This article explains how to design and implement hardware‑offloaded, IP‑ and port‑granular traffic shaping using Open vSwitch, OpenFlow meters, Mellanox HQoS, and DPDK, covering software limits, meter hierarchy, flow‑meter association, deletion workflows, RCU handling, and same‑host traffic considerations.
This article explains how traditional kernel network stacks struggle with high‑concurrency, low‑latency cloud data‑center workloads and introduces Baidu Intelligent Cloud’s Polar solution—Polar‑TCP and Polar‑RDMA—which combine user‑space DPDK drivers, a lightweight TCP stack, and an industrial RPC framework to achieve near‑RDMA performance while preserving compatibility with existing TCP ecosystems.
This article explains how traditional kernel network stacks struggle with modern cloud data‑center workloads and introduces Baidu Intelligent Cloud's Polar solution—Polar‑TCP and Polar‑RDMA—which combine user‑space DPDK drivers, a lightweight TCP stack, and an industrial‑grade RPC framework to achieve near‑RDMA performance while preserving ecosystem compatibility.
This article explains how Huawei Cloud's HyperRouter implements deterministic operations through a combination of L4/L7 load‑balancing co‑design, high‑performance data‑plane choices, self‑healing mechanisms, point‑to‑point architecture, Cell + Shuffle‑Sharding isolation, and user‑centric observability, providing a reproducible blueprint for reliable cloud services.
This article analyzes why VPC SNAT and Ops NAT gateways overlap, proposes a unified gateway design using LPM and classified hash tables, discusses rule handling, performance trade‑offs, and optimization techniques, and concludes that the solution greatly improves NAT performance while preserving flexibility.
This article explains the Cloud Connect Network (CCN) background, describes the V1 optimization that switched packet forwarding to a session model, outlines the resulting issues, proposes a route‑based session list redesign, presents performance test results, and discusses future enhancements for high‑performance cloud networking.
This article explains the challenges of using community OVS with Mellanox ConnectX‑6 NIC for hardware offload, details the design and implementation of Cx6 offload configuration, flow‑table match/action adaptation, Conn‑based large‑flow offload, and the coordination mechanism between hardware‑offloaded and software flow tables to achieve significant CPU savings while preserving security‑group and connection‑tracking correctness.
The paper presents Tencent’s TGW cloud gateway architecture, highlighting a 2.9× forwarding performance boost, lossless state migration within 4 seconds, sub‑minute fault detection, multi‑level fault‑tolerance mechanisms, and operational best practices that enable 100 % availability for massive online services.
This article explains the concept of real‑time operating systems, distinguishes soft and hard real‑time, and walks through Linux kernel factors such as clock tick, interrupt handling, scheduling classes, run‑bandwidth limits, and DPDK polling, providing concrete examples and practical tuning steps.
This article outlines the background, architectural design, implementation details, and future roadmap of an internet‑boundary cloud firewall built on DPDK, covering physical and software structures, traffic steering, rule matching logic, and planned enhancements such as logging and traffic analysis.
This article details the design, implementation, and performance evaluation of a user‑space MPTCP stack built on DPDK, showing how a layered, zero‑copy architecture and same‑core lock‑free forwarding can boost data‑center throughput by up to 100% while reducing latency by about 10%, all while remaining compatible with existing TCP applications.
At Netdev 0x19, ByteDance's STE team presented two technical talks: a NUMA‑aware MPTCP flow‑selection strategy that boosts Redis benchmark throughput by up to 30% and cuts tail latency by 6%, and a DPDK‑based user‑space MPTCP stack that halves latency and doubles throughput in data‑center tests.
At Netdev 0x19, ByteDance's STE team presented two talks—one on a NUMA‑locality‑aware MPTCP flow‑selection strategy that can raise throughput by up to 30% and lower tail latency by 6%, and another on a DPDK‑based user‑space MPTCP stack that reduces latency by nearly 10% and more than doubles throughput—showcasing practical performance gains for data‑center networking.
This article demonstrates how to use DPDK's Graph API for L3 packet forwarding, introduces a custom ACL node, and presents performance comparisons of different packet‑transfer mechanisms and batch sizes within the graph pipeline.
This article explains the design and implementation of a DPDK‑based domain interception gateway that analyzes HTTP/HTTPS traffic, uses TCP connection hijacking and DPDK’s KNI interface to block unregistered domains, detailing architecture, packet processing flow, and the trade‑offs of this approach.
dperf is an open‑source DPDK‑based network performance and load testing tool that can generate tens of millions of HTTP connections per second, achieve hundreds of Gbps throughput, and provides detailed statistics, with usage instructions, performance benchmarks, and configuration steps for testing environments.
