1 views collected around this technical thread.
To reduce resource contention and improve offline task reliability, this article examines the challenges of using Koordinator with Hadoop Yarn pods on Kubernetes, proposes real‑time resource reporting and task‑level eviction strategies, details community and custom solutions, and outlines future enhancements with Volcano.
HighService, Alibaba’s Pythonic AI service framework, accelerates large‑model inference and maximizes GPU utilization by separating CPU‑GPU processes, offering out‑of‑the‑box quantization, parallelism and caching, and dynamically reallocating idle GPUs across clusters through a master‑worker scheduler to keep online latency low while boosting offline throughput for diffusion and LLM workloads.
This article describes how OPPO's big‑data team transitioned from traditional IDC and EMR environments to a cloud‑native Kubernetes architecture, detailing the motivations, design principles, elastic scaling challenges, custom solutions, and future directions for large‑scale data processing on the cloud.
The paper presents Bilibili’s multi‑layer CDN edge‑node scheduling system, which groups heterogeneous nodes by quality and price, uses cost‑aware and resource‑aware heuristics—including maximum‑flow regional borrowing and contextual‑bandit utilization prediction—to allocate bandwidth per business, achieving a 43 % bandwidth reuse increase, 33 % coverage boost, and markedly lower stall rates.
Xiaohongshu’s cloud‑native team transformed its over‑90% containerized services by introducing resource‑pooled mixed deployment, fine‑grained unified scheduling, and an elastic container pool with global HPA and cluster autoscaling—driving 35% of resources to mixed use, tens of millions of daily core‑hours, and roughly 30% cost savings while preparing for hybrid‑cloud expansion and FinOps.
The award‑winning cloud‑native large‑scale cloud‑edge collaborative project, developed by Alipay, Zhejiang University, Xieyun Technology and Alibaba Cloud, delivers unified resource scheduling for millions of heterogeneous devices, achieving significant performance gains, extensive patents, papers, standards, and substantial economic benefits across multiple industries.
Virtualization resource scheduling algorithms, especially in OpenStack, address fragmented CPU allocation and uneven node utilization by dynamically consolidating VMs, employing NUMA-aware placement, and using resource scoring to trigger migrations, ultimately improving utilization, reducing costs, and enhancing performance in cloud environments.
This article explains how 360’s cloud platform uses a three‑layer architecture and Nova‑scheduler to manage thousands of servers and tens of thousands of VMs across multiple OpenStack clusters, detailing scheduling policies, resource‑pool handling, current challenges, and future improvement plans.
Kubernetes 1.31 adds an alpha‑stage ‘distribute-cpus-across-cores’ option to the CPUManager static policy, allowing CPUs to be spread across physical cores for better cache locality, reduced contention, and improved performance in multi‑core and performance‑sensitive workloads.
This article explains how Baidu processes internet‑scale content by applying deep AI‑driven understanding, detailing cost‑optimization, efficiency improvements, model‑service frameworks, resource‑scheduling systems, and batch‑compute platforms that together enable trillion‑level indexing and feature extraction.
The article analyzes low CPU utilization in pure online Kubernetes clusters, introduces mixed‑workload (online + offline) scheduling to improve resource efficiency, explains core techniques, kernel QoS features, and details Koordinator‑based implementations such as node resource reservation and scheduling adjustments.
Xiaohongshu’s cloud‑native platform adopted a four‑stage mixed‑workload scheduling strategy—reusing idle nodes, whole‑machine time‑sharing, normal mixed pools, and a unified scheduler (Tusker) that coordinates CPU, GPU and memory across Kubernetes and YARN—boosting average cluster CPU utilization from under 20 % to over 45 % and delivering millions of low‑cost core‑hours while preserving QoS for latency‑sensitive, mid, and batch jobs.
iQIYI’s mixed‑workload system colocates Spark/Hive big‑data jobs with online video services by running YARN NodeManagers inside Kubernetes, using an Elastic YARN Operator, Koordinator‑driven CPU oversubscription, and remote shuffle, boosting online CPU utilization from ~9 % to over 40 % and saving tens of millions of RMB annually.
The new tiered resource‑guarantee system for Didi’s elastic cloud containers defines S, A, and B priority levels with explicit over‑commit rules, upgrades OS, Kubernetes, kube‑odin, service‑tree, and CMP components, and thereby cuts CPU contention by up to 80%, reduces latency, improves scaling reliability, and lowers operational costs.
Didi's Elastic Cloud uses mixed deployment to co‑locate diverse services, employing tiered guarantees, custom Kubernetes scheduling, profiling, rescheduling, and isolation‑cluster techniques to boost utilization while preserving QoS, with a roadmap for broader automation and interference detection.
This article explores how Tencent integrates big data with cloud‑native technologies, detailing the evolution, opportunities, challenges, the peak‑range (FengLuan) architecture, engine and scheduling layers, mixed‑workload strategies, runtime optimizations, and future directions for large‑scale data platforms.
This article introduces Amiya, a self‑developed overcommit component that dynamically increases Yarn memory and vCore capacity on Bilibili's offline big‑data clusters, details its architecture, key implementation of overcommit, eviction and mixed‑deployment strategies, and evaluates its resource‑utilization impact.
This article examines the storage bottlenecks in large‑scale AI training, evaluates local‑disk and Alluxio‑based distributed caching strategies, proposes uniform cache eviction and replica‑aware global policies, and introduces the SiloD framework for coordinated compute‑storage scheduling to dramatically improve GPU utilization and overall cluster throughput.
The article describes Punica, a cloud‑native, function‑as‑a‑service platform that unifies content‑understanding inference services through a one‑stop portal and unattended operations, improving deployment speed, resource utilization, and reducing manual effort for AI model serving.
Punica provides a cloud‑native, one‑stop platform that unifies Baidu’s content‑understanding inference services, automates testing, resource provisioning, and monitoring, and enables unattended, self‑healing operations with dynamic scaling and GPU scheduling, cutting onboarding time by half and reclaiming hundreds of GPUs.