Critical weaknesses

Technical concerns

  • All results are based on simulations with synthetic workloads and carbon traces; there is no real-system evaluation and only limited external validation of the energy/carbon models.
  • The reliability objective (negative log of a product of independent link/node reliabilities) and the latency model (sum of components) may not capture tail latency or correlated failures; yet claims highlight “tail latency ≤ 22 ms,” which seems mismatched with the O4 definition.
  • The formal propositions, while correct, are relatively weak: Proposition 1 states a monotonic effect on an occupancy-weighted archive score due to a pruning heuristic; Proposition 2 is a straightforward Lipschitz bound from the Tchebycheff scalarization.

Experimental gaps or methodological issues

  • Missing comparisons to alternative diversity-preserving archivers such as solution-space bounded archives (e.g., HDAA/JDAA) or hypervolume-based archivers that might also mitigate collapse (see 2602.04745).
  • The RA step’s computational overhead scales with the size of the population-plus-archive set; runtime and complexity impacts under large archives are not reported.
  • Details on weight-vector design (distribution over the simplex, coverage of energy/carbon-dominant directions) and normalization across objectives are under-specified; HV/IGD reference points and normalization choices could affect conclusions.
  • The routing model fixes shortest paths and largely omits queueing/traffic dynamics; this may limit realism in transport energy/latency-carbon coupling.

Presentation issues

  • Hyperparameters for DDC (ϵ\epsilon RR, κ\kappa) and RA tie-breaking (∑k≤3λk≥0.5\sum_k \le 3 \lambda k \ge 0.5) are given but their sensitivity is not deeply explored; guidance for practitioners is limited.

Missing related work or comparisons

Questions for authors
  1. Can you provide a small-scale real-system validation (e.g., on a heterogeneous edge cluster with actual power/energy measurements and real carbon traces) to corroborate the simulation claims?
  2. How sensitive are results to ϵ\epsilon RR, κ\kappa, and the green-floor threshold ∑k≤3λk≥0.5\sum_k \le 3 \lambda k \ge 0.5? Please include a sensitivity study and practical guidance for setting these hyperparameters.
  3. Reliability is modeled as a product over nodes/links with independence assumptions. How robust are conclusions if correlations or stateful failures are introduced?
  4. How does fixing routing to shortest paths affect transport energy/latency trade-offs? Would adding routing decisions change the observed benefits or interact with DDC?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.