The paper presents Libra, a resource‑management system for Agentic RL post‑training that jointly optimizes training and rollout GPU allocation using a global planner, heterogeneous inference clusters, a causality‑driven multi‑level feedback queue, and an elastic hybrid pool, achieving up to three‑fold throughput gains and up to 2.5× faster reward convergence.
This article introduces three open‑source tools that improve GPU‑centric development: NBD VRAM, which turns GPU memory into Linux swap space; Reg Factory, a scheduler and monitor for multi‑GPU clusters; and vLLM Studio, a web UI for deploying and managing large‑model inference.
This article details how a leading autonomous‑driving company tackled multi‑tenant resource contention, low GPU utilization, and distributed task dead‑locks on a heterogeneous Kubernetes cluster by adopting Koordinator’s ElasticQuota, Reservation, Gang and Device‑Share features, achieving higher allocation rates, better fairness, and significantly improved GPU efficiency.
This article details how SF Tech's EffectiveGPU (EGPU) platform redesigns GPU resource management on Kubernetes, introducing fine‑grained memory and compute partitioning, priority‑based scheduling, Volcano integration, and monitoring pipelines to dramatically improve utilization and reduce hardware costs for AI workloads.
The article outlines why traditional data warehouses often become costly silos, introduces the four core capabilities of an intelligent data hub—unified ingestion, automated governance, dynamic resource scheduling, and service‑oriented output—and provides a practical four‑step roadmap to design, pilot, and operate such a hub.
This comprehensive guide walks you through constructing a fault‑tolerant HDFS high‑availability architecture, configuring dual NameNodes with ZooKeeper and JournalNode clusters, fine‑tuning YARN resource schedulers, implementing monitoring, automated failover testing, and performance optimization, all backed by real‑world production experiences and code examples.
SeamlessFlow, an industrial‑scale reinforcement‑learning training framework released by Kuaipilot, decouples trainer and agents via a novel data‑plane, introduces a tag‑based resource scheduler, and eliminates pipeline bubbles, achieving up to 100% token‑throughput boost and 62% reduction in overall training time across large‑model RL workloads.
The article explains how Apache RocketMQ has been re‑engineered for the AI era, addressing long‑running conversational workloads, scarce GPU resources, and multi‑agent workflow bottlenecks by introducing lightweight Lite‑Topic communication, intelligent resource scheduling, and cloud‑native architectural upgrades.
This article explains how Youzan transformed its Kubernetes clusters from static over‑commit scheduling to load‑balanced mixed workloads using Koordinator and the Longxi kernel, achieving higher CPU utilization, lower costs, and better resource management for both online and offline services.
In a detailed interview, Tencent Cloud OS chief architect Du Zhen explains how TencentOS Server has evolved over 15 years from an internal platform to a multi‑industry, AI‑optimized operating system, outlining its OS‑for‑AI and AI‑for‑OS strategies, performance‑focused scheduling innovations, SWAP redesign, migration solutions, ecosystem building, and future vision.
This article explains how effective production planning goes beyond task ordering to coordinate resources, align market‑production‑supply rhythms, and manage risks, offering a four‑step framework—forecasting, scheduling, collaboration, and system mechanisms—to achieve stable, value‑driven manufacturing outcomes.
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.
This article analyzes common GPU efficiency problems in enterprise AI compute platforms—such as low utilization, long fault‑resolution times, and limited performance gains—and presents three practical solutions: dynamic resource allocation, systematic fault‑tolerance, and system‑level tuning, illustrated with real‑world case studies.
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.
This article analyzes the three core technologies behind PaaS for AI—GPU resource management, node data optimization, and task scheduling—detailing their concepts, component architecture, critical workflows, technical advantages, and future challenges, while illustrating practical configurations with Kubernetes and Volcano examples.
This article explains GPU virtualization, comparing NVIDIA's software‑based vGPU and AMD's hardware‑based MxGPU, detailing their architecture, required hardware, licensing, performance indicators, resource scheduling strategies, slicing limits, and the advantages and drawbacks of each approach for virtualized workloads.
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.
Koordinator v1.4.0 introduces mixed K8s/YARN workloads, NUMA‑aware scheduling, CPU‑normalization, enhanced ElasticQuota with tree structures and non‑preemptible pods, cold‑memory reporting, QoS for non‑containerized applications, and a suite of bug‑fixes and performance improvements for enterprise Kubernetes clusters.
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.
