PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM-Based Multi-Agent Collaboration for Constrained Multi-Objective Container Placement

Suvarthi Sarkar, Abhishek Hazra, Akshat Mittal, Aryabartta Sahu

arXiv:2610.07030v1cs.DC

TL;DR

Power-aware container placement must balance affinity with power use under hard feasibility constraints, amid combinatorial and sequential decisions. The paper proposes a four-agent collaborative framework for placement. It reports that the framework outperforms heuristic and meta-heuristic baselines on power and affinity.

  • Problem

    Power-aware placement is combinatorial and sequential, with nonlinear power models and potentially conflicting power and affinity objectives.

    Source: §I · pp. 1–2

  • Method

    A four-agent framework uses objective-specific debate, Tolerance-Gated Arbitration, and post-placement refinement over shared deterministic candidate tables.

    Source: §I · p. 2

  • Results

    The framework outperforms existing heuristic and meta-heuristic baselines on both power consumption and affinity payoff.

    Source: §VII · p. 10

  • Takeaways & Limitations

    The theoretical results identify conditions under which the framework Pareto-dominates greedy heuristics on power and affinity.

    Source: §I · p. 2

    Reducing inference cost and latency remains a future-work target.

    Source: §VII · p. 10

Abstract

from arXiv · show

Container placement in data centers must simultaneously minimize power consumption and maximize affinity preferences, while satisfying multi-resource capacity and anti-affinity constraints. Traditional approaches typically rely on fixed rules, which lack adaptability to dynamic cluster states and are difficult to extend for adaptive decision-making. On the other hand, meta-heuristic methods, although more flexible, are often computationally expensive, slower and prone to getting trapped in local optima. In this work, we propose an adaptive and efficient approach based on an LLM-driven multi-agent collaboration framework, where four specialized agents operate in a closed-loop ReAct cycle at each placement step. A Power Consumption Agent and an Affinity Agent debate over competing objectives, while a Placement Agent resolves conflicts using Tolerance-Gated Arbitration (TGA). A Rearrangement Agent further refines decisions through the Monotone Improvement Rule (MIR) in a post-placement refinement phase. All the agents reason over deterministic, feasibility-filtered candidate tables provided by the environment, ensuring that every proposed action inherently satisfies hard constraints. Based on experimental evaluation using the Google Cluster Trace with a configuration of 100 applications and 25 machines, the proposed framework consumes 8.36% less power and achieves 39.30% higher affinity than the power-greedy baseline. Ablation studies confirm that each architectural component, the multi-round debate mechanism, the TGA in the Placement Agent, and the MIR in the Rearrangement Agent, contributes meaningfully to the overall performance, with removal of any single component degrading both power consumption and affinity satisfaction.

I. INTRODUCTION

Container placement must balance power consumption and affinity while meeting hard feasibility constraints, but fixed-rule and meta-heuristic approaches have limitations. The paper proposes a four-agent collaborative framework and positions it against prior placement and LLM-scheduling work.

I. INTRODUCTION

  • Placement is combinatorial, sequential, and governed by nonlinear power models, so early assignments can lock in weaker power–affinity trade-offs.The objectives can conflict, making Pareto-efficient trade-offs relevant alongside strict feasibility.Source: §I · pp. 1–2

II. LITERATURE REVIEW

  • Prior approaches trade adaptability or quality against efficiency: heuristics can be rigid, meta-heuristics slow or locally trapped, and learning-based schedulers require training and tuning,, –.ILP approaches can scale poorly with nonlinear power models or large clusters.Source: §I · p. 2, §II · p. 2

I. INTRODUCTION

  • Four specialized agents debate, arbitrate with Tolerance-Gated Arbitration (TGA), and refine placements with the Monotone Improvement Rule (MIR), using feasibility-filtered candidates.The Power Consumption and Affinity Agents debate; the Placement Agent arbitrates, while the Rearrangement Agent refines placements.Source: §I · pp. 1–2
  • Theoretical results bound TGA's power overhead relative to power-greedy and show TGA with MIR achieves affinity at least as high as any power-greedy heuristic.The paper identifies conditions for Pareto dominance on both power and affinity.Source: §I · p. 2
  • The framework achieves the best power–affinity trade-off across the evaluated heuristic, meta-heuristic, and alternative LLM approaches.The evaluation uses the Google Cluster Trace and includes ablations of debate rounds, candidate-list size, tolerance, and LLM choice.Source: §I · p. 2

II. LITERATURE REVIEW

  • Prior LLM schedulers and resource-management frameworks lack explicit separation of competing objectives or address neither hard-feasibility combinatorial optimization nor formal per-step quality guarantees –,.The paper adapts structured multi-agent debate to constrained placement optimization.Source: §II · pp. 2–3

III. PROBLEM FORMULATION

The formulation represents placement as successive assignments of container instances to machines, with resource demands, machine capacities, and affinity and anti-affinity data. It defines hard feasibility conditions and a nonlinear power model.

