Efficient resource management and optimization in cloud data centers require elastic scaling with bandwidth guarantees to support dynamic cloud workloads. Existing methods provide runtime guarantees through bandwidth allocation and path selection, or support growth through capacity reservation and elastic VM placement. The former lack joint compute and network reservations for growth; the latter have limitations in growth assumptions, capacity utilization, elasticity, or routing efficiency. We study elastic network planning under the hose model, aiming to place initial virtual machines (VMs) to maximize subsequent growth, termed elasticity, without relocating them. This problem is NP-hard even without growth. We therefore adopt a capacity-oriented provisioning strategy and propose D-slice, a polynomial-time maximum-flow-based framework. A Dslice is a directed acyclic graph (DAG)-shaped subgraph rooted at a switch or server and defines a hose-feasible VM-slot reservation vector together with link-bandwidth reservation. This reserved capacity provides the resource basis for elastic VM placement and guarantees hose-based traffic patterns within the slice.We further devise a flow-guided routing scheme based on the maximum-flow structure and an efficient short-cut routing scheme that reduces hop count without complex routing tables while preserving bandwidth feasibility. We also study the composition of multiple D-slices in multi-root spine-leaf DCNs, where compatible slices can be merged to increase supportable capacity, and extend the framework to a two-level spine-leaf metropolitan area network. Extensive simulation and testbed evaluations show the efficacy of D-slice and its flow-guided routing in hose-based DCNs.
The Autoscaling Drift Index (ADI) quantifies the mismatch between predicted and actual resource demands in cloud-native autoscaling, yet most ADI-based correction methods use static thresholds that cannot track workload regime shifts, causing delayed detection or frequent false alarms. We propose AdaptADI, a lightweight online Bayesian optimization framework that adaptively tunes ADI thresholds from observed drift patterns and correction outcomes. Threshold selection is modeled as a contextual multi-armed bandit using recent prediction errors, workload volatility, and scaling frequency as context, and optimized via a sparse Gaussian Process with Thompson Sampling for efficiency. Experiments on Google Cluster Trace and Alibaba Cluster Trace show AdaptADI reduces total provisioning error by 16–18% over the strongest learning-based baselines with under 2ms added latency per scaling decision, and adapts to regime changes by converging within 150 scaling decisions. A small-scale live deployment on a 5-node Kubernetes cluster corroborates these findings: AdaptADI attains the lowest SLO violation rate (0.064%, 11.6%–30.8% lower than fixed-threshold and context free Bayesian baselines) with no increase in mean replica count or p95 latency.
The recent success of artificial intelligence motivates many non-professional users to train their own models; those users often resort to cloud training services (aka Training-as-a-Service), seeking to obtain an accurate model at a modest cost. To that end, properly setting up the learning rate and batch size is crucial. While various Hyper-parameter Optimization (HPO) methods have been proposed in that regard, they largely act based on heavy-weight validation signals, being inefficient in the overall cost. We find that the model training process can be viewed as a two-dimensional voting process—with gradients for different iterations and from different samples; moreover, to attain cost-efficient training is to ensure that the gradient redundancy is within a proper range—which we find is generic across diverse models. Based on that insight, we introduce GR-Gauge, a general method that gauges the gradient redundancy to instruct HPO decisions like configuration searching and trial termination. Moreover, to improve overall training efficiency in shared clusters, we extend GR-Gauge to be resource-aware, by choosing the most appropriate batch size allowed within the healthy range. Extensive experiments demonstrate that GR-Gauge can help attain near-optimal accuracy, with a time reduction of over 80%; meanwhile, its resource-aware extension reduces the average job queuing time by 77% in a shared GPU cluster.
Resource allocation in large-scale cloud data centers remains a challenging problem due to high-dimensional resource heterogeneity and stochastic demand dynamics. Existing scheduling methods often incur substantial computational overhead in dense task–resource mapping spaces and are prone to unstable convergence under dynamic workloads. This paper proposes a Multi-Level Adaptive Grey Wolf Optimizer (MLA-GWO) to improve the scalability and robustness of cloud scheduling. Specifically, a state-aware sparse candidate mechanism with attention-inspired weighting is introduced at the mapping layer. Real-time VM resource pressure is used to screen and retain a compact Top-$K$ set of promising VM candidates, while deadline-sensitive service quality is handled separately through the ATD component of the scheduling objective. In addition, an adaptive leadership mechanism is developed at the search layer to regulate leader influence based on real-time fitness differences, thereby maintaining population diversity and improving convergence stability. Evaluations on the Alibaba Cluster-Trace-2017 dataset show that MLA-GWO reduces MS by 9.24% and ART by 11.5% at 1000 tasks, while maintaining stable recovery under partial VM failures. The results show that combining candidate sparsification with adaptive evolutionary search provides robust scheduling performance within the evaluated workload and VM pool ranges.
Nowadays, with the continuous growth of data generation and storage, individuals increasingly rely on Machine Learning as a Service (MLaaS) in cloud environments to mine the potential value of their data. However, while cloud-based MLaaS provides access to high-performance resources and high-quality services, protecting the privacy of sensitive data has emerged as an attractive challenge. In this paper, we design an efficient and secure MLaaS framework for k-means clustering over encrypted outsourced data. Unfortunately, most existing works on privacypreserving clustering rely on a strong, impractical assumption: all participating users are completely trusted and possess the same keys. This assumption fails to hold in real-world multi-user MLaaS scenarios, where users are mutually untrusted and cannot be granted access to the same keys. To address this gap, we propose a multi-user, secure and efficient k-means clustering scheme optimized with data packing and triangle inequality acceleration. Our scheme achieves high-efficiency clustering performance while rigorously guaranteeing the privacy of users’ raw data and final clustering results. Finally, we present a formal security analysis and theoretical computation complexity evaluation of the proposed scheme. Extensive experimental results further validate its practicality and efficiency in deployment scenarios.
