W EAVE : Efficient Co-Scheduling for Disaggregated RL Post-Training

Dynamic Compute and Network Orchestration for Disaggregated RL

This work builds Silverstone to orchestrate dynamically both compute and network in disaggregated RL, using a reconfigurable optical-electrical fabric called RFabric that achieves superior performance-cost efficiency at scale over static Fat-Tree networks.

HRL-TaskOpt: A Hierarchical Reinforcement Learning-Based Task Scheduling Framework for Multi-Cloud and Hybrid Environments

Cloud computing has emerged as a new paradigm, which entrusts task scheduling to ensure the satisfaction of stringent constraints on latency, energy, and resources for sustainably running real-time applications. State-of-the-art natural DRL-based scheduling solutions mainly rely heavily on DRL techniques and are either limited in scalability, adaptivity, or generality of workloads/infrastructures. State-of-the-art flat methods, such as DQN and actor–critic models, are not sufficiently effective at high levels of decision complexity and are not robust against varying system loads and task priorities. In this paper, we present HRL-TaskOpt, a novel Hierarchical Reinforcement Learning-based task scheduling framework that combines high-level global task offloading with millisecond-granularity local scheduling policies in an edge–cloud scenario. In the proposed framework, there are two levels of agents: a high-level policy that utilizes Proximal Policy Optimization (PPO) to select the optimal execution tiers (edge or cloud), and a low-level policy based on Deep Q-Networks (DQN) to manage scheduling within nodes (edge or cloud). Such decomposition enables HRL-TaskOpt to efficiently accommodate heterogeneous workloads and adapt to dynamically changing infrastructure. We use synthetically generated workloads that reflect the characteristics of real-world applications to demonstrate the effectiveness of our model and compare it with state-of-the-art models such as SA-DQN, DRL-DO, and GD-DRL. We experimentally demonstrate that HRL-TaskOpt achieves a time reduction of up to 21.4% for task completion (2,000-task workload vs. SA-DQN), an energy efficiency improvement of up to 20.3% (10,000-task workload vs. GD-DRL), and a task success rate improvement of up to 13.3% (averaged across baselines at the 2,000-task workload), compared to these models. In addition, robustness to resource failures and sensitivity to task-type diversity, demonstrated in images, validate the real-life usability of the model. HRL-TaskOpt offers a scalable and intelligent solution for adaptive task scheduling, making it an appealing candidate for deployment in next-generation edge–cloud continuum systems.

psRL: Efficient Training for Agentic AI via Training-Time Prefix Sharing

This paper proposes psRL (prefix sharing for RL), a new training system for agentic AI designed to exploit prefix redundancy among training samples, and introduces two novel prefix-sharing mechanisms that enable flexible, fine-grained workload distribution across GPU workers.

Scheduling for Green Computility via LLM-Guided Deep Reinforcement Learning

Green Computility scheduling in heterogeneous clusters requires balancing execution efficiency, deadline satisfaction, resource utilization, carbon emissions, and renewable energy usage under dynamic workload and energy conditions. Existing heuristic schedulers are efficient but often rely on fixed hand crafted rules, whereas reinforcement learning (RL) based schedulers may explore inefficiently in large and structured action spaces when no prior guidance is available. To address these challenges, we propose Green-LLM-DQN, a large language model (LLM) guided dueling double deep Q-network (D3QN) framework for trace-driven Green Computility scheduling. The framework uses an LLM to generate a structured and interpretable heuristic prior from the scheduling context, and incorporates this prior into a D3QN backbone for action scoring and selection. This design improves early stage exploration while preserving the adaptability of RL. Experiments on trace driven workloads show that Green-LLM-DQN achieves a competitive overall trade-off among return, carbon emissions, and green energy utilization than classical scheduling heuristics and learning based baselines.

Learning Workflow Scheduling on Multi-Resource Clusters

Deep Reinforcement Learning-based Dynamic TWT Scheduling for Heterogeneous Wi-Fi Networks

Target Wake Time (TWT) in IEEE 802.11ax enables significant power savings by coordinating station sleep schedules, but optimal TWT parameter selection under heterogeneous traffic remains an open challenge. We present a deep reinforcement learning framework for dynamic TWT scheduling that is protocol-compliant by design. The agent’s observation is restricted to 802.11ax buffer status reports, 802.11k station statistics, and AP-derived TWT metrics, ensuring direct deployment without firmware modifications. The system employs Proximal Policy Optimization (PPO) over a factored MultiDiscrete action space of 380 schedule-assignment combinations, trained end-to-end within ns-3 via shared-memory with episode-level process isolation for stable, reproducible training. We evaluate two architectures, MLP-PPO and LSTM-PPO, against an analytical M/D/1 baseline across three reward presets (energy, throughput, queue) over 50 seed-disjoint episodes with 16 stations from four heterogeneous device classes. Both RL agents substantially outperform the analytical baseline on all presets. On the energy preset, MLP-PPO reduces aggregate energy by 44% while improving throughput by 18% and reducing drops by 73%. On the queue preset, LSTM-PPO achieves 15% higher reward with 77% fewer drops by leveraging temporal correlations across beacon intervals. The two architectures exhibit complementary strengths. MLP-PPO excels on the energy-dominated preset where a stationary policy suffices while LSTM-PPO’s recurrent state captures multi-step queue-drain dynamics that transfer across network configurations, motivating future ensemble approaches.