This article details HuoLaLa's development of an in‑house Automatic Speech Recognition system, covering its architecture, VAD optimization, language‑model and hot‑word enhancements, punctuation restoration, task and resource scheduling, and the resulting improvements in accuracy and cost efficiency.
This article details Baidu Search's engineering practice for trillion‑scale content understanding, covering cost and efficiency challenges, model‑service framework, batch‑compute platform, resource‑scheduling system, HTAP storage design, and concrete optimization techniques such as multi‑process Python serving, dynamic batching, and two‑stage scheduling.
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.
The article explains Baidu’s cloud‑native AI resource scheduling across single‑ and multi‑GPU nodes, describes the PaddleFlow Kubernetes‑based workflow engine with its hierarchical queues, advanced scheduling algorithms, unified storage, and how these technologies improve GPU utilization, reduce fragmentation, and simplify AI task orchestration.
The article details how Tencent's Taiji machine‑learning platform reduces training costs and improves efficiency for large‑scale advertising models by leveraging cloud‑native mixed‑resource strategies—including online idle, offline elastic, and compute‑resource sharing—while maintaining high service stability through advanced scheduling, fault‑tolerance, and resource‑prediction techniques.
The interview explores the rapid rise of green computing, examining energy consumption of data centers, low CPU utilization, software‑centric optimization, cloud‑native scheduling, AI and big‑data workloads, and future technical and educational efforts needed to achieve sustainable, low‑carbon computing at scale.
ByteDance’s cloud‑native transformation details a layered technical system, multi‑year Kubernetes‑based evolution, unified multi‑cluster resource management, and hierarchical scheduling, illustrating how the company achieves high development speed, resource efficiency, and prepares for next‑generation serverless infrastructure.
This article explains how cloud‑native AI leverages container‑based architectures and advanced scheduling algorithms—such as resource queues, gang scheduling, bin‑packing, GPU topology‑aware and Tor‑aware dispatch—to improve resource and engineering efficiency, and introduces Baidu’s AI workflow engine PaddleFlow with its design, features, and deployment options.
This article explains why the financial sector is moving its big‑data workloads to cloud‑native platforms, compares cloud‑native systems with traditional Hadoop, describes deployment options such as Serverless YARN and Arcee Operator, and details the high‑performance GRO scheduler, agent, and ResLake resource‑lake architecture that together improve resource utilization, reduce costs, and ensure reliable, low‑latency processing for finance workloads.
This article details iQIYI's multi‑year journey of mixing offline and online workloads using Dragonfly and Koordinator, covering architectural evolution, key factors for successful co‑location, resource‑allocation strategies, the role of Anolis OS, pilot results, and future directions.
This presentation details how Guanyuan Data leverages Apache Spark within its self‑service analytics platform, covering product features, flexible deployment, resource isolation, performance challenges, architectural solutions, and future cloud‑native enhancements to support thousands of users and massive query workloads.
ByteDance’s cloud‑native computing team, led by Li Yakun, details how they transformed a Hadoop‑centric big‑data stack into a Kubernetes‑driven platform—customizing storage, middleware, and scheduling—to support petabyte‑scale workloads, achieve over 40% resource utilization, and sustain rapid product growth.
The article examines the inherent and accidental complexities of Kubernetes as a distributed cluster manager, discusses challenges in resource scheduling, infrastructure diversity, and operational overhead, and explores how cloud‑native solutions such as managed services, nodeless and serverless Kubernetes architectures attempt to reduce these complexities while introducing new trade‑offs.
vivo’s machine‑learning platform, built for its massive app‑store and e‑commerce ecosystem, streamlines data processing, model training, and deployment through quota‑based resource management, a custom ultra‑large‑scale TensorFlow‑vlps framework, OpenAPI‑driven training, and Jupyter‑integrated interactive development, boosting efficiency for billions of samples and features.
Bilibili’s Kubernetes cloud platform boosts server utilization by co‑locating latency‑sensitive online services with batch‑oriented offline jobs on the same nodes, using custom schedulers, extended resources, dynamic CPU/memory isolation, and a management console, achieving average CPU usage around 35 % and significant cost savings.
The article analyzes Meituan’s self‑designed Load Auto‑Regulator (LAR), detailing its tiered resource‑pool architecture, dynamic load‑to‑static‑resource mapping, and QoS mechanisms that together raise data‑center CPU utilization by 5‑10% while keeping online service quality stable, and discusses its deployment in online and mixed‑workload scenarios.