This article examines the Cloud Connect Network gateway architecture, identifies performance bottlenecks in the V1 packet‑forwarding process, proposes a session‑based optimization that reduces query complexity from O(n) to O(1), and validates the improvement with latency tests showing a drop from 500 ms to about 1 ms.
This article walks through Intel's high‑density scalable load balancer (HDSLB), detailing its layered software architecture, advanced elephant‑flow handling, fast/slow‑path separation, packet‑forwarding optimizations, and provides a deep code analysis of its configuration, initialization, and data‑plane job execution.
This article consolidates DPDK 17.11 source‑code notes to explain the library’s memory‑management subsystem, covering hugepage concepts, shared configuration mapping, NUMA‑aware allocation, and the custom allocator that enables high‑throughput packet processing on Linux.
This article explains the architecture and implementation of the virtio/vhost device model, details how QEMU creates and initializes virtio‑net‑pci devices, describes the vhost‑user communication protocol and its integration with DPDK and the Linux vhost‑net kernel driver, and provides practical command‑line examples and code snippets.
DPDK enables real‑time, high‑throughput packet handling on low‑cost commercial servers by bypassing the kernel, offering flexible processing models, and supporting a wide range of open‑source projects, while also presenting challenges and alternatives for modern networking workloads.
This article explains virtual network security mechanisms, including security groups, network ACLs, QoS policies, and quality‑assurance techniques, detailing implementations with Linux Bridge, iptables, OVS, conntrack, and DPDK, and discusses active probing and passive observation for network reliability.
DPDK enables low‑cost commodity servers to achieve high‑throughput, low‑latency packet processing by bypassing the kernel, offering flexible processing models, supporting user‑space networking stacks, and being adopted across telecom, cloud, and edge environments despite challenges such as debugging and CPU usage.
This article explains the origins of DPDK, describes its modular architecture and performance‑enhancing mechanisms such as UIO, hugepages, and CPU affinity, and reviews popular user‑space networking frameworks like F‑Stack and Seastar that leverage DPDK for high‑throughput cloud services.
dperf, Baidu’s open‑source DPDK‑based tool, delivers ultra‑high‑performance network testing by generating up to ten million HTTP connections per second, hundreds of gigabits of throughput, and detailed per‑second statistics, with step‑by‑step guidance on environment setup, compilation, and execution.
dperf, an open‑source Baidu tool built on DPDK, delivers ultra‑high network performance—supporting millions of HTTP connections per second, hundreds of Gbps throughput, and detailed statistics—making it ideal for load‑balancer, cloud VM, NIC, and CPU network testing, with step‑by‑step deployment instructions.
NVIDIA's new DOCA GPUNetIO library enables GPU‑initiated communication, allowing packets to be received directly into GPU memory, processed by CUDA kernels, and sent without CPU involvement, offering lower latency, higher scalability, and detailed pipeline examples including IP checksum, HTTP filtering, traffic forwarding, and 5G Aerial SDK integration.
The article analyzes the growing demands on network I/O, outlines Linux and x86 bottlenecks, and explains how DPDK’s user‑space bypass, UIO, PMD, and various optimization techniques such as HugePages, SIMD, and cache‑friendly design enable multi‑hundred‑million‑packet‑per‑second processing.
This article explains how to build a reliable, high‑performance load‑balancing platform by analyzing basic reliability requirements, introducing multi‑layer (DNS, L4, L7) balancing, comparing Direct‑Route, Tunnel and FULLNAT modes, and describing the VGW architecture, health‑check, fault isolation, redundancy and DPDK‑based performance optimizations.
This article investigates the performance degradation caused by DPDK memory fragmentation, explains the underlying heap and element management, presents detailed profiling and test results, and proposes two practical solutions that dramatically reduce allocation latency and improve overall system throughput.
This article analyzes the evolving demands of network I/O, identifies Linux/x86 bottlenecks, explains DPDK’s bypass architecture and UIO mechanism, and presents practical high‑performance coding and compilation optimizations such as HugePages, SIMD, poll‑mode drivers, and ecosystem tools for modern backend systems.
This article explains the need for reliable load balancing in large‑scale services, analyzes the problems of request distribution and fault isolation, and details the design of a three‑layer and four‑layer load‑balancing architecture—including DNS, Nginx, LVS, FULLNAT, and VGW—along with health‑check, redundancy, and performance optimization techniques.
