DRL-MADRL is competitive with the strongest heuristic under low contention and achieves the best SLA satisfaction at moderate and high contention, and the NumPy implementation requires approximately 80 KB per agent and sub-10 ms inference latency.
Abstract
Efficient task scheduling in large-scale distributed systems remains challenging because workloads are dynamic, resources are heterogeneous, and centralized schedulers introduce coordination bottlenecks. This paper presents DRL-MADRL, a lightweight decentralized multi-agent deep reinforcement learning scheduler for heterogeneous cloud-edge systems. Each agent maintains a per-node utilization-estimate vector updated by randomized pairwise gossip, enabling decentralized execution with O(1) peer contacts per update and O(N) payload per gossip event. The scheduler combines a NumPy actor-critic policy, priority-capacity assignment scoring, gossip-augmented observations, optional contention-adaptive score blending, and workload-feedback reward shaping. We formulate scheduling as a Decentralized Partially Observable Markov Decision Process and evaluate the method on a 100-node heterogeneous system with 1,000 tasks per episode across three contention regimes. Against seven baselines, including Random, Weighted Round-Robin, Priority-Min-Min, Priority-Capacity Heuristic, PPO, MADDPG, and MAPPO, DRL-MADRL is competitive with the strongest heuristic under low contention and achieves the best SLA satisfaction at moderate and high contention. At moderate contention, DRL-MADRL achieves 19.16 seconds average task completion time (ATCT) and 90.60 percent SLA satisfaction, improving SLA by 2.73 percentage points over PPO and 2.98 percentage points over PCH. The NumPy implementation requires approximately 80 KB per agent and sub-10 ms inference latency.
High-velocity workloads and intricate task dependencies inherent in distributed stream-processing systems pose a fundamental challenge to efficient resource allocation. Traditional heuristic and single-agent reinforcement learning (RL) schedulers frequently fail to recognize these complex network and data-flow interactions, leading to severe resource fragmentation and catastrophic tail latency spikes. In order to accomplish coordinated, low-latency scheduling, we propose a Topology-Aware Multi-Agent Reinforcement Learning (TAMARL) framework utilizing a Centralized Training and Decentralized Execution (CTDE) architecture. TAMARL allows distributed agents to optimize task placement across heterogeneous cluster nodes and prevent backpressure cascades by integrating topology-aware state representations. We evaluate TAMARL on a production-grade cloud-native stack leveraging Apache Flink and Kubernetes. Compared to state-of-the-art baselines across six demanding stress-test scenarios, experimental evaluations demonstrate that TAMARL improves Service Level Objective (SLO) attainment by 27% while reducing P99 tail latency by up to 68%. Additionally, TAMARL maintains stable, resilient performance under 90% cluster utilization while securing 95% network locality.
Sunday J. Awine, Jinwei Liu· IEEE International Conferenc...· 0 citations
This paper addresses the joint task offloading and resource allocation problem in multi-user MEC systems and proposes a decentralized control framework based on Multi-Agent Reinforcement Learning (MARL), which achieves lower total system cost and faster convergence than the full-local, full-offload, and heuristic baselines.
Youssef Oukissou, Mohamed Amine Meddaoui, Ayoub Belaidi et al.· International journal of Com...· 0 citations
A neural scheduling framework for distributed multi-robot task allocation, consisting of a multi-decoder graph attention model (MDGAM) policy model and a critic-free group relative multi-agent policy gradient (GRMAPG) training algorithm, which improves task-completion performance over existing heuristic and learning-based methods.
Allocating limited computation among concurrent learning tasks is difficult when each task must reach a target loss before a deadline but its required training effort is unknown. Existing approaches combine online loss prediction with adaptive resource allocation, yet commonly treat computation as continuously divisible throughput. We instead study a practical setting in which tasks arrive over time and computation is provided by discrete nodes. This setting introduces both uncertain demand and constrained sequential decisions. We propose MARA, which predicts future loss trajectories with conditional flow matching and coordinates compute nodes through a cooperative multi-agent autoregressive policy. A potential-based progress reward supplies intermediate training feedback while preserving the undiscounted task-completion objective. Across in-distribution, reinforcement-learning, and vision workloads, flow matching reduces remaining-resource prediction error relative to weighted least squares. At the scheduler's training load, MARA completes 63.46% of tasks on average, 8.54 percentage points above strong baseline Learning with Adaptive Resource Allocation (LARA), and remains ahead under unseen heavier workloads.
Hanye Zhao, Muning Wen, Yong Yu et al.· 0 citations
This paper proposes a multi-agent reinforcement learning (MARL) framework for TSN scheduling, where each TSN queue is modeled as an autonomous agent and the Heterogeneous-Agent Proximal Policy Optimization (HAPPO) algorithm is employed to explicitly model inter-agent dependencies and jointly optimize service delivery across queues.
Marcos Carvalho, Fatih Temiz, Shavbo Salehi et al.· 0 citations
Coordinating heterogeneous LLM agents under congested online settings is difficult because bursty task arrivals and limited per-agent capacity may induce hotspot overload and severe tail waiting time. This paper proposes an online scheduling method based on candidate-set contraction before assignment. Specifically, tasks are first routed through subscription matching to identify a task-relevant candidate pool, after which layered gating is applied to enforce capability feasibility, historical quality, and real-time load constraints. The remaining candidates are then ranked using a composite score that balances competence and load, with stable tie-breaking introduced to reduce assignment fluctuations under contention. We evaluate the method under a reproducible protocol with both regular and congested regimes. Across benchmarks covering code generation, arithmetic reasoning, and preference-based evaluation, the proposed approach preserves competitive task performance while reducing both mean and 95th-percentile waiting time in congested settings relative to representative linear, flat, and hierarchical baselines. The findings suggest that candidate contraction is a useful strategy for achieving more stable coordination in heterogeneous LLM-agent systems.
Weiqi Xie, Yuyao Li· Annual International Compute...· 0 citations