PaperlayerView exact revision on arXiv

Source-linked AI summary

IMPACT: Intent-driven Multi-agent Policy with Attention for SLO-guaranteed Microservice Migration in Cloud-edge Systems

Xinjin Li, Siru Tao, Shihan Yin, Yujian Long, Qingze Wang, Lu Cheng, Yeyang Zhou, Calvin Chang Liu, Yu Ma

arXiv:2609.35818v1cs.MA

TL;DR

Dynamic MEC conditions and partial observability make high-percentile SLOs difficult to guarantee, especially when migration and bandwidth are optimized separately. IMPACT coordinates those decisions through compact agent intents and double attention; across 5- and 20-edge settings, it reports lower mean latency and latency deviation while keeping energy close to the lowest-cost heuristic.

  • Problem

    Dynamic MEC conditions and partial observability complicate high-percentile SLO satisfaction, while existing approaches often optimize mean delay or separate migration from bandwidth control.

    Source: §I · p. 1, §II · p. 3

  • Method

    IMPACT uses cooperative multi-agent learning with compact intents for local SLO risk, migration urgency, and compute pressure, alongside double attention for coordination.

    Source: §I · p. 2

  • Results

    Across 5- and 20-edge settings, IMPACT cuts mean latency by 30-50% and latency deviation by 40-70%, with energy within 5% of the lowest-cost heuristic.

    Source: §I · p. 2

  • Takeaways & Limitations

    The reported results support intent-driven coordination as an approach to latency-guaranteed microservice migration with discrete bandwidth control in dynamic, partially observable edge environments.

    Source: §I · pp. 1–2

    The conclusion directs readers to supplementary materials for the detailed discussion of system limitations and future plans.

    Source: §VI · p. 6

Abstract

from arXiv · show

Ensuring strict tail-latency service-level objectives (SLOs) in dynamic mobile edge computing (MEC) systems remains challenging because user mobility, wireless fading, bursty workloads, and partial observability jointly undermine reliable cloud-edge orchestration. Existing microservice migration methods predominantly optimize average delay and often decouple migration from bandwidth control, leading to uncoordinated decisions, queue oscillation, and frequent high-percentile latency violations. To address this issue, we propose IMPACT, an intent-driven Agentic AI framework for cooperative microservice migration and bandwidth control in cloud-edge systems. Under centralized training with decentralized execution (CTDE), each edge cloud is modeled as an autonomous agent that encodes local SLO risk, migration urgency, and computational pressure into compact, semantic intent representations. IMPACT further introduces a double-attention mechanism that first selectively aggregates relevant peer intents for efficient inter-agent communication and then filters local observations to emphasize goal-relevant state information. This design enables robust coordination under partial observability and jointly optimizes service migration and discrete uplink bandwidth allocation. Extensive experiments in 5-edge and 20-edge scenarios show that IMPACT reduces mean latency by 30-50% and tail-latency deviation by 40-70% compared with state-of-the-art factorized multi-agent reinforcement learning (MARL) and heuristic baselines, while achieving near-zero SLO violation rates under tight thresholds and energy consumption close to the best heuristic baseline. These results demonstrate that intent-driven agentic coordination provides an effective and scalable solution for SLO-aware orchestration in complex cloud-edge intelligent systems.

I. INTRODUCTION

Dynamic mobility, fading, and bursty workloads make high-percentile SLOs difficult to meet when migration and bandwidth decisions are not coordinated. IMPACT combines intent-driven multi-agent coordination with a tail-aware MEC model; across 5- and 20-edge settings, it reports lower latency and deviation with energy near the lowest-cost heuristic.

I. INTRODUCTION

  • Mobility, fading, bursty arrivals, and partial observability make high-percentile SLO satisfaction difficult, motivating coordinated migration and bandwidth decisions across edge sites.The study focuses on migration and per-user bandwidth scheduling under limited edge resources and migration downtime.Source: §I · p. 1
  • The tail-aware MEC model couples wireless transport, M/M/1 compute queues, and migration downtime, using sliding-window P95 latency as the SLO metric.Source: §I · p. 2
  • IMPACT encodes local SLO risk, migration urgency, and compute pressure as compact intents, then uses double attention to fuse peer messages and local observations.Edge agents use local observations and sparse neighbor messages under CTDE, with SLO-shaped rewards and feasibility projection.Source: §I · p. 2
  • Across 5- and 20-edge settings, IMPACT cuts mean latency by 30-50% and latency deviation by 40-70%, with energy within 5% of the lowest-cost heuristic.The reported energy consumption is also below factorized MARL baselines.Source: §I · p. 2

II. RELATED WORK

Prior migration and reinforcement-learning approaches often emphasize mean performance or handle migration and bandwidth separately, while partial observability limits coordination. The paper frames these joint control decisions using models of wireless transmission, compute queues, migration penalties, and latency and energy objectives.