This article reviews the fundamentals of token‑bucket rate limiters, identifies precision, cascading compensation, and TCP‑loss sensitivity issues, and details two major improvements—port‑loan backpressure and the Carousel algorithm—while outlining future directions for more reliable network traffic shaping.
The article details how a multi‑layer load‑balancing platform (VGW) was designed—combining 7‑layer Nginx, 4‑layer LVS with FULLNAT, and 3‑layer network devices—to achieve business reliability, fault isolation via BGP‑announced VIPs, and high‑throughput performance using DPDK, while providing redundancy at server, link, and cluster levels.
The article examines the evolving demands on network I/O, outlines the limitations of traditional kernel‑based packet processing on Linux/x86, and explains how user‑space frameworks such as DPDK and its UIO/PMD architecture, along with optimization techniques like huge pages, SIMD, and cache prefetching, can achieve multi‑gigabit, million‑packet‑per‑second throughput.
DPDK’s libgraph, inspired by VPP’s vector packet processing, replaces traditional scalar packet handling with a graph‑based pipeline that improves i‑cache locality, reduces pointer copies, and enables flexible, high‑throughput networking across multiple CPUs and NICs, while introducing trade‑offs such as memory safety risks.
This article provides a comprehensive overview of DPDK, detailing its basic and optimization technologies, architectural components such as the Environment Abstraction Layer, core libraries, platform modules, poll‑mode drivers, hugepage memory management, and CPU‑affinity techniques that together enable high‑performance packet processing in NFV and SDN environments.
The article analyzes the growing demands on network I/O, explains Linux’s kernel bottlenecks, and shows how user‑space frameworks like DPDK and UIO can achieve multi‑hundred‑million packets per second through techniques such as huge pages, SIMD, zero‑copy, and careful CPU cache management.
This article examines the evolving demands of network I/O, analyzes Linux and x86 bottlenecks, explains DPDK's user‑space bypass architecture, and presents practical optimization techniques—including huge pages, SIMD, cache prefetching, and compile‑time tricks—to achieve multi‑gigabit packet processing rates.
This article provides a comprehensive overview of NVIDIA BlueField DPU technology, detailing its interface and driver architecture, work modes, kernel representation model, multi‑host capabilities, virtual switch implementations, scalable functions, RDMA support, security accelerators, and related performance‑enhancing features for modern cloud and data‑center environments.
This article provides a comprehensive technical overview of DPDK, covering its architecture, core libraries, platform modules, polling and CPU‑affinity techniques, huge‑page memory management, NUMA considerations, OS tuning steps, and integration with OVS for high‑performance packet processing.
This article explains how hardware architecture, Linux OS version, kernel configuration, OVS integration, memory management, NUMA awareness, and CPU micro‑architecture affect DPDK application performance and provides concrete tuning steps such as CPU isolation, service disabling, huge‑page setup, and optimized memory allocation.
This article explains DPDK’s open‑source data‑plane framework, detailing its architecture, core libraries, platform modules, poll‑mode drivers, huge‑page memory management, polling technique, and CPU‑affinity strategies that together eliminate kernel bottlenecks and dramatically improve packet‑forwarding performance.
This article provides a comprehensive overview of DPDK, covering its fundamental and optimization technologies, software architecture, core libraries, platform modules, poll‑mode drivers, huge‑page usage, polling techniques, and CPU‑affinity strategies for high‑performance packet processing in NFV environments.
This article presents a detailed technical overview of how ByteDance selected Kube-OVN for a mixed virtual‑machine and container networking scenario, describes the initial network design, identifies performance issues, and outlines three improvement plans including OVS‑DPDK, source‑route optimization, and a switch from Geneve to VXLAN.
This article examines real‑world load‑balancing implementations for ultra‑high traffic cases such as Alibaba's Double 11 shopping festival, China's 12306 railway ticketing system, WeChat's red‑packet service, and Douyin's Spring Festival gala, highlighting architectural principles, DPDK usage, ECMP routing, session synchronization, SET‑based sharding, and service‑mesh techniques.
The article describes how Trip.com built a high‑performance, software‑based layer‑4 load balancer (TDLB) using DPVS and DPDK, detailing its lock‑free session handling, user‑IP pass‑through, asynchronous logging, cluster session synchronization, resource isolation, configuration management via etcd/operator, health‑check strategies, and multi‑dimensional monitoring.
