A Resource-Aware Asynchronous PBT Framework with Dynamic LoRA for Heterogeneous Edge Clusters

Synchronous Distributed Training With Runtime-Adaptive Mechanisms in Hybrid Cloud Environments

The rapid growth of large datasets and AI models has necessitated scalable and powerful computing resources, driving the extension of training workloads to the cloud to balance cost and performance. Among distributed learning strategies, synchronous data-parallel training remains the most widely adopted approach due to its simplicity and stable convergence. However, heterogeneity between worker nodes and unreliability of network bandwidth pose substantial challenges. Since workers must synchronize gradients at every update step, even a single straggler can significantly degrade overall performance. In this paper, we present ASTRA, an auto-adaptive synchronous training framework tailored for hybrid cloud environments. ASTRA addresses these issues with three coordinated mechanisms: i) a network-topology detection module that clusters workers into groups by pairwise connection cost to enable hierarchical aggregation, where gradients are first aggregated within each group and then across groups, mitigating slow-link bottlenecks; ii) a load balancing module to efficiently distribute workloads among heterogeneous GPUs over time, reducing resource imbalance and straggler effects; and iii) an adaptive gradient compression module that dynamically adjusts compression levels based on network conditions to minimize inter-group communication overhead. Through automated adaptation to diverse system environments, our unified approach alleviates the burden of manual configuration for AI practitioners. Experimental evaluations demonstrate the effectiveness of the framework ASTRA, which achieves lower time-to-accuracy than a resource-heterogeneity-aware baseline and several compression-based frameworks, while preserving convergence quality and robustness across heterogeneous hybrid cloud environments.

Sparsity-Aware Scheduling for Efficient MoE-Based LLM Serving on Edge Servers

Co-Optimizing Request Scheduling and KV Caching for Edge LLM Serving

Deploying large language models (LLMs) on edge nodes enables low-latency and privacy-preserving inference, but faces severe resource constraints under high-concurrence workloads. While existing inference systems leverage intranode key–value (KV) caching to improve efficiency, they largely neglect the unique complexities of multinode edge environments. Specifically, reactive KV cache eviction policies suffer from temporal uncertainty, often discarding reusable KV caches prematurely, while the tight coupling between request scheduling and cache placement often leads to myopic decisions that exacerbate load imbalance and resource contention. To address these challenges, we propose a dynamic block-level paradigm that treats KV blocks as the fundamental units for caching and scheduling, enabling dynamic sharing, generation, and eviction of arbitrary-length prefixes. We present complete modeling of the spatiotemporal coupling between scheduling and caching under block-level granularity, capturing intricate interactions overlooked by prior work. Based on this model, we design an online joint optimization algorithm, which applies to general edge LLM serving scenarios. The algorithm decouples spatiotemporal dependencies via randomized rounding over per-slot subproblems, achieving a balance between real-time responsiveness and long-term system efficiency. Theoretical analysis establishes high-probability near-optimality guarantees, and extensive experiments show that our method reduces the average time to first token (TTFT) by up to 54.02% over existing baselines.

Edge-cloud collaboration-driven predictive modeling for high-performance computing centers.

Accurate coolant flow prediction is critical for active thermal management in high-performance computing (HPC) centers, yet it is inherently challenged by mixed-timescale dynamics and high-frequency workload surges. Existing deep learning methods often prioritize global accuracy on smoothed stationary trends, which may lead to phase delays during abrupt thermal transients. In addition, high-capacity architectures can introduce non-negligible computational overhead for latency-sensitive, resource-constrained edge controllers. To overcome these limitations, this study proposes a deployment-oriented edge-cloud collaboration (ECC) framework integrated with a transient-aware predictive architecture, named FS-Attention, designed to balance transient responsiveness, engineering deployability, and decision transparency. FS-Attention couples local feature synthesis, temporal-memory encoding, and attention-based temporal refinement to improve coolant-flow tracking under non-stationary operating conditions. Evaluations on the real-world Frontier supercomputer dataset show that the feature synthesis attention (FS-Attention) model achieves competitive full-year prediction accuracy, with a coefficient of determination (R²) of 0.8744 and a root mean square error (RMSE) of 0.0350. Under isolated critical thermal events (CTEs), FS-Attention obtains the lowest RMSE of 0.0757, slightly lower than the Temporal Fusion Transformer (TFT) and 5.61% lower than the standard Transformer. Platform-based profiling further shows an inference latency of 0.016 ms and a parameter size of 323.1 K, suggesting model-side compatibility with facility-side edge execution, while attention-shift analysis provides diagnostic evidence of temporally adaptive model behavior under dynamic thermal conditions.

Task-aware dynamic optimization for edge cloud systems

TADOF is a dynamic optimization framework that jointly performs energy-aware elastic scaling and task migration under drift-aware periodic modeling and burst detection and observed about 18–34% lower energy and 15–28% lower migration cost than threshold and A2C/TD3 baselines while preserving QoS.

ElastiCo: Elastic Configuration and Interference-Aware Orchestration for GPU Clusters

ElastiCo is presented, an elastic co-location framework that enables training and inference workloads to safely share GPUs through three integrated mechanisms that decomposes the resulting multi-resource allocation problem into per-job configuration selection subproblems via dynamic per-resource shadow prices.