In his Ant Group presentation at the 2022 China Computing Power Conference, He Zhengyu outlined how cloud‑native upgrades, time‑slice scheduling, AI‑driven capacity prediction, and mixed online‑offline deployment have dramatically improved server utilization, cut carbon emissions, and driven open‑source contributions toward sustainable computing.
AI elasticitydata center efficiencygreen computing
This article describes Huya's large‑scale offline‑online mixed deployment, detailing the low resource‑utilization problems, the time‑sharing and elastic scheduling solutions, the containerized architecture, multi‑datacenter isolation, heterogeneous resource handling, stability safeguards, and the resulting performance improvements and future directions.
ByteDance’s cloud‑native ecosystem combines a multi‑layered architecture, dynamic resource over‑provisioning control, hybrid online‑offline scheduling, and federated cluster management to boost container utilization from 23% to 63%, reduce costs by 40%, and support massive events like the 2021 Spring Festival Gala.
This interview follows 25‑year‑old Huawei Cloud engineer Tong Hao, who leveraged his competition‑honed algorithm skills to develop a universal resource‑re‑scheduling solver that fills “Tetris‑like” gaps in data‑center capacity, cutting operational costs by tens of millions of yuan while advancing cloud security and intelligent scheduling.
The article reviews the Global Information System Stability Summit and JD's technical architect Li Junliang's detailed case study on the engineering practices, observability, chaos engineering, and resource‑scheduling innovations that enable JD’s e‑commerce platform to handle sales‑peak traffic that spikes hundreds of times over normal load.
This article details the design and optimization of 58.com’s WPAI machine learning platform, covering background, training‑task scheduling, elastic inference scaling, offline‑online resource mixing, and model‑inference acceleration, and shows how these techniques collectively raised GPU usage by 51% and CPU usage by 38% while cutting costs.
Zhang's team at Zuoyebang details their journey to serverless Kubernetes virtual nodes, explaining how elastic scaling, fine-grained scheduling, and cost‑effective resource utilization transformed high‑peak online services, while addressing challenges in scheduling, observability, performance, and multi‑cloud resilience.
This article explains how ByteDance redesigned its machine‑learning platform using cloud‑native principles, detailing motivations, the shift from Yarn to Kubernetes, the implementation of PS‑Worker and AllReduce frameworks, unified operators, heterogeneous resource scheduling, elastic training, and future directions for large‑scale AI workloads.
Over three years Gaode Map halved overall latency by systematically identifying bottlenecks, applying reverse‑order targeted fixes, establishing forward‑order long‑term controls, and deploying adaptive resource scheduling, engine acceleration, H5 container enhancements, high‑performance components, and CI automation, resulting in sustainable core‑chain performance improvements and a better user experience.
Baidu’s mixed‑workload approach co‑locates offline batch jobs with latency‑sensitive online services on shared nodes, using a dynamic resource view, priority‑based scheduling, cpuset/NUMA isolation, eBPF policies, and predictive profiling, boosting CPU utilization above 40 % and saving billions of RMB in total cost of ownership.
This article analyzes Baidu's offline mixed‑tenant technology that combines online and offline workloads on the same physical servers, detailing the resource‑usage problems, dynamic resource views, priority schemes, isolation mechanisms, high‑performance scheduling, and future directions for cloud‑native clusters.
Over three years Gaode Map continuously optimized its full‑link performance, cutting core‑link latency by half through a systematic three‑step approach—identifying bottlenecks, reverse‑specialized problem solving, and long‑term forward control—resulting in dramatically improved user experience across diverse device tiers.
The article outlines a technical exchange hosted by 58.com AI Lab and Tianjin University that discusses high‑efficiency AI computing, resource‑aware scheduling for both online inference and offline training, and methods to mitigate GPU under‑utilization and gray‑interference in distributed deep‑learning platforms.
This article explains how Tencent Cloud EMR integrated Hadoop YARN with Kubernetes Pods to create a hybrid online‑offline deployment, implement elastic autoscaling and multi‑label resource allocation, and achieve several‑hundred‑percent improvements in CPU utilization while preserving cluster stability.
The article presents a GPU virtual‑sharing solution for AI inference workloads that isolates memory and compute resources via CUDA API interception, integrates with Kubernetes using the open‑source aliyun‑gpushare scheduler, and demonstrates doubled GPU utilization and minimal performance loss across multiple tests.
iQIYI created a custom GPU‑virtual‑sharing system that intercepts CUDA calls to enforce per‑container memory limits, rewrites kernel launches for compute isolation, and integrates with a Kubernetes scheduler extender, allowing multiple AI inference containers to share a single V100 with minimal overhead and more than doubling overall GPU utilization.