To meet growing business demands, 360’s virtualization team replaced the legacy overlay network with a host‑overlay VPC solution that decouples switches, supports private cloud isolation, leverages DPVS‑based gateways, and integrates monitoring, delivering high‑availability, scalable traffic handling across its 25 G data centers.
CKV+, Tencent’s next‑generation high‑performance NoSQL KV database compatible with Redis, Memcached and ASN, evolves from early CMEM architectures, adopts a multi‑tenant shared‑memory design built on Seastar, integrates DPDK for user‑space networking, and achieves up to ten million QPS with significant cost and latency improvements.
DPVS v1.9.0, the latest open‑source four‑layer network load balancer from iQIYI, now fully supports DPDK‑20.11 LTS, offers streamlined compilation, universal flow management, dynamic mbuf fields, enhanced per‑core scheduling and UDP health checks, and has been running stably in production for three months.
The article examines the evolution of Tencent Cloud’s F-Stack—a full user‑space networking framework built on DPDK, FreeBSD, and micro‑threading—detailing its architecture, supported versions, POSIX‑compatible API, real‑world deployments such as HttpDNS and DNSPod, performance gains over kernel stacks, and the roadmap for future enhancements.
This article describes 360's Cloud Connect Network solution, covering its background, application scenarios, design, server architecture, traffic forwarding mechanisms, CCSI traffic isolation, performance optimizations with DPDK, and future enhancements for multi‑VPC and IDC interconnectivity.
This article explains how NVMe technology, combined with Intel's DPDK/SPDK user‑space framework, transforms enterprise storage by reducing I/O latency, simplifying the software stack, and enabling high‑performance NVMe‑oF, FCP, and iSCSI services with detailed architectural diagrams and implementation enhancements.
This article explains how Intel’s DPDK and SPDK technologies enable high‑performance NVMe‑oF storage by leveraging user‑space drivers, RDMA, and modular architecture to reduce latency, improve I/O throughput, and address reliability challenges in modern distributed storage systems.
This article compares four classic Linux packet‑capture engines—libpcap (and its mmap variant), PF_RING, DPDK, and XDP—explaining their data‑flow architectures, copy operations, zero‑copy techniques, and typical use‑cases such as DDoS defense and high‑speed networking.
This article compares four classic Linux packet‑capture engines—libpcap, libpcap‑mmap, PF_RING, DPDK, and XDP—detailing their data‑path architectures, copy‑count reductions, zero‑copy techniques, and typical use‑cases such as DDoS defense and high‑speed trading.
This article explains the core principles of DPDK memory management, covering standard huge pages, NUMA node binding, direct memory access, IOMMU and IOVA addressing, custom allocators, and memory pools, and how these mechanisms together enable high‑performance packet processing on Linux systems.
This article summarizes essential concepts and practical optimization strategies for DPDK, covering architecture fundamentals, x86 performance analysis, virtualization overhead, NFV forwarding bottlenecks, huge‑page usage, polling modes, CPU affinity, and I/O virtualization techniques to achieve high‑throughput, low‑latency packet processing.
DiDi Cloud’s network team replaced CPU‑bound OVS‑DPDK processing with programmable NICs, achieving sub‑150 µs VM‑to‑VM latency and up to 9.5 Mpps on 64‑byte packets, while overcoming offload, driver, and flow‑table challenges through extensive patches and open‑source contributions.
This article examines the limitations of traditional L4 load balancers like LVS and explains how modern software solutions—leveraging DPDK, VPP, XDP/eBPF, and consistent‑hash algorithms such as Google Maglev and Facebook Katran—can achieve high‑throughput, scalable, and highly available L4 load balancing in data‑center environments.
This article explains the Linux network I/O architecture, detailing how TCP sends and receives data, the role of socket buffers, SKB structures, QDisc and ring buffers, and compares various zero‑copy techniques such as read/write, mmap/write, sendfile, splice, tee, DPDK and DMA‑gather approaches.
iQIYI built QNSM, a DPDK‑based full‑traffic analysis engine for complex multi‑zone network boundaries, delivering high‑performance (10+ Gbps per NIC) real‑time detection of DDoS, IDPS, and DLP, with modular pipelines, Kafka integration, and open‑source deployment across 22 clusters.
This article provides a comprehensive technical guide to DPDK and SPDK, covering their background, architecture, user‑space I/O bypass mechanisms, core performance optimizations such as HugePages, SIMD and cache management, and practical ecosystem recommendations for building high‑throughput network and storage services.