II. RELATED WORK

  • Migration changes which edge serves each user and affects uplink sharing and queue pressure, while bandwidth allocation shapes transmission time and queue lengths that can trigger future migrations.Optimizing the two levers in isolation can produce oscillation as migrations relieve one hotspot but create another.Source: §II · p. 2
  • Related work spans RL for offloading and migration, adaptive information filtering, and cooperative MARL, but existing MARL lacks explicit short-term intent representations and may assume centralized control or full observability.Source: §II · pp. 2–3
  • Existing approaches often optimize mean performance without explicit high-percentile latency guarantees or treat migration and bandwidth control in isolation; IMPACT optimizes both jointly through intent-driven cooperative MARL.Source: §II · p. 3

III. SYSTEM MODEL AND PROBLEM FORMULATION

  • The system model identifies service input and compute needs, edge CPU capacity, allocated cycles, uplink bandwidth, and transmission power, and represents transmission, compute queuing, migration penalties, and latency-energy objectives.The supplied model passages identify Shannon-based transmission, M/M/1 computation time, and propagation, pre-copy, and downtime components of migration penalty.Source: §III · p. 3

IV. INTENT-DRIVEN MULTI-AGENT COMMUNICATION FRAMEWORK

The section formulates edge-cloud orchestration as a cooperative Dec-POMDP under CTDE and specifies the observations and actions used to coordinate migration and bandwidth control. Its reward balances SLO performance with migration, energy, and messaging considerations.

A. Dec-POMDP Construction

  • Edge clouds are cooperative agents in a finite-horizon Dec-POMDP trained centrally and executed decentrally.The formulation uses a discount factor γ ∈ (0, 1].Source: A. Dec-POMDP Construction · p. 3
  • The global state combines normalized user-service latencies with per-edge CPU utilization, while each agent observes the latency matrix and its own utilization.A local history buffer can optionally provide recent observations for temporal context.Source: A. Dec-POMDP Construction · p. 3
  • Each agent can increase or decrease uplink bandwidth or migrate a service to another edge, with infeasible migrations masked out.Bandwidth is subsequently constrained by edge resources and per-link limits.Source: A. Dec-POMDP Construction · p. 3
  • The cooperative reward uses normalized SLO slack and violation fraction alongside migration and energy terms.The migration indicator records whether any migration occurred at that time step.Source: A. Dec-POMDP Construction · p. 3
  • Energy and intent-message costs are normalized, and their weights trade off those overheads against latency and SLO objectives.The reward is clipped to [−1, 0] for numerical stability.Source: A. Dec-POMDP Construction · pp. 3–4

B. Intent-Driven Multi-Agent Collaboration

IMPACT uses short-term intents and double attention to align peer communication and local state processing with agents’ goals. Regularization is intended to keep intent representations discriminative and temporally stable.

B. Intent-Driven Multi-Agent Collaboration

  • IMPACT addresses static attention structures by modeling agents’ short-term intents and aligning message and environmental fusion with current task goals.In this setting, agents are edge clouds choosing bandwidth adjustments or service migrations.Source: B. Intent-Driven Multi-Agent… · p. 4
  • A recurrent encoder maps each agent’s local history into a latent intent that represents its short-term goal for resource adjustment or service migration.The intent encoder uses a GRU and samples a latent variable from Gaussian parameters.Source: B. Intent-Driven Multi-Agent… · p. 4
  • Message attention uses an agent’s intent to select relevant teammate intents, reducing redundant communication.For microservice migration, relevant peers include edge clouds able to handle migration or resource adjustment.Source: B. Intent-Driven Multi-Agent… · p. 4
  • Environment attention uses aggregated intent to emphasize goal-related local features and suppress noisy or irrelevant observations.Relevant features can include resource usage, service requests, and migration feasibility.Source: B. Intent-Driven Multi-Agent… · p. 4
  • Entropy and continuity regularization are intended to preserve discriminable, temporally stable intents as conditions and task demands change.The overall objective balances task performance and intent quality while controlling energy and communication overhead.Source: B. Intent-Driven Multi-Agent… · pp. 4–5

A. Experiment Settings

The experiments use a custom MEC simulator with 20 mobile users, 5 or 20 edge clouds, and three microservice types. IMPACT uses CTDE, with centralized training and decentralized execution based on local observations and selectively shared intents.