III. PROBLEM FORMULATION

  • Placement is modeled as a sequence of assignments indexed by step t, over machines and unplaced container instances with resource-demand vectors.Each machine has capacities for CPU, memory, I/O, and bandwidth; instances have corresponding demand vectors.Source: §III · p. 3

A. Anti-affinity (hard constraint)

B. Affinity (soft preference)

  • The affinity matrix G encodes preferred instance–machine assignments, with violations reducing total payoff.G(i,m)=1 indicates that placing instance i on machine m is preferred.Source: B. Affinity (soft preference) · p. 3

C. Capacity (hard constraint)

D. Power Consumption model

  • The power model uses machine idle and peak power, total load, and utilization ratios to calculate power.The total power is formed from machine-level power terms.Source: D. Power Consumption model · p. 3
  • The load factor combines a cubic CPU-utilization term with linear contributions from the other resources; doubling CPU utilization raises its power component eightfold.The model sets α1 = 0.70, α2 = 0.10, α3 = 0.05, and α4 = 0.15 following,.Source: D. Power Consumption model · p. 3

E. Objectives

The framework minimizes power as its primary objective and maximizes affinity as its secondary objective, while a closed-loop agentic process makes feasible placement decisions from environment-maintained candidate tables.

E. Objectives

  • The framework minimizes power as the primary objective and maximizes affinity as the secondary objective over feasible placements.Feasibility requires satisfying the specified hard constraints.Source: E. Objectives · p. 3

IV. LLM-BASED REACT MULTI-AGENT PLACEMENT FRAMEWORK

  • At each step, agents reason over the current cluster state and produce a structured placement action, while the environment enforces hard constraints and updates metrics.The loop ends when all instances are placed, no feasible candidate remains, or the judge returns Stop.Source: §IV · p. 3, Figure 2 · p. 4

A. Environment-constructed Data Structures

  • The environment filters application–machine pairs for anti-affinity and resource capacity, then sorts feasible rows by incremental power and affinity.After placement, it updates the table to reflect changed utilization and removes the placed instance’s options.Source: A. Environment-constructed Data… · p. 4
  • Because agents select from the feasibility-filtered table, they cannot propose an infeasible placement.The candidate table shrinks as applications are placed and machines fill.Source: A. Environment-constructed Data… · p. 4

2) Scratchpad:

The scratchpad supplies recent trajectory context, while the agents use shared state and role-specific instructions to produce structured decisions.

2) Scratchpad:

  • The scratchpad records the current step, placement and swap counts, and recent placement records with instance, machine, affinity, and incremental power.It provides compact trajectory context to the agents.Source: 2) Scratchpad · p. 4

B. Agent Roles

  • Each of the four LLM-prompted agents receives shared state, role-specific instructions, and dynamic inputs, then outputs one structured action.Source: B. Agent Roles · p. 4

1) Power Consumption Agent:

  • The Power Consumption Agent reads the candidate table and scratchpad, then returns ranked placement recommendations with written arguments.From the second debate round onward, it also reads the Affinity Agent’s preceding argument.Source: 1) Power Consumption Agent · p. 4

2) Affinity Agent:

The Affinity Agent promotes affinity-preferred candidates, the Placement Agent arbitrates within a power-tolerance gate, and the Rearrangement Agent accepts only non-worsening refinements.

2) Affinity Agent:

  • The Affinity Agent ranks candidates with G(i, m) = 1 first, breaks ties by lowest incremental power, and uses power ranking alone if none exist.Across debate rounds, it adapts its ranking in response to the Power Consumption Agent’s feedback.Source: 2) Affinity Agent · pp. 4–5

3) Placement Agent:

  • Tolerance-Gated Arbitration selects among feasible candidates near the minimum-power option, prioritizing affinity and using debate recommendations to guide the final choice.At ε = 0, the rule reduces to power-greedy selection; increasing ε admits more affinity-preferred candidates.Source: 3) Placement Agent · p. 5

4) Rearrangement Agent:

  • The Rearrangement Agent accepts at most one Move or Swap after each placement, and only when power falls without affinity loss or affinity rises without a power increase.This Monotone Improvement Rule produces a feasible, monotonically improving refinement trajectory.Source: 4) Rearrangement Agent · p. 5

C. Debate Protocol

The paper contrasts fixed heuristics with adaptive decisions and analyzes how TGA and MIR bound power cost while preserving affinity relative to power-greedy placement.