After migrating to the cloud, enterprises must evaluate resource scale, cost pressure, and staffing before deciding whether to build their own scheduling system, and can choose among ECS, Dedicated Host, or private pool solutions, each with specific advantages, drawbacks, and suitable scenarios.
This article presents Kuaishou's large‑scale offline computing challenges and describes how the team customized YARN and built the Kwai scheduler to achieve multi‑threaded, pluggable resource scheduling for clusters of tens of thousands of nodes, supporting diverse workloads such as ETL, ad‑hoc queries, machine‑learning training, and real‑time Flink jobs.
This article explains how Facebook’s Twine system orchestrates containers across millions of servers, detailing its architecture, support for stateful services, cross‑data‑center control, elastic capacity handling, and the lessons learned from eight years of large‑scale operations.
This article examines the evolution of user data, computing power, and models, and presents the design principles, key architectural factors, and practical implementation techniques for building a full‑stack machine learning platform that supports large‑scale data processing, distributed training, and low‑latency online serving.
Gaode Map’s low‑cost, long‑term performance strategy for super‑apps combines an adaptive resource‑scheduling framework, full‑dimension monitoring, and closed‑loop control to cut startup time over 70%, shrink memory use 30% and binary size 20%, delivering up to three‑fold speed gains on low‑end devices while preserving development efficiency.
Suning Cloud leveraged micro‑scheduling, dynamic resource allocation, and AI‑driven security to handle the massive traffic of the 818 promotion, achieving a 10% increase in physical‑machine utilization while maintaining stability, cost efficiency, and robust protection against cyber attacks.
This article details Tongcheng‑Elong's journey from bare‑metal to a Kubernetes‑based cloud‑native platform, describing its architecture, the challenges of isolation, scheduling, resource utilization and promotion, and the engineering solutions—including custom scheduling, CPU binding, IP fixation, and over‑commit strategies—implemented to improve efficiency and reliability.
Alibaba Cloud’s award‑winning solutions address bursty peak‑load challenges by integrating container‑based hybrid deployment, intelligent scheduling, and resource isolation, enabling massive e‑commerce events, gene‑computing tasks, and national ticketing systems to achieve high performance, low cost, and near‑zero incremental investment.
Tencent Cloud has open‑sourced its server‑grade operating system kernel, TencentOS Kernel, which offers cloud‑optimized resource scheduling, enhanced container isolation, ARM64 hot‑patching, and performance‑security optimizations that boost CPU utilization and lower operating costs, extending the TencentOS family after the tiny IoT release.
This article details Meituan-Dianping's evolution from custom Docker‑based scaling to a Kubernetes‑driven, cloud‑native cluster management platform (HULK), describing its architecture, scheduler enhancements, Kubelet modifications, and resource‑optimization strategies for large‑scale operations.
This article details Meituan's experience optimizing the Hadoop YARN fair scheduler, covering background challenges, architectural components, resource abstractions, scheduling flow, performance metrics, a series of code‑level optimizations, stability strategies for production rollout, and future directions for large‑scale cluster scheduling.
This weekly note shares personal updates and a concise technical overview covering Yarn's resource scheduling, Hadoop's rack‑aware architecture, HDFS data flow, and practical solutions to the HDFS small‑file problem, along with links to further reading and upcoming work plans.
The article presents a detailed technical analysis of JD's Archimedes project, describing its evolution from JDOS 2.0 to a large‑scale container scheduling platform that dramatically improves resource utilization, deployment speed, and cost efficiency across JD’s data centers.
This article describes how the Liulishuo algorithm team designed and implemented a Kubernetes‑based training platform that addresses the iterative, data‑intensive, and resource‑dynamic characteristics of machine learning workloads by pooling resources, enabling rapid provisioning, and optimizing scheduling and storage.
360’s AI‑ops team shares a year‑long journey of turning massive operational data into intelligent solutions—covering background, their AIOps philosophy, practical modules like capacity forecasting, host classification, resource reclamation, smart MySQL scheduling, anomaly detection, alarm reduction, and root‑cause analysis—to dramatically improve cost, efficiency, and reliability.
This article describes how we identified underutilized CPU and memory resources in our company's servers, evaluated Kubernetes versus Apache Mesos, and built a non‑intrusive, Mesos‑based multi‑task scheduling system with dynamic resource reservation, monitoring, task isolation, and cluster‑wide observability, while addressing deployment challenges.