This article explains cloud‑native networking fundamentals, detailing the role of CNI, the high‑performance packet processing capabilities of DPDK and its UIO and poll‑mode drivers, and introduces the Contiv/VPP Kubernetes plugin that leverages VPP for efficient pod‑to‑pod communication.
This article provides an in‑depth technical overview of DPDK and SPDK, covering their background, the evolution of network I/O, Linux bottlenecks, user‑space I/O via UIO, poll‑mode drivers, performance‑optimizing techniques such as huge pages, SIMD, cache management, and the surrounding ecosystem and adoption.
The article analyzes the growing demands on network I/O, explains Linux and x86 bottlenecks, introduces DPDK’s user‑space bypass architecture and its core optimizations such as hugepages, poll‑mode drivers, SIMD, and CPU‑specific tuning, and finally discusses the DPDK ecosystem and practical considerations for backend developers.
DPVS is an open‑source, DPDK‑based user‑space load balancer that achieves line‑rate throughput by assigning each worker to a dedicated CPU core and NIC queue, provides lock‑free per‑CPU data structures and multiple forwarding modes including NAT64, and simplifies maintenance while supporting IPv4/IPv6 in large‑scale production deployments.
This article details the discovery, analysis, and resolution of a DPDK bug that caused a zero‑byte IP fragment to fill transmission queues, leading to packet loss and tx‑hang in UCloud's high‑availability ULB4 load balancer, and shares the tools and lessons learned for future debugging.
The article analyzes the evolving demands of network I/O, the limitations of traditional kernel‑based networking, and presents DPDK’s user‑space bypass architecture, UIO mechanism, and a series of low‑level optimizations—including HugePages, poll‑mode drivers, SIMD, and cache‑aware coding—to achieve multi‑gigabit packet processing performance on modern Linux servers.
By integrating OVS‑DPDK with VF‑based flow separation, stateless forwarding, and dual redundant processes, Didi Cloud’s engineering team achieved millisecond‑level hot upgrades—approximately 80 ms per VM—while maintaining high‑performance packet forwarding, scalability across hundreds of VMs, and rapid fault recovery.
GitHub’s GLB Director is an open‑source, layer‑4 load balancer designed for data‑center environments that scales a single IP across thousands of servers, uses ECMP, a stateless director layer, DPDK‑accelerated packet processing, and health‑check mechanisms to provide high‑availability without disrupting existing connections.
This article chronicles UCloud's evolution from early Open vSwitch‑based SDN to DPDK gateways, evaluates excluded VXLAN and OpenFlow solutions, and explains how adopting Barefoot's P4 programmable switches enabled higher performance, flexible control planes, sharding, and seamless grey‑release upgrades for its cloud network.
The article presents ContainerLB, a software‑defined L4 load‑balancing service built on Intel DPDK and X86 servers for JD.com, detailing its architecture, FULLNAT mode, high‑availability design, performance optimizations, and test results that demonstrate near‑line‑rate packet processing and scalable deployment in a container‑native data‑center.
This article presents the design, DPDK‑based implementation, performance optimizations, and test results of KDNS—a C‑language DNS server that replaces ContainerDNS's Go server, achieving near line‑rate processing, high availability, and significant latency improvements over traditional bind9 solutions.
The article describes the motivation, architecture, FullNat forwarding process, high‑availability mechanisms, and performance‑optimizing techniques of TVS, a custom four‑layer load‑balancing platform built to overcome LVS limitations and meet the company’s demanding traffic and reliability requirements.
This article details the evolution, architecture, deployment, monitoring, and performance optimization of Ultron—360’s internal OpenStack‑based virtualization platform—covering its three development stages, technical stack, automation with Ansible, advanced features like VXLAN and Ceph, and lessons learned from large‑scale operations.
This article details the HULK team's experience upgrading their private‑cloud OpenStack environment from Kilo to Mitaka, covering background, risk‑benefit analysis, preparation, step‑by‑step procedures, common pitfalls, and practical solutions, offering actionable guidance for similar OpenStack migrations.
Meituan’s self‑developed MGW replaces LVS with a DPDK‑based four‑layer load balancer that handles tens of gigabits per second and millions of concurrent connections by eliminating interrupts and locks, isolating control and data planes, and providing OSPF‑ECMP clustering with seamless failover and session synchronization.
This article details Meituan Cloud's evolution from a private, self‑developed cloud platform to a high‑performance public cloud, covering the motivations, architectural upgrades across physical links, host networking, gateways, controllers, and the adoption of overlay technologies like VXLAN to achieve industry‑leading scalability and reliability.