C. Debate Protocol

  • The debate protocol runs for N configurable rounds, with each agent producing ranked recommendations and arguments in every round.The transcript accumulates 2N blocks.Source: C. Debate Protocol · p. 5

A. Heuristics as Fixed-Rule Policies

  • Fixed-rule heuristics do not adapt their decision logic to changing cluster states and make assignments permanent once placed.The paper contrasts this with policies that revise decisions through moves or swaps.Source: A. Heuristics as Fixed-Rule Policies · p. 5

B. Why the Environment Demands Adaptive Rules?

  • Under the cubic power model, marginal placement cost grows quadratically with utilization, making a rule that works at low utilization potentially poor at high utilization.The cited sensitivity rises from 0.084 at utilization 0.2 to 1.344 at utilization 0.8, a 16× increase.Source: B. Why the Environment Demands… · p. 5

C. TGA as an Adaptive Policy

  • TGA adapts through a tolerance gate that widens as machines become more heavily loaded, while debate arguments reflect the changing candidate table and scratchpad.Source: C. TGA as an Adaptive Policy · p. 6

D. Formal Guarantees of Agentic Collaboration Architecture

  • TGA’s per-step power premium is bounded relative to the cheapest candidate, and MIR recoveries are subtracted from the accumulated overhead bound.Source: D. Formal Guarantees of Agentic… · p. 6
  • TGA with MIR guarantees total affinity at least as high as power-greedy placement, because TGA never chooses lower-affinity candidates and MIR does not reduce affinity.Source: D. Formal Guarantees of Agentic… · p. 6
  • If MIR recovers at least the power TGA spends on affinity, the approach achieves both lower power and higher affinity than greedy.The paper identifies this condition as a strict Pareto improvement.Source: D. Formal Guarantees of Agentic… · p. 6

E. Why Greedy Heuristic can not match the Agentic Collaboration Architecture?

Greedy placement cannot trade a bounded power increase for affinity or revise earlier assignments. TGA and MIR provide those two capabilities, respectively.

E. Why Greedy Heuristic can not match the Agentic Collaboration Architecture?

  • Greedy always minimizes immediate power, so it rejects affinity placements that cost even slightly more and loses those opportunities permanently.TGA can accept an affinity candidate when its power cost falls within the tolerance gate.Source: E. Why Greedy Heuristic can not match… · p. 6
  • Greedy assignments are final, whereas MIR can move an application to a cheaper machine after later placements increase its power cost.MIR can recover the placement's power cost through rearrangement.Source: E. Why Greedy Heuristic can not match… · p. 6

VI. EXPERIMENTAL EVALUATION

The evaluation examines placement quality, debate-parameter sensitivity, and token-cost efficiency across several LLM configurations.

VI. EXPERIMENTAL EVALUATION

  • The evaluation measures placement quality against existing baselines, sensitivity to debate parameters, and token-cost efficiency.The main experiments use Llama 3.1-8B; ablations and tuning use Llama 3.1 locally, with additional evaluations using DeepSeek-R1 and Claude models.Source: §VI · p. 7
  • Ablations and hyperparameter tuning use Llama 3.1 on a local GPU server through Ollama.The main experiments use Llama 3.1-8B.Source: §VI · p. 7
  • The architecture is also evaluated with DeepSeek-R1 and with Claude Opus 4.6 and Sonnet 4.6 in extended-thinking mode.The Claude evaluations use API-based inference.Source: §VI · p. 7

A. Dataset and Experimental Setup

In the Google Cluster trace evaluation, the proposed approach improves the power–affinity trade-off over the listed baselines. A four-application example traces agent arbitration and rearrangement.

A. Dataset and Experimental Setup

  • The setup uses Google Cluster workload data with synthetically sampled anti-affinity and affinity pairs, 100 applications, 25 machines, and three independent runs.Thirty percent of application–machine pairs are forbidden and 20% preferred; results are averaged over three runs with reported variance below 0.05% for power and 0.10% for affinity satisfaction.Source: A. Dataset and Experimental Setup · p. 7

B. Comparison against Existing Heuristic Approaches

  • 8.36% less power and 39.30% higher affinity than P-Greedy; the proposed approach also uses 7.12% less power and achieves 10.85% higher affinity than HOP-CAPM, and 4.26% less power and 1.83% higher affinity than CPAAP.The comparison uses P-Greedy as the base.Source: B. Comparison against Existing… · p. 7, Table I · p. 7