This article explains how we leveraged Apache Mesos to create a multi‑task scheduling system that maximizes idle CPU and memory on legacy CentOS machines without kernel upgrades, detailing architecture, deployment, monitoring, resource isolation, and remaining challenges.
Scheduling algorithms are a crucial component of cluster resource management systems, determining where containerized tasks run to ensure resource needs, high availability, fault tolerance, and cost efficiency across individual containers, applications, and entire data centers, while also supporting Alibaba’s global scheduling challenge.
This article details how 360's commercial advertising platform leverages Mesos and Docker containerization to solve data‑center migration, fault recovery, OS inconsistencies, and resource‑utilization challenges, describing the architecture, standardization, networking, storage, service discovery, and future plans.
This article explains Alibaba's co‑location (混部) technology that mixes online services and batch compute on the same physical servers, detailing its background, key characteristics, scheduling architecture, resource isolation mechanisms, cost‑saving formulas, and future roadmap, showing how it boosts utilization and reduces expenses.
This article analyzes the instability of Alibaba's test environment container provisioning, identifies root causes, and presents a comprehensive solution—including automatic container replacement, a buffer pool, and resource‑pool rationalization—that raised the container success rate to 99.9% and stabilized performance.
JD.com's Archimedes is a data center operating system that provides one-click scaling, intelligent scheduling, mixed deployment, and multi-site active-active capabilities, significantly improving resource utilization and reducing IT costs.
Alibaba’s mixed‑deployment (Co‑location) technology combines online services and batch compute tasks on shared physical resources, using priority‑based scheduling, resource isolation, and dynamic memory management to dramatically improve CPU utilization, cut infrastructure costs, and maintain service level objectives during peak traffic events.
This article describes the design, challenges, and operational practices of a cloud‑native elastic computing platform that reuses idle resources from production servers to support massive image compression, video transcoding, AI inference, and log processing while ensuring online services maintain performance, latency, and reliability.
Over nine years, Alibaba transformed its Double 11 e‑commerce platform from a centralized system to a highly elastic, cloud‑native architecture, employing distributed design, multi‑active regions, unified scheduling, containerization with Pouch, and hybrid deployment to dramatically cut costs and boost peak throughput.
The article details Alibaba's Taobao Personalization Platform (TPP) intelligent scheduling system, explaining its architecture, optimization algorithms, convergence logic, and performance results that dramatically improve CPU utilization and automate scaling during both regular operation and high‑traffic events like Double‑11.
The article explains how rapid internet growth has expanded data centers, why traditional operations fall short, presents a simple utilization formula, shows Alibaba’s mixed offline‑online scheduling experiment that raised server usage from 10% to over 40%, and announces an open dataset for academic research.
Tencent’s elastic computing platform replaces 24,000 physical servers with just 6,000 containers, delivering sustainable compute for billions of daily image compressions while also supporting video transcoding, Spark jobs, and AI workloads through dynamic resource isolation, named services, and intelligent scheduling.
This article explains the architecture of Docker Swarm, the reasons for running it on Apache Mesos, the integration process—including resource offers, task scheduling, and container creation—and discusses current limitations and future improvement directions.
This article details how Tencent Games built, customized, and continuously optimized a Kubernetes‑based container platform over three years to support tens of thousands of game containers, covering deployment modes, scheduler enhancements, network solutions, resource quotas, monitoring, storage, and the transition to micro‑service architectures.
This article examines resource allocation philosophies—auction, budgeting, and preemption—and compares the architectures, data models, and APIs of major schedulers such as Borg, Omega, Mesos, Kubernetes, and Alibaba’s Zeus, while also exploring sharing strategies, task classifications, utilization metrics, and predictive techniques for efficient resource management.
This article details China Mobile Zhejiang's journey from early virtualization to a full private‑cloud platform built on MESOS, covering why MESOS was chosen, the evolution of their cloud stages, DCOS implementation, automatic scaling, service discovery, and the operational benefits achieved.
This article compares the resource allocation philosophies, architectural designs, data handling, and API models of Borg, Omega, Mesos, Kubernetes, and Alibaba's Zeus, discussing auction, budgeting, preemption, sharing models, task types, utilization, prediction, and practical implementation details for large‑scale cloud native environments.
The article examines Alibaba's Zeus resource scheduling platform, detailing its background, problem analysis, container‑based virtualization, distributed architecture, strategies for improving resource utilization such as overselling and hybrid deployment, as well as stability measures and automation for large‑scale operations.