A. Experiment Settings

  • A centralized critic uses global state during training, while execution uses local observations and selectively shared intents.Reward shaping closes the loop with the dynamic environment.Source: Figure 3 · p. 5
  • The paper states that source code and full deployment experiments in reality and simulation are available to support reproduction.Supplementary materials and a GitHub link are to be provided upon publication under the double-blind review policy.Source: A. Experiment Settings · p. 5
  • The simulator has 20 mobile users, 5 or 20 edge clouds, and 3 microservice types, with a 300 ms latency SLO threshold.The evaluation compares IMPACT against Random, Greedy, VDN, QMIX, QPLEX, and MAIC.Source: A. Experiment Settings · p. 5

1) Convergence and Network Performance:

In the 20-edge setting, IMPACT outperforms the reported baselines on latency and latency deviation while keeping energy consumption close to Greedy. It also retains a higher-return convergence profile as the network scales, and the paper reports a similar strong trend in the 5-edge setting.

1) Convergence and Network Performance:

  • At 20 edges, IMPACT reaches a high-return plateau with small oscillations, while factorization baselines converge more slowly or attain lower final returns.QMIX has a prolonged unstable phase; VDN and QPLEX converge earlier but finish lower, and MAIC remains below IMPACT for most timesteps.Source: 1) Convergence and Network Performance · p. 5
  • IMPACT also performs strongly in the 5-edge setting, although detailed results are deferred to the supplementary material.Source: 1) Convergence and Network Performance · p. 5
  • 93.50 ms average latency is about 30% below Greedy and 46–50% below VDN, QMIX, and QPLEX; deviation is 0.047, reducing tail variability by 40–50%.Energy consumption is 3.42, close to Greedy at 3.40 and below factorized MARL methods at 3.67–3.69.Source: 1) Convergence and Network Performance · p. 5, Table I · p. 6

3) Ablation Experiments:

The supplied passages identify an ablation comparison of full IMPACT with variants that remove environment attention, message attention, or both. They also identify a separate figure showing the latency CDF in the 20-edge setting.

3) Ablation Experiments:

  • Fig. 6 compares full IMPACT with three ablated variants to examine the effects of the two attention components.The full model rises quickly, according to the accompanying text.Source: 3) Ablation Experiments · p. 6, Figure 6 · p. 6
  • Fig. 5 is identified as the CDF of service-migration average latency in the 20-edge setting.Source: Figure 5 · p. 6
  • The ablation labels denote removal of environment attention (e), message attention (m), or both (b).Source: Figure 6 · p. 6

2) CDF of Average Latency:

The 20-edge latency CDF favors IMPACT, with a steeper rise and high percentiles around 120–140 ms, while Greedy requires around 200 ms or more. The paper links its coordination design to compact peer intents, filtered local state, and coordinated migration and bandwidth decisions.

2) CDF of Average Latency:

  • The CDF captures central tendency and tail behavior; a left-shifted, steeper curve indicates lower latency and tighter dispersion.It directly reflects SLO percentiles such as P90 and P95.Source: 2) CDF of Average Latency · p. 6
  • In the 20-edge CDF, IMPACT reaches high percentiles at about 120–140 ms, compared with around 200 ms or more for Greedy.QMIX, QPLEX, and VDN cluster behind Greedy with wider spread, while Random has the heaviest tail and slowest rise.Source: 2) CDF of Average Latency · p. 6
  • IMPACT encodes SLO risk, migration urgency, and compute pressure in compact messages, then fuses peer intents with filtered local state.The paper says this supports timely migrations before queue buildup and coordinates them with discrete bandwidth adjustments.Source: 2) CDF of Average Latency · p. 6
  • Removing either attention component slows convergence and increases variance; removing both produces the slowest, noisiest learning.Late-stage returns are similar across variants, but full IMPACT achieves higher sample efficiency and lower variance.Source: 2) CDF of Average Latency · p. 6

VI. CONCLUSION

IMPACT is presented as an intent-driven communication framework for latency-guaranteed microservice migration in mobile edge computing. The conclusion reports performance gains over QMIX and identifies extensions for future work.

VI. CONCLUSION

  • Compared with QMIX, IMPACT accelerates convergence by over 30%, reduces P95 latency by up to 25% in light- and heavy-load scenarios, and decreases SLO violations by up to 40%.Source: §VI · p. 6
  • IMPACT achieves competitive migration costs and energy consumption relative to greedy baselines in dynamic environments with limited edge resources, bandwidth inertia, and migration cooldown.Source: §VI · p. 6
  • Future work will extend IMPACT to multi-tenant environments, incorporate migration techniques such as pre-warming, and validate it on real-world traces and edge prototypes.Source: §VI · pp. 6–7
  • A detailed discussion of system limitations and future plans is designated for supplementary material.Source: §VI · p. 6

Related papers

Loading 2609.35818v1…