C. Example for Illustration

  • In the four-application example, the agents debate placement options, TGA arbitrates, and MIR updates placements after each assignment.The example uses N=1, K=1, and ε=0.10; Table II presents application demands, machine state, affinity, and final placements.Source: C. Example for Illustration · p. 7, Table II · p. 8

1) Placement trace:

  • At t=2, MIR swaps app2 and app3, reducing power by 0.71 W; after t=3, another swap saves 2.21 W and gains one affinity point.The t=1 gate rejects the 6.86 W affinity recommendation because it exceeds the 5.46 W threshold; at t=3, the two policies have different candidate tables.Source: 1) Placement trace · p. 8

2) Final comparison:

The four-application comparison favors the proposed framework on both power and affinity. Architecture comparisons and tolerance tests report the trade-offs associated with coordination, refinement, and ε.

2) Final comparison:

  • 178.28 W and 2 affinity-satisfied placements for the proposed framework, versus 178.64 W and 1 for power-greedy.The proposed framework's example outcome is 0.36 W lower in cluster power and has one additional affinity point.Source: 2) Final comparison · p. 8

D. Comparison of Different LLM-based Architecture

  • The full architecture provides the best overall balance between power and affinity; simpler architectures use fewer tokens and LLM calls but have weaker placement quality.Table IV compares LLM-agentic architectures against P-Greedy.Source: D. Comparison of Different LLM-based… · p. 8, Table IV · p. 9
  • Specialized Power and Affinity agents coordinate through tolerance-based Placement Agent arbitration, while confidence scores alone do not match argument-based debate.The confidence-scoring variant improves on simpler baselines but does not match the full framework.Source: D. Comparison of Different LLM-based… · pp. 8–9
  • At low utilization, power curves are nearly identical, but the proposed architecture accumulates affinity earlier; later, rearrangement matters as cubic marginal costs rise.P-Greedy and Three Agents lack rearrangement, leaving early placements in low-cost slots as machines fill.Source: D. Comparison of Different LLM-based… · p. 9, Figure 3 · p. 9

E. Effect of Variation of Tolerance

  • ε=0.1 gives the best balance between power consumption and affinity satisfaction; both lower and higher tolerance settings degrade performance.ε=0 is overly rigid, while larger values admit a broader set of inferior candidates.Source: E. Effect of Variation of Tolerance · p. 9, Figure 4 · p. 9

F. Effect of Debate Rounds and Candidate List Size

The section examines debate-round and candidate-list trade-offs, then reports scalability and model comparisons, practical costs, and planned extensions.

F. Effect of Debate Rounds and Candidate List Size

  • The reference setting (N, K) = (3, 5) achieves the best overall trade-off; fewer debate rounds weaken decision quality, while additional rounds improve coordination.Increasing candidate-list size helps up to a moderate level, after which gains saturate and token usage continues to rise.Source: F. Effect of Debate Rounds and… · p. 9, Table V · p. 10

G. Scalability Analysis

  • The framework maintains the best overall performance across tested problem sizes, preserving its power–affinity trade-off as applications and machines increase.Rule-based heuristics remain limited by fixed local rules, while refinement-based methods improve results only partially.Source: G. Scalability Analysis · pp. 9–10, Table VI · p. 10

H. Comparison across Different LLMs

  • Claude-based variants achieve the strongest overall results, with Claude Opus best, while DeepSeek-R1 improves substantially over the Llama 3.1 baseline.The table compares models under N = 1 and K = 1, a reduced configuration not directly comparable to the full reference setting.Source: H. Comparison across Different LLMs · p. 10, Table VII · p. 10

I. Token Cost and Practical Overhead

  • The framework reduces power consumption by about 4.26% over the best existing heuristics in the 100-application, 25-machine setting.A 100-decision local run uses approximately 2M tokens and takes roughly 90 minutes; estimated token cost is about $0.5 for Llama 3.1-8B or $10 for Claude, versus around $100 in operational savings over a representative horizon. At larger scales, annual datacenter energy costs can range from around $1 million to tens of millions, while LLM scheduling costs remain a few hundred to low thousands of dollars.Source: §I · p. 10

VII. CONCLUSION

  • The four-agent framework outperforms existing heuristic and meta-heuristic baselines on power consumption and affinity payoff, while future work targets inference cost, latency, and dynamic workloads.Planned directions include distilling debate patterns into lightweight policies, prompt compression and response caching, and online rescheduling as applications arrive and depart.Source: §VII · p. 10
Loading 2610.07030v1…